1 LINKING PLANT DEMOGRAPHY, FOREST FU ELS, AND FIRE IN LONGLEAF PINE ( PINUS PALUSTRIS) SAVANNAS USING LIDAR REMO TE SENSING AND SIMULATION MODELING By EVA LOUISE LOUDERMILK A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL OF THE UNIVERSITY OF FLOR IDA IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY UNIVERSITY OF FLORIDA 2010
2 2010 Eva Louise Loudermilk
3 To my family, especially our latest addition, Kai
4 ACKNOWLEDGMENTS I thank m y family, friends, and colleagues for their endless support and encouragement. My parents and husband have always embraced my passion to continue my studies through graduate school. My son has provided me with daily doses of love and laughter. My advisor, Wendell P. Cropper, Jr. has had an immeasurable impact on my research, career, and life, providing continuous encouragement and humility throughout the years. Several of my committee members have been integral to my research, especially Bob Mitchell from the Jones Ecological Research Center. Sa bine Grunwald, Jasmeet Judge, a nd George Tanner have always provided time for me with us eful suggestions. Kevin Hier s and Joe OBrien have been fundamental collaborators, always available for research advice and insight. My lab mates, Doug, Jen, Nilesh, Denis, Todd, Sparkle, Helen, Luis and countless others along the way have been there to provide feedback on ideas, pe rsonal support, and endless entertainment. The School of Natural Resources and Environment and the School of Forest Resources and Conservation at the University of Florida prov ided logistic and fina ncial support throughout my graduate career. Special thanks go to Dr. Hu mphrey and Dr. White, directors of each school, respectively. I thank the Joseph W. Jones Ecolog ical Research Center and the Ordway-Swisher Biological Station for their support during research development, field work, and final stages of this work. I also thank the Na tional Center for Airborne Lase r Mapping at the University of Florida for providing the Mobile Terrestrial Laser Scanner and engineering expertise, with special thanks to Clint Slatton, Juan, Abhinav, and Heezin.
5 TABLE OF CONTENTS page ACKNOWLEDGMENTS ............................................................................................................... 4LIST OF TABLES ...........................................................................................................................8LIST OF FIGURES .......................................................................................................................10ABSTRACT ...................................................................................................................... .............12 CHAP TER 1 INTRODUCTION .................................................................................................................. 14Longleaf Pine Savannas: Fuels, Fire, and Plant Demography ................................................ 14Fuel and Fire Heterogeneity ...................................................................................................15Modeling Longleaf Savanna Demography ............................................................................. 18Study Overview ......................................................................................................................19Measuring Fuel Structure ................................................................................................ 19Examining Fuel and Fi re Relationships .......................................................................... 19Modeling Savanna Dynamics .......................................................................................... 20Study and Management Implications ..................................................................................... 202 GROUND BASED LIDAR (LIGHT D ETECTION AND RANGI NG): A NOVEL APPROACH TO QUANTIFY FINE SC ALE FUELBED CHARACTERISTICS ............... 23Introduction .................................................................................................................. ...........23Methods ..................................................................................................................................26Ground-LIDAR Instrumentation .....................................................................................26Individual Shrub Assessment ..........................................................................................28Field Assessment of Complex Fuelbeds ..........................................................................29Study site: Ichauway preserve ..................................................................................29Field plot data collection ..........................................................................................30Data Processing and Analysis .........................................................................................31Results and Discussion ........................................................................................................ ...34Application ................................................................................................................... ..........433 RELATING FINE-SCALE FOREST FUELBE D STR UCTURE AND TYPE TO FIRE BEHAVIOR USING GROUND-LIDAR, THERMAL IMAGING, AND REGRESSION-TREE MODELING ...................................................................................... 48Introduction .................................................................................................................. ...........48Fuel and Fire in the Longleaf Pine Ecosystem ................................................................ 48Measuring Fuelbed Structure ..........................................................................................50Relating Fuelbed to Fire Behavior ..................................................................................52Regression-Tree Modeling ..............................................................................................54
6 Methods ..................................................................................................................................56Ground-LIDAR (Light Detection a nd Ranging) Instrumentation ...................................56Study Site .........................................................................................................................58Data Collection ................................................................................................................58Data Processing ...............................................................................................................61Grouping of Plots ............................................................................................................65Regression-Tree Modeling ..............................................................................................65Results and Discussion ........................................................................................................ ...69Regression-tree models : collective data .......................................................................... 69Regression-tree models: fire groups ................................................................................ 72Regression-tree models : individual plots ........................................................................ 79Comparison of temperature models (Tmax vs. Q90) ...................................................... 85Comparison of residence tim e models (T300 vs. T500) .................................................86Evaluation of Potential Pitfalls ........................................................................................86Conclusions .............................................................................................................................89Implications .................................................................................................................. ..........904 SIMULATING LONGLEAF PINE AND HA R DWOOD DEMOGRAPHY, FOREST FUELS, AND FIRE USING A LATTICE APPROACH ....................................................... 93Introduction .................................................................................................................. ...........93Pine and Hardwood Dynamics ........................................................................................94Fuel and Fire Dynamics ...................................................................................................97Modeling Longleaf Pine and Hardwoods ........................................................................ 98Study Areas ...........................................................................................................................102Ichauway Preserve .........................................................................................................103Ordway Swisher Biological Station .............................................................................. 103Field Studies ..................................................................................................................104LLM (Longleaf Model): Model De velopment and Description ...........................................105Longleaf Pine Population ..............................................................................................107Fecundity ................................................................................................................107Growth ....................................................................................................................109Mortality ................................................................................................................. 109Competition index .................................................................................................. 112Hardwood Population .................................................................................................... 113Fecundity ................................................................................................................113Growth ....................................................................................................................114Mortality ................................................................................................................. 114Fire and Fuelbed Component ........................................................................................116Model Calibration and Evaluation .................................................................................119LLM: Model Evaluation and Discussion .............................................................................. 122Model Dynamics ...........................................................................................................122Fire, fecundity, and seed masting ........................................................................... 122Mortality among size classes ..................................................................................125Sensitivity analysis: Impacts from fire frequency and competition ....................... 128Evaluation among study sites .................................................................................133Model Vers atility ........................................................................................................... 136
7 Knowledge Gaps and Future Directions ........................................................................138Conclusions ...........................................................................................................................1405 CONCLUSIONS .................................................................................................................. 142Study Synthesis .....................................................................................................................142Measuring Fuel Structure .............................................................................................. 142Examining Fuel and Fi re Relationships ........................................................................ 143Modeling Savanna Dynamics ........................................................................................ 144Study and Management Implications ................................................................................... 145 APPENDIX A AUXILIARY INFORMATION FO R CART ANALYSIS .................................................. 149B AUXILIARY INFORMATION FOR LLM ......................................................................... 158LIST OF REFERENCES .............................................................................................................159BIOGRAPHICAL SKETCH .......................................................................................................171
8 LIST OF TABLES Table page 2-1 Manufacturer specifications of the gr ound-LIDAR (Light Detection and Ranging) instrum ent (Optechs ILRIS 36D) used in this study.. .......................................................272-2 ANOVA (analysis of variance) output an alyzing individual whole shrub volume (m3) estimates with three treatments .......................................................................................... 342-3 Linear regression re lating three volume (m3) estimation methods (LIDAR and two traditional methods assuming cylinder and sphere shape) to biomass (g) and leaf area ... 362-4 ANOVA output analyzing sect ional shrub 1) LIDAR volume (m3), 2) biomass (sqrtg), and 3) leaf area (sqrt-m2) estimates with three treatments ........................................... 392-5 Linear regression re lating LIDAR volume (m3) to biomass (g), and leaf area (m2) from sectional shrub data.. ................................................................................................. 403-1 Burn day mean wind speed (m/s) and direction. ................................................................ 603-2 Fuelbed [ground-LIDAR (Light Detection and Ranging) and point-i ntercept] and fire (FLIR Forward Looking Infrared) metrics created for the CART .................................. 633-3 Results of CART analysis at various multi-plot levels. ..................................................... 703-4 Descriptive statistics [mean (SD: standard deviation)] of fire response variables at the various multi-plot and i ndividual plot analyses. .......................................................... 723-5 Top five importance value (IV) rankings of input variables fro m CART analysis at various multi-plot level analysis. ..................................................................................... 733-6 Descriptive statistics [mean (SD)] of a representative set of LIDAR fuelbed metrics at the various multi-plot and indi vidual plot level analysis ............................................... 743-7 Percent abundance of each fuel type ..................................................................................753-8 Results of CART analysis for individual plots (n = 169 each) within the three fire groups, including those with x,y coor dinates included as inputs ....................................... 803-9 Top five importance value (IV) rankings of input variables fro m CART analysis at the individual plot level .....................................................................................................824-1 Input and calibrated parameters used for the LLM. ......................................................... 1074-2 Longleaf pine cone production categorized into four intensity levels for subadults and adults for mast and non-mast years in the LLM ....................................................... 108
9 4-3 Sensitivity analysis for the fire mortality parameter for longleaf pine (LP) and hardwoods (HW ). ........................................................................................................... 1314-4 Sensitivity analysis for the competition index (CI) parameter for longleaf pine ............. 1334-5 Height distributions estimated for Icha uway and Ordway from LIDAR data, field data, and simulation outputs from the LLM.. .................................................................. 1344-6 Population densities estimated for Icha uway and Ordway from LIDAR data, field data, and simulation outputs from the LLM. ................................................................... 134A-1 Date recorded and mean wind speed measur ed within each plot as the fire passed through.. ..................................................................................................................... ......149A-2 Comprehensive CART output of importance values (IVs) for each fuel metric within each fire model (Tmax, Q90, T300, T500) ...................................................................... 150B-1 Specifications of Optechs ALTM (Air borne Laser Terrain Ma pper) instrumentation for aerial LIDAR (Light Detection and Ra nging) data collection at the Ordway ........... 158B-2 Specifications of Optechs Gemini instrumentation for aerial LIDAR data collection at the Ichauway. .............................................................................................................. .158
10 LIST OF FIGURES Figure page 1-1 Conceptual diagram of longleaf pine-h ardwood population dynam ics and feedbacks associated with fuels and fire. ............................................................................................161-2 Fine-scale spatial heterogeneity of A) fuelbed structure (depth, m) and B) fire behavior (temperature C) recorded with in a longleaf pine-wiregrass system .................. 172-1 Output 3D (three-dimensional) point-clouds from the ground LIDAR (Light Detection and Ranging) system. ........................................................................................ 242-2 Variation of volume for two plant species (saw palmetto and wax myrtle), two sizes (large and small) and the methods used to calculate volume ............................................. 352-3 Linear relationships of le af area measurements of all tw elve individual whole shrubs .... 372-4 Mean volume and standard errors of vert ical sections (equal thirds) of two pineflatwoods shrub species. ....................................................................................................402-5 Linear relationships of LIDAR volume esti mates with A) leaf biomass and B) leaf area measurements of all twelve individual shrubs, divided into three equal vertical sections ...................................................................................................................... .........422-6 Linear correlation between modeled volumes calculated from ground LIDAR and point-intercept (P I) field sampling ..................................................................................... 432-7 Examples of spatial variation (empirical variogram) of fuelbed heights found within three 4 m x 4 m plots.......................................................................................................... 463-1 Fine-scale spatial heterogeneity of A) fuelbed structure (depth, m) and B) fire behavior (temperature in C) recorded within the longleaf pi ne-wiregrass system...........543-2 Example outputs of the A) ground-base d LIDAR (using the MTLS) and the B) thermal camera (using the FLIR) for a 4 m x 4 m forest fuelbed plot. .............................. 623-3 Distribution of grouped plots by each standardized mean fi re behavior variable used in the k-means cluster analysis .......................................................................................... 663-4 Distribution of grouped plots (k-means cl ustering) in relation to standardized maximum temperature ( C, Tmax) and residence time .................................................... 673-5 Regression-tree results using CART (C lassification and Regression Tree) for Q90 and T300 ............................................................................................................................713-6 Changes in fuelbed continuity among the three plots (4 m x 4 m) modeled individually .................................................................................................................. ......84
11 4-1 Conceptual diagram of longleaf pi ne-hardwood population dynam ics and LLM (Longleaf Model) components. ........................................................................................1064-2 Simulated growth curves for A) longleaf pine and B) hardwoods in two study sites (Ordway and Ichauway) using the LLM. ......................................................................... 1104-3 Spatial outputs from the LLM during a frequent fire regime ..........................................1234-4 Cumulative distribution functions of mean longleaf cone production over forty years. .1254-5 Simulated size-specific relative mortal ity of longleaf pines from competition, fire, and other natural causes ...................................................................................................1264-6 Simulated relative mortality of hardw oods from fire and other natural causes ............... 1284-7 Changes in relative abundance of A) la rger (> 10 m) and B) smaller (< 10 m) longleaf pine (LP) and hardwood (HW) tr ees across various fire probabilities ..............1304-8 Ichauway height frequency (%) distributions simulated by the LLM and estimated from field and LIDAR (Light Detection and Ranging) measurements ...........................135
12 Abstract of Dissertation Pres ented to the Graduate School of the University of Florida in Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy LINKING PLANT DEMOGRAPHY, FOREST FU ELS, AND FIRE IN LONGLEAF PINE ( PINUS PALUSTRIS) SAVANNAS USING LIDAR REMO TE SENSING AND SIMULATION MODELING By Eva Louise Loudermilk May 2010 Chair: Wendell P. Cropper Jr. Major: Interdisciplinary Ecology Fire is a dominant disturbance for many forested ecosystems and is driven by the vegetative fuels that provide th e combustible energy needed to carry a fire. Longleaf pine ( Pinus palustris ) savannas of the southeastern United States are among the most fire dependent on earth and when frequently burned provide the necessary understory surface fuelbed to sustain fire on short (1-5 year) return interval s. The relationship between fuel s, fire, and plant dynamics in these savannas is intricate, but is strongly regulated by the overstory structure through its control on fine-scale fuel variation. The relationship of within-fuelbed and fire heterogeneity has received little attention. The impacts that the fire regime (e.g., frequency, intensity) has on positive and negative feedbacks associated with plant community dynamics (i.e., competition, germination) has not been fully investigated as well. In this study, a remote sensing instrument namely ground-based LIDAR (Light Detection and Ranging) was used in a novel way to meas ure precise volume of individual plants and fuelbed plots at the sub-meter scale that allows for a more fundamental assessment of feedbacks between fuels, fire, and forest dynamics. The high resolution (< cm) and three-dimensional outputs of the LIDAR data provided information on plant structure not previously available. The
13 ground LIDAR fuel data were used in conjunction with in situ fuel data to predict fine-scale fire behavior using non-linear regr ession-tree analysis. A longleaf pine hardwood simulation model was created to link population level tree dynamics with fuel characteristics and stochastic fire regimes. This is the first known modeling work to simulate interactions between longleaf pine and ha rdwoods. Spatial components included seed dispersal (including pine seed ma sting), clonal spreading (for hardwoods), fuel dispersal and distribution, and competition from neighboring trees. Evaluations with in situ data were promising for two modeled longleaf pine site s. Tree distributions and community stability were examined with varying fire frequency. The model was especially useful in identifying scientific knowledge gaps associated with plant competition and facilitation, especially in relation to hardwood demography. The model u ltimately provided a foundation for studying fuel and fire heterogeneity infl uences on population dynamics.
14 CHAPTER 1 INTRODUCTION Longleaf Pine Savannas: Fuels, Fire, and Plant Demography Fire and the genus Pinus are inextricably linked ov er space and tim e (Agee 1998). Longleaf pine ( Pinus palustris Miller) forests of th e southeastern United States represent an archetype of a fire dependent ecosystem. These forests have among the shortest fire return intervals of any in the world, ty pically recurring every one to three years, occasionally extending to ten years on xeric sites (C hristensen 1981, 1988, Abrahamson and Hartnett 1990, Ware et al. 1993). Moreover, the understory of longleaf comm unities is extremely diverse (Kirkman et al. 2004), and this diversity is tightly coupled to frequent fire (M itchell et al. 2006). High fire frequency in longleaf pine forest s structures the vegetation and the fuels, perpetuating a low-intensity, surface fire regime w ith fuel beds dominated by pine needles, small shrubs, grasses, forbs and dead woody debris (O ttmar et al. 2003). Longleaf pine needles are especially important, providing an id eal fine fuel for frequent fire due to their high resin content and physical structure (Hendricks et al. 2002). Frequent fires suppress hardwoods, killing the aboveground portion of the plant (Williamson and Black 1981) and maintaining these species in low stature as understory shrubs (Jacqmain et al. 1999). Frequent fire maintains grasses and pines that, in turn, produce the fine fuels necessa ry for the low intensity regime. With decreased fire frequency or insufficient intensity, a woody midstory rapidly develops, dominated by fire sensitive trees and shrubs (Glitzenstein et al. 1995). Spatial distribution of the overstory influences fuels and fire behavior. Gaps in the overstory may create patches with in sufficient fine fuel for fire spre ad. This effect is exacerbated in larger gaps because pine needle accumulation declines exponentially with distance from a tree bole (Grace and Platt 1995). Moreover, fire freq uency and intensity are further reduced due to
15 the simultaneous increase in hardwood litter (W illiamson and Black 1981) and declines in pine/grass fuel loads (McGuire et al. 2001). Fi re intensity has been shown to be positively correlated to pine litter accumulation and negati vely correlated to hardwood litter (Williamson and Black 1981, Glitzenstein et al 1995, Ferguson et al. 2002). Increased sprouting of hardwoods coupled with decreased fire intensit y, results in a higher likelihood of hardwood survival in gaps. Surviv ing a single burn gives hardwoods a significant competitive advantage over pines, resulting in greater height growth, leaf area, and the development of thicker bark, whic h increases the probabi lity of surviving future fires. Pine seedlings are also more likely to establish in th ese forest gaps, where they experience less intense fires and increased light availa bility (McGuire et al. 2001, Broc kway et al. 2006, Pecot et al. 2007). With decreasing fire frequency, pine seedling survival might increase (Beckage and Stout 2000), but the seedlings are more likely to experience competition from faster growing hardwoods (Rebertus et al. 1989, Pecot et al. 2007). These dynamics of within-stand structural heterogeneity and their connection with fuels, fire behavior, and fire effect s have been difficult to conceptualize (Figure 1-1) and have yet to be fully investigated. Fuel and Fire Heterogeneity Fire regim es can be characterized by four ma in features: frequency, severity, seasonality, and heterogeneity (Gill 1975). Burn heteroge neity has received rela tively little attention, particularly at small spatial scales in freque ntly-burned, low intensit y fire regimes. Fire heterogeneity however, has been shown to be criti cal in regulating ecologi cal functions (nutrient cycles) and structure (forest dynamics and divers ity) where it has been studied, e.g. burned and unburned patch mosaics in intermediate or catastr ophic severity fire regimes (Clark et al. 2002). While patchiness (burned vs. unburned) is importa nt ecologically in ma ny fire regimes,
16 Figure 1-1. Conceptual diagra m of longleaf pine-hardwood popul ation dynamics and feedbacks associated with fuels and fire. variability in fire behavior with in burned areas has recently been investigated (Hiers et al. 2009) and is likely critical for pred icting post fire ecological effect s when species possess adaptations to survive fires. The coupli ng between the combustion envir onment of fire and vegetation response may be linked by understanding the role of fuels, which has been described as the ecology of fuels (Mitchell et al. 2009). This concept emphasizes the significance of fine-fuels and quantifying fuels and fire behavi or at the appropriate scale. What has historically been characterized as a chronic low-intensity surface fire regime burning through a homogenous fuelbed (Wade et al. 2000) has been found to vary highly at finescales (Figure 1-2) and in multiple dimensions (Hiers et al. 2009, Loudermilk et al. 2009). Fuelbeds (i.e., understory vegetation or surface fuels that carry fire) have ra rely been studied at
17 these scales, but variation at this scale likely drives fire intensity that in turn determines oak and pine demographics in the seedli ng and sapling size classes. Figure 1-2. Fine-scale spatial he terogeneity of A) fuelbed stru cture (depth, m) and B) fire behavior (temperature C) recorded within a longleaf pine-wiregrass system. These semivariograms were created from in situ data on fuelbed depth and fire temperature in 4 m x 4 m plots captured from LIDAR (Light Detection and Ranging) and thermal imaging remote sensing inst rumentation, respectively. The distribution and arrangement of fuels represent within-f uelbed variation, each with distinct bulk density, height, and composition which produce fine -scale burn heterogeneity. A synergy among the different types of fine fuel components of the fuelbed seems to exist, with pine needles interacting with grasses, forbs, and shrubs to form a horizontally continuous fuelbed with highly varying three dimensional stru ctures (Kirkman and Mitchell 2006). The structure of these fuels (or fuelbed itself) has been difficult to measure, due to the complex architecture of plants (e.g., spatial configuration, leaf area, size) and limited measurement methods (e.g., planar transects) (V an Wagner 1968). Advances in remote sensing techniques for measuring the complex fuel (using ground-LIDAR Light Detection and Ranging) and fire environment (using thermal im aging) provide for the opportunity to examine spatial fuelbed structure and fire behavior attributes as well as their potentia l relationships at the 0.0 0.1 0.2 0.00.51.01.52.0Semivariance of fuelbed depthDistance (m) 0 1000 2000 3000 4000 5000 6000 7000 0.00.51.01.52.0Semivariance of fire temperatureDistance (m) A B
18 sub-meter level. The fine-scale interactions between plants and this fire heterogeneity are especially important and poorly studied in th ese savannas. Unders tanding fuel, fire, and vegetation feedbacks is needed to predict forest dynamics. Modeling Longleaf Savanna Demography Fire effects on pine and hardwood dynam ics are influenced by spatial ecosystem structure and processes (e.g., fine-scale fuel variation, adult pine distribution, fire frequency and behavior), which occur at multiple scales. Modeling interactions between longleaf pine, hardwoods, and fire helps to test our understandi ng of the system and provides an opportunity to discover gaps in our current knowledge of these unique savannas. Several demographic modeling approaches have been applied to longlea f pine systems that incorporate some of these elements (Platt et al. 1988, Kaiser 1996). A limitati on of previous models is that they assume a certain level of intraand interspecies comp etition and spatial and temporal homogeneity. Longleaf pine systems however, can be very sensitive to changes across the landscape and through time, especially related to fire frequency. A spatially-explicit forest modeling scheme can significantly improve model realism and has been successful in many pl ant systems (Higgins et al. 200 1, He et al. 2004, Miller 2005). Several studies and modeling techniques that targ et spatial dynamics have been applied to longleaf pine and its biological community (Rathbun and Cressie 1994, Drake and Weishampel 2001, Loudermilk 2005, Loudermilk and Cropper 2007). An important missing component of these models is oak or hardwood dynamics. The modeling literature on southern hardwoods (found in longleaf pine savannas) is sparse and few have attempted to model their demogr aphy (Rebertus et al. 1989, Berg and Hamrick 1994), especially with explicit spatial and temporal in teractions with l ongleaf pine trees (Loudermilk and Cropper 2007). This is surprising, as their importance for shaping community
19 structure and interactions with fire have been recognized for some time (Heyward 1939, Wahlenberg 1946, McGinty and Christy 1977, Williamson and Black 1981). Study Overview This study focused on closely exam ining plantlevel and fuelbed structural characteristics in three dimensions, investigati ng the predictive power of these structural attributes with finescale fire behavior heterogeneit y, as well as developing a model of population level interactions of fuels, fire, and plant demographics. The sp ecific objectives for each chapter are as follows. Measuring Fuel Structure The objectives of this chapte r (2) were to describe a gr ound-based LIDAR approach to m easure fuel volume and loading. LIDAR was firs t used to measure indi vidual shrubs of two common species with contrasting life-forms found in pine flatw oods of the southeastern US coastal plain. Specifically, the hypothesis tested was that LIDAR volumes would be significantly less than those obtai ned from traditional means. Correlations of LIDAR volume measurement with mass and leaf area were ex amined. LIDAR characterization of complex fuelbeds, composed of many species of shrubs a nd herbaceous plants in the field, was examined. LIDAR measurements of fuelbed heterogeneity we re compared with traditional point intercept measures. This chapter provided the background necessary for using ground LIDAR to measure sub-cm three-dimensional fuel structure precisely a nd describe its implications for fine-scale fire research and inherent plant response. Examining Fuel and Fire Relationships The objectiv e of this chapter (3) was to exam ine the fine-scale relationships of fuelbed structure (from ground-LIDAR ) and fuel types (from in situ data) to fire behavior characteristics (from digital thermal imagery) across multiple forest fuelbed plots within a longleaf pinewiregrass ecosystem. More sp ecifically, the pred ictive power of various ground-LIDAR fuel
20 metrics coupled with fuel type data was analyz ed for determining surface fire temperature and residence time at sub-meter scales using regressi on-tree modeling. Relationships were examined at various grouped-plot levels and within repres entative individual plots. Another objective was to assess potential technical and modeling pitfalls as well as determine the implications for assessing fine-scale fire effects. This chapter integrated the id ea of measuring fine-scale fuelbed characteristics (i.e., chapter 2) with measurements of fire beha vior (e.g. using thermal imagery) and exemplified the importance fo r understanding this within-fire heterogeneity and implications for modeling forest dynamics. Modeling Savanna Dynamics The objective of this chapte r (4) was to develop a longleaf pine-hardwood stochastic sim ulation model, incorporati ng tree demography, plant competition, and fuel and fire characteristics with spatially and temporally exp licit components. Literature and data from two study sites were used to develop and evaluate the model with the goal to in corporate site specific calibration parameters for overall model versatilit y. Aerial LIDAR data and field measurements of population densities and height distributions were used for m odel evaluation. This model was used to identify scientific knowledge gaps of various population level ecosystem processes, specifically related to population structure, fu el and fire heterogene ity, hardwood demography, and plant competition and facilitation. This chap ter was the first to integrate the spatial and temporal dynamics of longleaf pine and ha rdwoods across fire regimes and provides a foundation for integrating fuel and fire heterogene ity characterist ics (i.e., chapter 2 and 3) for examining vegetation response a nd feedbacks across scales. Study and Management Implications The m anagement and restoration of natural l ongleaf pine forests re quires sophisticated understanding of the consequences and tradeoffs of applying techni ques such as harvest and fire
21 management. This study was intended to extend our understanding of fire and plant community interactions and to provide management insi ghts. The implications of this study include: Providing improved measurement techni ques (using ground-LIDAR) of vegetative fuelbed structural properties used for w ildland fire behavior prediction system software. Further understanding the relationship betw een fine-scale fuelbed heterogeneity and fire behavior, supporting the concept of the ecology of fuels and the wildland fuel cell concept. These relationships may be extende d to fire effects (e.g., understory plant mortality and divers ity) research at the same scale. Better understanding fuel, fire, and plant dynamics, especially related to longleaf pine and hardwoods demography. This in cludes the ability to assess spatial and temporal forest gap dynamics associated with hardwood encroachment, pine recruitment, and change in forest fuels. Better understanding of how a changing fire regime may impact longleaf pine ecosystem stability. Creating a modeling framework that may be expanded upon to implement ecological forestry silvicu ltural practices (e.g., singletree selection), ecosystem restoration and fire management prescr iptions, and impacts from catastrophic disturbances (e.g., hurricanes, wildfires). This could facil itate the understanding between dramatic changes to stand struct ure and ecosystem function to ultimately enhance long-term forest management planning
22 Creating a basis for future work to enha nce the understanding of the impacts of changing overstory structure on fine-scale fuel heterogeneity, fire behavior, and ultimately plant dynamics at multiple scales.
23 CHAPTER 2 GROUND BASED LIDAR (LIGHT DETECTI ON AND RANGING): A NOVEL APPROACH TO QUANT IFY FINE SCALE FUELBED CHARACTERISTICS Introduction Com ponents of fire behavior1, including ignition properties, rates of spread, and intensity are influenced by fuel loading, fuel depth, and thus density (DeBano et al. 1998). Fuel volume and loading are not only important in empirically understanding fire behavior and fire effects, but are drivers of models used to simulate fire behavior (Andrews and Queen 2001, Scott and Burgan 2005). Surface fuelbed characteristics have traditionally been quantified by both direct and indirect methods. Direct measurements commonly employed are tallies of down woody fuels along planar transects (Bro wn 1974), and coupled destructiv e biomass sampling, (i.e., clip plots) (Brown 1981). Indirect methods include visual cover estimates in plots or comparisons with photographs of known fuel loads/types (O ttmar et al. 2003, Keane and Dickinson 2007). These methods allow for stand level estimation of variables, such as fuel load, bulk density, and packing ratios that are then used to predict fire behavior (Reinhardt and Keane 1998, Andrews et al. 2004). Each method has significant limitations. Dir ect sampling is labor intensive often limiting sample size. Some techniques are not appropriate for all fuel types: Planar transects are not efficient at estimating fine fuels such as grasses. Indirect measures can be subjective, resulting in biased estimates. Furthermore, estimati ng volume for bulk density calculations relies on unrealistic simplifications; shrub or grass volumes are calculated by assuming the plants form simple geometric shapes such as a spheroid or cylinder ((Van Wagner 1968) Figure 2-1A). Such 1 Portions of the contents in the Chap ter have been published in the journal International Journal of Wildland Fire under the title, Ground-based LIDAR: a nove l approach to quantify fine-scale fuelbed characteristics (Loudermilk et al. 2009)
24 traditional volume measurement techniques ignore complex plant architecture. While these approaches are recognized to have inherent limitations, no alternativ es were previously available. Figure 2-1. Output 3D (thr ee-dimensional) point-clouds from the ground LIDAR (Light Detection and Ranging) system for: A) an individual saw palmetto shrub, and B) a 4 m x 4 m plot in a longleaf pine savanna fu elbed (max fuelbed height in plot: 2 m). Note that fuel volume of a plant is comm only measured by assuming a cylindrical or spheroid geometry (A). The color gradient represents height va riations, which aids in visualization of the 3D vegetation structure. Recent advances in laser rangi ng, or LIDAR (Light Detection and Ranging), technologies have enabled the successful m easurement of complex structures in the field with both high accuracy and precision (Hopkinson et al. 2004). LIDAR has been typically used in forestry for coarse scale remote sensing of forest canopy st ructures, estimating tree height distributions, canopy bulk density, and leaf area (Lefsky et al. 1999, Drake and Weishampel 2000, Hall et al. 2005, Roberts et al. 2005, Lee et al. 2009). Th ese airborne LIDAR approaches have been unsuited, however, in measuring understory vegetation, primarily because of the obstruction from the forest canopy and a horiz ontal resolution limited to the few decimeter scale. Modern ground-based LIDAR systems now have some of the strengths of airborne systems (in particular laser pulse rates of a few thousand Hz or more that enable high-resolution sampling), but they can attain sub-centimeter resolution and be positioned under the canopy to reduce the shadowing B 4 m 4 m 1 m 1.25 m A
25 effects of overstory tr ees (Slatton et al. 2004). The high-dens ity three-dimensional (3D) point data (more than 10,000 points per m2) obtained from such systems provide the precision needed to characterize fuelbeds, particularly by quantif ying fuel height distri butions and thus fuel volumes. One system in particular, the Mobile Terrestri al Laser Scanner (MTLS), is a static, stopand-scan laser scanner that covers a limited area, but captures data at the sub-cm level. The MTLS consists of Optechs ILRIS 36D (Intelligent Laser Ranging and Imaging System) ground based laser scanner (Lichti et al. 2002, Frhlich and Mettenleiter 2004), which was mounted on a lift atop a mobile platform (4 x 4 truck) by the National Center for Airborne Laser Mapping (NCALM) at the University of Florida. The M TLS is versatile in capturing details about the terrain at multiple angles. The ability to vary the LIDAR height and pointing angle using the MTLS allows significant reducti on in shadowing effects that may be found when using a LIDAR system on a tripod. Although ground-based LIDAR (also called terr estrial LIDAR) is a relatively new technique, there are several good examples of fo restry applications. Accurate tree and canopy metrics (e.g., timber volume, tree height and diam eter, gap fraction) have been successfully estimated using tripod mounted terrestrial LIDAR systems (Hopkinson et al. 2004, Watt and Donoghue 2005, Henning and Radtke 2006). Fine-scale leaf area (Lovell et al. 2003, Tanaka et al. 2004) and gap fraction (Danson et al. 2007) estimates of tree ca nopies have been correlated to that of hemispherical photographs. Small indi vidual trees were intensely measured using a voxel-based approach to determin e leaf area density at the mm3 scale (Hosoi and Omasa 2006). Ground-based LIDAR has also been used to unde rstand the impacts of instrument positioning on shadowing effects that influe nce tree canopy measurements (Van der Zande et al. 2006).
26 Although most terrestrial LIDAR systems used in forestry are stati onary, a portable LIDAR system has been developed to record 1 m scale canopy measurements while moving through a forest (Parker et al. 2004). To my knowledge, ground-based systems have yet to be reported as a means of surface fuel characterization. The objective of this study wa s to describe a ground-based LIDAR approach to measure fuel volume and loading. LIDAR was first used to measure individual shrubs of two common species with contrasting life-forms found in pine flatwoods of the southeastern US coastal plain. Specifically, the hypothesis tested was that LIDAR volumes would be significantly less than those obtained from traditional means. Correla tions of LIDAR volume measurement with mass and leaf area were examined. LIDAR characteri zation of complex fuelbeds, composed of many species of shrubs and herbaceous plants in th e field, was examined. LIDAR measurements of fuelbed heterogeneity were compared with traditional point intercept measures. Methods Ground-LIDAR Instrumentation The ILRIS ground-based LIDAR system uses a 1535 nm wavelength (near-infrared) laser with a pulse frequency of 2,000 Hz (or points per s econd), recording first or last returns of each laser pulse (user defined). The maximum field of view is 40 in both horizontal and vertical plane, although smaller fields of view can be sp ecified for a given scan. It can register laser returns from as little as 5 m away and out to a distance of 1500 m (at 80% target reflectivity). The particular ILRIS used in this work has a pan-tilt base on the MTLS, which allows for a 360 rotation in the horizonta l plane and roughly in the vertical plane (T able 2-1). The lift on the MTLS provides for a vertical adju stment of the scanner up to a he ight of about 9 m. Recording laser data from multiple positions around the target clearly will reduce information lost from shadowing effects, but an efficient imaging geom etry is desired to minimize the number of scans
27 required to sample each plot. The ability to adju st the instrument vertically and horizontally to the extent offered by the MTLS allows the user to achieve advantage ous scan positions and orientations in spite of constraints posed by th e terrain or nearby occl uding trees, which is essential to measuring precise fine-scale attributes of intricate plant stru ctures (Hosoi and Omasa 2006). Table 2-1. Manufacturer specifications of th e ground-LIDAR (Light De tection and Ranging) instrument (Optechs ILRIS 36D) used in this study. Note that the beam divergence value corresponds to a circular footprint diameter of 4 mm at an average range of 25 m. IFOV: Instantaneous Field of View, Aux FOV: Auxiliary Field of View, which refers to the range of motion of the t ilt and pan mounting for the LIDAR sensor. Texture refers to the radiometric values, which include a relati ve intensity of the return laser pulse and multispectral (RGB : Red Green Blue) values from a comounted digital camera. Specification type Specification value Range [m] 3-1500 @ 80% reflectance 3-800 @ 20% reflectance 3-350 @ 4% reflectance Range resolution [mm] 4 Azimuth, elevation resolution  0.00115 IFOV [Vertical Horizontal] 40 40 Aux FOV [Vertical Horizontal] -20 to 90 360 Laser type/color Infrared Laser wavelength [nm] 1,500 Scan rate [points/s or Hz] 2,000 Beam divergence  0.00974 Texture Intensity and RGB Weight [kg] 12 Dimensions (LxWxH) [cm] 32 32 22 Power supply 24 V DC Power consumption [Watts] 75 Point spacing in LIDAR point clouds will vary locally depending on the range from the laser to objects in the field of view and on the angular spacing of laser shots. For the ILRIS sensor, the pulse rate is fixed at 2000 Hz. So it is the angular separation of laser pulses that is used to achieve a desired point density, which is input via the ILRIS Controller software. To achieve this, the ILRIS pre-scans the user specified fi eld of view at a course resolution and
28 acquires an average range. The user then ente rs a desired linear point spacing into the ILRIS Controller software. This spac ing corresponds to the average lin ear separation between points on a hypothetical plane segment that spans the user sp ecified field of view, is located at the mean range, and is orthogonal to the sensors boresight Given this separation distance and the mean range, the necessary angular shot spacing is au tomatically calculated by the ILRIS Controller software. Under typical operati ng conditions, the linear point spacing is chosen to be anywhere from a few cm to 1 mm. High point densities come at the cost of increased memory to hold the data and increased time to complete the scan. For example, a 20 20 field of view centered at boresight with a mean range of 15 m and user sp ecified linear point spacing of 5 mm will yield an angular separation of 0.0189 require roughly 9 min to comple te, and result in just over 1.1 million points. The ILRIS collects: 1) x, y, z coordinate valu es with respect to the position of the laser sensor using a near-infrared wavelength (i.e., 1535 nm) laser or light pulse; 2) intensity values of the return; and 3) true color (Red Green Blue) va lues for each point obtained from an integrated and calibrated digital camera within the instru ment. A more complete description of the technology can be found in (Lich ti et al. 2002, Frhlich and Metten leiter 2004). Once an area is sampled by two or more scans, the point clouds from each scan are merged into a common coordinate frame using software that can accommoda te 3D spatial data (see Data Processing and Analysis). Inside the software environment, th e subset of the point cl oud that covers the region of interest can be isolated and the remaini ng points cropped out to minimize the computational burden and obtain statistics representative of the precise region of interest. Individual Shrub Assessment In May 2007, ground-LIDAR measurem ents were collected for two common southeastern US shrub species: saw palmetto ( Serenoa repens ) and wax myrtle (Myrica cerifera ). These
29 species are highly flammable a nd important wildland fuels (Wad e and Lunsford 1989). Using the ILRIS, twelve individually potted shrubs in two size classes (i.e., 0.5 m and 1 m in height, six for each species) were scanned within an enclosed building at the University of Florida. This provided an ideal setting for la ser-scanning with flat ground and minimal wind disturbance. Six plants were scanned at a time; using reference targets for s ubsequent LIDAR data processing (see Field Plot Data Collection). Three scans were taken per set of plants (6 scans total). After LIDAR acquisition, volume was recorded usi ng traditional field methods, by calculating geometric (cylinder and spheroid ) volumes using height and diameter measurements. To more fully analyze the structural variation, the shrubs were cut into three equally spaced vertical sections (or thirds) and measured leaf area using a LI-COR leaf area analyzer for each section. Biomass was dried and weighed for each section. The shrub LIDAR point clouds were also divided into thirds for volume estimation (see Data Processing and Analysis) to compare with the biomass and leaf area measurements. Field Assessment of Complex Fuelbeds Study site: Ichauway preserve The in-field portion of the st udy was perform ed at Ichauwa y, an 11,000 ha reserve of the Jones Ecological Research Center in southwestern Georgia, USA. Ichauway is located within the Plains and Wiregrass Plains subsections of the Lower Coastal Plain and Flatwoods section (McNab and Avers 1994). Ichauway has an extens ive tract of second-growth longleaf pine and has been managed with low intensity, dormant-seas on prescribed fires for at least 70 years, at a frequency of one to three years. The understory of the study area was primarily composed of wiregrass ( Aristida stricta ), many forb and prairie grass spec ies, as well as inter-dispersed hardwood shrubs (e.g., Diospyros spp., Prunus spp, Quercus spp., Sassafras albidum ). With
30 frequent fires, hardwoods are generally maintain ed at shrub size, occasionally reaching mature size. The specific study area ha d not burned for one year. Field plot data collection In spring 2007, a total of 26 geo-referenced 4 m x 4 m plots were established to measure fuelbed characteristics. The 4 m x 4 m area was chosen because it was large enough to capture heterogeneity at sub-meter scales, but small enough to support intensive sampling with minimal impact to the vegetation. A ladder was suspe nded horizontally across the plot to sample the interior with minimal disturbanc e to the vegetation. Spatially explicit point-intercept (PI) fuel data (169 samples, 0.33 m grid spacing) were r ecorded using a 5 mm graduated dowel within each plot. Only vegetation in physical contact with the dowel was measured. At each PI sample point, maximum fuelbed and litter depth (or height), as well as presence/absence of fuel and vegetation types were recorded. The geo-referencing and sampling intensity were used to capture the spatial variation of the fuelbed found within this small (16 m2) area and to relate to the sub-centimeter scale 3D laser data collected from the ground-based LIDAR. Within two weeks of field data collection, the MTLS collected ground-LIDAR data on all 26 plots. Prior to laser data collection, referen ce targets (consisting of a styrofoam ball on top of a metal rod, 0.5 1 m high) were placed at all fo ur corners of the plot. A double reference target (two Styrofoam balls on one metal rod) was used at the northwest corner of each plot to orient the plot for data processing. The MTLS was rest ricted to mapped roads and trails, as well as a buffer of 5 meters around each plot, to reduce vegetation disturbance. Th e ILRIS was lifted to a height of 7 m to capture a more aerial view of the plot to reduce shadowing effects and positioned to avoid tree bole or canopy obstruction. The ILRIS was set to a downward angle tilt of 25 from horizontal. A true color digital photograph was taken by the ILRIS for each plot, and used in the field to delineate the precise field of view for each scan. First-return laser pulses were
31 recorded with an input mean point spacing of 5 mm. Two scans were taken of each plot, from opposite sides, to mitigate shadowing effects and ensure more accurate and complete subcentimeter scale data for fuel plots. These two scans were merged in the processing stage to a single 3D point cloud data set. Data collecti on with the MTLS took approximately 20 minutes per plot. In comparison, field PI data collect ion lasted an average of two hours per plot. Data Processing and Analysis Initially, data processing involved converting th e collected laser data from binary to ASCII for mat. These raw data include a four column text file containing 3D orthogonal coordinates (x, y, z) and laser return intensity values for each of the sampled laser points. For this work, the Quick Terrain Modeler (Applied Imagery) and TerraScan (Terrasolid) software packages were used for processing the laser data, although the PolyWorks software package (InnovMetric) that is sold with th e ILRIS would also have been suitable. Of the two scans taken per plot, the first was horizontally rectified by compensating for th e original scanning geometry, since the instrument had a downward tilt of 25 w ith respect to horizontal. The reference targets were used to identify common points between th e two scans and create a common 3D coordinate frame for both scans (see (Wolf and Ghilani 1997) for details). The conformal transformation was then applied to the second s can to bring it to the same coor dinate frame of the first scan, combining them into a single spatially coherent da ta set. The merged data set was then rotated about the z-axis to orient the po int cloud in cardinal space. A constant value offset was added to the height values in order to set the ground point s to a value of zero. The digital image and the double target reference for the north west corner of the plot were especially helpful in this merging process and in orienting the plot in cardinal space. Si milar procedures were performed for the individual shrub data, using the top of the pot for ground reference. The individual shrubs (Figure 2-1A) and 4 m x 4 m plot areas (Figure 2-1B) we re clipped from the resulting
32 merged scans using the reference targets. Roughly 600,000 to 700,000 sample laser points were found within each 4 m x 4 m plot. Point-densitie s, volume estimates, and height distributions were measured using the merged LIDAR scans of each 4 m x 4 m plot and individual shrubs. It should be noted that when ranging to objects distributed in three dimensions, the actual 3D point spacing will necessarily be larger than the 2D spacing input by the user into the ILRIS Controller software for an equivalent range because of the added depth component. For example, the specification of a 5 mm linear spacing at the mean range might seem to suggest an average point spacing of 2.5 mm when two similar scans from different vi ewing directions are merged. But the actual 3D Euclidian separati on distance between points in all such merged scans for the forest plots was found to be roughly 1 cm on average, with a roughly 5 mm standard deviation. Volume was calculated using the LIDAR laser data as well as field measurements. LIDAR volume estimates were calculated in two ways. First, volume was calculated by determining the presence or absence of la ser points within each cm3 space (similar to the voxel approach in Hosoi and Osama 2006 and Van der Zande et al. 2006) for each whole potte d shrub and the three equally spaced vertical sections (or thirds) of each shrub. The process involved using a 1 cm3 3D window to move through each point cloud in the hor izontal and vertical directions, respectively. Every time a point(s) was found in the 3D window, 1 cm3 of volume was added to the sum for that shrub or section of shrub. Secondly, a surface plot (continuous rendering or TIN) was created using traditional Kriging techniques for both the LIDAR and PI grid field data (4 m x 4 m plots only) using fuel depth (h eight) values. The total volume (m3) found underneath this surface was calculated for each LIDAR and PI 4 m x 4 m dataset. Volume of the whole potted
33 shrubs was also calculated using field measurements (height and diameter) and applying common volumetric formulas for both a cylinder and spheroid. A paired one-way t-test was us ed to assess differences betw een whole-plant shrub volumes calculated by LIDAR (LIDAR volume) and by traditional field methods (cylindrical and spheroid volume). The hypothesis tested was th at traditional field methods would overestimate individual shrub volume compared to LIDAR and that the cylindrical vol ume would be larger than the spheroid volume estimates. An ANOVA (analysis of variance) was used to assess if volume estimates of the whole shrubs differed when affected by the following three factors: 1) volume estimation technique (LIDAR and traditiona l field methods: cylinder and spheroid); 2) shrub species (saw palmetto and wax myrtle); and 3) size of plan t (large and small). Additional ANOVA tests were used to assess the effects of size, species, and thei r interactions on whole shrub biomass, leaf area, LIDAR volume, and volume estimated by traditional field methods. ANOVA was used to assess the differences in LIDAR volume, biomass, and leaf area estimates of the sectional shrub data with shrub volume split into thirds using three treatments: 1) shrub species (saw palmetto and wax myrtle); 2) bottom, middle, top third section of plant; and 3) shrub size (large, small). Least-squares simple linear regression te sted the relationships among LIDAR volume, biomass, and leaf area of the sectional shrub data as well as the individual whole shrubs. A paired two tailed t-test was used to compare the 4 m x 4 m plot LIDAR volume estimates and PI volume. Regression analy zed how LIDAR and PI volume estimates varied over a range of fuelbed volumes. All statistical assumptions were met and Type I error was set at 0.05 for all tests. The dependent variables (i.e., leaf area and biomass) were transformed using the square root (sqrt) function for shrub regre ssion analysis. The coefficient of determination
34 (R2) and root mean square error (RMSE) were used to assess regression m odel fit. To analyze the spatial variability within the 4 m x 4 m pl ots, empirical variograms of the PI and LIDAR datasets were created. The Surfer 8 program (G olden Software Inc.) was used to create the variograms with the following se ttings: omni-directional lag tolerance, maximum lag tolerance of 3 m, 25 lags, and a 0.4 m maximu m lag width for plot smoothing. Results and Discussion Ground-based LIDAR was able to capture precisely defined volum es of fuels at both the individual shrub scale and within complex he rbaceous fuelbeds (Figure 2-1). While ground based LIDAR has been used to accurately define the volume of overstory trees (Hopkinson et al. 2004), this is the first application of ground-based LIDAR aime d downward to assess fuel characteristics. There were discrepancies found between the three vo lumetric measurement techniques of individual shrubs as well as differences associated with plant size and species (Table 2-2). Compared to LIDAR, both trad itional methods of measuring volume (i.e., Table 2-2. ANOVA (analysis of variance) output analyzing individual whole shrub volume (m3) estimates with three treatments: A) vol ume estimation method (LIDAR, cylinder, spheroid), B) plant size (large, small) and C) shrub species (wax myrtle, saw palmetto) Source term DF Sum of squares Mean square F-ratio p Power ( = 0.05) A: method 2 0.1557 0.0779 75.59 0.0000* 1.00 B: size 1 0.2549 0.2549 247.48 0.0000* 1.00 AB 2 0.1109 0.0554 53.82 0.0000* 1.00 C: species 1 0.0401 0.0401 38.98 0.0000* 1.00 AC 2 0.0220 0.0110 10.67 0.0005* 0.97 BC 1 0.0283 0.0283 27.44 0.0000* 0.99 ABC 2 0.0155 0.0077 752.00 0.0029* 0.91 S 24 0.0247 0.0010 Total (adjusted) 35 0.6521 Total 36 Term significant at = 0.05, DF = degrees of freedom, p = p-value cylindrical and spheroid calculati ons) of individual shrubs were si gnificantly larger than LIDAR estimates (Figure 2-2; p < 0.0038, n = 12). Not surprisingly, th e cylindrical measurements were
35 larger than the spheroid measurements ( p = 0.0029, n = 12). The extent of the discrepancy between LIDAR and field measurements varied with species and size because of variation in the distribution of leaf area. Larger plant sizes resulted in greate r differences of volumes between sampling methods (Figure 2-2). LIDAR shrub volume did not significantly vary by species. Species was significant for traditional volume methods, biomass, and leaf area ( p < 0.006, n = 12). Figure 2-2. Variation of volume for two plant spec ies (saw palmetto and wax myrtle), two sizes (large and small) and the methods used to calculate volume (LIDAR vs. traditional sampling: cylinder and spheroid). Note the broken y-axis to see smaller differences in LIDAR volume estimates. Additionally, LIDAR volumes of individual whole shrubs were more strongly correlated to leaf area and biomass than traditionally estimated volumes (Table 2-3, Figure 2-3). The difference between methods was especially signi ficant when relating vo lumes to leaf area [R2 = 0.81 vs. 0.41]. Biomass and volume relationships were strong (R2 = 0.86, 0.96) regardless of method, although LIDAR volumes had the strongest relationship (R2 = 0.96). LIDAR volume
36 (R2 = 0.81) was more representa tive of leaf area than the traditional methods (R2 = 0.41) at this scale because LIDAR was better able to capture the spatial dimension of the plant leaves that characterizes leaf area. Furthermore, the LIDAR was able to measure the vertical arrangement of the plant, where much of the variation of plant structure occurs (Figure 2-1, 2-4). The cylinder and sphere volumes were most likely ove r-simplified representations of true plant leaf area and architecture, especially alo ng the vertical axis (Figure 2-1). Table 2-3. Linear regression relating three vol ume (m3) estimation methods (LIDAR and two traditional methods assuming cylinder and sphe re shape) to biomass to biomass (g) and leaf area (m2) from individual whole shr ubs (n = 12). Model used: f (x) = bo + b1(volume) Dependent variable Independent variable R2 RMSE Sqrt(leaf area) LIDAR volume 0.81 11.33 Sqrt(leaf area) Cylinder volume 0.41 19.9 Sqrt(leaf area) Sphere volume 0.41 19.9 Sqrt(leaf biomass) LIDAR volume 0.96 0.59 Sqrt(leaf biomass) Cylinder volume 0.86 1.06 Sqrt(leaf biomass) Sphere volume 0.86 1.06 The discrepancy between LIDAR and traditionally estimated volume resulted from two factors. This disparity can be attributed to va riation in shrub structure and the assumed geometry used to calculate traditional vol ume (Figure 2-1A), as well as uneven distribution of leaf area within a shrub (Table 2-4). By segregating individual shrubs into thirds (n = 36) and using LIDAR, the distribution of leaf area was captured. As such, sect ion-specific volume estimations of each species were evident (Figure 2-4). Subsequent analysis shows that species, plant size, and plant section significantly interacted to in fluence LIDAR volume, biomass, and leaf area (Table 2-4). Using the section data across all sp ecies and sizes, a strong linear relationship was found between LIDAR volume and biomass (R2 = 0.83) and leaf area (R2 = 0.70, Table 2-5, Figure 2-5.). Although relationships were stronger using the w hole shrubs (Table 2-3) the difference in shrubs sizes (small, large) caused a clumping effect with the data points driving the
37 A B Figure 2-3. Linear relationships of leaf area measurements of all twelve individual whole shrubs with A) LIDAR volume and B) traditiona l volume estimates assuming a cylinder shape. strength of the regression results (Figure 2-3) This suggests that knowledge of the vertical arrangement of the plant as well as size is critical for accurately estimating plant volumes. The accuracy of the LIDAR volume calculations (u sing a 3D moving window, see Data Processing and Analysis) can be supported by the fact that th e laser point-density dist ributions are directly impacted by plant structure and size (Figure 2-1) when shadowing effects are minimized (i.e., using multiple scan angles). Shrub 3D spatial va riability (related to volume, biomass, and leaf R = 0.81 0 20 40 60 80 100 120 00.0050.010.0150.020.0250.03Sqrt(leaf area (cm2))LIDAR volume (m3) R = 0.41 0 20 40 60 80 100 120 00.20.40.6Sqrt(leaf area(cm2))Cylinder Volume (m3)
38 area) typically increases with si ze. This stresses the need for more accurate volume estimates of complex understory fuelbeds, wher e shrub or other plant sizes vary and where volumes of larger shrubs (> 0.5 m in height) may be more severely overestimated than smaller shrubs when using traditional methods. These results could be cr itical for the successful application of groundbased LIDAR for surface fuels assessments and represent a unique capability of LIDAR vis--vis traditional methods. For example, applying th is ground-based LIDAR approach for estimating volumes at a larger plot or management unit level is foreseeable with the versatility of the MTLS (e.g., truck and instrument mobility, vertical and horizontal angular positioning, ranging up to 1,500 m), especially within the open mid-st ory of this savanna-type woodland. Overall, the use of this small cubic space (1 cm3) to calculate volume for the individual shrubs proved advantageous for several reason s. First, the voxel approach minimized the possibility of overestimating volume because the overlapping laser points that may have been collected from the same plant canopy element (i.e., from merging of scans) were represented as a single volumetric unit (cm3) rather than a raw point count. Hosoi and Osama (2006) concluded that the ability to capture all of the plant canopy elements fully and evenly through the use of several scans and scanning angles outweighs the possibi lity of overestimating volume, and such overestimation is minimized by using a small voxel. Without multiple scans, volume may be severely underestimated because of the significant loss of laser data on the shadowed side of the plant, which may require further statistical pro cedures to correct (Van der Zande et al. 2006). This technique also provided the ability to measure volume at various scales of plant size as well as estimate volume, leaf area, and biomass for vari ous portions of the plant. This is especially useful when plants are larger or highly comple x in structure or shape and assumptions about plant geometry (i.e., cylinder, sphe roid) become less reliable.
39 Table 2-4. ANOVA output analyzing s ectional shrub 1) LIDAR volume (m3), 2) biomass (sqrtg), and 3) leaf area (sqrt-m2) estimates with three treatments: A) plant species (wax myrtle, saw palmetto), B) plant section (botto m, middle, top), and C) plant size (large, small) 1) LIDAR volume Source term DF Sum of squares Mean square F-ratio Prob level Power ( = 0.05) A: species 1 4.70E-07 4.70E-07 0.28 0.603 0.080 B: section 2 8.83E-05 4.42E-05 26.12 <0.001* 1.000 AB 2 1.81E-05 9.07E-06 5.36 0.012* 0.791 C: size 1 4.68E-04 4.68E-04 276.55 <0.001* 1.000 AC 1 2.58E-07 2.58E-07 0.15 0.699 0.066 BC 2 5.45E-05 2.72E-05 16.11 <0.001* 0.999 ABC 2 1.60E-05 8.01E-06 4.74 0.019* 0.737 S 24 4.06E-05 1.69E-06 Total (adjusted) 35 6.86E-04 Total 36 2) Biomass Source term DF Sum of squares Mean square F-ratio Prob level Power ( = 0.05) A: species 1 2.15 2.15 4.51 0.044* 0.532 B: section 2 50.53 25.27 53.10 <0.001* 1.000 AB 2 4.11 2.06 4.32 0.025* 0.695 C: size 1 67.34 67.34 141.51 <0.001* 1.000 AC 1 0.39 0.39 0.82 0.375 0.140 BC 2 8.64 4.32 9.08 0.001* 0.956 ABC 2 2.35 1.17 2.47 0.106 0.447 S 24 11.42 0.48 Total (adjusted) 35 146.93 Total 36 3) Leaf area Source term DF Sum of squares Mean square F-ratio Prob level Power ( = 0.05) A: species 1 1199.63 1199.63 40.28 <0.001* 1.000 B: section 2 4544.19 2272.09 76.29 <0.001* 1.000 AB 2 191.04 95.52 3.21 0.058 0.557 C: size 1 4840.19 4840.19 162.52 <0.001* 1.000 AC 1 21.10 21.10 0.71 0.408 0.128 BC 2 757.72 378.86 12.72 <0.001* 0.992 ABC 2 272.65 136.32 4.58 0.021* 0.721 S 24 714.78 29.78 Total (adjusted) 35 12541.30 Total 36 Term significant at = 0.05
40 Figure 2-4. Mean volume and standard errors of vertical sections (equal thirds) of two pineflatwoods shrub species (number of samples = 36; 12 plants, 3 sections each). Table 2-5. Linear regressi on relating LIDAR volume (m3) to biomass (g), and leaf area (m2) from sectional shrub data. Model used: f (x) = bo + b1(LIDAR Volume). Dependent variable n R2 RMSE Sqrt(leaf biomass) 36 0.83 0.87 Sqrt(leaf biomass), large shrubs only 18 0.78 0.96 Sqrt(leaf biomass), small shrubs only 18 0.60 0.56 Sqrt(leaf area) 36 0.70 10.51 Sqrt(leaf area), large shrubs only 18 0.62 12.02 Sqrt(leaf area), small shrubs only 18 0.33 8.03 n = number of samples At the fuelbed scale (4 m x 4 m plots), th e volume estimates from traditional point intercept sampling and LIDAR were linearly correlated (R2 = 0.48; Figure 2-6) and were not significantly different from one another (p = 0.12). The slope of th e regression line was not significantly different from a 1: 1 relationship [CI (confidence inte rval) for slope ranged from 0.39 to 1.01; CI for intercept ranged from -0.5 to 2.87]. This suggests that fuelbed volume may be estimated at sub-meter scales using either met hod with similar results. The lack of variation 0.000 0.002 0.004 0.006 0.008 0.010 0.012 TopMiddleBottom Plant SectionVolume (m3) Saw Palmetto Wax Myrtle
41 captured in the model may be explained by th e difference in sampling intensity (0.33 m vs. ~0.005 m) between the PI data and LIDAR data, respectively. The aggregation of fuels within a fuelbed (and thus variance) influences the relative accuracy of traditional point intercept sampli ng when compared to LIDAR. The empirical variograms of the 4 m x 4 m plots illustrated that the spatial distribution of fuelbed heights within a relatively small area (16 m2) was highly variable at multiple scales. This was illustrated by the changing slope within the variogram plot (Figure 2-7). High spat ial variation at very small scales was observed (< 1 m lag distance). These small scale patterns in variation can be observed within the variograms and were more ev ident within the LIDAR data compared to the PI data (Figure 2-7A, D). Some plots displayed a distribution of spatial variance (e.g., Figure 27B) indicating abrupt changes in fuelbed heights or with a hole in the variogram. These plots consisted of a very low continuous fuelbed of gra sses and forbs, with a few large inter-dispersed shrubs (Figure 2-7E). The hole in the semivariogram represented the discontinuity of heights between these different fuel types. Other plots had a more linear distribution of spatial variance (Figure 2-7C), suggesting more c onsistently independent heights w ithin the dataset. These fuels were more randomly distributed with cluste ring only found within fuels found near the ground, namely forbs, some grasses and pine litter (Figur e 2-7F). The LIDAR data was determined more reliable than the PI data for measuring surface fuel height distributions for two reasons. First, the LIDAR data create a much lower nugget effect (erro rs of spatial variation or measurement) than the PI data. Second, the LIDAR data were more se nsitive to subtle change s in spatial heights at fine-scales. This was seen by co mparing the (non-spatial) variance ( 2) and spatial variance (variogram) of fuelbed heights within a plot. The higher the 2 the more likely the PI data detected spatial changes in height, while the LIDA R data continued to detect changes at smaller
42 2 values [e.g., compare Figure 2-7B (larger variance) to Fig. 2-7 A,C (smaller variance)]. Not surprisingly, this discrepancy is mainly becau se of the high concentration of LIDAR points compared to those that can be obtained through PI methods. It is important to note that the field effort to obtain this degree of accuracy through LI DAR was considerably less than that of point intercept for all plots in this study. A B Figure 2-5. Linear relationships of LIDAR volume estimates with A) leaf biomass and B) leaf area measurements of all twelve individual shrubs, divided into three equal vertical sections (n = 36). 0 10 20 30 40 50 60 70 80 00.0040.0080.0120.016 LIDAR Volume (m3)Sqrt(Leaf Area (cm2))R 2 = 0.70 0 1 2 3 4 5 6 7 8 9 00.0040.0080.0120.016LIDAR Volume (m3)Sqrt(Leaf Biomass (g))R 2 = 0.83
43 Figure 2-6. Linear correlation between modele d volumes calculated from ground LIDAR and point-intercept (PI) field sa mpling. Solid Line: calculat ed regression line. Dotted line: 1:1 line. Application Ground-based LIDAR e fficiently capt ured fine scale fuelbed char acteristics related to plant structure, specifically height a nd volume. Its precision compared to traditional methods is explained by LIDARs ability to capture the ac tual vegetation structure versus an assumed geometry. For discrete shrubs, ground-based LIDAR can accurately measure volumes which can be used to derive biomass and leaf area. This is particularly useful in shrubby fuelbeds, such as pocosin, chaparral or southeastern pine flatw oods, where destructive harvesting of fuels for biomass is logistically difficult. In similar eco system types where prescribed burning is less frequent (hence larger fuel buildup) or in larger plots or management units, this laser-technology may still be applicable for estimating fuelbed vol ume. The surface grid technique of estimating volume beneath a 3D rendering (used for the 4m x 4m plots) would be especially useful. The size of the area that can be captured with the MTLS is only bound by time for data collection and processing and hard drive space. Shadowing effects however, seem to be the limiting factor LIDAR Volume (m 3 ) 024681 0PI Volume (m 3 ) 0 2 4 6 8 10 R 2 = 0.48
44 when capturing distinct spatial plant structure and may become more difficult to obtain where fuels are thicker or denser. This can be minimized with multiple scans (particularly from elevated positions), careful positioning of the MTLS, and thorough processing techniques (merging of scans and removing trees). Costs were quite minimal ($350 per day for data collection with the MTLS) considering the nominal data acquisition time for a large amount of information. For reference, all the shrub data was collected in less than one day, while all 26 4 m x 4 m fuelbed pl ots were collected in two days. Each fuelbed plot took approximately 15 to 30 minutes each plus travel time to each plot. For comparison, point-intercept data collection took 2 to 3 weeks. Although data processing was quite involved (w ith various software packages and engineering expertise required) for the LIDAR data, the procedure was similar to aerial LIDAR processing and was included in the cost. The time needed for data en try from (4 m x 4 m plot ) field recordings and LIDAR data processing was approximately the sa me (~ one hour per plot). Ultimately, the LIDAR data required more expertise in understand ing, recording, and processing of the data than the point-intercept approach, but the outputs were an order of magnitude more precise, which may provide for detailed information to relate to fine-sca le fire behavior. Fire behavior prediction system software is sensitive to fluctuations in parameters associated with fuelbed height measurements such as calculating volume and fuel density (Andrews and Queen 2001, Scott and Burgan 2005). Vo lume is especially difficult to measure, compared to mass estimates in many fuelbeds, becau se of the complex structure. As shown in this study, LIDAR provides an opportunity for im proving measurements of fuelbed properties that drive fire behavior and possibly enhance th e accuracy of fire behavior prediction models. Moreover, LIDAR has the capacity to measure th ose characteristics with a precision and at a
45 finer scale than is afforded by traditional me thods. The degree to which such fine scale heterogeneity is critical to fire behavior and fire effects is not currently known but can only be tested by the development of methods with the ca pacity for such measures. This approach to using ground-based LIDAR is promising in this as pect, as both measurem ents of the physical structure of fuelbeds and discrete fuel types can be sampled nondestructively and analyzed in a spatially explicit context. By understanding how fuel structure and fine-scale volume vary among fuels, other fuel characteristics, such as surface area to volume ratios, packing ratio, fuel continuity, and patchiness can be examined in detail for effects on fire behavior.
46 Figure 2-7. Examples of spatial variation (empir ical variogram) of fuelbed heights found within three 4 m x 4 m plots recorded by poin t intercept (PI) field sampling and ground LIDAR. Here it is illustrated: A) how spa tial variation may only be evident using the LIDAR data, B) how strong correlations be tween LIDAR and field sampling may be found where spatial variation is large compared to other plots (0.45 vs. 0.12 m2 for B vs. A, respectively) and strong fluctuations of spatial cont inuity (i.e., low fuels with inter-dispersed shrubs) across the plot and C) how an overall low (< 0.06 m2) height variation (compared to A,B) may only be evident using the LIDAR data, where fuels may be continuous past a range of 1 m. LIDAR imagery (D-F) provide a 3D visual representation of fuelbed height distributions and height (m) statistics associated with each variogram (A-B), respectively. Note scale is different between graphs and color gradients in the LIDAR imagery are relative to vertical height distributions within each plot. 0.00 0.04 0.08 0.12 0.16 0.00.51.01.52.02.53.0Variogram (m2)Lag Distance (m) 0.00 0.10 0.20 0.30 0.40 0.50 0.00.51.01.52.02.53.0Variogram (m2)Lag Distance (m) PI data LIDAR data Max ht = 2.5 m 2 = 0.26 E Max ht = 1.8 m 2 = 0.11 D A B
47 Figure 2-7. Continued. 0.00 0.02 0.04 0.06 0.08 0.00.51.01.52.02.53.0Variogram (m2)Lag Distance (m) PI data LIDAR data Max ht = 1.5 m 2 = 0.09 F C
48 CHAPTER 3 RELATING FINE-SCALE FOREST FUELBE D STR UCTURE AND TYPE TO FIRE BEHAVIOR USING GROUND-LIDAR, THERMA L IMAGING, AND REGRESSION-TREE MODELING Introduction Fuel and Fire in the Longleaf Pine Ecosystem Longleaf pine savannas of the s outheastern U.S. are linked to frequent low-intensity f ires (Agee 1998). Longleaf pine and many of the understo ry grass and forb species depend on fire to reduce competition from faster growing shad e-tolerant plants and create an ideal soil environment for successful regeneration (Myers 1990) The grassy understory is relatively low in stature (0.5 1.5 m) creating a bed of fine fu els that carry a low-inte nsity fire through the forest. Although this fuelbed defined here as the understory vegetation or surface fuels that carry fire (DeBano et al. 1998) seems homogeneous at first glance, fuel types and structure are diverse at fine scales (i.e., within a few meters ). The dynamics between this heterogeneous fuel environment and the regulation of future vegetati on by fire has been referred to as the ecology of fuels (Mitchell et al. 2009). This concept empha sizes the importance of fuels as a link between the combustion environment and vegetation respons e. More specifically, it targets a need for understanding the ecological significance of fine-fuels (espec ially pine-litter) and quantifying fuels and fire behavior at the appropriate scale (Hiers et al. 2007, Mitchell et al. 2009). The appropriate scale is ultimately determined by the spatial range of their i nherent heterogeneity or variation (Hiers et al. 2009, Loudermilk et al. 2009). The variation wi thin the fuelbed is characterized by vegetation fuel types, structure, biomass, and condition (e.g., moisture content). Fuel types in this st udy included wiregrass (e.g., Aristida stricta, A. beyrichiana, Sporobolus junceus ), other grasses, forbs, ferns, woody shrubs ( Quercus falcata, Q. incana, Q. laevis, Persimmon spp., Rubus spp.), pine litter, shrub litter, and various downed woody fuels.
49 One fuel (or non-fuel) type in particular is ch aracterized by patches of bare-soil, most often represented by turned up soil mounds from ground-dwelling pocket gophers ( Geomys pinetis ) and less frequently by gopher tortoises ( Gopherus polyphemus ). Pocket gophers can be abundant, creating a bare soil mound every few mete rs and found scattered across a longleaf pine stand (Simkin and Michener 2005). This unique fuel bed characteristic may act as a small fire break inadvertently influenci ng plot-level fire behavior (Hiers et al. 2009). Fuelbed fires are also influenced by vegetation structure, which is a product of plant size, height, shape, leaf area, or spat ial distribution. These structural attributes represent fuel volume and biomass, hence the amount of energy availabl e to burn. Their spa tial size and leaf area distribution directly influences their energy loss rate as a fire passes through (Ryan 2002). Wiregrass distribution is especially pertinent in this structure. The understory is often noted as having a sea of wiregrass, where its bunchy pl ant form creates a wave-like, but continuous fuelbed not often exceeding 1.5 m in height. Ta ller woody shrubs are often peppered throughout this grassy understory, further al tering fuelbed structur al properties. Pine-needle litter cast from the overstory becomes intertwined with the spikelike wiregrass blades to form a mesh fuelbed. This creates an oxygen rich environment with ea sily ignitable and continuous fine fuels, important for fire spread (Myers 1990). Fuelbed depth (or vegetation height) is an important structural component that represents abundance and continuity of fuels. Depth may range from only a few centimeters to about two meters in a frequent fire regime. Fuel types vary along this height gr adient. Leaf litter may cover the first few cm of the sandy soil bed, while grasses, small shrubs, and forbs may be found as intermediate sized fuels. Taller shrubs and longleaf seedlings may break up the fuelbed further altering fuelbed structural charact eristics. Perched leaf litter litter draped on top of and
50 hanging within the plant fuelbed is also found along this height gradient. Shrubs taller than about two meters have most likely escaped the le thal fire flames (Rebertus et al. 1989), where they are no longer a part of the su rface fuels, hence unavailable as combustible energy. As such, estimating fuelbed structure components is complex, but essential for relating to fire behavior at the same scale. Measuring Fuelbed Structure Vegetation s tructure is one of the more difficult aspects of the surface fuelbed to measure due to the complexity in plant architecture. Surface fuelbed characteristics have traditionally been quantified by both direct and indirect me thods. Direct measurements commonly employed are tallies of down woody fuels along planar transects (Brown 1974), and coupled destructive biomass sampling (i.e., clip plots) (Brown 1981). Indirect methods include visual cover estimates in plots or comparisons with photographs of known fuel loads or types (Ottmar et al. 2003, Keane and Dickinson 2007). These methods allow for stand level estimation of variables, such as fuel load, bulk density, and packing ratios that are then used to predict fire behavior (Reinhardt and Keane 1998, Andrews et al. 2004). Each method has significant limitations. Dir ect sampling is labor intensive often limiting sample size. Some techniques are not appropriate for all fuel types: Pl anar transects are not efficient at estimating fine fuels such as grasses. Indirect measures can be subjective, resulting in biased estimates. A plants architecture is of ten simplified into general geometric shapes such as a spheroid or cylinder (Van Wagner 1968) to estimate volume for bulk density calculations. While these approaches are recognized to have inherent limitations, no alternatives were previously available. Recent advances in laser ranging, or LIDAR, technologies have enabled the successful measurement of complex structures in the field with both high accuracy and precision
51 (Hopkinson et al. 2004). LIDAR ha s been typically used in fore stry for coarse scale remote sensing of forest canopy struct ures, estimating tree height di stributions, canopy bulk density, and leaf area (Lefsky et al. 1999, Drake and Weis hampel 2000, Hall et al. 2005, Roberts et al. 2005, Lee et al. 2009). These airborne LIDAR approaches have been unsuited, however, in measuring understory vegetation, primarily because of the obstruction from the forest canopy and a horizontal resolution limited to the few decimeter scale. Modern ground-based LIDAR systems now have many of the strengths of airborne syst ems (in particular lase r pulse rates of a few thousand Hz or more that enable high-resoluti on sampling), but they can attain sub-centimeter resolution and be positioned under the canopy to re duce the shadowing effects of overstory trees (Slatton et al. 2004). The high-density three-di mensional (3D) point data (more than 10,000 points per m2) obtained from such systems provide the needed precision to characterize fuelbeds, particularly by quantifying fuel height distributions, fuel dens ity and volume measurements. These data also provide the means to develop a multitude of LIDAR based fuel metrics based on the vertical and horizontal distribution of vegetation. One system in particular, the Mobile Terrestri al Laser Scanner (MTLS), is a static, stopand-scan laser scanner that covers a limited area, but captures data at the sub-cm level. The MTLS consists of Optechs ILRIS 36D (Intelligent Laser Ranging and Imaging System) ground based laser scanner (Lichti et al. 2002, Frhlich and Mettenleiter 2004), which was mounted on a lift atop a mobile platform (4 x 4 truck) by the National Center for Airborne Laser Mapping (NCALM) at the University of Florida. The MTLS captures details about the terrain at multiple angles. The ability to vary the LIDAR height and pointing angle us ing the MTLS allows significant reduction in shadowi ng effects that may be found when using a LIDAR system on a tripod.
52 Ground-based LIDAR (also called terrestrial LIDAR) has only recently been used in forestry applications. Accurate tree and canopy metrics (e.g., timber volume, tree height and diameter, gap fraction) have been successfully estimated using tripod mounted terrestrial LIDAR systems (Hopkinson et al. 2004, Watt and Donoghue 2005, Henning and Radtke 2006). Finescale leaf area (Lovell et al. 2003, Tanaka et al. 2004) and gap fraction (Danson et al. 2007) estimates of tree canopies have been correlated to that of hemispherical photographs. Small individual trees were measured using a voxel-based approach to determine leaf area density at the mm3 scale (Hosoi and Omasa 2006). Ground-based LIDAR has also been used to understand the impacts of instrument positioning on sh adowing effects that influence tree canopy measurements (Van der Zande et al. 2006). A lthough most terrestrial LIDAR systems used in forestry are stationary, a portable LIDAR system has been de veloped to record 1 m scale canopy measurements while moving through a forest (P arker et al. 2004). Only recently has groundLIDAR been used to characterize the surface fuelbed, demonstrating the importance of measuring and accounting for plant structure in volume estimates as well as understanding the spatial variability of fuelbed heights at small scales (Loudermilk et al. 2009). Relating Fuelbed to Fire Behavior Fuelbed structural com ponents, e.g., plant heights, height variation, volume, volume distribution, are important drivers of fire behavior from the plan t to landscape level (DeBano et al. 1998, Mutlu et al. 2008). Assessing relationships between fuels, fire, and fire effects are, however, generalized to the sta nd or landscape level often assu ming homogeneity at any finer scale. This assumption ignores the spatial hetero geneity of fuels and fire behavior within the fuelbed that influences overall plant community dynamics (Hiers et al. 2009). Quantifying the fire environment has been focused on visual es timates of flame length (Hoffmann and Solbrig 2003, Kennard et al. 2005, Robertson and Ostertag 2007) which is subjective, may be obscured
53 by smoke (Fernandes et al. 2000), and unrepresentative of the h eating environment needed for fuel combustion (Hiers et al. 2009). Thermocoup les or temperature-sensiti ve paints have been used for point-source measures of fire intensity allowing for interpolation to larger scales. Although these measures have been useful, estima tes of fire intensity have been compromised because of issues with instrument measurement lags causing erroneous low combustion threshold temperatures (Iverson et al. 2004, Kennard et al. 2005). Furthermore, the recordings are spatially limited and therefore intricate measurements of heterogeneity are lost. Thermal videography (i.e., FLIR Forward Looking Infrared imaging system) is a promising technique, which overcomes some of the aforementioned shortcomings by recording precise temperatures in a spatially and temporally explicit context. Its thermal signature range is similar to remote sensing sensor s that collect data beyond the visible light spectrum, yet specific for fire temperature readings (wavelengths 7.5 to 13 m). FLIR has been used for many years by the United States Department of Agriculture Forest Service for airborne fire detection and fire mapping over large areas, bene fitting from its thermal sens itivity and smoke penetration capabilities (Krausmann and Hicks 1996). Here, the FLIR has been used to analyze fire line conditions and determine safe and efficient strategies to combat wildfires. It has also been used to measure fire plumes of chaparral shrubs in a laboratory setting (S un and Weise 2003). Just recently has the system been used to measure surf ace fires at the plot level (Hiers et al. 2009). Ground-based LIDAR and FLIR technologies have the means to capture the variability of fuelbed structure (continuity) a nd fire behavior at very hi gh-resolutions (Figure 3-1). Interestingly, fuel and fire heterogeneity varies along the same distance gradient (~ 1 to 1.5 m) (Mitchell et al. 2006). After about 1 m, the fuels and fire behavior become relatively homogeneous, illustrating little to no properties of spatial autocorrela tion (Hiers et al. 2009,
54 Loudermilk et al. 2009). The relationship between the fuelbed structure and specific fuel types to fire behavior properties at this fine scale has been unknown a nd unachievable thus far due to the bottleneck of limited measurement capabilities and difficulty in assessing their multifaceted relationships. Figure 3-1. Fine-scale spatial he terogeneity of A) fuelbed stru cture (depth, m) and B) fire behavior (temperature in C ) recorded within the longleaf pine-wiregrass system used in this study. These modeled (exponen tial) semivariograms were created from in situ data on fuelbed depth and fire temperatur e in 4 m x 4 m plots captured from LIDAR (Light Detection and Ranging) and FLIR (Forward Looking Infrared) remote sensing instruments, respectively. Regression-Tree Modeling The relationships between fuel structure (fuelb ed depth, leaf and stem distribution, surface to volum e ratio, continuity) and the fire environment (combustion properties, weather) are complex and non-linear. In a longle af pine savanna, fires are of relatively low intensity with typical flame heights below a few meters (Robertson and Ostertag 2007 ). Slightly taller shrubs (> 2 to 3 m) may escape a passing fire and fu el combustion may only oc cur beneath or nearby the shrub canopy, consuming surface fuels, such as bunch grasses, smaller shr ubs, and leaf litter. Burn characteristics within this area may not co rrelate highly with typical fuel measurements (e.g., biomass) because of the patchiness of the bur n. These intricate relationships are difficult to 0.0 0.1 0.2 0.00.51.01.52.0Semivariance of fuelbed depthDistance (m) 0 1000 2000 3000 4000 5000 6000 7000 0.00.51.01.52.0Semivariance of fire temperatureDistance (m) A B
55 quantify and traditional multivariate regression appro aches used to explain their interactions or develop predictive models may not be appropria te for such multifaceted relationships (Grunwald et al. 2009). Regression tree modeling (also known as tre e-based modeling) provides an alternative approach to multivariate regression, designed speci fically for complex non-linear relationships. Satisfying typical regression assumptions relate d to input variables, residuals, and sample distributions is essentially unnecessary in tree-based models, due to their ability to partition the predicted variation of th e independent variables into autono mous predictors (Breiman et al. 1984). The CART (Classification and Regression Tree) so ftware (v. 5.0, Salford Systems, San Diego, CA) follows this partitioning of variation in a binary recursive approach, where parent nodes (representing all or a portion of the unexplaine d variation) are always split into two child nodes. These two nodes represent th e splitting of the variation typically into lower and higher portions based on the associated predictor value. These child nodes are recursive as they can become parent nodes and repeat the process to eventually create rela tively homogenous (low deviation) terminal nodes. The means of the va lues within each terminal node are the models predictors (Grunwald et al. 2009). CART can be used to either categorized features (similar to land cover classes in remote sensing) or to pred ict continuous variables (similar to regression) using categorical or continuous input variables for both. Tree-based modeling has been extensively us ed in environmental studies to unravel relationships between large numbers of independent variables that may all contribute in part to a particular environmental or landscape phenomeno n. They have been especially important for predicting soil attributes, such as organic ca rbon (Vasques et al. 2009) and phosphorus loading (Grunwald et al. 2009) as well as assessing hydrological trends, such as rainfall-runoff modeling
56 (Solomatine and Dulal 2003) and understand ing water level-discharge relationships (Bhattacharya and Solomatine 2005) Tree-based modeling is a promising tool in ecological studies, used successfully for cla ssification applications of forest land cover predictions (Moisen and Frescino 2002), vegetation mapping under various climate scenarios (P rasad et al. 2006), as well as determining abundance distributions of so ft coral taxa (De'ath and Fabricius 2000). To date, regression trees have not been used for assess ing forest fuel and fire behavior relationships. The objective of this study was to examine the fine-scale relationships of fuelbed structure (from ground-LIDAR) and fuel types (from in situ data) to fire behavior characteristics (from FLIR) across multiple forest fuelbed plots within a longleaf pine-wiregrass ecosystem. More specifically, the predictive power of various ground-LIDAR fuel me trics coupled with fuel type data were analyzed for determining surface fire temperature and residence time at sub-meter scales using regression-tree modeling. Relations hips were examined at various grouped-plot levels and within representati ve individual plots. Another objective was to assess potential technical and modeling pitfalls as well as determine the implications for assessing fine-scale fire effects. Methods Ground-LIDAR (Light Detection a nd Ranging ) Instrumentation The ILRIS ground-based LIDAR system uses a 1535 nm wavelength (near-infrared) laser with a frequency of 2,000 Hz (pulses per second), recording first or last returns of each laser pulse (user defined). The ma ximum field of view is 40 in both horizontal and vertical plane, although smaller fields of view can be specified for a given scan. The system can register laser returns from 5 to 1500 m away (at 80% target refl ectivity). The ILRIS used in this work has a pan-tilt base on the MTLS, which allows for a 360 rotation in the horizontal plane and roughly in the vertical plane. The lif t on the MTLS provides for a vertical adjustment of the scanner
57 up to a height of about 9 m. Recording laser data from multiple positi ons around the target will reduce information lost from sha dowing effects, but an efficient imaging geometry is desired to minimize the number of scans required to sample each plot. The ability to adjust the instrument vertically and horizontally to the extent offered by the MTLS allows the user to achieve advantageous scan positions and orientations in spite of constraints posed by the terrain or nearby occluding trees, which is essential to measuring precise fine-scale at tributes of intricate plant structures (Hosoi and Omasa 2006). Point density in LIDAR point clouds will vary locally depending on the range from the laser to objects in the field of view and on the angular spacing of laser shots. For the ILRIS sensor, the pulse rate is fixed at 2000 Hz. Th e angular separation of laser pulses is used to achieve a desired point density, wh ich is input via the ILRIS Cont roller software. To achieve this, the ILRIS pre-scans the user specified field of view at a course resolution and acquires an average range. The user then enters the desi red linear point spacing into the ILRIS Controller software. This spacing corresponds to the average linear separation between points on a hypothetical plane segment that spans the user speci fied field of view, is located at the mean range, and is orthogonal to the sensors boresight Given this separation distance and the mean range, the necessary angular shot spacing is au tomatically calculated by the ILRIS Controller software. Under typical operati ng conditions, the linear point spacing is chosen to be anywhere from a few cm to 1 mm. High point densities come at the cost of increased memory to hold the data and increased time to complete the scan. For example, a 20 20 field of view centered at boresight with a mean range of 15 m and user sp ecified linear point spacing of 5 mm will yield an angular separation of 0.0189 require roughly 9 min to comple te, and result in just over 1.1 million points.
58 The ILRIS collects: 1) x, y, z coordinate valu es with respect to the position of the laser sensor; 2) intensity values of the return; and 3) true color (Red Green Blue) values for each point obtained from an integrated and calibrated di gital camera within the instrument. A more complete description of the technology can be found in (Lic hti et al. 2002, Frhlich and Mettenleiter 2004). Once an area is sampled by tw o or more scans, the point clouds from each scan are merged into a common coordinate fr ame using software that can accommodate 3D spatial data. The subset of the point cloud that covers the region of interest can be isolated and the remaining points cropped out to minimize the computational burde n and obtain statistics representative of the prec ise region of interest. Study Site The site was located at Ichauway, an 11,000 ha reserve of the Joseph W Jones Ecological Research Center in southwestern Georgia, USA. Ichauway is cl assified within the Plains and Wiregrass Plains subsections of the Lower Co astal Plain and Flatwoods section (McNab and Avers 1994). Ichauway has an extensive tract of second-growth longleaf pine and has been managed with low intensity, dormant-season prescrib ed fires for at least 70 years, at a frequency of one to three years. The unde rstory of the study area was pr imarily composed of wiregrass ( Aristida stricta ), many forb and prairie grass species, as well as inter-dispersed hardwood shrubs (e.g., Diospyros spp., Prunus spp, Quercus spp., Sassafras albidum ). With frequent fires, hardwoods are generally maintained at shrub si ze, occasionally reaching mature size. The study area had not burned for one year. Data Collection In spring 2007, a total of 20 georeferenced 4 m x 4 m plots were established to measure fuelbed characteristics. The 4 m x 4 m area was chosen because it was large enough to capture heterogeneity at sub-meter scales, but small enough to support intensive sampling with minimal
59 impact to the vegetation. A ladder was suspe nded horizontally across the plot to sample the interior with minimal disturbanc e to the vegetation. Spatially explicit point-intercept (PI) fuel data (169 samples, 0.33 m grid spacing) were r ecorded using a 5 mm graduated dowel within each plot. Only vegetation in physical contact with the dowel was measured. At each PI sample point, maximum fuelbed and litter depth (or height), as well as presence/absence of fuel and vegetation types were recorded. The georeferenci ng and sampling intensity were used to capture the spatial variation of the fuel bed found within this small (16 m2) area and to relate to the subcentimeter scale 3D lase r data collected from the ground-based LIDAR. Within two weeks of field data collection, the MTLS was used to collect ground-LIDAR data on all 20 plots. Prior to laser data coll ection, reference targets (consisting of a styrofoam ball on top of a metal rod, 0.5 to 1 m tall) were placed at all four corners of the plot. A double reference target (two Styrofoam balls on one metal rod) was used at the northwest corner of each plot to orient the plot data for processing. Th e MTLS was restricted to mapped roads and trails, as well as a buffer of 5 meters around each plot to reduce vegetation disturbance. The ILRIS was lifted to a height of 7 m to capture an aerial view of the pl ot to reduce shadowing effects and avoid tree bole or canopy obstruction. The ILRIS was set to a downward angle tilt of 25 from horizontal. A true color digital photograph was ta ken by the ILRIS for each plot, and used in the field to delineate the precise field of view for each scan. First-return laser pulses were recorded with an input mean point spacing of 5 mm. Two scans were taken of each plot, from opposite sides, to mitigate shadowing effects and ensure more accurate and complete sub-centimeter scale data for fuel plots. These two scans were merg ed in the processing stage to a single 3D point cloud data set. Data collection with the MTLS took approximately 20 minutes per plot. In comparison, field PI data collection last ed an average of two hours per plot.
60 In each plot, fire behavior was recorded using a FLIR S60 digital thermal imagery system. The camera and operator were positioned on a boom about 7 m above the ground and 10 m from the plot edge. Thermal images were captured as the controlled fires bur ned through the plots at 0.25 Hz. A 0.5 x lens was attached to the camer a to increase the field of view and record temperatures for the entire plot simultaneously. The camera has two temperature range settings (-40 to 500 C or 300 to 1500 C); the latter (higher) temperatur e range was used here. The measurements were corrected for air temperature, relative humidity, distance from target, and the emissivity was set at 0.96. A flare was placed at each plot corner and ignited immediately before the fires to provide surveyed reference points fo r post data processing. Further details are in Hiers et al. (2009). All plots were burned during three prescribed burns conduc ted at an operational scale (about 70 ha each) using strip head fires on Fe bruary 23, 27 and March 16, 2007. A strip head fire was ignited 5 m upwind of each plot with over 100 m separating strip heads from downwind strips when plots were ignited. Plots within a burn unit were burned and sampled from downwind to upwind. Each operational strip head fire was allowed to burn through the plots undisrupted by additional lines of fire (Hiers et al. 2009). M ean wind speed and temperature were quite similar between days, ranging from 1.47 to 2.19 m/s and 19.4 to 22.9 C, respectively (Table 3-1). Plot level wind speeds were record ed during the fire (Table A-1). The mean plot level wind speed values ranged from 1.1 to 2.9 m/ s. Relative humidity on the other hand was quite different between days, ranging from 14.9 to 65.9 %. Table 3-1. Burn day mean wind speed (m/s) and direction from wind sensors next to each plot as well as mean temperature and relative humid ity (RH) from a local weather station. Burn date Wind speed Wind direction Temp. (C ) RH (%) No. plots February 23, 2007 1.80 South 19.36 14.88 5 February 27, 2007 1.47 Southwest 22.88 42.28 7 March 16, 2007 2.19 Southwest 22.06 65.86 8
61 Data Processing The presence or absen ce data for each fuel type collected in the field (PI data) were represented as categorical binary variables (1 = present, 0 = ab sent) for each spatially explicit point (33 cm spacing) within each plot. These bi nary fuel type variables were used as (11) independent variables in the statistical analyses (e.g., CART). The point-intercept data across all plots were converted to distan ce matrices for use in cluster analysis (Hiers et al. 2009) Euclidean distances were used for continuous vari ables (fuelbed height, l itter height), whereas asymmetric Jaccard distances were used for the 0 or 1 binary variables (presence or absence of longleaf pine litter, etc). The Jaccard distance measured the over lap of presence (similarity) of various fuel types, while Euclidean distance meas ured the degree of continuity of fuelbed or litter depths. Cluster trees we re then calculated us ing the centroid method, and the number of clusters was selected based on the Cubic Clustering Criterion and Pseudo-F value traces (Everitt 1980). In the case of multiple local maxima for these criteria, the cluster groups were compared with known plot characteristics to select the final number of cl usters (SAS Institute Inc. 2003). This process is similar to development of fuelbe ds at larger scales (Dimitrakopoulos 2002). The (11) clusters [detailed in Hier s et al. (2009)] represented a vari ety of fuel types ranging from patches of bare ground and coarse woody debris to complexes of pine litter, shrubs, and grasses. The most common fuels clusters were mixed gram inoids, wiregrass with shrubs, and wiregrass with perched longleaf pine litter. The Quick Terrain Modeler (Applied Imager y) and TerraScan (Terrasolid) software packages were used for processing the laser data Of the two scans taken per plot, the first was horizontally rectified by compensa ting for the original scanning ge ometry, since the instrument had a downward tilt of 25 with re spect to horizontal. The referen ce targets were used to identify common points between the two scans and create a common 3D coordinate frame for both scans
62 (see (Wolf and Ghilani 1997) for details). The second scan was adjusted to fit the same 3D coordinate frame as the first, which combines them into a single spatially consistent data set. The digital image and the double target reference for th e northwest corner of the plot were especially helpful in this merging process and in orienting the plot in cardinal space. The 4 m x 4 m plot areas were clipped from the resulting merged scans using the reference targets. Roughly 600,000 to 700,000 sample laser points were found with in each 4 m x 4 m plot (Figure 3-2). A B Figure 3-2. Example outputs of the A) ground-based LIDAR (using the MTLS) and the B) thermal camera (using the FLIR) for a 4 m x 4 m forest fuelbed plot. The color gradient in A represents the vegetation height distribution (maximum vegetation height = 4 m), which aids in visualization of the 3D vegetation structure. Blue and green represent lower vegetation, leaf litter and ground points, while yellow and red represent taller vegetation. The ground-LIDAR data were scaled down to th e spatial resolution (33 cm x 33 cm) of the PI data. This is a commonly used approach for aerial LIDAR data, wh ere LIDAR metrics were calculated within equally spaced vertical sections or bins (e.g., pi xels) (Goetz et al. 2007). Using MATLAB (The MathWorks, Inc.) 13 LIDAR related metrics were determined based on the distribution of ground-LIDAR lase r points and intensity values found within each bin (Table 32). LIDAR height (z) values calculated within each bin included mean, maximum, variance, kurtosis, skewness, distribution ratio (mean/max), sum, and spread (max-min). The various
63 height distribution values (e.g., skewness, kurtosis) represented unique structural features of the fuels not represented by the more common attributes (e.g., mean, max) and test their relationship Table 3-2. Fuelbed [ground-LIDAR (Light Detection and Ranging) and point-interce pt] and fire (FLIR Forward Looking Infrared) metrics created for the CART (Classification and Regression Tree) analysis. Metrics were determined w ithin 33 cm x 33 cm cells within 20 4 m x 4 m forest plots. All heig ht values are in meters and point-intercept data are categorical. All in tensity and point density valu es are normalized across all plots. Fuelbed metrics Fire metrics Ground-LIDAR Point-intercept FLIR x, y coordinatesa Fuel Clusters (1-11)bMaximum temperature C Mean height Fuel types (0/1): 90th quantile temperature C Maximum height 1 and 10 hour fuels (HR 10) Residence time above 300C Variance of height 100 and 1,000 hour fuels (HR 100) Residence time above 500C Kurtosis of height Pine litter Skewness of height Oak litter Height distribution ratio (mean/max) Perched pine litter Sum of heights Perched oak litter Spread of heights (max min) Wiregrass Point density (0-1) Other grasses Mean LIDAR intensity (0-1) Forbs Int sq r t [sum of sqrt(LIDAR intensity)] Shrubs Int q dr t [sum of qdrt(LIDAR intensity)] Volatile shrubs Intln r [sum of LIDAR intensity] Bare soil (no vegetation/fuel) a used for plot level analysis only; b used for multi-plot (collective and grouped) analysis only, HR = hour, sqrt = square root, qdrt = quadratic with the fire metrics. Laser point density values and LIDAR in tensity values were normalized across all plots by dividing each bins value by the maximum value across all plots. Three additional intensity variables were created to represent the intera ction of volume of fuel (i.e., laser point density) and type (leaf surface area) or condition (moisture) of the fuel (i.e., LIDAR intensity). LIDAR intensity values have been used to differentiate between broad and narrow leaved trees species and been correlated with fuel moisture (Garc a et al. 2010), but the relationship with surface fine-fuels is unknown. One variable was simply the sum of the normalized LIDAR intensity values ( intlnr = inti/intmax, i = individual laser point) within each bin
64 (linear relationship between intensity value a nd point density), while the other two were normalized (as above) and scaled by the quadratic (Equation 3-1) a nd square root (Equation 3-2) function and summed individually within each bin (i.e., intqdrt, intsqrt respectively). This scaling represented a range of degrees that the backscat tered pulse may be influenced by the volume or density of fuels found within each 33 cm x 33 cm area. The quadratic function suppressed the influence that the amount or si ze of fuel (point density) had on LIDAR intensity compared to the linear relationship. The opposite was true for the square root function. (3-1) (3-2) The FLIR fire images were analyzed usi ng FLIR Systems ThermaCAM Researcher Pro software version 2.7 (Figure 3-2). Each image was then transformed into a tab-delimited ASCII file where each cell represented pixel temperature. These files were again transformed into TIFF files for post-processing, with the temperature data in these files captured as pixel color values. The TIFF files were rectified and georeferenced using image processing software (Geomatica). Summary statistics over time (represented by the multiple image captures by the camera) were computed for each pixel in the original therma l imagery data, including maximum temperature (Tmax), 90th quantile temperature (Q90), residence time above 300C (T300), and residence time above 500C (T500). The Q90 temperature data we re used to analyze the thermal imagery because the lowest temperature measured on th e cameras high-temperature setting was 300C; thereby, Q90 distinguished unburned patches from sm oldering fire (Hiers et al. 2009). The two residence time variables potentially segregated the smoldering effect s of fire into low (T300) and higher (T500) smoldering intens ities. These resulting FLIR datasets of temperature and
65 residence time were also scaled down (from ab out 4 cm x 4 cm) to the 33 cm x 33 cm spatial resolution of the PI (and now ground-LIDAR) datase t by averaging each of the four fire metrics within each bin. Grouping of Plots All twenty plots were grouped based on their fi re behavior m etrics using a k-means cluster analysis. Each plot was represented by their st andardized mean value fo r each of the four fire metrics. The number of clusters was chosen based on the Cubic Clustering Criterion and Pseudo F statistic. There were clearly three distinct clusters among th e plots, representing three levels of fire intensity (Figure 3-3, Fi gure 3-4). The first group represented the middle range of temperatures and residence times corresponding to nine of the twenty plots. The second cluster represented the higher range of temperatures and residence times, while the third cluster represented the lower range, corres ponding to five and six plots ( out of twenty), respectively. This plot grouping provided for a streamlined data analysis approach to analyze relationships between the fuel and fire environment within similar fire behavior ranges There was no distinction between the three fire groups based on particular day of burn or plotlevel wind speed (range = 1.79 1.85 m/s). Regression-Tree Modeling The analys is regressed the LIDAR metrics a nd fuel types with each of the four fire behavior metrics using least squares regression-tree modeling in CART. CART analysis was run for the collective data (20 plots) each grouped dataset [mid-range (9 plots), high-range (5 plots), and low-range (6 plots)], and three randomly chos en plots within each group (3 plots). Spatial coordinates (x,y) were incorporated into one fire metric model within each of the three individual plots to assess the influence of sp atial configuration of a plot level fire. A total of 31 CART simulations were run overall. Committees tree s (CT) in ARCing mode (Adaptive Resampling
66 Mid-rangeHigh-rangeLow-rangeTmax -3 -2 -1 0 1 2 3 Mid-rangeHigh-rangeLow-rangeT300 -3 -2 -1 0 1 2 3 Fire Behavior Group Mid-rangeHigh-rangeLow-rangeQ90 -3 -2 -1 0 1 2 3 Fire Behavior Group Mid-rangeHigh-rangeLow-rangeT500 -3 -2 -1 0 1 2 3 Figure 3-3. Distribution of grouped plots by each standardized mean fire behavior variable used in the k-means cluster analysis. Tmax: maximum temperature (C), Q90: 90th quantile temperature, T300: residence tim e above 300C, T500: residence time above 500C. and Combining) were used as the primary method of applying regression tree modeling in CART. This was chosen due to the complexity of the data, large number of possible predictor variables, the non-linear relati onships between LIDAR and FLIR metrics, as well as CT model robustness over other possible CART modes (i.e., single-tree, committee-tree in bootstrap aggregation ) (Grunwald et al. 2009). The strength of running the CT in Arcing mode is its ability to run multiple (often hundreds of) regr ession-trees and learn from each tree, based on
67 probability of bootstrap selection. This ultimate ly builds a stable CT model where predicted values are averaged across the multiple versions of the predictors. Detailed descriptions of each regression tree methodology, includi ng statistical aspects are found elsewhere (Breiman et al. 1984, De'ath and Fabricius 2000, Prasad et al. 2006, Grunwald et al. 2009). T500 -2-10123Tmax -3 -2 -1 0 1 2 Mid range High range Low range Figure 3-4. Distribution of gr ouped plots (k-means clustering) in relation to standardized maximum temperature ( C, Tmax) and residence time (number of four second intervals) above 500 C (T500). One-hundred trees were run per CT simulati on, while incorporating the ten-fold crossvalidation procedure in CART to test predic tion performance and optimize tree size (hence control over-fitting). An op timal tree was found by finding the one with the minimum crossvalidation relative error. Parent node and terminal node sample si ze was limited to ten and three, respectively and node complexity was scaled ba sed on sample size, all further optimizing tree size and reducing run-time. For each model, the coefficient of determination (R2), root mean square error (RMSE), residual prediction devi ation (RPD), cross-validation relative error (CVRE), re-substitution relative error (RRE ), and number of terminal nodes (Tn) were recorded for each model. CVRE, RRE, and Tn are specific to regression-tree modeling and useful for
68 analyzing model fit of CT models. RRE is the estimate of the overall relative error produced for a tree among various possibl e tree sizes (i.e., varying Tn). The re-substitution error produced for each node during the splitting process aids in findi ng the best split. CVRE is the relative crossvalidation error of a tree among va rious possible tree sizes. The CV RE and RRE for the best fit models (lowest RRE, CVRE and associated Tn) were presented. RPD is the ratio of the standard deviation of the validation or reference sample set (SDr) and the standard error of prediction (SEP, Equation 3-3). In this case, the RPD (Equ ation 3-4) takes into acc ount the variability of the referenced samples used in the ten-fold cr oss-validation (i.e., the obs erved values) scaled by the SEP and useful to compare predictive abil ity across models with differing units (i.e., temperature and residence time). For more de tailed information on CVRE, RRE, RPD, or CART see (De'ath and Fabricius 2000, Grunwal d et al. 2009, Vasques et al. 2009). (3-3) (3-4) The CART relative importance rankings (cal led importance values (IV) hereafter) were used to assess the ability of the fuel metrics to predict each of the fire me trics. While model fit and error statistics provided ove rall model robustness measures, th ese IVs allowed us to assess the relative explanatory power of the metrics in a heterogeneous fire environment. CART measures each variables ability to simulate a gi ven tree and its role as a surrogate for other variables within that tree. This is calculated across nodes, totale d, and scaled relative to the best performing variable (valued at 100), creating tree specific IVs ra nging from 0-100. IVs for each of the 31 model runs, including the collective, gr ouped, and individual plot data were recorded (Table A-2).
69 Results and Discussion Regression-tree models: collective data The relationships between fine-scale fuelbed and fire behavior characteristics were strong, especially w hen the fire environment was categor ized a priori (Table 3-3, Figure 3-5). Although other studies have used both ground (Hopkinson et al. 2004, Loudermilk et al. 2009) and aerial LIDAR (Lefsky et al. 1999, Hall et al. 2005) as well as intensive field inventory data (Brown 1974, 1981) to characterize forest fuels and structure, this analysis is the fi rst to relate complex fuelbed structure and types to spec ific fire attributes at such a fine scale (sub-meter). Applying the collective data (all 20 plots) in CART provided a general desc ription of potential associations between the fuelbed and fire. The lo w to intermediate model strength (R2 = 0.22 0.54), particularly in comparison to the grouped or in dividual analysis was not surprising, as variation in weather and fuel conditions that drive stand (i.e ., plot to plot) changes in fire behavior most likely caused significant noise in th e collective datase t. Within this extensive dataset (n = 3380), the distribution of temperatures was restricted by the lower li mit (300C) setting on the FLIR camera. This caused a left-skewed distribution of points, or a clumpi ng effect towards 300C mainly impacting Q90 predictions (Figure 3-5D). A sim ilar effect occurred with T500, where many values were at or close to zero. A more uniform dispersion was often found with Tmax and T300 in the various datasets (Table 3-4). The collect ive model for T300 (R2 = 0.54, RMSE = 1.91, RPD = 1.46) included effects of fuel structure and type on fire residence time across multiple fuelbed plots regardless of between plot variation in weather and fuel conditions. Fuelbed height and height distribution metrics were the strongest drivers of both temperature and residence time in the collective models, while LIDAR intensity metrics were moderate drivers (Table 3-5). The fuel clusters played an importa nt role, more so than the (binary) fuel types themselves, although in general they were moderate ly ranked according to their respective IVs.
70 Table 3-3. Results of CART analysis at vari ous multi-plot levels using the ten-fold crossvalidation procedure. Response variable s are the fire metrics, namely Tmax (Maximum temperature in C), Q90 (90th Quantile Temperature in C), T300 (residence time above 300C), and T500 (residence time above 500C). Residence times (T300, T500) are in number of four second intervals. Collective data (20 plots, n = 3380) Response variable Tn CVRERRE R2 RMSE RPD Tmax 3300.540 0.124 0.43 64.37 1.24 Q90 3220.390 0.090 0.22 63.62 1.11 T300 3230.317 0.062 0.54 1.91 1.46 T500 3410.296 0.063 0.38 1.61 0.90 High-range fire group (5 plots, n = 845) Response variable Tn CVRERRE R2 RMSE RPD Tmax 1360.572 0.098 0.80 23.52 2.25 Q90 1320.422 0.075 0.88 24.24 2.86 T300 950.177 0.051 0.88 0.97 2.92 T500 1220.316 0.081 0.83 0.60 2.41 Mid-range data (9 plots, n = 1521) Response variable Tn CVRERRE R2 RMSE RPD Tmax 2360.536 0.114 0.78 26.85 2.14 Q90 2180.518 0.117 0.79 22.54 2.16 T300 2190.397 0.082 0.83 0.91 2.40 T500 2430.596 0.141 0.78 0.50 2.08 Low-range data (6 plots, n = 1014) Response variable Tn CVRERRE R2 RMSE RPD Tmax 1280.455 0.105 0.83 33.30 2.43 Q90 1400.280 0.076 0.86 8.80 2.30 T300 1520.323 0.087 0.80 0.90 2.18 T500 1500.429 0.094 0.79 0.37 2.11 Tn = no. of terminal nodes, CVRE = cr oss-validation relative error, RRE = resubstitution relative error, R2 = coefficient of determination, RMSE = root mean square error, RPD = residual prediction deviation As expected, the individual fuel types did not play as significant a role in predicting the collective fire metrics, mainly due to the binary na ture of the variables. Despite this, particular fuel types were notable, such as deciduous oak and pine litter, wiregrass, graminoids, and shrubs.
71 Wiregrass was a strong predictor for temperature, while pine litter was a strong predictor for residence time among the fuel types. Figure 3-5. Regression-tree resu lts using CART (Classification and Regression Tree) for Q90 and T300 using the collective data (20 pl ots, n = 3380, A,D), the high-range fire group (5 plots, n = 845, B,E), and individual plots within the high-range fire group (n = 169, C, F). Note the change in scale for the indivi dual plots (C,F). R = 0.54 0 5 10 15 20 25 30 0102030Predicted T300Observed T300 R = 0.22 300 400 500 600 700 300400500600700Predicted Q90Observed Q90 R = 0.88 0 5 10 15 20 25 30 0102030Predicted T300Observed T300 R = 0.88 300 400 500 600 700 300400500600700Predicted Q90Observed Q90 R = 0.68 4 6 8 10 12 4681012Predicted T300Observed T300 R = 0.80 350 450 550 650 350450550650Predicted Q90Observed Q90 A B C D E F
72 Generally, strong fuel metrics were found within the top-five ranked IVs or with IVs above 65, while moderate fuel metrics were found within the 6 to 10 ranked IVs or with IVs between 30 and 65 and weak fuel metrics were found within the IV rankings below 10 or with IVs below 30. The LIDAR fuel metrics (continuous variable s) and the fuel cluste r metric (categorical variable) were often strong to moderate drivers, while many of the individual fuel metrics (e.g., wiregrass, pine litter; binary) were often within the weaker rankings. Table 3-4. Descriptive statistics [mean (SD: standard deviation) ] of fire response variables at the various multi-plot and individual plot anal yses (no. of plots are in parenthesis). Temperature (Tmax, Q90) are in C. Residence time (T300, T500) are in number of four second intervals. Tmax Q90 T300 T500 Collective data (20) 544 (80) 368 (70) 6.1 (2.8) 2.0 (1.4) High-range data (5) 611 (53) 447 (69) 8.3 (2.8) 3.5 (1.4) Mid-range data (9) 548 (57) 364 (49) 6.2 (2.2) 1.9 (1.0) Low-range data (6) 482 (81) 310 (20) 4.1 (2.0) 0.9 (0.8) High-range individual plot 605 (53) 504 (40) 6.7 (1.1) 2.6 (0.9) Mid-range individual plot 569 ( 55) 370 (51) 6.9 (2.4) 2.2 (1.1) Low-range individual plot 507 ( 82) 307 (13) 3.8 (1.3) 0.9 (0.7) Regression-tree models: fire groups The grouped fire m odels (low, middle, and highrange data) performed considerably better (R2 = 0.78 to 0.88, RPD = 2.11 to 2.92, Table 3-3) than the collective data (and ma ny of the individual plots). This demonstrated the importance of ch aracterizing the fire environment and the ability of CART to model this relationship across mu ltiple plots albeit the high fire and fuel heterogeneity within groups and individual plot s (Table 3-4, 3-6, 3-7). All of the models exhibited strong mode l fit, with high R2 values (R2 = 0.78 to 0.88), low error (e.g., RRE = 0.051 to 0.141), and RPD values above two (RPD = 2.11 to 2.92). Residual error (RMSE) was considerably lower (often half) of what was observe d for the collective dataset. The trends that
73 Table 3-5. Top five importance value (IV) rankings of input variables fr om CART analysis at various multi-plot level analysis. Temper ature (Tmax, Q90) are in C. Residence time (T300, T500) are in number of four second intervals. Response Variables Collective data (20 plots, n = 3380) Top five ranked fuel metrics (IV) Tmax Max ht (100), Spread of hts (95.41), Ht variance (90.07), Mean ht (82.0), Ht kurtosis (79.7) Q90 Ht dist. Ratio (100), Sum of hts (96.1), Mean ht (93.8), Ht skewness (86.9), Ht kurtosis (80.1) T300 Mean ht (100), Ht dist. Ratio (70.5), Spread of hts (66.2), Max ht (66.2), Height skewness (60.7) T500 Mean ht (100), Spread of hts (97.2), Ht variance (92.5), Ht dist. Ratio (92.0), Max ht (87.9) High-range fire group (5 plots, n = 845) Tmax Spread of hts (100), Max ht (96.3), Sum of hts (81.2), Mean ht (81), Point density (75.5) Q90 Fuel clusters (100), Mean ht (73.3), Point density (61.7), Sum of hts (54.8), Spread of hts (53.8) T300 Mean LIDAR intensity (100), Ht dist. Ratio (66.5), Fuel clusters (66.3), Int-lnr (40.8), Int-qdrt (39.4) T500 Sum of hts (100), Mean ht (97.9), Ht dist. Ratio (96.7), Point density (69.4), Max ht (66.0) Mid-range fire group (9 plots, n = 1521) Tmax Mean ht (100), Ht variance (96.5), Sum of hts (82.1), Ht kurtosis (81.9), Ht skewness (78.3) Q90 Spread of hts (100), Max ht (85.3), Ht variance (79.1), Ht dist. Ratio (74.4), Ht skewness (74.2) T300 Mean ht (100), Ht variance (78.2), Ht kurtosis (71.0), Ht dist. Ratio (45.6), Spread of hts (42.6) T500 Mean ht (100), Ht variance (96.8), Int-sqrt (88.4), Mean LIDAR intensity (82.8), Spread of hts (80.2) Low-range fire group (6 plots, n = 1690) Tmax Mean LIDAR intensity (100), Mean ht (72.1), Ht skewness (72.1), Spread of hts (68.5), Fuel clusters (64.8) Q90 Ht variance (100), Max ht (86.3), Spread of hts (83.6), Mean ht (81.1), Ht dist. Ratio (59.8) T300 Ht kurtosis (100), Ht skewness (86.0), Mean ht (78.5), Int-qdrt (76.1), Sum of hts (72.9) T500 Mean ht (100), Ht variance (73.2), Max ht (73.0), Spread of hts (70.1), Sum of hts (62.9)
74 were muted in the collective data set (driving the error a nd lack of model fit) especially for Q90, were disaggregated in the fire groups and individual plots (Figur e 3-5). There was a trend in model fit from strongest in the high-range fire group and lowest in the middle-range group. The discrepancy may be associated with differences in fire metric variability across groups. The strongest models in the high-range group (Q90, T300, T500) had the highest variability among fire metrics across groups (Table 3-4), while the Tmax model in the low-range group had the highest temperature variability (SD = 81 C compared to 53 and 57 for the high and middle range group, respectively) across groups and strong model fit (RPD = 2.43). The prominent Q90 model fit in this group was driven by extremel y low error (RMSE = 8.8), where low variability did not hinder model strength. Descriptive statis tics revealed little varia tion within and between LIDAR metric datasets (Table 3-6), explaining little in predictiv e capabilities. These values however, illustrated the coarser-scale homogeneity traditional measured. Table 3-6. Descriptive statistics [m ean (SD)] of a representative set of LIDAR fuelbed metrics at the various multi-plot and individual plot level analysis (no. of plots are in parenthesis). LIDAR mean in tensity and point density we re normalized (0-1) across all plots. All height values are in meters. Mean fuelbed depth (LIDAR height) was one of the most significant drivers for all models (including collective, grouped, and individual plot datasets), signifying its value for fire Maximum height Height variance Mean height Mean LIDAR intensity LIDAR point density Collective data (20) 0.97 (0.40) 0.06 (0.10) 0.35 (0.18) 0.53 (0.10) 0.14 (0.07) High-range data (5) 0.92 (0.31) 0.05 (0.05) 0.31 (0.14) 0.56 (0.06) 0.15 (0.06) Mid-range data (9) 0.98 (0.42) 0.06 (0.10) 0.36 (0.19) 0.53 (0.10) 0.14 (0.07) Low-range data (6) 0.98 (0.44) 0.07 (0.13) 0.38 (0.19) 0.52 (0.11) 0.14 (0.07) High-range individual plot 1.03 (0.25) 0.04 (0.03) 0.39 (0.12) 0.55 (0.08) 0.17 (0.05) Mid-range individual plot 0.98 (0.42) 0.06 (0.10) 0.36 (0.19) 0.53 (0.10) 0.14 (0.07) Low-range individual plot 0.98 (0.44) 0.07 (0.13) 0.38 (0.19) 0.52 (0.11) 0.14 (0.07)
75 temperature and residence time predictions. Twen ty-five of the 31 models simulated (including the individuals plots with x, y variables) had mean LIDAR height with in the top-five ranked variables. Other height and fuelbed metrics va ried by group in response to changes in fuelbed continuity. Variability in fuel bed depth (e.g., LIDAR height variance, skewness, kurtosis) was a strong predictor for the middle and low-range fire models, while sum of hei ghts, point densities, and fuel clusters were more prominent in the high-range fire models. In the two lower fire intensity groups, fuelbed depth varied more with in cells and plots than the high-range fire group (Table 3-6). The skewness and kurtosis metrics would account for fuels that were distributed towards the shorter or taller vegetation or those that were excessively peaked (or lacking a peak) around the mean height, respectively. This disrupti on in fuelbed continuity may cause variability in radiative and convective heat transfer rate s between neighboring fuels at small horizontal scales (< 1 m), influencing fire intensity values (DeB ano et al. 1998). As fuel becomes more spatially uniform both horizontall y and vertically (re presented by point density and sum of Table 3-7. Percent abundance of each fuel type as well as number of point-intercept fuel clusters within each multi-plot or individual (indiv.) plot analysis. Number of plots is in parenthesis. PI fuel type % abundance No. PI WG Shrubs Forbs Gram. Vol. shrubs HR 10 HR 100 P. pine litter P. oak litter Pine litter Dec. oak litter Bare soil Collective data (20) 11 51 5 40 30 4 10 2 37 3 51 20 9 High-range data (5) 9 58 6 33 20 2 12 1 47 3 69 20 7 Mid-range data (9) 11 41 4 47 34 3 9 2 34 3 50 22 8 Low-range data (6) 8 59 6 35 34 6 10 3 32 3 38 17 13 High-range indiv. plot na 78 3 29 18 4 7 1 41 1 69 9 7 Mid-range indiv. plot na 64 1 18 27 6 14 2 50 1 78 7 4 Low-range indiv. plot na 80 12 39 40 3 6 3 20 9 15 34 15 No. PI: Number of PI clusters, WG: Wiregr ass, Gram.: Graminoids, Vol. shrubs: Vola tile shrubs, HR: hour fuels, P: Perched, Dec.: Deciduous
76 heights), so do heat transfer m echanisms driving small-scale fire intensity. Fuel types however, may vary substantially between neighboring cells (i.e., sensitivity of fuel cluster metric), influencing the variability seen in Q90, T300, a nd T500 within the high-range fire group (Table 3-4). Maximum height and spread was more impor tant for temperature than residence time within all three fire groups. Maximum fuelbed height within each cell re presents the amount of fuel available to burn, directly impacting fire temperature (Whelan 1995). Residence time is influenced by bulk density and specific fuel type s, such as downed woody debris (e.g., branches, pine-cones), with minimal influe nce from maximum fuelbed height within a cell. For instance, a cell with mixed shrubs, wiregrass and leaf litter may be relatively tall (~1.5 to 2 m), but burns readily with little smoldering o ccurring. Another cell may have the same general fuel make-up plus some downed woody debris that may ember for some time. Temperatures may be similar within these two cells (with cons tant weather and fuel condition s), but residence time may be quite different. LIDAR intensity metrics were moderate to significant drivers among models. Mean intensity was the most important metric (#1 in IVs) for T300 and Tmax in the high and lowrange fire groups, respectively (Table 3-5). The intsqrt and intqdrt were significant drivers for T500 (#3 IV, middle-range group) and T300 (#4 IV, low-range group), respectively. Otherwise, all intensity variables had moderate IVs with no distinct trends between models or groups. There were several inconsistencies with using the various LIDAR intensity metrics among the grouped fire models. LIDAR intensity was more often more important for residence time than temperature among the fire groups (i.e., within to p-five IVs in half the models, Table 3-5), although mean intensity was the top (#1 IV) me tric for Tmax (low-range group). The four
77 LIDAR intensity metrics generally clumped together (i.e., acted well as surrogates for each other), especially as they decr eased in importance ranking. In so me cases, there was little to no distinction between intensity variables (Table A-2). This was inconsistent with T500 (middlerange), where intsqrt and mean intensity were both highl y ranked (no. 3 and 4) and clumped together. It was unc lear whether creating three scaled LIDAR intensity metrics (e.g., intqdrt) was beneficial, especially because LIDAR point density exhibited high surrogate potential to some or all of the scaled intensity metrics. On the ot her hand, there were cases where point density acted independently from the scaled intensity va lues (i.e., Q90, high-range; T300, middle and low range). The difficulty in using LI DAR intensity values in relation to fuels is that intensity values represent target surface reflectan ce coupled with size of the in tercepted surface. These are difficult to separate and infer upon individually (Riano et al. 2004). A lthough the interpretation of LIDAR intensity (or those coupled with point density) for model predictions was somewhat ambiguous, the variables clearly ha d a role (quite significant in some cases) in representing a unique aspect of the fuelbed a nd predicting fire behavior. The point-intercept fuel cluster metric was a uniquely significant model driver among all fire group models. Within the high-range fire group, the fuel cluster metric was the most significant driver for Q90 (#1 IV ) and #3 (IV) for T300 (Table 3-5). Their considerable model strength (RPD = 2.86 2.92) coupled with their si gnificant correlation with the fuel clusters supports the findings in Hiers et al. (2009), where Q90 was more se nsitive to the fuel clusters than the other three fire metrics. This also illustrated the link between the heterogeneous fuelbed captured by the fuel clus ters and the complex fuel environment. This may have been most evident within higher fire intensity plots. Here there was a stronger distinction of fuels to represent variability in fire behavior, especially associated with smoldering effects. Conversely,
78 the fuel clusters played a stronger role for Tmax than Q90 in the middle (# 7 and # 14 IV, respectively) and low-range (#5 and #14 IV, resp ectively) fire groups. Here, the impacts of fuelbed continuity were diminished and the fuel clusters coupled with structure influenced temperature variability. The fuel clusters were well dispersed among the fire groups, with 8-11 of the clusters found in each gr oup. Those found in one group and not another were the clusters that only consisted of a handful of samples that may have ex isted in one plot alone. As such, there was no distinction of specific fuel clusters with a particul ar fire group. The categorical fuel cluster variables were nonetheless signifi cant among the fire group models and often out performed many of the continuous LIDAR variables. Presence of particular fuel types was distinguishable between the fire groups, although they were less prominent than at the collective le vel. The most prominent fuel type metrics within each model ranged in IVs from 5.4 to 27.0 and 20.1 to 34.9 in the fire groups and collective models, respectively. Wiregrass was a comparatively stronger driver in the high-range fire groups (maximum IV = 21.6), while the results were mixed for the other two groups. Bare soil was a dominant fuel metric for Tmax fo r both high (IV = 9.0) and middle-range (IV = 13.3) fire groups as well as T500 (IV = 6.4, low group). It would seem residence time would be more sensitive to areas of bare soil th an temperature (as seen in the i ndividual plots below). On the other hand, maximum temperature re adings could be influenced by bare soil in especially high intensity fire areas, where fluctuations in temp erature were most likel y higher (Table 3-4). Otherwise the relative importance between single fuel types was unclear, with some strong influences from graminoids, perche d pine and oak litter, and forb s (Table A-2). Although pine litter (including perched litter) was less significant a driver, it was interesting to note that the percent abundance of each was di rectly related to the associated fire group. For instance, the
79 abundance of pine litter and perched pine litter was highest in high-range fire group (e.g., 69% and 47%, respectively compared to 38 to 50% and 32 to 34%, respectively for other groups, Table 3-7). The variability in pine litter abundance was most likely less influential at the grouped level because the variability was dissipa ted by clustering the plots by fire intensity. Relative to other fuel types, pine-litter a bundance and distribution dominantly impacts fire intensity within this system, with their high surface area to volume ratio and high-resin content (Hendricks et al. 2002). This fuel type however, cannot be a ccurately measured with LIDAR instrumentation because of its implicit structure within the fuelbed and compaction near the ground, supporting the need to incorporate fuel types or part icular biomass measurements into this type of fire behavior analysis. As surf ace leaf litter, it mats th e ground within only a few centimeters, making it virtually indistinct to ground measurements, even using ground-based LIDAR. As perched pine needles, it is essentiall y draped over and intert wined with other plants, becoming a part of their architecture. Some of this was most likely incorporated into the volume or point density estimates measured from groundLIDAR, but the disparity was still unclear. The clustering of fuels (i.e., using the fuel cl uster metric) was ultimately the most effective approach to capturing the complex heterogeneity in fuel types for analyzing with surface fire behavior. Regression-tree models: individual plots Using the in dividual plots, examining model fit within a particular small-scale fire event was possible, minimally influenced by changes in local weather or fuel state. The mean wind speed values within the three individual plots ranged between 1.2 and 2.0 m/s. All models were rather strong (R2 = 0.60 to 0.84, RPD = 1.66 to 2.27, Table 3-8), while results were intermediate and more variable between the collective and gro uped models. The chosen individual high-range plot generally performed better than the middle or low-range group, similar to the grouped data
80 findings. Interestingly, although th e variation in fuels was simila r to the grouped data, model fit varied more and was often below that of the grouped data: Only five of the fifteen model runs had RPD values above two. The disparity may be explained by the ability of the model to handle extreme values or outliers. When other plots had comparable outliers, especially within a characteristic fire group, then the associated variability was smoothed, driving correlations and improving predictive models. Hence, stronger m odels may be evident using grouped data of particular fire traits, while i ndividual plots may vary consider ably depending on the local fuel, weather, and fire environment. Table 3-8. Results of CART analysis for indivi dual plots (n = 169 each) within the three fire groups, including those with x,y coordinate s included as inputs. Temperatures (Tmax, Q90) are in C. Residence times (T300, T500) are in number of four second intervals. High-range individual plot Response variable Tn CVRE RRE R2 RMSE RPD Tmax 27 0.457 0.064 0.81 23.08 2.27 Q90 25 0.513 0.115 0.80 17.74 2.26 T300 24 0.485 0.112 0.68 0.64 1.78 T300 (x,y) 25 0.442 0.111 0.74 0.58 1.95 T500 29 0.598 0.104 0.76 0.44 2.03 Middle-range individual plot Response variable Tn CVRE RRE R2 RMSE RPD Tmax 25 0.566 0.119 0.74 27.95 1.95 Q90 22 0.645 0.127 0.64 30.34 1.66 Q90 (x,y) 21 0.523 0.162 0.72 27.10 1.86 T300 25 0.475 0.109 0.60 1.54 1.58 T500 27 0.438 0.115 0.61 0.72 1.59 Low-range individual plot Response variable Tn CVRE RRE R2 RMSE RPD Tmax 25 0.404 0.101 0.79 37.02 2.20 Tmax (x,y) 25 0.321 0.086 0.84 33.20 2.45 Q90 19 0.406 0.092 0.75 7.80 1.71 T300 24 0.608 0.191 0.70 0.74 1.81 T500 13 0.665 0.276 0.76 0.33 2.01 Tn = no. of terminal nodes, CVRE = cro ss-validation relative error, RRE = resubstitution relative error, R2 = coefficient of determination, RMSE = root mean square error, RPD = resi dual prediction deviation
81 The predictive power of each fuel metric was less evident between individual plot models than among the fire groups (Table 3-9). As not ed earlier, mean fuelbed depth (LIDAR height) was one of the most significant drivers: Nine of the twelve individual mode ls had mean height in the top-five of the IV rankings. In the low-rang e plot, it ranked within the top three IVs for all models. Point density was a strong driver for all models in the high-range fire plot (within top five ranked IVs), relatively cons istent with the grouped result s. Point density (and sum of heights) played a stronger role in the two lower intensity rang ed plots than within the grouped results. This suggests that plot level fuel continuity may be more sensitive at the plot level. As such, the height distribution metrics (e.g., height variance, skewness, and kurtosis) were not as significant at the plot level as they were at the multi-plot level for the middl e and low-range data. The two models for T500 were however, consistent with the middle and low ranges, in relation to point density and height distribution metrics. Fuelbed con tinuity was quite distinguishable between plots (Figure 3-6). The high-range plot displayed fuels th at had little to no disruptions in fuelbed structure and heights across the plot (Table 3-7A). The fuelbed becomes more patchy within the middle-range plot (Figure 3-6B), while large variations in height as well as patchy fuels were found in the high-range plot (Figure 3-6C ). This further supp orts that changes in spatial fuelbed structure had an apparent impact on fire behavior with in and between plots. LIDAR intensity metrics were moderate to hi gh predictors across all plots (Table 3-9), signifying their importance as a fuel metric and use for fire behavior modeling. Seven of the twelve individual models (without x,y) had at least one of the LI DAR metrics within the top five ranked IVs. All three (x,y) individual models ha s a LIDAR intensity value within the top five ranked IVs. Unfortunately, deciphering the valu e of LIDAR intensity in association with a particular fire metric was still difficult.
82 Table 3-9. Top five importance value (IV) rankings of input variables from CART analysis at the individual plot level. Te mperature (Tmax, Q90) are in C. Residence time (T300, T500) are in number of four second intervals. Response Variables High-range plot (5 plots, n = 845) Top five ranked fuel metrics (IVs) Tmax Point density (100), Int-sqrt (81.8), Int-lnr (81.3), Sum of hts (60.1), Intqdrt(57.6) Q90 Spread of hts (100), Sum of hts (87.2), Ht dist. Ratio (77.2), Point density (76.8), Int-lnr (66.8) T300 Point density (100), Int-lnr (91.8), Int-qdrt (85.9), Int-sqrt (82.8), Mean ht (46.4) T300 (x,y) Ht variance (100), Point density (97.3), Int-qdrt (89.9), Spread of hts (79.9), Y coordinate (78.9) T500 Ht dist. Ratio (100), Ht kurtosis (86.7), Point density (79.6), Mean ht (67.9), Ht variance (67.5) Mid-range plot (9 plots, n = 1521) Tmax Sum of hts (100), Mean ht (91.3), Ht skewness (66.2), Point density (64.9), Ht variance (44.9) Q90 Max ht (100), Mean LIDAR intensity (93.9), Mean ht (88.3), Spread of hts (79.4), Sum of hts (77.7) Q90 (x,y) Y coordinate (100), Mean LIDAR intensity (90.7), Ht skewness (71.6), Max ht (68.3), Spread of hts (66.1) T300 Ht dist. Ratio (100), Point density (96.3), Spread of hts (82.4), Mean ht (76.4), Int-lnr (74.3) T500 Ht variance (100), Ht dist. ratio (76.0), Ht skewness (74.4), Ht kurtosis (71.2), Spread of hts (65.7) Low-range plot (6 plots, n = 1690) Tmax Mean ht (100), Spread of hts (87.0), Max ht (77.9), Ht kurtosis (69.0), Ht skewness (60.6) Tmax (x,y) X coordinate (100), Ht kurtosis (58.8), Ht skewness (57.7), Spread of hts (49.2), Int-lnr (48.8) Q90 Mean ht (100), Point density (65.5), Ht variance (62.8), Sum of hts (59.7), Max ht (57.8) T300 Mean ht (100), Sum of hts (67.4), Mean ht (62.6), Int-sqrt (60.1), Intqdrt (56.1) T500 Ht skewness (100), Mean ht (98.4), Ht kurtosis, Mean LIDAR intensity (61.9), Ht dist. Ratio (53.0) Analysis at the plot level reve als the significance of particular fuel types influencing fire behavior mechanisms not relevant in larger scale analysis. For in stance, bare soil, a very small portion of the overall fuel types (< 1% abundance) played an impor tant role for residence time for those plots containing bare so il patches (high-range and low-range plots, Table A-2). For the high-range plot, the bare-soil metric (#8 IV) was especially strong for T300, outperforming
83 various LIDAR height metrics. The 10 and 100hour fuel metrics were more pronounced at the plot level as well, also demons trating significant importance for both Q90 and residence time. Ten-hour fuels were relevant for Q90 (middle ra nge plot, IV = 14.1) and T300 (high-range plot, IV = 15), while 100-hour fuels were relevant for Q90 (IV = 19.3) and T300 (IV = 8.9) in the low-range plot. The significance of ten or 100 hour fuels were not found for the grouped analysis. Otherwise, wiregrass, forbs, various li tter types, and graminoids played relevant roles similar to the multi-plot analyses. Spatial configuration (x,y coordinates) of the fire proved to be an important predictor of fire behavior, demonstrated by the improvements ma de by each of the three chosen (x,y) models (Table 3-8, initial RPD = 1.66 to 2.20 and improve d RPD = 1.86 to 2.45). The improvement was especially prominent for initially weaker models (e.g., Q90, RPD improved from 1.66 to 1.86 and R2 from 0.64 to 0.72). Here wind direction or neighborhood fuel combustion properties most likely influenced fire beha vior more than fuel structure and type. One of the spatial coordinates was found in the top five IVs for all three (x,y) models (Tab le 3-9), which changed the predictive ability (IVs) of so me of the other variables. This caused a spatial masking of certain fuel characteristics. For instance, the significance of bare soil in the T300 model was diminished in the T300 (x,y) model [i.e., IV chan ged from # 8 (35.8) to # 15 (19.9)]. Similarly, wiregrass and graminoids were diminished (to a lesser degree) in the Q90 (x,y) and Tmax (x,y) models. This also caused other fuel characteristics to be revealed or increase in importance. For example, deciduous oak litter was simply not a predictor for Q90 until the x,y values were included [i.e., IV ranking (value) changed from #17 (0) to #12 (17.1)]. Here it became the most significant fuel type metric. The same occurred for forbs in the Tmax (x,y) model. Similarly, some LIDAR metrics, such as height variance in the T300 model signif icantly changed in
84 A B C Figure 3-6. Changes in fuelbed continuity among the three plots (4 m x 4 m) modeled individually, namely A) the high-range plot (max height: 1.5 m), B) the middle-range plot (max ht: 1.5 m), and C) the low-range plot (max ht: 3.9 m). Note the loss in horizontal and vertical connec tivity and increase in patchiness of fuels from the highrange to low-range plot (A to C). The co lor gradient represen ts vegetation height distribution, which aids in visualization of the 3D vegetation structure. importance [from # 12 (21.0) to #1 (100)] for T3 00. On the other hand, some metrics were consistent in their signif icance (e.g., point density and intqdrt for T300, mean LIDAR intensity for Q90). These results signify the importance of incorporating the spatial configuration of fire
85 movement (spatial autocorrelation) when assessing the relationships between fuels and fire and the potential pitfalls when interpreting correlations. It is important to note that comparing importa nce values between models, principally those with varying terminal nodes and input variable s (e.g., fire group vs. individual plot) should be done carefully. The IVs are rela tive rankings specific to a given tree. These rankings may change dramatically by includi ng or excluding just one input variable, as seen by comparing the multi-plot groups to the individual plots. This changes relationships between variables and their predictive power within the model. Comparison of temperature models (Tmax vs. Q90) Tm ax outperformed Q90 in th e collective datase t (RPD = 1.24 vs. 1.11, respectively) and in all three individual plots (RPD = 1.95 to 2.27 vs. 1.66 to 2.26, respectively), but the opposite was true for the three fire groups. In the coll ective dataset, the discrepancy was most likely attributed to the clumping effect that occurr ed with Q90 (discussed earlier), due to the constrained lower limit on temperature. The fire groups appeared to allow enough variability to be expressed within the Q90 te mperatures across several plots, and were not restricted to variability within one plot. Q90 was crea ted to distinguish between unburned patches and smoldering fire (Hiers et al. 2009) This may be most evident across several plots with similar fire characteristics (hence, gr ouped plots), where there may be sufficient, yet distinguishable variability to model Q90. The comparison between the Tmax and Q90 models in the low-range fire group was however somewhat unclear, as the R2 and RMSE were lower for Q90, but the RPD was higher. This disparity was caused by th e extremely low range of temperatures across the low Q90 dataset (300 to 450 C) that suppressed the RPD valu e. In comparison, the middle and high-range fire groups had Q90 temperatur e ranged up to about 550 a nd 620C, respectively.
86 There were little similarities in fuel metric association betw een Tmax and Q90 at any level of analysis. The discrepancy may be explained by what the two temperature measurements represented. Tmax represented th e upper limit of temperature found within an area, regardless of how long it burned. On the other hand, Q90 diff erentiated more precisely the unburned and smoldering vegetation, representing a coupling betw een temperature and residence time (Hiers et al. 2009). This likely link between fuel types with temperatur e was best illustrated by the particularly strong model for Q90 (high-range group, R2 = 0.88, RPD = 2.86), where the fuel clusters were the most significant predictor. Both Tmax and Q90 were nonetheless important fire metrics to utilize as response variables with distinct uses and fuel metric associations. Comparison of residence time models (T300 vs. T500) The fuel m etrics were able to predict T300 better than T500 for th e grouped (RPD = 2.18 2.92 vs. 2.08 2.41, respectively) a nd collective (RPD 1.46 vs. 0.90, respectively) datasets. T300 performed especially well (Figure 3-5) in the high-range fire group (RPD = 2.92). Model fit between T300 and T500 was quite similar at the plot level analysis (Tab le 3-8). There were more similarities in fuel metric associations between residence time models than temperature models, as the two metrics were distinguishable simply by smoldering intensity. Mean fuelbed height was a strong predictor in almost all re sidence time models, while other metrics were similar among particular multi-plot or individual plot models (e.g., fuel clusters, point density, height distribution ratio). Us ing temperature thresholds was deemed important for estimating various residence time intensities. Evaluation of Potential Pitfalls Potential pitf alls that may influence model fit and intrinsic fuel metric relationships include those associated with non-modeled fire behavi or drivers, instrument ation limitations, and complexity of the regression-tree modeling appro ach. Typical fire behavior simulation modeling
87 incorporates a multitude of fuel characteristics as well as fuel and w eather conditions (Andrews et al. 2004), many of which were not used in this study. These models are designed for stand to landscape level fire behavior however, where cha nges in weather and fuel conditions are most influential. Local and micro-scal e wind patterns and other weather variables as well as variation in fuel conditions (e.g., fuel moisture) were most certainly influential, esp ecially at the collective plot level. This was demonstrated by the low to moderate model fits (R2 = 0.22 to 0.54). Much of the plot-to-plot variability was however, most likely accounted for by the clustering of plots into fire groups. Within some of the plots, there were presumably significant wind direction shifts and intensity changes. Capturing these fine-scale dynamics was beyond the scope of this study, although the x,y coordinates were to some extent representative of this phenomenon. There was likely little to no change in fuel mo isture during the 3 to 5 minute fires, which was minimal time for significant relatively humidity changes (if any) to influence fuel moisture content. This virtually eliminates the need to incorporate fuel moisture conditions for plot level fire behavior analysis. Furthermore, the variatio n associated with fuel moisture content was also represented by the fuel type metrics, including the fuel clusters (at the multi-plot level). Fuel biomass estimates were not included in the mode l and it was not clear whether results (i.e., model fit) would considerably differ. Alt hough biomass cannot be directly measured by groundbased LIDAR, certain metrics may be representati ve of biomass. A particular fuels (groundbased LIDAR) laser point-density drives volum e estimation. These volumes can be highly correlated with biomass (and leaf area) estimate s (Loudermilk et al. 2009). Coupled with the various fuel type metrics, these LIDAR fuel metrics may be a proxy for fuel biomass. Furthermore, it may be unreasonable to perform intensive biomass (and fuel moisture) sampling to perhaps only increase the output model fits in crementally. It may prove useful however, to
88 test the significance of these potential drivers, such as within plot wind patterns, fuel biomass, and fuel moisture conditions to assess their relative associations with fine-scale fire behavior. LIDAR has been used successfully to measur e fuel structure metrics, but it has been limited in distinguishing between vegetation types. The distinction has mainly been driven by variation in structure. For instance, height thresholds have been used to classify the understory vegetation and tree overstory (R iano et al. 2003). Tree type s have been differentiated by individual tree canopy shape and size or when c oupled with remotely sensed imagery (Popescu et al. 2004). Differentiating by t ype in a complex fuelbed with physically intermingled plants and plant parts seems inherently difficult. LIDAR intensity values have been used to distinguish between fuel types by variation in leaf area a nd vegetation height (Su and Bork 2007, Wing et al. 2010), but no study has been found to distinguis h between vegetation types within the surface fuelbed. The ambiguity of the LIDAR intensity values was confounded by the difficulty in determining the direct link to a specific fuel char acteristic. The (current ) inability of LIDAR to distinguish between fuels and measure important fuel types, such as pine-l itter was unfortunate. For this study, the integration of fuel type s eased this shortcoming. A study linking LIDAR intensity values with specific fuelbed types or characteristics would be highly beneficial. The temperature restrictive settings and camera overhead positioning may have hindered the fire data captured by the FLIR. These relati vely low intensity prescribed burns were only restricted by the low temperature setting (300 C) on the camera, which impacted some of the Q90 models. The camera viewpoint was not direc tly over the fire, possibly skewing the spatial temperature readings. The readings were taken from the downwind side of the plot however, minimizing camera angle effects. In the end, the ability to use the FLIR for fine-scale fire behavior characterization outweighed these limitations.
89 There was some difficulty in interpreting predictability of the particular fuel metrics using this integrative regression-tree approach, mainly because of the large tree outputs and uncertainty with comparing IVs within and between mode ls. The CART outputs (especially in ARCing mode) resemble the black box approach similar to other non-linear complex models (e.g., neural networks). The inputs are related to th e response variable(s) by di stributing the variation throughout this black box, then providing a measure of model fit. Interp reting this black box can be challenging. Fortunately, regre ssion-trees have been found to be more easily interpreted than neural networks (McBratney et al. 2003), as the variation can be traced throughout the tree regardless of size. There are also alternative tree-based modeling approaches that may prove useful for reducing issues with model over-fitting and interpretation, especially while using the multiple regression-tree (ARCing) approach (To ng et al. 2003, Prasad et al. 2006). Ultimately, this study illustrated that the relationships between the fuelbed and fire environment was multifaceted and more simple linea r approaches would not suffice. Conclusions This study was the first integration of hi gh-resolution L IDAR and thermal imagery to assess fine-scale relationships of the forest surface fuelbed and fire behavior characteristics in longleaf pine woodlands. Based on the results, grouping the plots based on similar fire characteristics provided the most optimal models in terms of fit and predictive capability. Grouping the plots suppressed the impacts from plot-to-plot variation in fuel, fire and weather conditions that were not explicitly modeled (i.e., fuel moisture, fire in tensity, wind and relative humidity, respectively). The strongest model fit was found in the hi gh-range fire groups, specifically for Q90 and T300, where the fuel clusters were highly significant predictors. This illustrates the link between fuelbed heterogeneity and the associated complexity of the fire environment. Although grouped datasets were most promising in terms of model fit, the
90 collective dataset and individuals plots had their advantages. The collective models provided an overall assessment of the potential predictive power and drivers of fire behavior across plots. The individual plots revealed significance of specific fuel types (e.g., bareground, deciduous litter) as well as the spatial configuration of fire, which influences the prediction strength of some of the fuel metrics. Impacts from individual fuel types were more distinct at the collective level and individual plot level than within the fire groups. Mean fuelbed height wa s overall a significant variable, signifying its importance for temperature a nd residence time predictions. Otherwise, significance of other fuel metrics (i.e., heights, height distribu tion metrics, point density, and LIDAR intensity metrics) varied by model. Am ong the grouped and individual plots, the fuel metrics demonstrated the importance of fuel con tinuity (both horizontal and vertical) in driving fire behavior. The LIDAR metric s were often strong predictors, but it was still unclear how the LIDAR intensity metrics characterized the fuelbed and influenced fire behavior response. The grouped data presumably segregated the fuel ty pes that influenced fi re intensity, not only suppressing IVs among fuel types (compared to the collective data), but bringing out the sensitivity of the fuel clusters. Small-scale infl uences from distinct fuel types were only found at the plot level, signifying the importance of cap turing and examining the relationships of fuels and fire within one particular sm all-scale fire event. Pine-lit ter presence was strongest at the collective level, while grouping the pl ots suppressed this sensitivity. Implications This study is one step closer to understanding the com plex responses of vegetation to fire behavior and further establishing the concept of the ecology of fuels (M itchell et al. 2009) and the wildland fuel cell concept (Hiers et al 2009). Both address the difficulty of past approaches to connect fuel heter ogeneity to fire behavi or and fire behavior to fire effects,
91 specifically associated with long leaf pine woodlands. This study related fuel heterogeneity to fire behavior and has implications for extending knowledge for fine-scale fi re effects research. For instance, the fine spatial scale at which th e multitude of plant sp ecies coexist (up to 50 species per m2) that have been found in this system (Walker and Peet 1983, Kirkman et al. 2004) has been suggested to be driven by fire behavior hetero geneity at the same scale (Mitchell et al. 2006). As understory fuels, several species ha ve been documented as even modifying fire behavior (Rebertus et al. 1989). It has also been suggested that fire effects at the small scale include individual plant recruitment and death (Mitchell et al. 2006, Thaxton and Platt 2006), but the coupled implications on the la rger scale plant population is unknown. This study also has implications associated with ecosystem management, especially related to ecological forestry silvicultural prescriptions. Fine-fuels, in this case pine-needle cast, drive fire dynamics within a longleaf pine system by supplying a continuous flammable fuelbed. The removal of the overstory through timber extractio n disrupts (among other things) the subsequent distribution and abundance of pine litter, contributing to nega tive feedbacks between fuel continuity and fire behavior. The degree of change to the sm all-scale heterogeneity in the fuelbed is relatively unknown and could be examin ed through the techniques in this study. This could have implications on natural pine rege neration response to subsequent vegetative competition (i.e., established hardwoods) in newly cr eated forest gaps (i.e., from individual or group-selection techniques). The competitive (or possible facilitation) interactions between established pine seedlings and hardwood sprouts may be determin ed by the spatial and temporal changes in fuels (impacted by overstory removal) th at ultimately drive changes in fire behavior (Mitchell et al. 2006, Pecot et al. 2007). This new approach to relating fuel heterogeneity and
92 fire behavior has proved complex, yet promising especially for implementing with multi-level fire effects research.
93 CHAPTER 4 SIMULATING LONGLEAF PI NE AND HAR DWOOD DEMOGRAPHY, FOREST FUELS, AND FIRE USING A LA TTICE APPROACH Introduction Spatially and tem porally exp licit forest models provide a means to better understanding ecosystem processes and underlying mechanisms identifying knowledge gaps, as well as contributing to long-term management pl anning (Botkin 1993). Longleaf pine (Pinus palustris ) savannas, in particular, are important to model because they are complex systems with intricate interactions between plants and fr equent fire that are often over simplified. Furthermore, land managers routinely battle hardwood invasion on degraded sites and may struggle with understanding the interactions between longleaf pine, hardwoods, forest fuels, and their application of prescribed fire (Provencher et al. 2001). Applyi ng a modeling scheme with these demographic and environmental components can illuminate the causes and underlying mechanisms regulating ecosystem dynamics. Longleaf pine systems have been studied for over a century, but a comprehensive model of its natural vegetation dynamics ha s yet to be created. This is in part due to the complexity associated with fire and the slow long-term cha nges that impact the system over time. Longleaf pine is a long-lived slow grow ing species (up to 400 to 500 year s), which makes data collection and analysis of tree demography difficult (Platt et al. 1988). Seed produc tion is highly variable over time with seed-masting intervals ranging be tween 7 to 10 years (Myers 1990) or longer. Stand density, site productivity, and genetics also influence an individuals cone production (Boyer 1990). Only a few known data sources exist with observations of cone and seed production across the region (Boyer 1998). Also, th ere are not many reference sites available to gather information on the natural dynamics of the system. Most extant longleaf forests are
94 degraded, and even the most pri stine are likely second growth forests (Platt et al. 1988, Hartnett and Krofta 1989). Tree competition and its impacts on biological community succe ssion is a slow, but continuously changing, complex process. The fee dback mechanisms associated with fire are especially intricate. Tree litter and unde rstory vegetation influences fuel type, distribution, and continuity and subsequent fire behavior. Fee dbacks occur with fire frequency, intensity, and spread impacting vegetation demography, including longleaf pine recruitment success and hardwood survival. Fire in itself is a highly co mplex mechanism which is difficult to understand, with influences not only from the vegetati on makeup and condition, but weather and climate conditions as well. Applying a modeling scheme with these importa nt processes allows us to simulate and visualize them through time. Furthermore, one can analyze potential impacts from landscape change scenarios (e.g., altering fire frequency) without experimental changes to land management that are widely believed to lead to degradation of native longl eaf pine savanna sites. Pine and Hardwood Dynamics Tree dynam ics within longleaf pine sava nnas including survival, competition, and interactions with fire are critic al regulators of forest dynamics. Competition has been used to explain this system in terms of interand intr aspecies competition as well as indirect interactions related to vegetative fuel and fire characteristics. The physiognomy of longleaf pine savannas is usually characterized by the generally large separati on distance (often over 10 m) between adult longleaf pines, and recruits in gaps (Gagnon et al. 2004). The highly dispersed overstory may be a product of op timizing resource utilization (hi gh light environment, low soilnutrient and droughty soil conditions) coupled wi th a unique ability to withstand and even perpetuate fires.
95 Intra-species competition between longleaf pine adults on their recruits is an important process (Myers 1990, Grace and Platt 1995). Light availability may limit longleaf growth and survival (McGuire et al. 2001, Pecot et al. 2007). Needle cast from a dult trees indirectly influences nearby established seedlings by variat ions in fire intensit y (Grace and Platt 1995). Seedlings closer to adults tend to experience hotter fires (accumulation of highly flammable pine needle-litter); decreasing their chan ce of survival after fire. On the other hand, other more firesensitive species may be suppressed or remove d, decreasing inter-specie s competition effects, which may facilitate seedling surv ival. Seedlings farther from a dults or within forest gaps experience less intense fires, increasing their chan ce of survival after fire (Beckage and Stout 2000), but are more likely to experience competition from faster growing hardwoods (Rebertus et al. 1989, Pecot et al. 2007). The adult pines may also facilitate seedling survival by providing shade during drought conditions by potentially reducing plant temperatures and water-stress (Myers 1990, Pecot et al. 2007). The inter-species competition between hardwoods and longleaf pine is an important management issue in longleaf pine savannas. Ha rdwoods are often consider ed pests because of their vigorous growth, re-sprouting after fire and strong competition with longleaf pine. Hardwoods are a fundamental part of this syst em however, provide food for forest animals [e.g., fox squirrel ( Sciurus niger)] and generally co-exist well with longleaf pine in a natural and consistent fire regime (Jacqmain et al. 1999). With a frequent fire regime, hardwoods are f ound as shrub-like plants scattered throughout the grass and forb understory and they may be concentrated in ar eas or patches that experience little or less intense fires. These are often in forest gaps, where longleaf seedlings may also have established. Hardwoods are suppressed closer to gap edges or near adu lt longleaf pines for two
96 reasons. First, there is more inter-species competition for belo w-ground resources (soil nutrients and water) between hardwood sprouts and adult pi nes. Light competition does not seem to be a significant driver of competi tion between many southeast hard wood shrubs and longleaf pine adults (Pecot et al. 2007). Fire intensity is often higher near a dult pines, where needle-litter accumulation is high, causing higher top-kill probabil ity near adult pines. Top-killed hardwoods are likely to re-sprout many times, with a s eemingly endless amount of underground reserves (Williamson and Black 1981, Rebertus et al. 1989, Guerin 1993). Longleaf pines seedlings are vulnerable to comp etition from hardwoods as they allocate resources for tap-root growth and do not initia te height growth for many years (Myers 1990). During this time, hardwoods use water and nutri ents; grow vigorously and spread through rhizomes, possibly shading-out pine seedlings over just a few year s. Fuel type and structure changes in these hardwood patches as well, reduci ng fuel continuity and he nce fire intensity and spread (Mitchell et al. 2006). If pine-needle cast and grasses cannot carry fire into these hardwood patches, hardwood growth is further enhanced and they may continue to suppress fires and possibly pine seedling growth and survival. Forest gap size is therefore important when understanding these intraand inter-species competition dynamics. Ideal gap size has been de bated (Brockway and Outcalt 1998, Pecot et al. 2007) in terms of optimal pine recruitment grow th and survival as well as hardwood shrub stature. It ultimately depends on site conditions and land management goals. With fire suppression over several decades, the ecosystem ma y pass a successional threshold and transition into a hardwood dominated forest-type, such as turkey oak barrens, oak scrub, or hardwood hammock (Myers, 1990). Although longleaf pine adults may live through decades, possibly centuries of fire suppression, the forest cannot perp etuate itself within 1 to 2 decades after fires
97 cease (Hartnett and Krofta 1989) because their larg e seeds cannot penetrate the understory brush and reach bare-mineral soil (Rebertus et al. 1989, Myers 1990, Glitzenstein et al. 1995). If a fire does occur after a long period of fire suppression, higher than usual mortality rates of the remaining adults may occur. Fine roots that ha ve grown into the accumulated duff layer at the base of a pine are susceptible to lethal temperatures from duf f smoldering (Varner et al. 2005). Fuel and Fire Dynamics Longleaf pine forests are inextricably linked to fire dynam ics (Agee 1998) and represent an archetype of fire dependent ecosystems. They ar e known to have among the shortest fire return intervals of any forest system in the world, with fire return times ranging from 1 to 5 years, extending to 10 years on some xeric sites a nd above 10 years in mixed-species forests (Christensen 1981, Bridges 1989, Abrahamson a nd Hartnett 1990, Ware et al. 1993). This high fire frequency struct ures the vegetation and thereby the fuels, perpetuating a lowintensity surface fire regime with a diverse but continuous fuelbe d (Hiers et al. 2009) dominated by flammable pine needles, small shrubs, grasses, forbs, and small diameter dead woody debris (Ottmar et al. 2003). Longleaf pine needles ar e an especially important component of the fuelbed with their high resin content and physic al structure (Hendric ks et al. 2002). The continuous flammable grassy understory and pine n eedle cast creates a fuel bed that carries a lowintensity fire across the landscape. Wiregrass (e.g., Aristida stricta, A. beyr ichiana, Sporobolus junceus ) often carpets the understory in undisturbed savannas and creates a unique fuelbed stru cture that perpetuates fire. The long, thin leaf blades (~0.5 to 1 m in height) intertwine with fallen pi ne needles creating an oxygen rich environment with easily ignitable dry fuels. Burning also promotes fresh wiregrass regrowth and blooming. It is relatively sensitive to shade from taller competing plants (i.e., longleaf pine and hardwoods) (Mulligan et al. 2002).
98 The hardwoods shrubs that are inter-dispersed in this fuelbed also add to the leaf and woody stem fuels, with some oak species having similar leaf litter burn characteristics as longleaf pine (Kane et al. 2008). With a natural fire frequency, hardwoods remain in the shrub state, often below just a few mete rs in height (Guerin 1993). If fire frequency or intensity is reduced, these hardwoods may grow rapidly to develop a woody midstory (Williamson and Black 1981, Provencher et al. 2001). This positive feedback loop creates a competitive environment supporting hardwoods, while suppressing pine recruitment and other light sensitive plant species (Kirkman et al. 2004). A less flam mable environment is created with the lack of fire facilitating plant species (grasses, low-shrubs ) and lack of pine-needle litter. This positive feedback loop may explain forest gaps where oak domes or hardwood patches will develop due to the lack of fuel continuity and fire intensity (Williamson and Black 1981, Guerin 1993). Ultimately, the competitive interactions among plant species have complex effects on fire dynamics that structure this community. Modeling Longleaf Pine and Hardwoods Modeling interactions between longleaf pine, hardwoods, a nd fire helps to test our understanding of the system and provides an opp ortunity to discover gaps in the current knowledge pool of these unique savannas. Existing longleaf pine models typically simulate timber production, population dynamics, or spat ial ecology. Stem wood growth has been modeled mainly for evenor uneven-aged pine plantations as well as some self-regenerating sites. These growth and yield equations aid plantation managers in estimating timber revenues. The equations are based on site index, stand density, stand age, and tree heights (Boyer 1983, Farrar 1985, Kush et al. 2006). Growth and yield models may include density effects (Quicke et al. 1994) and climatic factors as well (Meldahl et al. 1998).
99 Basal area, growth, and surviv al equations for individual long leaf pine trees have been developed from an extensive database of naturally regenerated even-aged stands (Quicke et al. 1994, 1997). Although important for plantation management, these equations may not be applicable for management goals that include ma intaining or enhancing ecosystem services. For this, one must take into account the complex in teractions and processe s that influence pine establishment, growth, and survival, as well as exchanges with the other species in the community. Three matrix population models have been used to represent the si ze-specific demography of longleaf pines. (Kaiser 1996) created a size -class matrix population model of longleaf pine from a mixed pine-oak forest in Texas. Both height and dbh (diameter at breast height) size classes were used to represent growth, survival, and fecundity. Three height classes were used for seedlings under breast height (1.47 m), while se ven dbh classes were used for the larger trees, up to a > 50 cm class. Another size-clas s matrix population model was developed from a representative old-growth longleaf pine woodland in Georgia (Pla tt et al. 1988). Thousands of trees were measured to generate growth and su rvival rates in eight dbh size classes ranging from 2 to 10 cm up to a > 70 cm class over four years, but fecundity estimates were not made. Platts model was extended to incorporate a seedling size class (0 to 2 cm ), fecundity estimates, densitydependent seedling survival and growth, and a va riable fire regime (Cropper and Loudermilk 2006, Loudermilk and Cropper 2007). A limitation of the standard matrix population model is that it assumes that demographic parameters (e.g., growth) are constants, independent of tree density. Also, longleaf pine systems are very sensitive to spatial dynamics associated with fuel distribution and fire, which cannot be captured using this modeling approach. Loudermilk and Cropper (2007) were able to
100 incorporate a fire regime however, by creating an alternate fire-suppressed longleaf pine matrix by simply eliminating fecundity during fire suppr ession. The fire-inheren t and alternate firesuppressed matrix model could be run in any sequen ce, but restricted to f our year intervals. Matrix population models have no explicit spatial or temporal hetero geneity associated with fire suppression and intensity (e.g., fuel accumula tion, distribution) or tree demography (e.g., competition, succession). A spatially-explicit forest modeling scheme can significantly improve model realism and has been successful in many pl ant systems (Higgins et al. 200 1, He et al. 2004, Miller 2005). Several spatial modeling techniques have been applied to longleaf pi ne and its biological community. Beginning at the fine scale, (Rat hbun and Cressie 1994) illustrated the importance of spatial structure within a longleaf pine population by modeling individual demographic process (birth, growth, and surv ival) using space-time equations. Each process was modeled for juvenile, sub-adult, and adult pine dbh size clas ses and was directly impacted by distance to or intensity of nearby sub-adult and adult tree competition as well as tree size. At a much larger scale, the LANDIS model (Mladenoff et al. 1996) was used to assess broad spatial and temporal succession dynami cs for a longleaf pine system (Loudermilk 2005, Loudermilk and Cropper 2007). Long-term 10 year age-cohort (presence/absence) dynamics of longleaf pine and three competitive oak species were simulated with two fire disturbance regimes using ecosystem-scale inputs. At an intermediate-scale, the Savanna Game model incorporated spatial dependence in a lattice-based longleaf pine model (Drake and Weishampel 2001). Cells (5 m x 5 m spatial resolution) consis ting of one longleaf pine each at a particular height were simulated yearly over a 25 ha area. This model accurately simulated the height structure and multifractal dimensions of a longleaf pine savanna when compared to in situ data
101 collected from aerial LIDAR (Light Detection and Ranging). Each cell interacted with its neighbors through seed production and dispersal, intra-species competition, fuel distribution (pine litter), and fire. Growth and mortality were influenced by tree size, competition, and fire intensity. A resource availability function was used to simulate intra-species competition: Heights of nearby competing trees limited resource availability. Fi re spread and intensity were a function of accumulated pine leaf litter. This modeling approach allows one to simulate complex space-time interactions using relativ ely basic probability-based functions. An important component of many longleaf pine savannas that has not been incorporated into these models is oak or ha rdwood dynamics. The modeling literature on southern hardwoods (found in longleaf pine savannas) is sparse and their demography has rarely been simulated. Age-height and height-dbh growth relationships of turkey oak ( Quercus laevis ) and sand post oak ( Q. margaretta ) were modeled alongside longleaf (Greenberg and Simmons 1999). A lifetable of turkey oak ag e-cohorts (0 to 10 up to 41 to 50 yrs.) was developed for a turkey oak dominated sandhill community (McGinty and Chri sty 1977). Although growth and survival may be inferred from these models, their spatial inte ractions with other species and fire are not explicit. Location of oaks in relation to nearby adult longleaf pines affects oak growth (Pecot et al. 2007) and survival after fire (Williamson and Black 1981), and recently the importance of incorporating basal spro uting in models has been rec ognized (Dietze an d Clark 2008). (Rebertus et al. 1989) applied this spatial dependence to turkey oaks in a longleaf pine sandhill community with survival and re-sprouting ability a function of oak size an d distance and size of the nearest adult longleaf pine. (Berg and Hamrick 1994) studied oak vegetative propagation
102 ability and modeled clonal propagation distance of two southern oak species ( Q. laevis, Q. Margaretta ). The LANDIS longleaf pine model (Loudermilk 2005, Loudermilk and Cropper 2007) is the only lattice-based spatial mode l incorporating oak dynamics with longleaf that we know of. Here, longleaf pine and three oak species (Q. laevis, Q. geminate, Q. hemisphaerica ) were used to simulate successional dynamics and the fire regime. With a 10 year time step and 20 m2 cell size, this model was not able to simulate finer-s cale interactions between plants, specifically competition feedbacks between hardwood sprouts a nd pine seedlings in relation to fire. There is clearly a need to incorporate hard wood dynamics into a spatially and temporally explicit longleaf pine de mographic model. The study objective was to develop a longleaf pinehardwood stochastic simulation model, incorporating tree demography, plant competition, and fuel and fire characteristics w ith spatially and temporally explicit components. Literature and data from two study sites were used to deve lop and evaluate the m odel with the goal to incorporate site specific calibration parameters fo r overall model versatility. Aerial LIDAR data and field measurements of population densities a nd height distributions were used for model evaluation. This model was used to identify scientific knowledge gaps of various population level ecosystem processes, specifically rela ted to population struct ure, fuel and fire heterogeneity, hardwood demography, a nd plant competition and facilitation. Study Areas Two longleaf pine savanna study areas were us ed for m odel simulation and evaluation. They vary in site characteristics, hardwood species composition, and management history. Utilizing two sites provides the opportunity to deve lop a more flexible model that may be used for a variety of longleaf pine savannas across the southeastern U.S. with minor calibration adjustments built into the model.
103 Ichauway Preserve One set of model param eters was estimated for the longleaf pine savanna community at Ichauway, an 11,700 ha reserve of the Jones Ec ological Research Center in southwestern Georgia, USA. Ichauway is located within the Plains and Wiregrass Plai ns subsections of the Lower Coastal Plain and Flatwoods section (McNab and Avers 1994). The humid subtropical climate (Christensen 1981) has a mean annual precipitation of 131 cm, occurring throughout the year. Mean daily temperatures ra nge from 21 to 34C in the summer and 5 to 17C in the winter. The soils are classified as excessively drained soils of the Orangeburn (fine-loamy, siliceous, thermic Typic Paleudults) and Wagram (loamy, siliceous, thermic Aren ic Paleudults) series (Battaglia et al. 2002). Ichauway has an extensive tract of sec ond-growth longleaf pine and has been managed with low intensity, dormant-season prescribed fires for at least 70 years at a frequency of about one to three years. The unde rstory of the study area was primarily composed of wiregrass ( Aristida stricta ) with many forb and prairie grass species. The overstory is almost entirely dominated by longleaf pine, with mature hardwoods making up a minor component (Palik and Pederson 1996). The hardwood species composition can be quite diverse, including Quercus incana, Q. falcata, Q. laevis, Q. stella ta, Q. virginiana, Q. hemisphaerica, Diospyros spp., Prunus spp, Sassafras albidum Ordway Swisher Biological Station Another set of m odel parameters was estimat ed for the longleaf pine savanna community at the Ordway-Swisher Biologica l Station (Ordway) a 3,800 ha rese rve in north-central Florida, USA. It is managed by the University of Floridas Department of Wildlife Ecology and Conservation. The humid, warm temperate clim ate has annual temperatures and precipitation averaging 20C and 143 cm, respectively. The soils in the study area were classified as very deep, excessively drained soils on sandy upla nds of the Candler (hyperthermic, uncoated
104 Lamellic Quartzipsamments) and Apopka (loamy, siliceous, subactive, hyperthermic Grossarenic Paleudults) series. Ordway has a large amount of second-growth longleaf pine as well and has been intensively managed for the past 30 years with prescribed fire. The current fire frequency is 2 to 5 years, with some areas reaching more than 10 years. The understory of the study area is primarily composed of wiregrass ( Aristida stricta ) with many forb and prairie grass species. The overstory is generally domina ted by longleaf pine, although th ese sandy hills are abundantly populated by turkey oaks ( Q. laevis ). Past bouts of fire-suppressi on allowed much of the turkey oak population to thrive and grow into the mid and overstory in much of the area. Although some areas have essentially become turkey oak barrens (Myers 1990) these were not found within the specific study area within the rese rve. Other oak or hardwood species found intermittently throughout the site include Q. virginiana, Q. geminata, Q. hemisphaerica, and Q. margaretta. Field Studies To im prove estimates of hardwood and pine s eedling survival after fire, a monitoring study was conducted at both Ordway and Ichauway. Twen ty five seedlings were planted in 16 random plots at Ordway and 27 random plots (4 m x 4 m) at Ichauway, both in January 2007. Prescribed burns were performed in March 2008 and June 2009 at Ichauway and Ordway, respectively. Seedling survival and size (RCD: r oot collar diameter) were measured after each s ites burn. In addition, hardwood survival and tr ee height (after the burn) was recorded at the Ordway across six randomly chosen forest gaps ranging in si ze from 0.05 to 0.18 ha. Within the 27 plots at Ichauway, hardwood survival, complete mortality ( not top-killed), stem density (of re-sprouts) were recorded after the burn. Clonal spread was also monitored for one year within the 4 m x 4 m plots. The results were used to improve model representation of height thresholds and rates of
105 hardwood fire survival. Seedling fire survival was simply used as an additional in situ dataset with which to evaluate with model outputs. LLM (Longleaf Model): Model Development and Description The longleaf pine and hardwood population model (LLM) developed for this study was a spatially and temporally explicit stochastic model. A lattice-based approach was used, similar to cellular automata (Silv ertown et al. 1992), where trees within cells (5 m x 5 m) interacted with nearby cells (trees) based on adjacency and distance. The model may also be classified as an Individual Based Model (DeAngelis and Mooij 2005) with each tree in the population simulated over time. The python programming language (v. 2.5.4, Python Software Foundation) was used for model development. A reasonably comp act area (125 m x 125 m; 1.56 ha) was modeled using a torus shaped landscape, eliminating possibl e edge effects. Spatia l interactions included seed dispersal or clonal rhizomatous spreadi ng (fecundity), interand intra-species plant competition impacts on growth and mortality, as well as effects of tree density on fine-fuel distribution and accumulation (Figure 4-1). The LLM had a one year time step, with yearly probabilities for fire occurrence a nd pine seed masting events. Both longleaf pines and hardwoods could exist in each cell. During the reproductive stage, up to ten trees of each type (20 total) could establ ish in an empty cell, wi th individual mortality incidences causing a natural thinning effect as the cell or trees within a cell age. Growth, in height (m), occurred at the cell scale, where all trees of a tree type within a cel l grew and aged together. Mortality for an indi vidual tree could be from comp etition, fire, or other natural causes. All trees were classifi ed as juveniles, subadults, a nd adults similar to Drake and Weishampel (2001). Model growth and population size distributions were driven by the site index input for each tree type, representative of site-specific characteristics (Table 4-1).
106 Figure 4-1. Conceptual diagram of longleaf pine-hardwood population dynamics and LLM (Longleaf Model) components. A random landscape consisting of longleaf pi nes (0.1 to 30 m in height) and hardwood sprouts (0.5 m size only) was used to initialize the model. The in itial longleaf pine population was the same for every run, which consisted of 72 trees 10 to 30 m and 190 trees below 10 m in height. The hardwood community was newly created for every run, which consisted of randomly placed hardwoods 0.5 m in height acro ss 10% (~63 cells) of the landscape. Model dynamic stability was reached at about year 100 to 150, characterized by lower amplitude oscillation. This transient be havior is common for models, wh ere the inherent stochasticity associated with the initial conditions is worke d out (Haefner 1996), where representative model processes can be examined. Calibration and sensitivity analysis of parameters is more appropriate after transient dynamics driven by initial conditions have dissipated.
107 Table 4-1. Input and calibrated parameters used for the LLM (Longleaf Model) and for two longleaf pine savanna study sites (Ichauway and Ordway) across two distinct fire regimes. LLM fire probabilities used ar e in parenthesis by each site name. CI = competition index; TDI = tree density index. Frequent fire regime Infrequent fire regime Ichauway (0.35) Ordway (0.25) Ichauway (0.05) Ordway (0.05) Parameters LP HW LP HW LP HW LP HW Site index SI (m) 35.0 25.0 25.0 20.0 35.0 25.0 25.0 20.0 Asymptotic height AH (m) a 45.5 32.5 32.5 26.0 45.5 32.5 32.5 26.0 CI parameter 0.07 0.095 0.075 0.10 0.074 0.08 0.075 0.07 TDI parameter 0.26 0.26 0.26 0.26 0.26 0.26 0.26 0.26 Fire mortality parameter 0.99 0.66 0.95 0.70 0.99 0.66 0.95 0.70 Adult size (m) b 23 17 17 13 23 17 17 13 Subadult size (m) c 11 8 8 7 11 8 8 7 aasymptotic height = SI x 1.3, badult size = AH x 0.51, csubadult size = AH x 0.25 Longleaf Pine Population Fecundity Seed and cone production of longleaf pine can vary highly between and within episodic m asting events that may be due to climatic, bi ological, and adaptation factors (Boyer and White 1989, Boyer 1998, Pederson et al. 1999, Koenig and Knops 2005). Masting is an important feature of longleaf pines, and is included as a stochastic element of the model based on data collected for over 40 years across the southeas tern U.S. (Boyer 1998). Each year had a 0.15 probability of being a mast year where only adult and subadult trees produced 40 to 125 cones and 10 to 70 cones, respectively (Table 4-2). During non-mast years, adults and subadults produced 1 to 40 cones and 1 to 20 cones, respec tively. Cone production was further varied by incorporating four cone production intensity leve ls for mast and non-mast years (Table 4-2). Juveniles did not produce any cones. In the LL M, all trees produced 32 seeds per cone (Boyer 1990).
108 Seeds were dispersed based on distance from a focal cell with subadult or adult trees and whether it was a mast year or not, similar to (Drake and Weishampel 2001). If it was a mast year, seeds were dispersed three cells and two ce lls from the focal cell for adults and subadults, respectively. If it was a non-mast year, seeds were dispersed two and one cell from the focal cell for adults and subadults, respectivel y. More seeds were dispersed in cells closer to the focal cell and in equal proportions within the ring of interest. For instance, if the focal cell (adult) produced 64 seeds in a non-mast year, each cell in the first closest ring of cells (8 cells total) received four seeds and each cell in the second closest ring of cells (16 cells total) received two seeds each. Table 4-2. Longleaf pine cone pr oduction categorized into four in tensity levels for subadults and adults for mast and non-mast years in th e LLM. All trees produced 32 seeds per cone. Cone production intensity level Mast year 1 2 3 4 Subadult 10 to 20 20 to 30 30 to 50 50 to 70 Adulta 40 to 50 50 to 70 70 to 90 90 to 125 Non-mast year Subadultb 1 to 5 5 to 10 10 to 15 15 to 20 Adulta,b 1 to 10 10 to 20 20 to 30 30 to 40 a Boyer (1998) and data from Boyer; b Platt et al. (1988) Germination rates are highly dependent on the ab ility of the large s eeds to reach the soil bed, free from immediate competition (Myers 1990). With fire as a means to remove understory vegetation and accumulated dead biomass, fire frequency essentially drives germination probability. Therefore seed germination ( GLP, Equation 4-1) was a function of time since fire ( tsft) at time t and a maximum germination probability ( Gp = 0.03). Fecundity was also restricted during fire suppression to simulate the transition of community st ate (i.e., longleaf pine savanna
109 to hardwood forest). When fires were suppres sed for twenty years in the LLM, germination was set to zero, representing a permanent change from the savanna state. (4-1) Growth Growth is often m odeled as function of potential growth with modifi ers (Quicke et al. 1994). For the LLM, longleaf pine height growth ( HtLP, Equation 4-2) within a cell was asymptotic and a function of age ( AgeLP). This potential growth was modified or suppressed based on the competitive environment ( CILP) of adjacent trees (see Competition Index). The asymptotic height ( AHtLP) was the maximum for every tree in the longleaf pine population. The asymptotic height was determined by a stands site index of longleaf pine tr ees (Table 4-1). This was estimated based on data from two natural longl eaf pine stands with qu ite different site index values (Ichauway: 35, Ordway: 25). This allows model application across areas that vary in productivity and other environmental factors (soil characteristics). Variation in site index provided distinct simulated growth cu rves between sites (Figure 4-2). (4-2) Furthermore, a dormant tag was implemented for each newly established seedling to represent the commonly found gra ss stage (Myers 1990). Each seedling remained in the grass stage for 1 to 10 years (average 5), while initia ting height growth using the equation above. For simplicity, all trees within a cell grew or remained in the grass stage together (i.e., same height values within a cell). Mortality Three m ortality functions were created for longleaf pine within the LLM; death from competition, natural causes, and fire. Each tree within a cell was evaluated individually for
110 A B Figure 4-2. Simulated growth curves for A) l ongleaf pine and B) hardwoods in two study sites (Ordway and Ichauway) using the LLM. mortality. Longleaf pines are sensitive to competition, especially when young (Pecot et al. 2007) and this competition may be a function of dist ance to adult trees (Brockway and Outcalt 1998) and understory vegetation (McGuire et al. 2001). A negative logistic function (Equation 4-3) was used to represent mortality from competition ( McLP). This function was dependent on tree size through the use of the competition index ( CILP: see Competition Index below), where adult trees were minimally influenced by competitive mortality effects. (4-3) 0 5 10 15 20 25 30 35 050100150200250Longleaf height (m) Ordway Ichauway 0 5 10 15 20 050100150200250Hardwood height (m)Age (years)
111 There are various natural causes of mortality be sides mortality from competition and fire in longleaf pine systems that are not explicitly m odeled in the LLM. To account for this, a size distribution of mortality rates were estimated from four year censu s data (Platt et al. 1988) and associated models on seedling mortality (Cropper and Loudermilk 2006, Loudermilk and Cropper 2007) to implement a mortality function for other natural causes. There are typically two peaks in mortality rate s across size classes; one peak for the small size classes (e.g., disease) and an other one in the large size cla sses (e.g., lightning, windthrow). This creates a bathtub shaped mortality curve. Two natural mortality functions were developed by relating mortality rates and th eir associated size classes usi ng Curve Expert (v. 1.38, Hixson, TN). The first four (~0 to 23 m height) and la st five size classes (~23 to > 34 m height) in (Loudermilk and Cropper 2007) were used for th e young and older natural mortality functions, respectively. The young natural mortality function created for longleaf pine (MnLPy, Equation 44) was best fit with a Weibull function. Here mortality rates ranged from approximately 0.1 to 0.004 y-1, with trees from germinants (0.1 m), through middle sizes, and up to the model adult size. The older natural mortality function created for longleaf pine (MnLPo, Equation 4-5) used a logistic function. (4-4) (4-5) Older mortality rates ranged from about 0.001 to 0.03 y-1 (max) for all adult sized longleaf pines. This peak in adult mortality seems to be common in longleaf pine woodlands (Palik and Pederson 1996, Outcalt 2008) and modeled similarly in (Kaiser 1996). Both mortality functions were a function of input tree height ( HtLPy for those below adult size or HtLPo for those adult size or larger).
112 Longleaf pine mortality from fire is dependent on tree size and fuel and fire characteristics that drive fire intensity (Grace and Platt 1995). For simplicity, longleaf pine fire mortality ( MfLP, Equation 4-6) was modeled following (Dra ke and Weishampel 2001), where fuel ( Fuel ) and tree size (HtLP) determine fire mortality probability. Here, fuel acts as a surrogate for fire intensity. Fuel was the summed biomass values for pine litter ( BLP), hardwood litter ( BHW), and wiregrass ( BWG) accumulated within each cell since the last fire. (See Fire and Fuel Component for biomass and fire details). Fire intensity affected mainly smaller pines, while larger pines were more apt to survive a fire with minimal influen ce from intensity levels. The longleaf pine fire mortality parameter (FPLP) was a value used for calibration w ith values ranging from zero to one (Table 4-1). (4-6) Competition index The com petition index ( CILP: Equation 4-7) was directly impacted by tree size within the focal cell and its neighboring trees with output values ranging from zero to one. A value closer to one simulated higher competition while those cl oser to zero simulated lower competition. The CI parameter (CPLP) and the asymptotic height value ( AHtLP) were used for model calibration (Table 4-1). The CPLP typically ranged from 0.05 to 0.1. With a lower CPLP value, growth increased and mortality from competition decreased (see Model Dynamics). (4-7) The sum height value (SumHt 1LP, Equation 4-8) was a func tion of tree heights of neighboring cells. Within a tw o cell radius (dista nce range: 2.5 m to 18.6 m), all adult and subadult trees or those larger than the focal cell were summed, incl uding both longleaf and hardwood trees. Each sum height value was adjusted by the number of trees (CtLP or CtHW) and
113 maximum number ( MaxLP or MaxHW) allowed within a cell fo r each tree type. Similar approaches have been used successfully in latt ice type models to represent neighborhood plant competition (Drake and Weishampel 2001). The sa me functions for competition (Equation 4-7) and sum height (Equation 4-8) we re used for hardwoods as well. (4-8) Hardwood Population Fecundity Fecundity of southern hardwoods is a produc t of acorns from mature trees and clonal rhizomatous sprouting (Abrahamson and Layne 2003) In a frequently burned longleaf pine savanna, the main reproducti on pathway is through clonal sprouting from young non-acorn baring shrubs (Guerin 1993). A lthough mature hardwoods are pres ent, they are few and most likely contribute little to acorn born trees. The LLM mode led fecundity based essentially on clonal sprouting, although this co uld be representative of seed germination from a nearby adult tree. Hardwood clonal spread distance has been found to be within 5 to 10 m. Genetic clonal pairs of two southern oak species ( Q. laevis, Q. Margaretta ) were frequently found within 5 m of each other, some ranging up to 10 m (Berg and Hamrick 1994), and found across many size classes (up to ~20 cm dbh). In the LLM, all hard woods were able to spread (average distance 7 m) to one randomly chosen empty (of other hardwoods) adjacent cell per time step. Clonal spread rate and vigor is especi ally difficult to determine and im portant controls are unknown. In the field monitoring study at Ichauway, an average of ten hardwood sprouts were found within a 16 m2 area plots, and no new clonal sprouts were found over one year, suggesting a relatively low spread rate. Using this info rmation for the LLM, 1 to 10 (randomly chosen) clones could spread into one adjacent cell of a hardwood at a probability spread rate of 0.05.
114 Growth Growth rates of southern hardwoods are not well known, but probably follow a sim ilar pattern in asymptotic growth as other tree species (Greenberg and Simmons 1999). To simulate hardwood height growth, the same gr owth pattern as longleaf pine was used for the larger trees. A distinct difference between hardwood and longleaf height growth occurs in the younger years. Young hardwood sprouts grow vigorously for seve ral years (assuming no disturbance) and may reach several meters in height quickly (Rebertu s et al. 1989). As such, young hardwoods in the LLM grew by 0.5 m per year for ages 1-3. Thereafter, hardwood growth ( HtHW, Equation 4-9) followed the same function as longleaf pine. Th e hardwood parameters are distinct from the longleaf pine equation (Equation 2-2) with unique asymptotic height and competition index values (Table 4-1). The ability for hardwoods to reach 1.5 m in three years was an advantage over longleaf pine in the mode l, especially with the implem entation of the longleaf dormant grass stage. For comparison, it took longleaf pine seedlings an average of 5.6 years (SD standard deviation = 2.1 years) to reach 1.5 m in height. (4-9) Mortality In a frequent fire regim e, complete mortality from fire is significantly less in southern hardwoods than for longleaf pine, as most sma ll or young hardwoods are top-killed and re-sprout from their root base. When larger, they have th ick fire-resistant bark to minimize effects from fire (Myers 1990). Hardwoods under 2 m tend to be top-killed (Williamson and Black 1981, Guerin 1993), most likely associated with flame heights for these low-in tensity fires (Robertson and Ostertag 2007). Similar results were found in this studys fi eld observations, where almost all (97%) of hardwoods below 2 m were top-killed and those th at survived were mainly above 2 m [similar to
115 (Guerin 1993)], suggesting an identi fiable threshold. Those that survived fire and were below 2 m were found in large forest gaps (> 0.25 ha), wh ere the altered fire environment hindered fire spread and intensity creating unburned patches within the understory. Using these field observations, most hardwoods below 2 m in the LLM were top-killed, with a 0.02 probability of complete survival and a 0.006 probability of complete mortality. For comparison, Williamson and Black (1981) reported a 0.03 probability of turk ey oak survival of those below 2.0 m, and Rebertus et al. (1989) modeled a 0.05 yearly mort ality probability for tu rkey oak sprouts. Those hardwoods top-killed in the LLM re-sprouted to a size of 0.5 m in one year and age was reset to one. This re-sprouting could cont inue over time in the model if the hardwood remained below 2 m (continuously top-killed). It has been noted that the supply of underground reserves seems endless for continuous regrowth after being top-killed (Williamson and Black 1981, Guerin 1993). For logistical ease, all hard woods below 2.0 m in a cell had the same fate regarding fire mortality, survival, or top-kill. All other mortality functions (for taller hardwoods and longleaf pine) checked individual trees for mortality. Hardwoods above 2.0 m are not often top-killed in a frequent fire regime. They may, however experience delayed mortality. These tr ees may be damaged by fi re, but its dominant apical meristem may still be intact. This apical domi nance may suppress sprouting from the base, while using its carbohydrate re serves for allocation to crown growth. True tree mortality may not occur for several years, even with some weak re-sprouts that may die along with the tree (Rebertus et al. 1989). The mi d-size classes have been found to experience the highest fire mortality (even if delayed) because of their incr eased susceptibility to bark or crown damage and reduced ability to re-sprout. As such, higher fire mortality rates were simulated in the middlesize classes, compared to those below 2 m (e .g., ~0.01 for < 2 m, 0.10 for 2-5 m). In the LLM,
116 the hardwood fire mortality ( MfHW, Equation 4-10) function was sim ilar to longleaf pine and only used for hardwoods over 2.0 m in height. The fuel ( Fuel ) and height ( HtHW) values followed the same description as for the longleaf fire mo rtality function. The hardwood fire mortality parameter ( FPHW) was a value used for calibration, ranging from zero to one (Table 4-1). (4-10) Southern hardwoods in longleaf pine savanna s may experience other natural causes of mortality (e.g., disease, fungal infections, inse cts, longevity) that ar e not explicitly modeled here. For simplicity, natural mortality ( MnHW: Equation 4-11) was modeled as a function of height ( HtHW). This linear function was restricted to a maximum probability of 0.05, applied only to trees above 2 m in height. This also ensures that hardwoods reach a reasonable maximum age (approximately 200 years in the LLM). Southern hardwoods have not been known to experience mortality directly from interspecies competition in a fire-m aintained longleaf community as their competitive abilities (e.g., nutrient uptake efficiency, shade tolerance) dom inate when fire frequency is reduced or eliminated entirely from the sy stem (Hartnett and Krofta 1989, Glitzenstein et al. 1995). As such, hardwood mortality from competition was not modeled here. Hardwood growth suppression from competition was simulated usi ng the same functional relationship as longleaf (Equation 4-7, 4-8). (4-11) Fire and Fuelbed Component Frequent fire (1-7 yrs.) is considered to be necessary in m ost longleaf pine savannas, and many managed systems have prescribed burn plans th at follow this regime (Glitzenstein et al. 2003). In the LLM, fire occurrence was stochastic with a yearly probability of 0.35 and 0.25 for
117 Ichauway and Ordway, respectively. For this stu dys purpose, fire size was equal to the extent (1.56 ha). Every cell burned completely and varied in intensity as a function of the accumulation of fuel biomass. If a fire o ccurred, all fuel was consumed (i.e ., reset to zero) in the entire study area. The fuelbed in longleaf savannas almost entirely consists of fine-fuels that feed and carry a fire, and their distribution is dependent on the spatial structure and density of the overstory (Mitchell et al. 2006). Pine a nd hardwood litter accumulation is more abundant under or near the tree and in higher tree density patches (Robe rtson and Ostertag 2007), while wiregrass is typically more abundant in areas with less oversto ry density due to increased light availability (Mulligan et al. 2002). In the LLM, fuel accumulation (kg/cell) was mode led for three of the main fine-fuel types found in longleaf pine savannas: pi ne needle leaf litter, hardwood leaf litter and wiregrass. The pine litter (BLP, Equation 4-12) and hardwood litter ( BHW, Equation 4-13) biomass functions were linear and created from literature values (Robertson and Ostertag 2007). (4-12) (4-13) Time since fire ( tsft) at time t and the tree density indices ( TDI ) (see descriptions below) ultimately drive biomass values and accumulation rates. Biomass accumulation began the year after a fire occurred and continued until another fi re occurred and biomass was reset to zero. The simulated longleaf pine litter biomass were in th e range of values for other studies with similar pine densities (Gresham 1982, Holder 2000). Pine litter accumulated much faster than hardwood litter in the model, attributable to the greater production of pine adul ts relative to the smaller, but also abundant shrub-size hardwoods that pr oduce less overall litter biomass (Robertson and
118 Ostertag 2007). This is reasonable assumpti on for a frequently burned savanna where this difference in tree size distribution betw een tree types is regulated by fire. The wiregrass biomass (BWG, Equation 4-14) accumulation function was asymptotic and was based on literature values (Mulligan et al. 2002). An asymptotic biomass value ( ABWG) limited fuel accumulation and was consistent with a decomposition based wiregrass equilibrium that may occur a few years after fire. The ABWG was estimated at 6.25 kg per cell, based on data from (Mulligan et al. 2002). (4-14) All three biomass functions were infl uenced by the tree density index (TDI Equation 415), a function of the density and size of nei ghboring trees. This function was used for both longleaf pines and hardwoods independently. The TDI parameter ( DP ) was used to calibrate for biomass accumulation. In the LLM, DP for both tree types was 0.26, w ith typical ranges from zero to one. A higher TDI value corresponded to either higher pine or hardwood litter biomass and less wiregrass abundance. TDI for wiregr ass was simply the summation of hardwood TDI and longleaf TDI. SumHt 2 (Equation 4-16) was created spec ifically for the TDI equation and was a function of tree heights of neighboring cells for either longleaf pines or hardwoods, including the focal cell. Within a two cell radi us, all adult and subadu lt trees were summed for either tree type. (4-15) (4-16) Each height value was adjusted by the number of trees ( Ct ) and maximum number ( Max) allowed within a cell for each tree type, similar to SumHt1 (Equation 4-8). The height of the focal cell ( Htf) was also included, regardless of size to ensure biomass accumulation from all size
119 classes within the focal cell. The Htf value was multiplied by three to simulate the stronger influence from biomass accumulation from the trees in the focal cell, compar ed to adjacent cells. Model Calibration and Evaluation Calibra tion defines the relationships within and between model functions essential for simulating realistic forest processes (Haefner 1996) and is particularly useful for modeling sitespecificity. The main variables used in the LLM calibration process includ ed the site index, CI parameter ( CPLP), fire mortality parameter ( FPLP and FPHW), and tree density index parameter ( DP ). Each site has unique parameters for both longleaf pine and hardwoods (Table 4-1). Using these parameters, the LLM was calibrated to main tain a subadult and adu lt longleaf pine (i.e., those trees > 10 m height or approximately 11 cm dbh) density of ~125 trees/ha ( SD) over time, mainly by adjusting the CI and fire mortalit y parameter. For comparison, densities (per ha) have been found to range widely, from 50 to 209 [trees > 5 cm dbh (Greenberg and Simmons 1999, Outcalt 2008)], 190 to 233 [trees > 10 cm dbh, (Penfound and Watkins 1937)], and 43 to 90 [trees > 20 cm dbh, (Grace and Platt 1995)]. Adult hardwood densities were calibrated to maintain a very small density (< 20/ha) when simulating a frequent fire regime. Seedling, juvenile, and subadult populations were not direc tly calibrated. The stand index was not calibrate d per se, but simply an input parameter that was site and tree type specific. The input site index values were estimated from aerial LIDAR and literature values. The canopy height model of the LIDAR dataset for each study area was used to approximate longleaf pine maximum canopy heig hts (35 m: Ichauway, 25 m: Ordway). Hardwood site index values were determined from literature values of measured tree height growth. Hardwoods were much less abundant as adults for each study site and were difficult to discern from the LIDAR datasets. From literature estimates, hardwoods at Ordway (i.e., more xeric type areas) had a site index of 20 m (H arlow 1990), while at Ichauway (i.e., more mesic
120 area) the site index was set at 25 m (Bela nger and Krinard 1990). Potential growth characteristics in the LLM were driven by these site index values (Figure 4-2). The CI parameter ( CPLP or CPHW) was a complex variable used to calibrate several model components, including growth characteristic s (associated with s ite index), longevity, competition, and population densities, wh ile the fire mortality parameter ( FPLP or FPHW) mainly influenced population densities by changes in fire mortality rates. Longleaf pine growth and competition were optimized by, for example, lowe ring the longleaf CI parameter to improve height growth and reduce mortality sensitivity from competition. Mortality from competition and fire was calibrated by assessing the change in population densities over time and later evaluated with independent LIDAR and field me asurements (see Model Dynamics). The CI parameter also influenced tree longevity by driv ing tree growth rates a nd competitive responses. The target longevity age for l ongleaf pine and hardwoods were approximately 200-400 years and 100-200 years, respectively. The fire mortality parameter (FPLP or FPHW) was calibrated in association with fire probability to balance longleaf and hardwood populat ion densities without affecting growth (i.e., through the CI parameter). The TDI parameter (impacting fuel accumulation) was used alongside the fire mortality para meter to monitor fire mortality impacts on population numbers, but the LLM was very insensit ive to changes in this para meter (see Model Dynamics). Ultimately, the same TDI parameter was used for all models (Table 4-1). The LLM was evaluated with independent in situ and aerial LIDAR data from both study sites to compare their estimated tree population densities and hei ght distributions. LIDAR data were collected on January 2007 and March 2008 for Ordway and Ichauway, respectively. At Ordway, the Optech ALTM (Airborne Laser Terrain Mapper) 1233 was used with a pulse
121 frequency of 33 kHz, recording first and last re turns (Table B-1). At Ichauway, the Optech Gemini was used with a pulse frequency of 125 kHz, recording first, second, third, and last returns (Table B-2), re sulting in denser laser point clouds than with the ALTM 1233 system. Twenty randomly selected 125 m x 125 m plots within each study site were created in a Geographic Information System (ArcGIS 9.2, ESRI Redlands, CA) to examine tree populations within the LIDAR dataset and in the field. An algorithmic technique, previously developed for slash pine plantations (Lee et al. 2010), was used to detect tree positions of canopy trees above 10 m in height using the raw LIDAR point data within each of the 20 plots separately. The algorithm takes a top-down approach to identifying treetops. It starts by finding the highest point in each of the segregated LIDAR datasets. The poin ts proximal (i.e., 3 m horizontally) to the first treetop are then removed from the LIDAR data set, resulting in a new subset. This process is repeated to identify additional treetops in progres sively smaller subsets and terminates when all treetops are found. To reduce er rors of commission, the algor ithm was restricted to only identifying trees above 10 m in height and at l east 3m apart. Once all trees were detected, tree heights were estimated by an adaptive multi-s cale filtering procedure (K ampa and Slatton 2004). Field data were subsequently collected at both field sites to compare with the LIDAR tree estimates and outputs from the LLM. Here, longleaf and hardwood population densities, including those below 10 m in height could be m easured. These were not estimated by using the LIDAR dataset. A fixed area stratified grid sampling approach (25 m spacing) was used to collect field data in five of the 20 plots measur ed by the aerial LIDAR for each site. At each sample point (16 per plot), heights were recorded for all adult tr ees (both longleaf and hardwoods) within a 10 m radius. Abundance of a ll longleaf seedlings or sapling below 10 m as well as hardwoods 2 to 10 m tall wa s also recorded at each samp le point. Hardwoods below 2 m
122 tall were not recorded to reduce bias and erro r from discerning individuals from vigorously sprouting plants (after fire) in dense areas. Longleaf pine and hardwood height and age distributions were recorded over ten LLM simulations to a ssess with the LIDAR and field estimates. LLM: Model Evaluation and Discussion Model Dynamics Fire, fecundity, and seed masting A frequent fire regim e creates an ideal environment for longl eaf pine seed germination and establishment. In the LLM, pine seedlings were able to establish in forest gaps, where no adult pines where present (Figure 4-3A, B). If hardwoods were present (Figure 4-3C), seedlings were more susceptible to competition. This is repres entative of the change in fire environment in forest gaps, where the lack of pine leaf litter may reduce fire intensity and promote hardwood survival and continued height growth. In this particular output (Figure 4-3C), hardwoods established in a gap area about 0.06 ha, the size of a small gap opening that could be from a dispersed retention or small aggregate rete ntion (i.e., single-tree selection) harvesting prescription (Palik et al. 2003). There is significant variation in observed hardwood survival in longleaf savannas. At Ordway, hardwood surviv al (mean = 120 per ha, stems above 2 m) was found in all six forest gaps ranging from 0.05 to 0.18 ha in size. At Ichauway, hardwood survival (mean = 20 per ha, stems above 2 m) was found in gaps ranging from 0.10 to 0.30 ha in size. The difference between sites was likely due to the disturbance history, current site condition, and current population structure. This (0.06 ha) gap size has been found to be sufficient for pine regeneration (Pecot et al. 200 7), where parent trees f acilitated survival of nearby seedlings.
123 A B C Figure 4-3. Spatial outputs from the LLM during a frequent fire regime illustrating A) longleaf pine tree height (m) distribution, B) the number of pines within each cell, and C) hardwood tree height (m) distri bution. The pink ci rcles indicate where pine seedlings established in a forest gap after a seed ma sting event. The orange circles indicate where seedlings were susceptible to competition from hardwood encroachment in a forest gap. Each colored square is a 5 m x 5 m cell (25 x 25 cells, 1.56 ha extent).
124 In the LLM, a 0.06 ha gap would be in the ne xt adjacent cell, less than 5 m separation distance. This was well within the exclusionary zone suggested by Brockway and Outcalt (1998). It is important to note that although se edlings may have successfully established nearby an adjacent parent pine tree in the LLM, s eedling growth and mort ality was significantly impacted by competition nearby these adults. The LLM realistically represents the complex spatial interactions that occur between longl eaf pine and hardwoods in relation to fire, particularly with respect to spatial patterns of pine regeneration success and competition. Successful pine seedling establishment is most likely to occur fo llowing a seed masting event (Brockway et al. 2006). As such, the te mporal dispersion of seeds, specifically the synchrony between a seed masting event and fire occurrence was also examined using the LLM. The synchrony of seed mast years (0.15 prob.) and fire years (0.35 prob.) was examined by monitoring their occurrences over ten 100 year runs. They synchronized an average of 6% (SD = 2%). An average of 15% of mast years a nd 6% of fire years were synchronized. It was reasonable to conclude that th is 6% synchrony was adequate fo r replenishing the population, as the reproductive success need only be minimal to ensure overall success of this long-lived species. Output cone production of the LLM was evaluate d with respect to long-term field data. Ten forty-year simulations of the LLM (masti ng probability = 0.15) were run to compare with Boyers (1998) forty-year collection of regiona l cone production data. Adult cone production in the LLM was averaged for each year (n = 10 per year). Cone production means (LLM and Boyers in situ data) were compared using a two samp le two-tailed equa l variance t-test ( = 0.05, n = 40). The results support that the LLM and in situ means were not significantly different from each other ( p = 0.76). Their similar pooled ch aracteristics and comparable
125 cumulative distributions (Figure 44) concluded that the LLM was consistent with observed cone production estimates. Figure 4-4. Cumulative distributi on functions of mean longleaf c one production over forty years for both Boyers regional in situ data (Boyer 1998) and simulated output from the LLM. Mortality among size classes Longleaf pine and hard wood relative annual mortality rates were compared across size classes to assess how each form of mortality impacts the populati on at various life stages within the LLM. Relative mortality between natural ca uses, competition (longleaf only), and fire was calculated within each size class for each fire regime. This was done for both longleaf and hardwoods. The Ichauway LLM was used exclusively here (for consistency) to compare mortality between a fire-frequented (0.35 fire probability) and fire-suppressed (0.05 fire probability) regime. Similar results however, we re found with the Ordway LLM. The results were consistent with known pine and hardwood mortality response. Within a frequent fire regime (Figure 4-5A), longleaf pine seedlings and saplings were more susceptible to fire than larger trees. Competition was more prominent in the mid-size classes where the trees were less susceptible to fire and competition may be more influential. In these (sapling to subadult) size 0 200 400 600 800 1000 1200 0510152025303540Cumulative Cone ProductionTime (yrs.) LLM Boyer
126 classes, trees may be seen clumped together in similar aged cohorts (most likely in a forest gap and from a prior masting event). A B Figure 4-5. Simulated size-specific relative mortality of longleaf pine s from competition, fire, and other natural causes for A) a frequent fi re regime and B) a fire suppressed fire regime. Relative mortality between natural causes, competition, and fire was calculated within each size class for each fire regime. Results were similar for both study sites (Ordway and Ichauway). 0.0 0.2 0.4 0.6 0.8 1.0 1311182327303234>34Relative mortality Competition Natural Fire 0.0 0.2 0.4 0.6 0.8 1.0 1311182327303234>34Relative mortalityHeight (m) size class
127 The midsized trees are tall enough to escape lethal fire temperatures, and are less susceptible to disease or infec tion that the younger (brown spot needle blight) and older trees (heart rot) may experience. They may begin to thin-out and eventually become more spatially dispersed as their older neighbors. At higher de nsities, competition may be a stronger cause of mortality than fire or natural causes. Seedlings were quite sensitive to competition as well (0 to 1 m size class). The larger trees were highly resistant to (low -intensity frequent) fire and competition, but perished from natural causes of mortality, such as old age, heart rot, windthrow, or lightning. With simulated fire suppression competition was the dominant driver of mortality among all but the very largest size classes (Figure 4-5B). Pine seedlings and saplings were negatively influenced by faster growing more shade tolera nt hardwoods. The largest trees were least susceptible to competition and may live on for ma ny years, eventually dying from other natural causes or a catastrophic fire event. Fire mo rtality was variable across size classes and the occasional fire with large amounts of accumulated fuel s killed even large pines. Little to no fire mortality was found in the 0 to 1 m size class be cause there were few seedlings to kill, as germination was ultimately eliminated in the LLM (after 20 years of fire suppression). Competition had such a strong impact on the grow th function (Equation 4-2) that trees never exceeded 34 m in height. Hardwoods were more susceptible to fire in the smaller size classes (Figure 4-6A). The smallest (1.5 m size class) hardwoods survived most fires by re-sprouting afterwards. As such, these top-killed hardwoods were not included here, minimizing rates of mortality by fire in this size class (although they were higher than natural morality rates). Larger hardwoods were more susceptible to natural mortality than by fire, via the linear natural mortality function (Equation 4-
128 11). These larger trees have thick bark that can withstand these understory fires. During fire suppression, relative fire mortal ity rates decreased, raising the relative mortality from other natural causes (Figure 4-6B). A B Figure 4-6. Simulated relative mortality of hardwoods from fire and other natural causes across various size classes for A) a frequent fire regime and B) a fire suppressed regime. Relative mortality between natural causes and fire was calculated within each size class for each fire regime. Results were similar for both study sites (Ordway and Ichauway). Sensitivity analysis: Impacts from fire frequency and competition To assess th e influences from fire and competition on population structure within the LLM, a sensitivity analysis was performed by varying the fire probability as well as associated 0.0 0.2 0.4 0.6 0.8 1.0 1.535101520Relative mortality Natural Fire 0.0 0.2 0.4 0.6 0.8 1.0 1.535101520Relative mortalityHeight (m) size class
129 calibration parameters [i.e., fire mortality parameter ( FPLPf), CI parameter ( CPLP) and TDI parameter ( DP )]. A range of fire probabilities (i.e., 0.05, 0.10, 0.15, 0.20, 0.25, 0.30, 0.35) were used to assess relative changes in population densities and community structure. Relative abundance between hardwoods and longleaf pines was calculated within each of the two size classes for a particular fire probability. Thr ee values of each calibration parameter (e.g., fire mortality parameter) were used to examine mode l sensitivity. These values consisted of the calibrated or base value and two values associated with a 20% ch ange in that parameter. For instance, the base or calibrated value of the fire mortality parameter for hardwoods was 0.66. The 20% increase and decrease of that value was 0.792 and 0.528, respectively. For each fire probability and parameter value, the LLM was rep licated 100 times to capture the stochastic variability across simulations. Each simulati on was run for 500 years and tree densities were recorded for the last 300 years, reducing the e ffects from transient dynamics when comparing model simulations. Mean densities were calculate d for each time step, each replicate, and across replicates for both longleaf pi ne and hardwoods in two size classes (< 10 m and > 10 m). For consistency, only the Ichauway model was used (e.g., fire probability = 0.35, Table 4-1). Changes in fire probability ( by 0.05) had dramatic effects to relative population densities of longleaf pines and hardwoods, especially below a 0.25 probability (Figure 4-7). This was evident across size classes. These results sugges t that within this particular system, a fire probability of 0.15 (mean fire return interval = 6.7 years) or above may sustain a longleaf pine dominated overstory (Figure 4-7A ). At this fire frequency (0.15), the system may be in transition, where its ecosystem resilience was high ly sensitive. Simulated fire probabilities below 0.15 led to a successional shift in community type with greater hardwood survival and therefore greater competitive reduction of pines.
130 A B Figure 4-7. Changes in relative abundance of A) larger (> 10 m) and B) smaller (< 10 m) longleaf pine (LP) and hardwood (HW) trees across various fire probabilities (0.05 to 0.35) using the LLM. Relative abundance between hardwoods and longleaf was calculated within each size class fo r a particular fire probability. The few longleaf pines remaining in low fire fr equency simulations were larger older trees that were less sensitive to competition. Recr uitment was suppressed however, and the entire longleaf population would eventua lly disappear without intensive restoration efforts. Higher mean fire return rates (2.8 to 5 yrs., 0.20 to 0.35 probabilities, Figure 4-7) was consistent with long-term ecosystem resilience, as negative feedbacks (e.g., hardwood suppression, dominance 0.00 0.20 0.40 0.60 0.80 1.00 0.050.10.150.220.127.116.11Relative abundance HW > 10 m LP > 10 m 0.00 0.20 0.40 0.60 0.80 1.00 0.050.10.150.18.104.22.168Relative abundance Fire probability HW < 10 m LP < 10 m
131 of both larger and smaller longleaf pines) main tained longleaf savanna community stability. Interestingly, hardwoods were rarely eliminated from the system, regardless of fire frequency. As fire probability is only one of the important controls of the fire regime in the LLM. The fire mortality parameter ( FPLP or FPHW) impacted fire intensity (i .e., mortality rates), hence influencing population numbers for a given fire pr obability. Populations were quite sensitive to a 20% change in the fire mortal ity parameter for both longleaf pine and hardwoods (Table 4-3). As expected, population densities for those trees above 10 m changed in accordance with their change in fire mortality parameter (i.e., higher mortality = lower densities). This was also the same for hardwoods below 10 m. Interestingl y, longleaf pine densitie s below 10 m increased with a larger fire mortality parameters for both tree types. This was most likely a model response of seedling establishment in increasingly open cells containing no (recently dead from fire) trees, while fire frequency (0.35 probability ) and longleaf fecundity functions were held constant among model runs. Table 4-3. Sensitivity analysis for the fire mo rtality parameter for longleaf pine (LP) and hardwoods (HW). All values are mean (s tandard error) across 100 LLM simulations. Fire mortality rate Population density (per ha) Fire mortality parameter (LP) LP HW LP > 10mLP < 10m HW > 10mHW < 10m 0.59 (-40%) 0.25 0.15 175 (1.7) 677 (8.3) 1 (0.4) 31 (2.9) 0.79 (-20%) 0.43 0.15 151 (2.4) 700 (11.5) 4 (0.7) 65 (4.4) 0.99 (base) 0.54 0.14 124 (2.5) 811 (19.6) 9 (1.0) 117 (6.1) Fire mortality parameter (HW) LP HW LP > 10mLP < 10m HW > 10mHW < 10m 0.528 (-20%) 0.54 0.11 100 (1.8) 675 (12.4) 33 (1.0) 333 (5.4) 0.66 (base) 0.54 0.14 124 (2.4) 828 (15.2) 9 (0.7) 119 (5.1) 0.792 (+20%) 0.53 0.16 140 (2.1) 898 (12.8) 1 (0.3) 16 (2.0) Hardwood spread may not have responded as quickly because of the shorter spread distance (i.e., one cell as opposed to longleafs 1 to 3 cells) and restricted spread rate (0.05 probability of establishing in one adjacent cell). Although the fire mortality parameters only
132 impacted mortality rates of their respective tree type, both populations were significantly affected regardless of direction of change or parameter modified. Hardwoods responded positively (higher densities), as fire mortality increasingly impact ed longleaf populations, while longleaf responded similarly when hard wood mortality rates increased. An overall fire mortality rate of 0.54 may se em high, but mortality was largely for newly established seedlings. For comparison, this studys field observations found 24 to 93% of planted seedlings [mean (SD) RCD: 0.8 cm (0.3)] died from fire. Overall seedling mortality rates have been found to be rather high in th e literature as well, i. e., 62 to 66% (Boyer 1963, Croker and Boyer 1975, Jones et al. 2003, Palik et al. 2003), and modeled as such (Cropper and Loudermilk 2006, Loudermilk and Cropper 2007). Population densities were reasonable for a fire-maintained longleaf pine savanna across mortality parameter values: Large tree densi ties ranged from100 to 175 longleaf pines per ha and 1 to 33 hardwoods per ha. Ultimately, this mortality parameter was useful for finetuning how fire intensity influenced population abunda nce within the LLM, and could be used to calibrate for site-specific condi tions or known population estimates. Along with variation in fire regime and fire intensity, competitive interactions within these systems are intricate, impacting growth patterns and population structure. In the LLM, mortality from competition was also quite sensitive to 20% changes in the longleaf CI parameter (Table 44). As expected, longleaf pine densities re sponded negatively to increased competition, while hardwoods responded positively to an increase in the longleaf CI. Similar changes occurred in both size classes. Longleaf mortality rates (from competition) at least doubled for each 20% increase in value, although density values only changed by 20%. As with the fire mortality parameter, all of these populati on densities were reasonable for a fire-maintained system: Larger
133 tree densities ranged from 106 to 156 for longlea f and 3 to 16 for hardwoods. Although the population density was sensitive to changes in competition intensity, the community structure (e.g., longleaf savanna to hardwood forest) remained similar to the base CI value case. The CI parameter was also useful for calibrating compe tition intensity effects on the tree populations Table 4-4. Sensitivity analysis for the competition index (CI) parameter for longleaf pine (LP). Mean LP and hardwood (HW) population dens ities [means (standard error)] as well as competition mortality rate (for LP) was assessed across 100 LLM simulations for each CI parameter. Population density CI parameter Comp. mort. rate LP > 10 m LP < 10 m HW > 10 m HW < 10 m 0.056 (-20%) 0.02 156 (2.2) 1538 (20.0) 3 (0.4) 68 (4.3) 0.070 (base) 0.05 129 (1.7) 1599 (11.9) 7 (0.4) 115 (5.7) 0.084 (+20%) 0.10 106 (1.4) 1534 (13.3) 16 (0.5) 190 (5.9) The TDI parameter for both longleaf and ha rdwoods was quite insensitive, with no significant changes to population de nsities with a 20% change to th e parameter. This was due to the opposing effect that tree dens ity has on tree leaf-litter pro duction and wiregrass accumulation in the LLM. Although leaflitter accumulated nearby trees, wiregrass grew more readily in forest gaps or in low tree density. Evaluation among study sites The LLM simulated height distributions and population densities c onsistent with two longleaf pine savanna sites (Table 4-5, 4-6). LLM heights w ithin and among sites were closely related to field and LIDAR estimates, while popula tion densities were slightly underestimated in comparison with the field data. The close relationship between hei ght distributions among datasets and the LLM was evident overall and by tree type. The LLM slightly overestimated height for longleaf pines above 10 m at Ordway and underestimated hardwood heights above 10 m at Ichauway (Table 4-5). The simulated height frequency distributi ons (trees above 10 m) were representative of populations that fluctu ate through time (Figure 4-8). The comprehensive
134 height distribution across ten LLM runs illustra ted a bi-model distribution of heights (Figure 48A), where the in situ data (for Ichauway) demonstrated a population dominated by larger, older canopy trees (Figure 4-8B). Among runs however, the distributions varied widely, where some were comparable to the in situ findings (Figure 4-8C) and others representing a less mature forest dominated by a younger cohort of trees (Figure 4-8D). Both Ordway and Ichauway LLM models revealed these fluctuati ons. These complex size frequenc y distributions have been found in other longleaf pine woodlands (Platt et al. 1988, Hartnett and Krof ta 1989, Gilliam and Platt 1999) and other forests (Wright et al. 2003). Table 4-5. Height dist ributions estimated for Ichauway and Ordway from LIDAR data, field data, and simulation outputs from the LLM. Longleaf pine (LP) and hardwood (HW) mean (SD) heights were recorded for trees above and below 10 m. Ichauway height distribution (m) All > 10m LP > 10 m LP < 10 m HW > 10 m HW < 10 m LIDAR data 25.2 (3.1) Field data 22.8 (4.8) 22.8 (4.8) 18.4 (4.6) LLM output 22.2 (6.7) 22.5 (6.6) 2.4 (1.1) 13.5 (3.0) 2.4 (1.3) Ordway height distribution (m) All > 10m LP > 10 m LP < 10 m HW > 10 m HW < 10 m LIDAR data 15.0 (3.4) Field data 14.3 (3.0) 14.6 (3.0) 12.1 (1.6) LLM output 18.2 (5.2) 18.5 (5.1) 4.0 (1.8) 12.3 (1.9) 4.0 (1.9) Table 4-6. Population densities estimated for Ichauway and Ordway from LIDAR data, field data, and simulation outputs from the LLM. Longleaf pine (LP) and hardwood (HW) mean (SD) densities were recorded for trees above and below 10 m. Ichauway population de nsity (per ha) All > 10m LP > 10 m LP < 10 m HW > 10 m HW 2-10 m LIDAR data 86 (18) Field data 145 (95) 144 (94) 49 (139) 3 (2) 115 (268) LLM output 125 (23) 120 (24) 81 (97) 5 (3) 56 (52) Ordway population density (per ha) All > 10m LP > 10 m LP < 10 m HW > 10 m HW 2-10 m LIDAR data 101 (18) Field data 171 (84) 156 (83) 304 (310) 15 (9) 113 (139) LLM output 110 (26) 104 (25) 339 (363) 6 (4) 71 (51)
135 Figure 4-8. Ichauway height frequency (%) dist ributions simulated by the LLM and estimated from field and LIDAR (Light Detection and Ranging) measurements. The A) comprehensive outputs (10 model runs) from the LLM illustrated a bi-modal distribution, while the B) in situ estimates demonstrated a population dominated by larger sized trees. Indivi dual runs illustrated fluctu ations in population size distributions, namely C) a less mature forest and D) a more mature forest, which was more comparable to the Ichauway in situ estimates (B). Note the difference in scale of frequency between graphs. For both sites, simulated population densitie s were underestimated for longleaf pines above 10 m, while hardwood values were comparable compared to field estimates (Table 4-6). Although there were some differen ces between densities among the small tree densities for both tree types, observed densities in longleaf fore sts are highly variable (Greenberg and Simmons 1999, Gagnon et al. 2004) and the simulated values we re within the one standard deviation of the 0 1 2 3 4 5 6 7 8 9 10141822263034Frequency (%) 0 2 4 6 8 10 12 14 10141822263034Frequency (%) 0 2 4 6 8 10 12 14 16 18 101418222630Frequency (%)Size class (m) Field LIDAR 0 2 4 6 8 10 12 14 16 18 10141822263034Frequency (%)Size class (m) A B C D
136 field estimates. The variability may be an artifact of past and current management regimes (e.g., changes in fire frequency, ranching, timber harves ting) as well as differe nces in recruitment and longleaf seedling growth. Population densitie s were underestimated with the LIDAR data, mainly attributable to tree omission by the algorith m (i.e., restricted 3 m distance between trees). Many trees within the smaller size classes (~10 to 15 m) were missed, as they were found either clumped together (within 3 m) or under the canopy of older larger trees when measuring them on site. The similarities between the in situ and LLM height distributi ons and population densities were remarkable (Table 4-5, 4-6) considering the minimal data us ed for potential height growth (site index). The only data us ed to calibrate heights were the maximum canopy height of the LIDAR dataset (not the height distributions) an d literature values associated with typical hardwood height estimates (see LLM: Model Development and Description). Population density estimates are quite variable in the literature as well. Su rprisingly, the small longleaf (< 10) densities (Table 4-6) were found to be seve ral orders of magnitude different between sites and modeled as such using the LLM without any calibration. Model Versatility The LLM is the first m odel to simulate both longleaf pine and hardwood demography in a fire-maintained ecosystem. This approach is ba sed on critical population ch aracteristics such as pine seed masting and tree density driven competition effects. The spatially explicit outputs were especially useful for examining community forest gap dynamics associated with hardwood encroachment and seedling establishment. The model was also able to demonstrate temporal fluctuations in population size st ructure that may occur naturally (Cropper and Loudermilk 2006) and most likely driven by seed masting events. After a seed mast, a rather large age-cohort may establish, while little successful regeneration may o ccur between mast years. As these trees age,
137 this bump in population density is observed across size classes (despite mortality) demonstrating oscillations in size structure through time. Natural disasters, such as hurricanes may dr ive these fluctuations as well as time lag effects associated with density dependence. Many larger trees may be downed by wind during storms creating pockets of potentia l regeneration sites. This coupled loss of overstory and possible increase in single-aged seedlings w ould dramatically impact the populations size distribution and succeeding fluctu ations in structure. There is also potential for us ing this model to further study this ecosystems community resilience or stability, by examining potential successional thresholds associated with the managed fire regime. The LLM clearly simulate s longleaf pines sensitivity to long intervals between fires, where a longleaf savanna may tran sition to a hardwood forest. Examining this transition zone of ecosystem resilience has been especially difficult to measure and understand, when seedlings may still establish periodically, but hardwoods have begun to dominate the midstory. With fire suppression, the composition of the understory changes, pine-needle effects on fire intensity are altered and wiregrass (and many other plant species) fails to reproduce. Although the adult pines continue to survive at low fire probability, recruitment limitation slowly reduces population size. Using the LLM, one could assess the restor ation potential of an area, perhaps determining the ideal fire frequenc y and time needed to transition back to a more stable community state. Using a model such as the LLM would provi de the opportunity to examine simulated temporal changes in detail without altering intact savanna sites and monitoring long-term changes. Forest models are often site-specific, with non-linear sensitivity to small changes in (sometimes many) parameters and designed for small scale analyses[ e.g., (Botkin 1993)].
138 Although these models have their application, a general model for longleaf pine savannas could be applied across a wide range of sites with di fferent management histories and compositions. The scale and relative simple i nputs of the LLM worked well for simulating two contrasting longleaf pine savanna sites in this study. Tr ee demography and fuels were modeled at a scale small enough to minimize influences from soil, w eather, or seasonal vari ability, while having the ability to assess individual plant mortality. The LLM scal e was large enough to examine changes in population structure and forest ga ps dynamics, without requiring finer-scale processes, such as soil nutrient cycling and fine-root dynamics. Moreover, simple inputs (e.g., site index, fire frequency) were used to initialize the model. Knowledge Gaps and Future Directions Models are often tim es most useful for de termining scientific knowledge gaps of the system at the scale being modeled (Botkin 1993). Although many can visualize the physiognomy of longleaf pine savannas, with their open park-like understory, lack of mid-story and highly dispersed overstory of adult pines, th e pristine structure of the forest is not completely understood due to the lack of true re ference sites (Platt et al. 1988). As in many forest ecosystems, structural va riation is large, bu t the full range of plausible longleaf pine population structure, fire interactions, and regeneration characteristics may not be known. This modeling approach provides a tool to examine th e consequences of altera tion of critical longleaf ecosystem processes (Figure 4-1). There is limited understanding of longleaf popul ation dynamics within the middle size classes associated with growth, mortality (from competition, fire, and other causes), as well as thinning rate. Seedling survival and growth has been intensively studied (Brockway and Outcalt 1998, McGuire et al. 2001, Palik et al. 2003, Gagnon et al. 2004, Pecot et al. 2007), overstory (adult) mortality characteristic s (Palik and Pederson 1996, Outcal t 2008) and plantation growth
139 rates (Quicke et al. 1994) have been addressed, but there are fe w studies specifically addressing the intermediate life st ages. Studying the dynamics within these middle size classes may provide for a better understand ing of longleaf comm unity variation. In the LLM, fuels and fire intens ity were homogeneous at the 25 m2 scale and no heterogeneity associated with burned and unburned patches were modeled. Recently, fuel and fire heterogeneity has been recogn ized at very fine (sub-meter) spatial scales in this ecosystem (Hiers et al. 2009, Loudermilk et al. 2009), well within the simulated cell size (5 m x 5 m). It has also been theorized that this fine-scale heterogeneity may driv e some degree of population level dynamics (Mitchell et al. 2009). Applying within ce ll heterogeneity and the ability of cells to burn or remain unburned (change of fuel continuity and fire spread) coul d provide some insight in feedbacks associated with fuel and fire heterogeneity coupled w ith hardwood and seedling (survival and competition) response, especially in forest gaps. It is un clear whether a modeling approach implementing this heterogeneity would change the dynamics of the model. Hardwood demography and interactions with lo ngleaf have had little attention in the literature, with studies mainly focusing on longl eaf regeneration. Knowledge gaps associated with hardwood demography in these fire-maint ained systems included limited understanding of clonal spread rate and vigor, re-sprouting longevity (after multiple top-kills), competitive effects of adult longleaf pines on ha rdwood growth, and size-specific mo rtality. Clonal hardwoods are difficult to characterize with standard de mographic modeling, perhaps requiring better understanding of the thick rhizomatous root structure (Berg and Ha mrick 1994). Underground carbohydrate reserves seem to indefinitely suppl y hardwood re-sprouting ability after fire. Identifying a longevity threshold or limit of th ese reserves would prove beneficial for managing hardwood encroachment.
140 It has been found that adult pines suppress growth of nearby hardwoods via the belowground competition (Pecot et al. 2007). The competitive intensity is poorly understood, especially across various sites with differing soil conditions. Furt hermore, in a frequent fire system longleaf pines and hardwood can co-exist Longleaf seedlings and hardwood sprouts are often found together in the understory, suggestin g a possible facilitation interaction between species. Although hardwoods and adult longleaf pines may shade-out pine seedlings, reducing growth and possibly survival, this shady environment may fa cilitate survival by lowering temperatures and reducing water stress during extended droughty conditions (Myers 1990, Pecot et al. 2007). There is little work that quan tifies hardwood mortality rates acr oss size classes in relation to fire, competition (if any), or other natural cause s. Most focus on fire mortality is in the small size classes (Rebertus et al. 1989, Robertson and Ostertag 2007). Quantifying these competitive or possible facilitation effects could prove usef ul for more accurately modeling their population dynamics. Ultimately, understanding hardwood dynamics is fundamental in assessing population success of longleaf pine savannas. Conclusions This is the first sim ulation model to integr ate longleaf pine and hardwood dynamics across space and time. The LMM proved versatile with relatively simple input parameters and calibration procedures for multi-site use. The LLM simulated realistic spatial gap dynamics associated with seedling and hardwood establishm ent as well as their re lative interand intraspecies competition. The cone production function, an importa nt aspect of longleaf pine dynamics, is closely tied to long-term regional observations. The LLM simulated changes in causes of mortality in association with tree si ze and fire frequency probability. It clearly demonstrated the change in population community structure (i.e., longlea f savanna to hardwood
141 forest) as fire probability was reduced, in cluding transition communities where ecosystem stability was highly sensitive (i.e ., degraded longl eaf site). Many scientific knowledge gaps were recognized. Southern hardwood demography in this system in general is relative unknown, including clonal sprouting vigor an d longevity, intensity of adult pine competition on hardwood growth, as well as detailed competitive or facilitative interactions between hardwood sprouts and pi ne seedlings. Targeting fuelbed and fire heterogeneity may provide for a multi-level understanding of fire behavior feedbacks on vegetation response. This modeling approach supports the notion that understanding hardwood dynamics within longleaf systems is fundament al to understanding population scale processes important for scientific advancement a nd sound ecosystem management practices.
142 CHAPTER 5 CONCLUSIONS Study Synthesis This study has contributed to the science of l ongleaf pine ecosystem s by: 1) demonstrating a robust approach to precisely measuring fine-s cale (sub-meter) vegetative fuelbed structural characteristics, namely volume, and finding stro ng correlations with leaf area and biomass (Loudermilk et al. 2009), 2) illu strating how fine-scale fuelbed at tributes (structure and type) drive variation in fire behavior and 3) coupling hardwood and l ongleaf pine demographics in a comprehensive individual tree based simulation model. This work emphasizes: 1) the complexities of within-fuelbed and fire heter ogeneity and the potential implications on fire effects, which have just begun to be rec ognized (Thaxton and Pla tt 2006, Hiers et al. 2009, Mitchell et al. 2009), 2) the valu e of understanding the interactions of hardwoods in longleaf pine savannas in relation to fi re frequency, as well as 3) th e importance of defining knowledge gaps of plant species-specific demography and in ter-species relationships within an ecosystem. Following are the principal conclusions of this study. Measuring Fuel Structure In this chapter (2), ground-based LID AR (L ight Detection and Ranging) was used to measure fuel volume at both the individual shrub scale and within complex herbaceous fuelbeds. As hypothesized, traditional methods of measur ing volume (i.e., cylindrical and spheroid calculations) of individual shrubs were significantly larger than LIDAR estimates. LIDAR volumes of individual whole shrubs were more st rongly correlated to leaf area and biomass than traditionally estimated volumes. The distribution of leaf area within a shrub was more precisely captured by the LIDAR: Section-specific (i.e., shr ub split into thirds) volume estimations of each species were evident, with str ong correlations with plant leaf area and biomass. This suggests
143 that knowledge of the vertical arra ngement of the plant as well as si ze is critical for accurately estimating plant volumes. Spatial distribution of heights within th e fuelbed a relatively small area (16 m2) was highly variable both vertically and horizontal ly, supporting the need to understand plant and fuelbed structure in multiple dimensions. The ground LIDAR was more reliable than traditional methods for measuring surface fuel structure and hei ght distributions due to its ability to capture the true architecture of plants (v s. an assumed geometry) and its se nsitivity to subtle changes in spatial heights at fine-scales. This fine-scale fuel heterogeneity may be critical for examining fire behavior at the same level. Examining Fuel and Fire Relationships The study in this chapter (3 ) was the first integration of high-resolution LIDAR and therm al imagery to assess fine-scale relationships of the forest surface fu elbed and fire behavior characteristics in longleaf pine woodlands. Ba sed on the results, grouping fuelbed plots based on similar fire characteristics provided the best models in terms of fit and predictive capability. Grouping the plots suppressed the impacts from plot-to-plot variation in fuel, fire and weather conditions that were not explicitly modeled (i.e., fuel moisture, fire in tensity, wind and relative humidity, respectively). The strongest model f it was found in the high-range fire groups, where the clustered fuel types were highly significant predictors. Mean fuelbed height was overall a significant measure of vegetation structure, signifying its importance for temperature and residence time predictions. Otherwise, significance of other LIDAR measured fuelbed attributes (i.e., heights, height distribu tion metrics, point density, and LIDAR intensity metrics) varied by model. Fu el continuity (both hor izontal and vertical) influenced fire behavior across grouped and individual plots. The gr ouped models segregated the fuel types that influenced fire intensity (e.g., pine litter), while the individual plot level
144 analysis revealed significance of specific fuel types (e.g., bare soil, deci duous litter) as well as the spatial configuration (x,y coordinates) of fire. Modeling Savanna Dynamics This chapter (4) presents the first sim ula tion model (i.e., LLM Longleaf Model) to integrate longleaf pine and hardwood dynamics across space and time. The LMM was versatile with relatively simple input parameters and calib ration procedures for multi-site use. The LLM simulated realistic spatial gap dynamics associated with seedli ng and hardwood establishment as well as their relative interand intra-species competition. Th e cone production function, an important aspect of longleaf pine dynamics, is closely tied to long-term regional observations. The LLM simulated changes in mortality associated with tree size and fire frequency probability. It clearly demonstrated the ch ange in population community structure (i.e., longleaf savanna to hardwood forest) as fire probability was re duced, including transition communities where ecosystem stability was compromised (i.e., degraded longleaf site). Many scientific knowledge gaps still exist. So uthern hardwood demography in this system in general is poorly known, including clonal spr outing vigor and longevity, intensity of adult pine competition on hardwood growth, as well as deta iled competitive or facilitative interactions between hardwood sprouts and pine seedlings. St udies of fuelbed and fire heterogeneity may provide a basis for multi-level understanding of fire behavior feedbacks on vegetation response. This modeling approach is based on and s upports the notion that understanding hardwood dynamics within longleaf systems is fundament al to understanding population scale processes and implementing sound ecosystem management practices. The LLM provides a foundation for integrating fuel and fire hete rogeneity characteristics (i.e., chapter 2 and 3) for examining vegetation response and feedbacks across scales.
145 Study and Management Implications Fire behavior prediction system software is sensitive to parameters associated with fuelbed height measurements, used for calculating vo lume and fuel density (Andrews and Queen 2001, Scott and Burgan 2005). Volume is especially difficult to measure, compared to mass estimates in many fuelbeds, because of the complex structure. As shown in this study, LIDAR provides an opportunity for improving measurements of fuelbed properties (i.e., chapter 2) that drive fire behavior and possibly enhance the accuracy of fire behavior prediction models. Moreover, LIDAR has the capacity to measure those characteristics w ith a precision and at a finer scale than is afforded by traditional methods. Ground-based LIDAR is promising, as both measurements of the physical structure of fuelbeds and discrete fuel types can be sample d non-destructively and analyzed in a spatially explicit context. Understanding how fuel struct ure and fine-scale volum e varies among fuels, other fuel characteristics, such as surface area to volume ratios, packing ratio, fuel continuity, and patchiness can ultimately impact fire behavior. Coupling fuelbed heterogeneity a nd fire behavior (i.e., chapter 3) is a promising approach to understanding the complex responses of vegeta tion to fire behavior and further establishing the concept of the ecolo gy of fuels (Mitchell et al. 2009) a nd the wildland fu el cell concept (Hiers et al. 2009). Both concep ts address the difficu lty of past approaches to connect fuel heterogeneity to fire behavior a nd fire behavior to fire effect s, specifically associated with longleaf pine woodlands. This study related fu el heterogeneity to fi re behavior and has implications for extending knowledge for fine-scale fire effects research. For instance, the fine spatial scale at which the multitude of plant species coexist (up to 50 species per m) that have been found in this system (Walker and Peet 1983, Ki rkman et al. 2004) has been attributed to fire behavior heterogeneity at the same scale (Mitch ell et al. 2006). As understory fuels, several
146 species have been documented as even modifying fire behavior (Rebertus et al. 1989). It has also been suggested that fire effects at the sm all scale include individu al plant recruitment and death (Mitchell et al. 2006, Thax ton and Platt 2006), but the coupled implica tions on the larger scale plant population is unknown. Fire frequency, intensity, and spread influe nce plant interactions within the system, especially between longleaf pine and hardwood trees. Although these tree interactions have been known and studied for decades, a comprehens ive model of the dynamics had not been established. Creating a stochas tic simulation model (i.e., chapter 4, LLM) of fuels, fire and longleaf and hardwood demography was used to test our understanding of the system. The LLM has implications for longleaf pine ma nagers, specifically related to ecological forestry silvicultural prescriptions, fire mana gement, ecosystem, restoration, and effects from catastrophic disturbances (e.g., w ildfires, hurricanes). The mode l provides the opportunity to apply various management plans specific to a pa rticular longleaf pine forest and monitor longterm changes. A harvesting function could be ad ded to simulate explicit harvesting regimes or retention treatments (Franklin et al. 1997, Palik et al. 2003). Ti mber extraction disrupts (among other things) the subsequent di stribution and abundance of pine litter, contributing to negative feedbacks between fuel continuity and fire be havior often encouraging hardwood encroachment in these harvested areas. Although these newly created gaps are prospe cts for pine regenera tion, understanding the relationship between these new pi ne recruits and encroaching hardwoods is difficult. The competitive (or possible facilitati on) interactions between pine seedlings and hardwood sprouts may be determined by the spatial and temporal changes in fuels (impacted by overstory removal) that ultimately drive changes in fire behavior (Mitchell et al. 2006, Pecot et al. 2007). The LLM
147 could simulate regeneration potenti al related to gap (extraction) size, interactions with hardwood sprouts, coupled with harvesting intensity and frequency. A range of fire frequencies in the LLM can be applied (similar to the approach in this study) to a particular longleaf pi ne site to assess ecosystem stabil ity and restoration potential. For instance, if a site is degraded or reachi ng a threshold state transitioning into a hardwood community, the LLM could be used to determine th e ideal fire frequency to enhance restoration efforts. One could also determine if and how l ong it would take a particular fire frequency to restore the ecosystem. The LLM can furthermore be enhanced by simulating catastrophic disturbances that may destroy a large portion of the adult pine populat ion. Wildfires and hurricanes have been known to be detrimental to longleaf pine systems and could unexpectedly alter management plans (Palik and Pederson 1996, Outcalt 2008). Using the LLM, one could assess the long-term implications of these landscape level disturbances associ ated with changes to the fuelbed, hardwood encroachment and community stab ility, as well as restoration pot ential (e.g., fire frequency and duration). This could facilitate the understanding between drama tic changes to stand structure and ecosystem function to ultimately enhance long -term forest management planning (Palik and Pederson 1996). Implementing fuel and fire heterogeneity with in the now homogeneous 5 m x 5 m cells of the LLM would provide for an approach to assess multi-level fire feedbacks and plant interactions. Fuelbed structure is regulated by th e overstory because of changes to fine-fuels and subsequent vegetation response at many spatial and temporal sc ales (Mitchell et al. 2009). Rather than building a landscape (1 -2 ha) of heterogeneous fuels at the sub-meter level (33 cm x
148 33 cm), a distribution of fuelbed attributes may be created within the (i.e., 5 m x 5 m) cells that vary along a gradient of overstory densities (and ti me since overstory distur bance, e.g., harvest). Overstory canopy structure or density can be determined by coupling aerial LIDAR and field data (Hall et al. 2005). There is potential to interpol ate the fine-fuelbed structure distributions (from ground-LIDAR) to the land scape level by co-krigging with the spatial overstory structure (from aerial LIDAR). This would streamline data manipulation, processing, and model simulation time as well as provide for an optimization approach for up-scaling the highly complex ground-LIDAR data. These fuelbed structural distributions can be used as inputs and calibration parameters for the LLM. A quantitative understanding of the im pacts of fine-scale fire behavior on plant demography and biodiversity is still an inherent limitati on (Mitchell et al. 2009). Field experiments for better understandin g these fire effects (Thaxton a nd Platt 2006) would be critical for further developing and evaluating the LLM. This would ultimately advance our scientific understanding of longleaf pine eco system dynamics across scales. This new approach to relating the relationship between fuel heterogeneity and fire behavior in this study has proven complex, yet promisi ng, especially for implementing with multi-level fire effects research. Modeling the intricate relationships of longleaf pines and hardwoods was an essential step to understandi ng the dynamics within the system. There is great potential for implementing the LLM with more detailed fire beha vior analysis, especially in association with various fire management a nd silvicultural practices.
149 APPENDIX A AUXILIARY INFORMATION F OR CART ANALYSIS Table A-1. Date recorded and m ean wind speed measured within each plot as the fire passed through. Plot Fire Group Wind speed (m/s) Date 1 Middle-range 1.73 2/23/2007 2 Middle-range 1.43 2/27/2007 3 Middle-range 1.26 2/27/2007 4 Middle-range 2.45 2/23/2007 5 Middle-range 1.81 2/23/2007 6 Middle-range 1.66 2/27/2009 7 Middle-range 2.88 3/16/2007 8 Middle-range 2.31 3/16/2007 9 Middle-range 1.15 2/27/2007 1 High-range 1.64 2/27/2007 2 High-range 1.41 3/16/2007 3 High-range 1.46 2/23/2007 4 High-range 2.39 2/27/2007 5 High-range 2.03 3/16/2007 1 Low-range 1.65 2/23/2007 2 Low-range 1.56 3/16/2007 3 Low-range 1.34 2/27/2007 4 Low-range 1.89 3/16/2007 5 Low-range 2.18 3/16/2007 6 Low-range 2.41 3/16/2007
150 Table A-2. Comprehensive CART output of import ance values (IVs) for each fuel metric within each fire model (Tmax, Q90, T300, T500). Each table is for each multi-plot and individual plot level. Fuelbed metrics ar e in descending order of importance for each fire model. Complete dataset (20 plots) Tmax Q90 T300 T500 Fuel metric IV Fuel metric IV Fuel metric IV Fuel metric IV Max ht 100.00 Ht distribution ratio 100.00 Mean ht 100.00 Mean ht 100.00 Spread of hts 95.41 Sum of hts 96.10 Ht distribution ratio 70.47 Spread of hts 97.22 Ht variance 90.07 Mean ht 93.79 Spr ead of hts 66.21 Ht variance 92.48 Mean ht 82.01 Ht skewness 86.94 Max ht 66.21 Ht distribution ratio 91.99 Ht kurtosis 79.65 Ht kurtosis 80.14 Ht skewness 60.65 Max ht 87.90 Mean LIDAR intensity 74.06 Spread of hts 74.20 Ht variance 60.05 Int-sqrt 73.33 Point density 65.59 Ht variance 70.18 Ht kurtosis 51.32 Ht skewness 71.02 Ht skewness 64.20 Max ht 68.29 Mean LIDAR intensity 47.27 Point density 58.72 Sum of hts 57.09 Fuel clusters 65.54 Point density 42.29 Int-lnr 55.89 Fuel clusters 55.75 Point density 60.50 Fuel clusters 42.18 Mean LIDAR intensity 55.60 Ht distribution ratio 52.81 Mean LIDAR intensity 53.55 Int-lnr 42.16 Ht kurtosis 54.48 Int-lnr 48.78 Int-sqrt 45.97 Sum of hts 35.89 Sum of hts 45.99 Int-sqrt 45.45 Int-lnr 44.89 Int-sqrt 30.56 Fuel clusters 41.90 Int-sqrt 39.26 Int-sqrt 39.46 Int-sqrt 27.09 Int-sqrt 40.19 Wiregrass 26.09 Wiregrass 23.69 Pine litter 20.10 Pine litter 34.85 Volatile shrubs 11.88 HR 10 8.07 Dec. oak litter 13.36 Graminoids 30.69 Dec. oak litter 8.18 Graminoids 6. 74 Shrubs 11.51 Shrubs 25.70 Perched oak litter 5.46 Dec. oak litter 6.58 Graminoids 10. 74 Perched oa k litter 13.11 HR 10 5.36 Perched pine litter 6.02 Perched pine litte r 9.74 Bare soil 6.03 Pine litter 4.72 Pine litter 5. 23 HR 10 9.11 HR 100 4.44 HR 100 4.57 Perched oak litter 4.15 Wire grass 6.37 Perched pine litter 4.19 Graminoids 4.47 Shrubs 3.65 Volatile shr ubs 3.32 HR 10 3.58 Perched pine litter 3.85 Forbs 3.62 HR 100 1.94 Forbs 3.21 Forbs 3.73 Bare soil 2.99 Fo rbs 1.51 Volatile shrubs 2.21 Bare soil 2.48 Volatile sh rubs 1.73 Perche d oak litter 0.90 D ec. oak litter 2.18 Shrubs 2.37 HR 100 0.12 Bare soil 0.71 Wiregrass 1.80
151 High-range fire group (5 plots) Tmax Q90 T300 T500 Fuel metric IV Fuel metric IV Fuel metric IV Fuel metric IV Spread of hts 100.00 Fuel clusters 100.00 Mean LIDAR intensity 100.00 Sum of hts 100.00 Max ht 96.30 Mean ht 73.25 Ht distribution ratio 66.50 Mean ht 97.85 Sum of hts 81.16 Point density 61.73 Fuel clusters 66.29 Ht distribution ratio 96.66 Mean ht 81.00 Sum of hts 54.79 Int-lnr 40.79 Point density 69.38 Point density 75.52 Spread of hts 53.84 Int-sqrt 39.37 Max ht 66.02 Int-sqrt 75.42 Ht skewness 46.94 Spread of hts 37.23 Spread of hts 61.56 Int-lnr 74.05 Mean LIDAR intensity 46.07 Point density 35.91 Mean LIDAR intensity 57.73 Ht variance 71.06 Max ht 44.99 Int-sqrt 35.50 Ht variance 54.96 Int-sqrt 60.89 Int-lnr 42.94 Max ht 32.14 Fuel clusters 53.70 Ht skewness 51.28 Ht distribution ratio 42.07 Mean ht 30.94 Int-sqrt 47.56 Fuel clusters 48.03 Int-sqrt 41.74 Ht variance 23.81 Int-lnr 42.90 Ht kurtosis 47.43 Ht variance 39.94 Ht skewness 23.27 Ht skewness 36.71 Ht distribution ratio 43.53 Ht kurtosis 38.01 Sum of hts 23.06 Int-sqrt 28.61 Mean LIDAR intensity 36.73 Int-sqrt 36.59 Wiregrass 21.55 Ht kurtosis 17.09 Perched pine litter 9.73 Gram inoids 17.52 Ht kurtosis 14. 79 Perched oa k litter 10.58 Bare soil 9.05 Wiregrass 16.93 Perc hed pine litter 12.81 Wiregrass 10.46 Pine litter 8.76 Forbs 11.55 D ec. oak litter 5.10 Forbs 6.49 Wiregrass 7.25 Shrubs 7.35 HR 100 1.65 Shrubs 5.04 Forbs 3.29 Dec. oak litter 7.13 HR 10 1.48 Pine litter 4.88 Graminoids 2.41 Pine litter 6.50 Pine litter 1.10 Dec. oak litter 4.15 Dec. oak litter 1.65 Perched pine litter 2.93 Graminoids 0.55 Graminoids 2.30 Volatile shrubs 1.06 Bare soil 1.58 Shrubs 0. 49 HR 10 2.00 Shrubs 0.41 Volatile shrubs 0.86 Forbs 0.23 Perched pine litter 0.98 HR 10 0.25 HR 10 0.44 Perched oak litter 0.23 Volatile shrubs 0.13 HR 100 0.12 Perched oak litter 0.11 Ba re soil 0.19 Bare soil 0.01 Perched oak litter 0.00 HR 100 0.00 Volatile shrubs 0.00 HR 100 0.00
152 Middle-range fire group (9 plots) Tmax Q90 T300 T500 Fuel metric IV Fuel metric IV Fuel metric IV Fuel metric IV Spread of hts 100.00 Fuel clusters 100.00 Mean LIDAR intensity 100.00 Sum of hts 100.00 Max ht 96.30 Mean ht 73.25 Ht distribution ratio 66.50 Mean ht 97.85 Sum of hts 81.16 Point density 61.73 Fuel clusters 66.29 Ht distribution ratio 96.66 Mean ht 81.00 Sum of hts 54.79 Int-lnr 40.79 Point density 69.38 Point density 75.52 Spread of hts 53.84 Int-sqrt 39.37 Max ht 66.02 Int-sqrt 75.42 Ht skewness 46.94 Spread of hts 37.23 Spread of hts 61.56 Int-lnr 74.05 Mean LIDAR intensity 46.07 Point density 35.91 Mean LIDAR intensity 57.73 Ht variance 71.06 Max ht 44.99 Int-sqrt 35.50 Ht variance 54.96 Int-sqrt 60.89 Int-lnr 42.94 Max ht 32.14 Fuel clusters 53.70 Ht skewness 51.28 Ht distribution ratio 42.07 Mean ht 30.94 Int-sqrt 47.56 Fuel clusters 48.03 Int-sqrt 41.74 Ht variance 23.81 Int-lnr 42.90 Ht kurtosis 47.43 Ht variance 39.94 Ht skewness 23.27 Ht skewness 36.71 Ht distribution ratio 43.53 Ht kurtosis 38.01 Sum of hts 23.06 Int-sqrt 28.61 Mean LIDAR intensity 36.73 Int-sqrt 36.59 Wiregrass 21.55 Ht kurtosis 17.09 Perched pine litter 9.73 Gram inoids 17.52 Ht kurtosis 14. 79 Perched oa k litter 10.58 Bare soil 9.05 Wiregrass 16.93 Perc hed pine litter 12.81 Wiregrass 10.46 Pine litter 8.76 Forbs 11.55 D ec. oak litter 5.10 Forbs 6.49 Wiregrass 7.25 Shrubs 7.35 HR 100 1.65 Shrubs 5.04 Forbs 3.29 Dec. oak litter 7.13 HR 10 1.48 Pine litter 4.88 Graminoids 2.41 Pine litter 6.50 Pine litter 1.10 Dec. oak litter 4.15 Dec. oak litter 1.65 Perched pine litter 2.93 Graminoids 0.55 Graminoids 2.30 Volatile shrubs 1.06 Bare soil 1.58 Shrubs 0. 49 HR 10 2.00 Shrubs 0.41 Volatile shrubs 0.86 Forbs 0.23 Perched pine litter 0.98 HR 10 0.25 HR 10 0.44 Perched oak litter 0.23 Volatile shrubs 0.13 HR 100 0.12 Perched oak litter 0.11 Ba re soil 0.19 Bare soil 0.01 Perched oak litter 0.00 HR 100 0.00 Volatile shrubs 0.00 HR 100 0.00
153 Low-range fire group (6 plots) Tmax Q90 T300 T500 Fuel metric IV Fuel metric IV Fuel metric IV Fuel metric IV Mean LIDAR intensity 100.00 Ht variance 100.00 Ht kurtosis 100.00 Mean ht 100.00 Mean ht 72.11 Max ht 86.26 Ht skewness 86.02 Ht variance 73.20 Ht skewness 72.11 Spread of hts 83.57 Mean ht 78.52 Max ht 72.95 Spread of hts 68.54 Mean ht 81.07 Int-sqrt 76.06 Spread of hts 70.05 Fuel clusters 64.77 Ht distribution ratio 59.84 Sum of hts 72.92 Sum of hts 62.90 Ht variance 64.55 Mean LIDAR intensity 53.10 Point density 64.54 Ht kurtosis 56.02 Sum of hts 63.60 Point density 52.54 Fuel clusters 62.96 Ht skewness 51.15 Ht kurtosis 62.47 Ht skewness 51.58 Ht distri bution ratio 61.98 Ht distribution ratio 42.89 Ht distribution ratio 51.60 Ht kurtosis 50.56 Max ht 58.09 Mean LIDAR intensity 42.11 Max ht 50.32 Sum of hts 50.14 Int-lnr 56.09 Point density 37.57 Point density 49.87 Int-lnr 41.05 Mean LIDAR intensity 55.08 Int-sqrt 36.42 Int-sqrt 37.52 Int-sqrt 40.56 Ht variance 53.42 Int-lnr 33.50 Int-lnr 34.26 Int-sqrt 37.33 Int-sqrt 39.22 Fuel clusters 24.72 Wiregrass 27.09 Fuel clusters 34.15 Spread of hts 38.17 Int-sqrt 23.46 Int-sqrt 24.29 Perched oa k litter 20.04 Perched pine litter 16.35 HR 10 12.07 Pine litter 9.86 Graminoids 19.01 Dec. oak litter 10.86 Bare soil 6.35 Forbs 8.01 Pine litter 9.73 Shrubs 10.71 Shrubs 5.87 Shrubs 5.76 Forbs 7.88 Forbs 10.60 Wiregrass 5.61 Volatile shrubs 4.08 Wiregrass 7.60 Wi regrass 10.31 Perche d pine litter 4.92 Dec. oak litter 3.63 Dec. oak litter 6.43 Graminoids 10.18 Graminoids 3.82 Graminoids 3.27 Shrubs 4.30 Perched oak litter 5.12 Perche d oak litter 3.01 Bare soil 2.89 Perched pine litter 1.94 HR 10 1.31 Pine litter 2.24 HR 10 1.78 Bare soil 1.55 Pine litter 1.25 Volatile shrubs 1.72 Perched oak litter 1.74 HR 10 0.72 Volatile shrubs 1. 25 Dec. oak litter 1.03 HR 100 1.07 Volatile shrubs 0.47 Bare soil 0.13 Forbs 0.75 Perched pine litter 0.83 HR 100 0. 05 HR 100 0.06 HR 100 0.66
154 High-range fire group individual plot Tmax Q90 T300 T500 Fuel metric IV Fuel metric IV Fuel metric IV Fuel metric IV Point density 100.00 Spread of hts 100.00 Point density 100.00 Ht distribution ratio 100.00 Int-sqrt 81.84 Sum of hts 87.24 Int-lnr 91.83 Ht kurtosis 86.71 Int-lnr 81.27 Ht distribution ratio 77.23 Int-sqrt 85.98 Point density 79.56 Sum of hts 60.14 Point density 76.79 Int-sqrt 82.80 Mean ht 67.93 Int-sqrt 57.58 Int-lnr 66.80 Mean ht 46.36 Ht variance 67.53 Mean LIDAR intensity 55.52 Mean ht 65.20 Sum of hts 42.86 Spread of hts 56.19 Ht variance 53.99 Int-sqrt 59.07 Ht distribution ratio 40.23 Sum of hts 52.76 Ht skewness 51.49 Int-sqrt 56.13 Bare soil 35.78 Ht skewness 49.18 Spread of hts 45.01 Mean LIDAR intensity 54.92 Ht kurtosis 34.65 Max ht 45.97 Max ht 40.79 Ht variance 54.48 Spread of hts 30.59 Int-sqrt 32.73 Mean ht 33.65 Ht skewness 54.15 Ht skewness 29.41 Int-lnr 27.13 Ht distribution ratio 22.75 Max ht 50.27 Ht variance 20.98 Mean LIDAR intensity 27.11 Ht kurtosis 22.45 Ht kurtosis 32.99 Mean LIDAR intensity 18.05 Int-sqrt 14.40 Wiregrass 12.76 Perched pi ne litter 1.92 Max ht 15. 38 Dec. oak litter 8.27 Forbs 12.49 Dec. oak litter 0.15 HR 10 14.98 Shrubs 3.25 Volatile shrubs 7.62 Pine litter 0.13 Dec. oa k litter 4.28 Perche d pine litter 3.16 HR 10 1.93 Graminoids 0.09 Graminoids 2. 01 Pine litter 1.79 Perched pine litter 1.05 Bare soil 0.00 Volatile shrubs 1.18 Bare soil 0.81 Bare soil 0.00 Forbs 0.00 Forbs 0.00 Forbs 0.00 Dec. oak litter 0.00 HR 10 0.00 HR 100 0.00 Graminoids 0.00 Graminoids 0.00 HR 100 0.00 Pine litter 0.00 HR 10 0.00 HR 100 0.00 Perched oak litter 0.00 Perched oak litter 0.00 HR 100 0.00 Pine litter 0.00 Shrubs 0. 00 Perched pine litter 0.00 Perched oak litter 0.00 Perched oak litter 0.00 Volatile shrubs 0.00 Shrubs 0. 00 Volatile shrubs 0.00 Shrubs 0.00 Wiregrass 0.00 Wiregrass 0.00 Wiregrass 0.00
155 Middle-range fire group individua l plot Tmax Q90 T300 T500 Fuel metric IV Fuel metric IV Fuel metric IV Fuel metric IV Sum of hts 100.00 Max ht 100.00 Ht distribution ratio 100.00 Ht variance 100.00 Mean ht 91.33 Mean LIDAR intensity 93.91 Po int density 96.31 Ht distribution ratio 76.02 Ht skewness 66.21 Mean ht 88.29 Spread of hts 82.42 Ht skewness 74.44 Point density 64.91 Spread of hts 79.39 Mean ht 76.41 Ht kurtosis 71.07 Ht variance 44.94 Sum of hts 77.69 Int-lnr 74.25 Spread of hts 65.65 Ht kurtosis 42.71 Ht variance 69.78 Int-sqrt 72.65 Point density 55.13 Spread of hts 41.23 Ht skewness 52.11 Max ht 58.97 Max ht 45.65 Ht distribution ratio 33.31 Ht distribution ratio 43.76 Int-sqrt 56.17 Mean ht 44.79 Int-lnr 31.29 Point density 40.32 Ht variance 54.13 Sum of hts 44.07 Max ht 29.95 Ht kurtosis 38.04 Sum of hts 53.17 Int-lnr 36.04 Mean LIDAR intensity 29.37 Int-lnr 29.68 Mean LI DAR intensity 47.95 Mean LIDAR intensity 29.94 Int-sqrt 27.99 Int-sqrt 23.04 Ht skewness 27.75 Int-sqrt 27.26 Int-sqrt 16.84 Wiregrass 22.90 Forbs 27.48 Int-sqrt 20.93 Wiregrass 14.51 HR 10 14.12 Ht kurtosis 24.75 Wiregrass 5.27 Dec. oak litter 10.07 Pine litter 12.62 Perched pine litte r 14.96 Grami noids 2.68 Pine litter 9.75 Int-sq rt 1.77 Wiregrass 8. 07 Pine litter 0.42 Bare soil 7.85 Bare soil 0.00 HR 10 1.56 Bare soil 0.00 Perched pine litter 0.62 Dec. oak litter 0.00 Bare soil 0.00 Dec. oa k litter 0.00 Forbs 0.00 Forbs 0.00 Dec. oak litter 0.00 Forbs 0.00 Graminoids 0.00 Graminoids 0.00 Graminoids 0.00 HR 10 0.00 HR 10 0.00 HR 100 0.00 HR 100 0.00 HR 100 0.00 HR 100 0.00 Perched oak litter 0.00 Pi ne litter 0.00 Perc hed oak litter 0.00 Perched oak litter 0.00 Pe rched pine litter 0.00 Pe rched oak litter 0.00 Pe rched pine litter 0.00 Shrubs 0.00 Shrubs 0.00 Shrubs 0.00 Shrubs 0.00 Volatile shrubs 0.00 Volatile shrubs 0.00 Volatile shrubs 0.00 Volatile shrubs 0.00
156 Low-range fire group individual plot Tmax Q90 T300 T500 Fuel metric IV Fuel metric IV Fuel metric IV Fuel metric IV Mean ht 100.00 Mean ht 100.00 Mean LI DAR intensity 100.00 Ht skewness 100.00 Spread of hts 87.01 Point density 65.49 Sum of hts 67.40 Mean ht 98.37 Max ht 77.86 Ht variance 62.77 Mean ht 62.60 Ht kurtosis 76.03 Ht kurtosis 69.05 Sum of hts 59.71 Int-sqrt 60.06 Mean LIDAR intensity 61.98 Ht skewness 60.62 Max ht 57.80 Int-sqrt 56.07 Ht distribution ratio 53.01 Ht distribution ratio 57.87 Int-lnr 44.13 Ht skewness 55.15 Ht variance 49.73 Ht variance 56.21 Mean LIDAR intensity 41. 20 Ht variance 53.32 Sum of hts 36.46 Mean LIDAR intensity 54.52 Spread of hts 41.13 Int-lnr 47.88 Point density 33.20 Point density 27.37 Ht distribution ratio 38.15 Point density 47.07 Spread of hts 25.33 Int-lnr 20.75 Ht skewness 33.59 Spread of hts 38.14 Int-lnr 25.11 Int-sqrt 18.96 Ht kurtosis 33.50 Max ht 29.46 Int-sqrt 22.86 Int-sqrt 18.43 Int-sqrt 30.90 Ht kurtosis 28.92 Int-sqrt 18.16 Sum of hts 14.36 Int-sqrt 28.40 Graminoids 26.31 Max ht 13.01 Graminoids 10.96 HR 100 19.31 Bare soil 18.92 Perched pine litter 12.85 Dec. oak litter 2.43 Bare soil 1.56 Ht distribution ratio 13.24 Bare soil 12.46 HR 10 1.78 Shrubs 0.82 HR 100 8.91 Forbs 8.18 Perched pine litter 1.55 Wiregrass 0.53 Shrubs 7. 13 HR 10 2.25 Pine litter 1.49 Dec. oak litter 0.00 Wiregrass 6.93 Perched oak litter 0.46 Forbs 0.87 Forbs 0.00 Dec. oa k litter 6.44 Dec. oak litter 0.00 Wiregrass 0.19 Graminoids 0.00 Forbs 3.13 Graminoids 0.00 Bare soil 0.00 HR 10 0.00 Pine litter 2.28 HR 100 0.00 HR 100 0.00 Pine litter 0.00 HR 10 0.00 Pine litter 0.00 Perched oak litter 0.00 Perche d oak litter 0.00 Perched oa k litter 0.00 Sh rubs 0.00 Shrubs 0.00 Perched pine litter 0.00 Perched pine litter 0.00 Volatile shrubs 0.00 Volatile shrubs 0.00 Volatile shrubs 0.00 Volatile shr ubs 0.00 Wiregrass 0.00
157 Individual plots using x,y c oordinates T300 (x,y) Q90 (x,y) Tmax (x,y) Fuel metric IV Fuel metric IV Fuel metric IV Ht variance 100.00 Y coordina te 100.00 X coordinate 100.00 Point density 97.25 Mean LIDAR intensity 90.65 Ht kurtosis 58.80 Int-sqrt 89.93 Ht skewness 71.60 Ht skewness 57.73 Spread of hts 79.94 Max ht 68.33 Spread of hts 49.15 Y coordinate 78.87 Spread of hts 66.10 Int-lnr 48.79 Sum of hts 65.43 Ht kurtosis 60.93 Int-sqrt 47.79 Int-lnr 61.69 Ht distribution ratio 58.79 Ht distribution ratio 45.96 Max ht 59.40 X coordinate 43.26 Point density 45.41 Ht kurtosis 57.46 Mean ht 40.41 Max ht 44.03 Int-sqrt 54.30 Point density 30.34 Int-sqrt 36.26 Ht skewness 37.17 Ht variance 29.08 Ht variance 32.14 X coordinate 36.50 Dec. oa k litter 17.09 Sum of hts 25.41 Mean ht 29.43 Sum of hts 15.97 Mean ht 25.05 Mean LIDAR intensity 27.08 Wiregrass 10.18 Forbs 18.74 Bare soil 19.94 Graminoids 6.71 Y coordinate 13.90 Wiregrass 6.07 Perched pine litter 3.88 Mean LIDAR intensity 13.64 Perched pine litter 5.31 Forbs 2.06 Dec. oak litter 6.68 Pine litter 4.73 Bare soil 0.00 Perched oak litter 4.85 Ht distribution ratio 4.08 HR 10 0.00 Pine litter 3.48 Forbs 2.14 HR 100 0.00 Bare soil 1.35 HR 10 1.85 Pine litter 0. 00 Graminoids 0.00 Dec. oak litter 0.00 Perched oak litter 0.00 HR 10 0.00 Graminoids 0.00 Shrubs 0.00 HR 100 0.00 HR 100 0.00 Volatile shrubs 0.00 Perched pine litter 0.00 Perched oak litter 0.00 Intlnr 0.00 Shrubs 0.00 Shrubs 0.00 Int-sqrt 0. 00 Volatile shrubs 0.00 Volatile shrubs 0.00 Int-sq rt 0.00 Wiregrass 0.00
158 APPENDIX B AUXILIARY INFORMATION FOR LLM Table B-1. Specifications of Optechs ALTM (Airborne L aser Terrain Mapper) instrumentation for aerial LIDAR (Light Detection and Rangi ng) data collection at the Ordway study site. Specification Description Laser Nd:YAG, 1.064 micrometers Pulse frequency 33000pps Operating altitude 330 to 2000 m Range accuracy 2 cm (single-shot) Intensity data typica lly 8 bit grey tones Number of returns recorded per laser pulse 2 Scanner design Oscillating mirror Scanner range 20 degrees from nadir Scanner frequency 030 Hz Data recording hard disk Table B-2. Specifications of Optechs Gemini instrumentation for aerial LIDAR data collection at the Ichauway study site. Specification Description Operating altitude 150 4000 m, Nominal Horizontal accuracy 1/5,500 x altitude (m AGL); 1 sigma Elevation accuracy 5 30 cm; 1 sigma Range capture Up to 4 range measurements, including 1st, 2nd, 3rd, last returns Intensity capture 12-bit dynamic range for all recorded returns, including last returns Scan FOV (Field of View) 0 50 degrees; programmable in increments of degree Scan frequency 0 70 Hz Scanner product Up to Scan angle x Scan frequency = 1000 Roll compensation degrees at full FOV more under reduced FOV Pulse rate frequency 125 kHz Position Orientation System Applanix POS/AV 510 OEM includes embedded BD950 12-channel 10Hz GPS receiver Laser wavelength/class 1047 nanometers / Class IV (FDA 21 CFR) Beam divergence nominal ( full angle) Dual divergence 0.25 mrad (1/e) or 0.80 mrad (1/e)
159 LIST OF REFERENCES Abraham son, W. G., and D. C. Hartnett. 1990. Pi ne flatwoods and dry prairies. Pages 103-149 in R. L. Myers and J. J. Ewel, editors. Ecosyste ms of Florida. University of Central Florida Press, Orlando, FL. Abrahamson, W. G., and J. N. Layne. 2003. Long-te rm patterns of acorn production for five oak species in xeric Florida uplands. Ecology 84:2476-2492. Agee, J. K. 1998. Fire and pine ecosystems. Pages 193-218 in D. M. Richardson, editor. Ecology and biogeography of Pinus. Cambridge University Press. Andrews, P., C. Bevins, and R. Seli. 2004. Behave Plus fire modeling system, version 3.0: User's guide. RMRS-GTR-106WWW, USDA Forest Serv ice Rocky Mountain Research Station, Ogden, Utah. Andrews, P. L., and L. P. Queen. 2001. Fire modeling and information system technology. International Journal of Wildland Fire 10:343-352. Battaglia, M. A., P. Mou, B. Palik, and R. J. Mitchell. 2002. The effect of spatially variable overstory on the understory light environment of an open-canopied l ongleaf pine forest. Canadian Journal of Forest Research 32:1984. Beckage, B., and I. J. Stout. 2000. Effects of repeated burning on species richness in a Florida pine savanna: A test of the intermediate disturbance hypothesis. Journal of Vegetation Science 11 :113-122. Belanger, R. P., and R. M. Krinard. 1990. Quercus falcata michx., southern red oak. Pages 12401247 in R. M. Burns and B. H. Honkala, ed itors. Silvics of North America, vol. 2, Hardwoods. USDA Forest Service Agricultural Handbook 654, Washington, D.C. Berg, E. E., and J. L. Hamrick. 1994. Spatial and genetic structure of two sandhills oaks: Quercus laevis and Quercus margaretta (Fagaceae). American Journal of Botany 81:714. Bhattacharya, B., and D. P. Solomatine. 2005. Neur al networks and m5 model trees in modelling water level-discharge relationship. Neurocomputing 63:381-396. Botkin, D. B. 1993. Forest dynamics: An ecological model. Oxford University Press, Inc., New York, NY. Boyer, W. D. 1963. Development of longleaf pine seedlings under parent trees. USDA Forest Service, Southern Forest Experi ment Station, New Orleans, LA. Boyer, W. D. 1983. Variations in height-overage curves for young longleaf pine plantations. Forest Science 29:15-27.
160 Boyer, W. D. 1990. Pinus palustris Mill., longleaf pine. Pages 405-412 in R. M. Burns and B. H. Honkala, editors. Silvics of North Ameri ca, vol. 1, Conifers. USDA Forest Service Agricultural Handbook 654, Washington, D.C. Boyer, W. D. 1998. Long-term changes in flower ing and cone production by longleaf pine. Ninth biennial southern silvicultural research conference. USDA Forest Service, Southern Research Station, Clemson, SC. Boyer, W. D., and J. B. White. 1989. Natura l regeneration of longleaf pine. Pages 94-113 in Symposium on the Managment of Longleaf pine. USDA Forest Service, Southern Forest Experiment Station, New Orleans, LA. Breiman, L., J. H. Friedman, R. A. Olshen, a nd C. J. Stone. 1984. Classi fication and regression trees. Wadsworth, Inc., Belmont, CA. Bridges, E. L. 1989. Longleaf pine communities of the west gulf coastal plain. Natural Areas Journal 9 :246-263. Brockway, D. G., and K. W. Outcalt. 1998. Gapphase regeneration in lo ngleaf pine wiregrass ecosystems. Forest Ecology and Management 106:125-139. Brockway, D. G., K. W. Out calt, and W. D. Boyer. 2006. L ongleaf pine regeneration ecology and methods. Pages 95-133 in S. Jose, E. J. Jokela, and D. L. Miller, editors The longleaf pine ecosystem: Ecology, silviculture, and rest oration. Springer Science+Business Media, LLC, New York, NY. Brown, J. K. 1974. Handbook for inventorying downed woody material. GTR-INT-16, USDA Forest Service, Intermountain Research Station, Ogden, Utah. Brown, J. K. 1981. Bulk densities of non-uniform surface fuels and their application to fire modeling. Forest Science 27:667-683. Christensen, N. L. 1981. Fire regimes in southeastern ecosystems. Pages 112-136 in H. A. Mooney, T. M. Bonnicksen, N. L. Christensen, J. E. Lotan, and W. A. Reiners, editors. Fire regimes and ecosystem properties. USDA Forest Service, Washington, D.C. Christensen, N. L. 1988. Vegetation of the southeastern coasta l plain. Pages 317-363 in M. G. Barbour and W. D. Billings, editors. North Am erican terrestrial vegetation. Cambridge University Press, New York. Clark, J. S., A. M. Gill, and P. Kershaw. 2002. Spa tial variability in fire regimes: Its effects on recent and past vegetation. Pages 125-144 in R. A. Bradstock, J. E. Williams, and A. M. Gill, editors. Flammable Australia: The fire regimes and biodiversity of a continent. Cambridge University Press, Cambridge, U.K. Croker, T. C., and W. D. Boyer. 1975. Regene rating longleaf pine na turally. USDA Forest Service, Southern Forest Experi ment Station, New Orleans, LA.
161 Cropper, J. W. P., and E. L. Loudermilk. 2006. Th e interaction of seedling density dependence and fire in a matrix population model of longleaf pine (Pinus palustris ). Ecological Modelling 198 :487-494. Danson, F. M., D. Hetherington, F. Morsdorf, B. Koetz, and B. Allgower. 2007. Forest canopy gap fraction from terrestrial laser scanni ng. Geoscience and Remote Sensing Letters, IEEE 4:157-160. De'ath, G., and K. E. Fabricius. 2000. Classification and regression trees : A powerful yet simple technique for ecological data analysis. Ecology 81 :3178-3192. DeAngelis, D. L., and W. M. Mooij. 2005. I ndividual-based modeling of ecological and evolutionary processes. A nnual Review of Ecology and Evolutionary Systematics 36:147-168. DeBano, L. F., D. G. Neary, and P. F. Ffolliott. 1998. Fire effects on ecosystems. John Wiley & Sons, Inc., New York, NY. Dietze, M. C., and J. S. Clark. 2008. Cha nging the gap dynamics paradigm: Vegetative regeneration control on fore st response to disturban ce. Ecological Monographs 78:331347. Dimitrakopoulos, A. P. 2002. Mediterranean fuel m odels and potential fire behaviour in Greece. International Journal of Wildland Fire 11:127-130. Drake, J. B., and J. F. Weishampel. 2000. Multifr actal analysis of canopy height measures in a longleaf pine savanna. Forest Ecology and Management 128:121-127. Drake, J. B., and J. F. Weishampel. 2001. Simulati ng vertical and horizontal multifractal patterns of a longleaf pine savanna. Ecological Modelling 145:129-142. Everitt, B. 1980. Cluster analysis. 2nd editi on. Heineman Educational Books, London, England. Farrar, R. M. J. 1985. Volume and growth predictions for thinned even-aged natural longleaf pine stands in the east gulf area. Research Paper SO-220, USDA Forest Service, Southern Research Station New Orleans, LA. Ferguson, S. A., J. E. Ruthford, S. J. McKay, D. Wright, C. Wright, and R. D. Ottmar. 2002. Measuring moisture dynamics to predict fi re severity in longleaf pine forests. International Journal of Wildland Fire 11:267-279. Fernandes, P. M., W. R. Catc hpole, and F. C. Rego. 2000. Shrubl and fire behaviour modelling with microplot data. Canadian Journal of Forest Research 30:889-899. Franklin, J. F., D. E. Berg, D. A. Thornburgh, an d J. C. Tappenier. 1997. A lternative silvicultural approaches to timber harvesting: Variable retention harvest systems. Pages 111-139 in K. A. Kohm and J. F. Franklin, editors. Creating a forestry for the twenty-first century: The science of ecosystem management. Island Press, Washington D.C.
162 Frhlich, C., and M. Mettenleiter. 2004. Terrestrial laser scanning new perspectives in 3D surveying. Pages 7-13 in Laser-scanners for forest and landscape assessment. International Archives of Photogrammetry, Remote Sensing and Spatial Sensing, Freiburg, Germany. Gagnon, J. L., E. J. Jokela, W. K. Moser, and D. A. Huber. 2004. Characteristics of gaps and natural regeneration in mature longleaf pine flatwoods ecosystems. Forest Ecology and Management 187:373-380. Garca, M., D. Riao, E. Chuvieco, and F. M. Danson. 2010. Estimating biomass carbon stocks for a Mediterranean forest in central Spain us ing lidar height and intensity data. Remote Sensing of Environment 114:816-830. Gill, A. M. 1975. Fire and the Australian flora: A review. Australian Forestry 38 :4-25. Gilliam, F. S., and W. J. Pla tt. 1999. Effects of long-term fire exclusion on tree species composition and stand structure in an old-growth Pinus palustris (longleaf pine) forest. Plant Ecology 140:15-26. Glitzenstein, J. S., W. J. Platt, and D. R. Streng. 1995. Effects of fire regime and habitat on tree dynamics in north Florida longleaf pi ne savannas. Ecological Monographs 65:441-476. Glitzenstein, J. S., D. R. Streng, and D. D. Wade. 2003. Fire frequency effects on longleaf pine ( Pinus palustris Miller) vegetation in South Caro lina and northeast Florida, USA. Natural Areas Journal 23 :22-37. Goetz, S., D. Steinberg, R. Dubayah, and B. Bl air. 2007. Laser remote sensing of canopy habitat heterogeneity as a predictor of bird species richness in an eastern temperate forest, USA. Remote Sensing of Environment 108:254-263. Grace, S. L., and W. J. Platt. 1995. Effects of a dult tree density and fire on the demography of pregrass stage juvenile longleaf pine ( Pinus palustris mill.). Journal of Ecology 83:75-86. Greenberg, C. H., and R. W. Simmons. 1999. Ag e, composition, and stand structure of oldgrowth oak sites in the Fl orida high pine landscape: Implications for ecosystem management and restoration. Natural Areas Journal 19:30-40. Gresham, C. A. 1982. Litterfall patterns in mature loblolly and longleaf pine stands in coastal South Carolina. Forest Science 28 :223-231. Grunwald, S., S. H. Daroub, T. A. Lang, and O. A. Diaz. 2009. Tree-based modeling of complex interactions of phosphorus loadings and envi ronmental factors. Science of the Total Environment 407:3772-3783. Guerin, D. N. 1993. Oak dome clonal structure and fire ecology in a Florida longleaf pine dominated community. Bulletin of the Torrey Botanical Club 120:107-114.
163 Haefner, J. W. 1996. Modeling biological systems: Principles and applications. Chapman & Hall, New York. Hall, S. A., I. C. Burke, D. O. Box, M. R. Kaufmann, and J. M. St oker. 2005. Estimating stand structure using discrete-return lidar: An example from low density, fire prone ponderosa pine forests. Forest Ecology and Management 208:189-209. Harlow, R. F. 1990. Quercus laevis Walt., turkey oak in R. M. Burns and B. H. Honkala, editors. Silvics of North America, vol. 2, Hardwoods. USDA Forest Service Agricultural Handbook 654, Washington, D.C. Hartnett, D. C., and D. M. Krof ta. 1989. Fifty-five years of postfire succession in a southern mixed hardwood forest. Bulletin of the Torrey Botanical Club 116:107-113. He, H. S., B. Z. Shang, T. R. Crow, E. J. Gu stafson, and S. R. Shifley. 2004. Simulating forest fuel and fire risk dynamics across lands capes--LANDIS fuel module design. Ecological Modelling 180 :135-151. Hendricks, J. J., C. A. Wilson, and L. R. Boring. 2002. Foliar litter position and decomposition in a fire-maintained longleaf pine wiregr ass ecosystem. Canadian Journal of Forest Research 32 :928-941. Henning, J. G., and P. J. Radtke. 2006. Gr ound-based laser imaging for assessing threedimensional forest canopy structure. Photogr ammetric Engineering and Remote Sensing 72:1349-1358. Heyward, F. 1939. The relation of fire to stand composition of longleaf pine forests. Ecology 20:287-304. Hiers, J. K., J. J. O'Brien, R. J. Mitchell, J. M. Grego, and E. L. Loudermilk. 2009. The wildland fuel cell concept: An approach to characterize fine-scale variation in fuels and fire in frequently burned longleaf pine forests. International Journa l of Wildland Fire 18:315325. Hiers, J. K., J. J. O'Brien, R. E. Will, and R. J. Mitchell. 2007. Forest floor depth mediates understory vigor in xeric Pinus palustris ecosystems. Ecological Applications 17:806814. Higgins, S. I., D. M. Richardson, and R. M. Cowling. 2001. Validation of a spatial simulation model of a spreading alien plant pop ulation. Journal of Applied Ecology 38:571-584. Hoffmann, W. A., and O. T. Solbrig. 2003. The role of topkill in the differential response of savanna woody species to fire. Forest Ecology and Management 180:273-286. Holder, C. D. 2000. Geography of Pinus elliottii Englem. and Pinus palustris Mill. Leaf lifespans in the southeastern USA. Journal of Biogeography 27:311-318.
164 Hopkinson, C., L. Chasmer, Y.-P. Colin, and P. Treitz. 2004. Assessing forest metrics with a ground-based scanning lidar. Canadian Journal of Forest Research 34:573-583. Hosoi, F., and K. Omasa. 2006. Voxel-based 3-d m odeling of individual trees for estimating leaf area density using high-resolution portable scanning lidar. Geoscience and Remote Sensing, IEEE Transactions on Ge oscience and Remote Sensing 44:3610-3618. Iverson, L. R., D. A. Yaussy, J. Rebbeck, T. F. Hutchinson, R. P. Long, and A. M. Prasad. 2004. A comparison of thermocouples and temperatur e paints to monitor spatial and temporal characteristics of landscape-scale prescribed fires. International J ournal of Wildland Fire 13:311-322. Jacqmain, E. I., R. H. Jones, and R. J. Mitche ll. 1999. Influences of frequent cool-season burning across a soil moisture gradient on oak commun ity structure in longleaf pine ecosystems. American Midland Naturalist 141:85-100. Jones, R. H., R. J. Mitchell, G. N. Stevens, and S. D. Pecot. 2003. Controls of fine root dynamics across a gradient of gap sizes in a pine woodland. Oecologia 134:132-143. Kaiser, L. C. 1996. The future of longleaf pine in a mixed pine-oak forest (Big Thicket, Texas). Rice University, Houston, TX. Kampa, K., and K. C. Slatton. 2004. An adaptive multiscale filter for se gmenting vegetation in ALSM data. Pages 3837-3840 in IEEE Geoscience and Remote Sensing Symposium, Anchorage, AK. Kane, J. M., J. M. Varner, and J. K. Hiers. 2008. The burning characteri stics of southeastern oaks: Discriminating fire facilitators fr om fire impeders. Forest Ecology and Management 256:2039-2045. Keane, R. E., and L. J. Dickinson. 2007. The phot o-load sampling technique: Estimating surface fuel loadings from downward-looking photogr aphs of synthetic fuelbeds. RMRS-GTR190, USDA Forest Service, Rocky Mountain Research Station, Fort Collins, CO. Kennard, D. K., K. W. Outcalt, D. Jones, a nd J. J. O'Brien. 2005. Comparing techniques for estimating flame temperature of pr escribed fires. Fire Ecology 1:75-84. Kirkman, L. K., P. C. Goebel, and B. J. Pali k. 2004. Predicting plant sp ecies diversity in a longleaf pine landscape. Ecoscience 11:80-93. Kirkman, L. K., and R. J. Mitchell. 2006. Conservation management of Pinus palustris ecosystems from a landscape perspective. Applied Vegetation Science 9 :67-74. Koenig, W. D., and J. M. H. Knops. 2005. The myst ery of masting in trees American Scientist 93:340-347.
165 Krausmann, W. J., and P. C. Hicks. 1996. The a pplication of forward looking infrared imaging systems in wildland fire suppression. Pages 355-362 in Remote Sensing, People in Partnership with Technology: Proceedings of the Sixth Fore st Service Remote Sensing Applications Conference. American Societ y of Photogrammetry and Remote Sensing, Denver, CO. Kush, J. S., J. C. G. Goelz, R. A. Williams, D. R. Carter, and P. E. Linehan. 2006. Longleaf pine growth and yield. Pages 251-267 in S. Jose, E. J. Jokela, and D. L. Miller, editors. The longleaf pine ecosystem: Ecology, silv iculture, and restoration. Springer Science+Business Meida, LLC, New York, NY. Lee, H., K. C. Slatton, B. E. Roth, and W. P. Cropper. 2009. Predicti on of forest canopy light interception using three-dimensional airborne lidar data. International Journal of Remote Sensing 30:189-207. Lee, H., K. C. Slatton, B. E. Roth, and W. P. Cropper. 2010. Adaptive cluste ring of airborne lidar data to segment individual tree crowns in ma naged pine forests. International Journal of Remote Sensing 31:117 139. Lefsky, M. A., W. B. Cohen, S. A. Acker, G. G. Parker, T. A. Spies, and D. Harding. 1999. Lidar remote sensing of the canopy structure and biophysical properties of Douglas-fir western hemlock forests. Remote Sensing of Environment 70:339-361. Lichti, D. D., S. J. Gordon, and M. P. Stew art. 2002. Ground-based laser scanners: Operation, systems and applications. Geomatica 56:21-33. Loudermilk, E. L. 2005. Spatial and demographic modeling techniques applied to the longleaf pine ( Pinus palustris ) ecosystem of north central Flor ida. University of Florida, Gainesville, FL. Loudermilk, E. L., and J. W. P. Cropper. 2007. Multi-scale modeling of longleaf pine ( Pinus palustris ). Canadian Journal of Forest Research 37 :2080. Loudermilk, E. L., J. K. Hiers, J. J. O'Brien, R. J. Mitchell, A. Singhania, J. C. Fernandez, W. P. Cropper, and K. C. Slatton. 2009. Ground-based lid ar: A novel approach to quantify finescale fuelbed characteristics. International Journal of Wildland Fire 18:676-685. Lovell, J. L., D. L. B. Jupp, D. S. Culvenor and N. C. Coops. 2003. Using airborne and groundbased lidar to measure canopy structure in Australian forests. Canadian Journal of Remote Sensing 29:607-622. McBratney, A. B., M. L. Me ndonca Santos, and B. Minasny. 2003. On digital soil mapping. Geoderma 117 :3-52. McGinty, D. T., and E. J. Christy. 1977. Turk ey oak ecology on a Georgia sandhill. American Midland Naturalist 98 :487-491.
166 McGuire, J. P., R. J. Mitchell, E. B. Moser, S. D. Pecot, D. H. Gjerstad, and C. W. Hedman. 2001. Gaps in a gappy forest: Plant resources longleaf pine re generation, understory response to tree removal in longleaf pine sava nnas. Canadian Journal of Forest Research 31:765-778. McNab, W. H., and P. E. Aver s. 1994. Ecological subregions of the United States: Section descriptions. USDA Forest Service, Washington, DC. Meldahl, R. S., J. S. Kush, J. N. Rayamajhi, and J. Robert M. Fa rrar. 1998. Productivity of natural stands of longleaf pine in relation to competition and climatic factors. Pages 231254 in R. A. Mickler and S. Fox, editors. The pr oductivity and sustaina bility of southern forest ecosystems in a changing envir onment. Springer-Verlag, New York, NY. Miller, J. 2005. Incorporating spatial dependence in predictive vegetation models: Residual interpolation methods. Professional Geographer 57 :169-184. Mitchell, R. J., J. K. Hiers, J. O'Brien, and G. Starr. 2009. Ecological fo restry in the southeast: Understanding the ecology of fu els. Journal of Forestry 107:391-397. Mitchell, R. J., J. K. Hiers, J. J. O'Brien, S. B. Jack, and R. T. Engstrom. 2006. Silviculture that sustains: The nexus between silviculture, fre quent prescribed fire and conservation of biodiversity in longleaf pine forests of the s outheastern United States. Canadian Journal of Forest Research 36 :2724-2736. Mladenoff, J. D., G. E. Host, J. Boeder, and T. R. Crow. 1996. LANDIS: A spatial model of forest landscape disturbance, succe ssion, and management. Pages 175-179 in M. F. Goodchild, L. T. Steyaert, B. O. Parks, C. A. Johnston, D. R. Maidment, and M. P. Crane, editors. GIS and environmental m odeling: Progress and research issues. GIS World Books, Fort Collins, Colorado. Moisen, G. G., and T. S. Fr escino. 2002. Comparing five mode lling techniques for predicting forest characteristics. Ecological Modelling 157:209-225. Mulligan, M. K., L. K. Kirkman, and R. J. Mitchell. 2002. Aristida beyrichiana (wiregrass) establishment and recruitment: Implicatio ns for restoration. Restoration Ecology 10:6876. Mutlu, M., S. C. Popescu, and K. Zhao. 2008. Sensitiv ity analysis of fire behavior modeling with lidar-derived surface fuel maps. Forest Ecology and Management 256:289-294. Myers, R. L. 1990. Scrub and high pine. Pages 150-193 in R. L. Myers and J. J. Ewel, editors. Ecosystems of Florida. University of Central Florida Press, Orlando, FL. Ottmar, R. D., R. E. Vihnanek, and J. W. Mathey. 2003. Stereo photo series for quantifying natural fuels. Volume VIa: Sand hill, sand pine scrub, and hardwood with white pine types in the southeast United States with supplemental sites for volume VI. PMS 838, National Wildfire Coordinating Group, National Interagency Fire Center, Boise, Idaho.
167 Outcalt, K. W. 2008. Lightning, fi re and longleaf pine: Using na tural disturbance to guide management. Forest Ecology and Management 255:3351-3359. Palik, B., R. J. Mitchell, S. Pecot, M. Ba ttaglia, and M. Pu. 2003. Spatial distribution of overstory retention influences resources and growth of longleaf pine seedlings. Ecological Applications 13:674-686. Palik, B. J., and N. Pederson. 1996. Overstory mort ality and canopy disturbances in longleaf pine ecosystems. Canadian Journal of Forest Research 26:2035-2047. Parker, G. G., D. J. Harding, and M. L. Be rger. 2004. A portable li dar system for rapid determination of forest canopy stru cture. Journal of Applied Ecology 41:755-767. Pecot, S. D., R. J. Mitchell, B. J. Palik, E. B. Moser, and J. K. Hiers. 2007. Competitive responses of seedlings and understory plants in longleaf pine woodlands: Separating canopy influences above and below ground. Ca nadian Journal of Forest Research 37:634648. Pederson, N., J. S. Kush, R. S. Meldahl, and W. D. Boyer. 2000. Longleaf pine cone crops and climate: A possible link. Pages 255-258 in Tenth Biennial Southern Silvicultural Research Conference, Shreveport, LA. Penfound, W. T., and A. G. Watkins. 1937. Phytos ociological studies in the pinelands of southeastern Louisiana. Am erican Midland Naturalist 18:661-682. Platt, W. J., G. W. Evans, and S. L. Ra thbun. 1988. The population dynamics of a long-lived conifer ( Pinus palustris ). The American Naturalist 131:491-525. Popescu, S. C., R. Wynne, H., and J. A. Sc rivani. 2004. Fusion of small-footprint lidar and multispectral data to estimate plotlevel volume and biomass in deciduous and pine forests in Virginia, USA. Forest Science 50:551-565. Prasad, A., L. Iverson, and A. Liaw. 2006. Newer classification and regre ssion tree techniques: Bagging and random forests for eco logical prediction. Ecosystems 9:181-199. Provencher, L., B. J. Herring, D. R. Gordon, H. L. Rodgers, K. E. M. Galley, G. W. Tanner, J. L. Hardesty, and L. A. Brennan. 2001. Eff ects of hardwood reduction techniques on longleaf pine sandhill vegetation in nort hwest Florida. Restoration Ecology 9:13-27. Quicke, H. E., R. S. Meldahl, and J. S. Ku sh. 1994. Basal area growth of individual trees: A model derived from a regional longleaf pine growth study. Forest Science 40:528-542. Quicke, H. E., R. S. Meldahl, and J. S. Kush. 1997. A survival rate model for naturally regenerated longleaf pine. Southern Journal of Applied Forestry 21:97-101. Rathbun, S. L., and N. Cressie. 1994. A space-time survival point process for a longleaf pine forest in southern Georgia. Journal of the American Statistical Association 89:1164-1174.
168 Rebertus, A. J., G. B. Williamson, and E. B. Moser. 1989. Longleaf pine pyrogenicity and turkey oak mortality in Florida xeric sandhills. Ecology 70:60-70. Reinhardt, E., and R. E. Keane. 1998. Fofem a first order fire effects model. Fire Management Notes 58:25-28. Riano, D., E. Meier, B. Allgower, E. Chuviec o, and S. L. Ustin. 2003. Modeling airborne laser scanning data for the spatial generation of cr itical forest parameters in fire behavior modeling. Remote Sensing of Environment 86:177-186. Riano, D., F. Valladares, S. Condes, and E. Chuvieco. 2004. Estimation of leaf area index and covered ground from airborne laser scanner ( lidar) in two contrasting forests. Remote Sensing of Environment 124:269-275. Roberts, S. D., T. J. Dean, D. L. Evans, J. W. McCombs, R. L. Harrington, and P. A. Glass. 2005. Estimating individual tree l eaf area in loblolly pine pl antations using lidar-derived measurements of height and crown dime nsions. Forest Ecology and Management 213:5470. Robertson, K. M., and T. E. Ostertag. 2007. Effects of land use on fuel characteristics and fire behavior in pinelands of s outhwest Georgia. Pages 181-191 in 23rd Tall Timbers Fire Ecology Conference: Fire in Grassland and Shrubland Ecosystems, Tall Timbers Research Station, Tallahassee, FL. Ryan, K. C. 2002. Dynamic interactions between fo rest structure and fire behavior in boreal ecosystems. Silva Fennica 36:13-39. Scott, J. H., and R. E. Burgan. 2005. Standard fire behavior fuel models: A comprehensive set for use with Rothermel's surface fire sp read model. RMRS-GTR-153, USDA Forest Service, Rocky Mountain Research Station, Fort Collins, CO. Silvertown, J., S. Holtier, J. Johnson, and P. Dale. 1992. Cellular automaton models of interspecific competition for space--the effect of pattern on process. The Journal of Ecology 80:527-533. Simkin, S. M., and W. K. Michener. 2005. Faunal soil disturbance regime of a longleaf pine ecosystem. Southeastern Naturalist 4:133-152. Slatton, K. C., M. Coleman, W. Carter, R. Shre stha, and M. Sartori. 2004. Control methods for merging ALSM and ground-based laser point clouds acquired under forest canopies. Pages 96-103 in 4th Internati onal Asia-Paci c Environmental Remote Sensing Symposium. SPIE. Solomatine, D. P., and K. N. Dulal. 2003. Model trees as an alternative to neural networks in rainfall-runoff modelling. Hydrological Sciences Journal 48 :399-411. Su, J. G., and E. W. Bork. 2007. Characterization of diverse plant communities in aspen parkland rangeland using lidar data. Applied Vegetation Science 10 :407-416.
169 Sun, L., and D. R. Weise. 2003. Experimental stu dy on fire behavior of live chaparral fuels. 2nd International Wildland Fire Ecology and Fire Management Congress, Orlando, FL. Tanaka, T., H. Park, and S. Hattori. 2004. Measur ement of forest canopy structure by a laser plane range-finding method: Improvement of ra diative resolution a nd examples of its application. Agricultural and Forest Meteorology 125:129-142. Thaxton, J. M., and W. J. Platt. 2006. Small-scal e fuel variation alters fire intensity and shrub abundance in a pine savanna. Ecology 87:1331-1337. Tong, W., H. Hong, H. Fang, Q. Xie, and R. Perkins. 2003. Decision forest: Combining the predictions of multiple independent decision tree models. Journal of Chemical Information and Computer Sciences 43:525-531. Van der Zande, D., W. Hoet, I. Jonckheere, J. van Aardt, and P. Coppin. 2006. Influence of measurement set-up of ground-based lidar for de rivation of tree struct ure. Agricultural and Forest Meteorology 141:147-160. Van Wagner, C. E. 1968. The line intersect method in forest fuel sampling. Forest Science 14:20-26. Varner, J. M. I., D. R. Gordon, F. E. Putz, and J. K. Hiers. 2005. Restorin g fire to long-unburned ( Pinus palustris ) ecosystems: Novel fire effects and consequences for long-unburned ecosystems. Restoration Ecology 13:536-544. Vasques, G. M., S. Grunwald, and J. O. Sickman. 2009. Modeling of soil organic carbon fractions using visible-near-infrared spectroscopy. Soil Sc ience Society of American Journal 73 :176-184. Wade, D. D., B. L. Brock, P. H. Brose, J. B. Grace, G. A. Hoch, and W. A. Patterson. 2000. Fire in eastern ecosystems. Pages 53-96 in J. K. Brown and J. K. Smith, editors. Wildland fire in ecosystems: Effects of fire on flora. US DA Forest Service, Rocky Mountain Research Station, Ogden, UT. Wahlenberg, W. G. 1946. Longleaf pine: Its us e, ecology, regeneration, protection, growth, and management. Charles Lathrop Pack Forestry Foundation, Washington, D.C. Walker, J., and R. K. Peet. 1983. Composition and species diversity of pine-wiregrass savannas of the green swamp, North Carolina. Vegetatio 55 :163-179. Ware, S., C. Frost, and P. D. Doerr. 1993. Sout hern mixed hardwood forest: The former longleaf pine forest. Pages 447-493 in W. H. Martin, S. G. Boyce, and A. C. Echternacht, editors. Biodiversity of the southeastern United States: Lowland terrestrial communities. John Wiley and Sons, New York. Watt, P. J., and D. N. M. Donoghue. 2005. Measuri ng forest structure with terrestrial laser scanning. International Jour nal of Remote Sensing 26:1437-1446.
170 Whelan, R. J. 1995. The ecology of fire. Camb ridge University Press, Cambridge, UK. Williamson, G. B., and E. M. Black. 1981. High te mperature of forest fires under pines as a selective advantage over oaks. Nature 293:643-644. Wing, M. G., A. Eklund, and J. Sessions. 2010. Appl ying lidar technology for tree measurements in burned landscapes. Internati onal Journal of Wildland Fire 19 :104-114. Wolf, P. R., and C. D. Ghilani. 1997. Adjustment computations: Statistics and least squares in surveying gis. John Wiley & Sons, Inc., New York, NY. Wright, S. J., H. C. Muller-Landau, R. Cond it, and S. P. Hubbell. 2003. Gap-dependent recruitment, realized vita l rates, and size distributions of tropical trees. Ecology 84:31743185.
171 BIOGRAPHICAL SKETCH Eva Louise Louderm ilk was born a first genera tion American from a Swedish family. She received a Bachelor of Science degree from the University of Florida in Animal Science in 2001. After gaining extensive field and laboratory experience, she continued her education in 2003 with the School of Forest Resources and Conserva tion at the University of Florida to receive her Master of Science degree in 2005. She immedi ately started her Doctor of Philosophy degree with the School of Natural Resources and Environment (Interdisciplinary Ecology) at the University of Florida. She is married to Er ick and has a son Kai, born during her time as a Doctor of Philosophy graduate student.