Ecosystems (2008) 11: 45–60 DOI: 10.1007/s10021-007-9106-z

Patterns of Forest Damage in a Southern Mississippi Landscape Caused by Hurricane Katrina John A. Kupfer,,* Aaron T. Myers, Sarah E. McLane, and Ginni N. Melton Department of Geography, University of South Carolina, 709 Bull Street, Room 127, Columbia, South Carolina 29208, USA

ABSTRACT Understanding and predicting the ways in which large and intense hurricanes affect ecosystem structure, composition and function is important for the successful management of coastal forest ecosystems. In this research, we categorized forest damage resulting from Hurricane Katrina into four classes (none, low, moderate, heavy) for nearly 450 plots in a 153,000 ha landscape in southern Mississippi, USA, using a combination of air photo interpretation and field sampling. We then developed predictive damage models using single tree classification tree analysis (CTA) and stochastic gradient boosting (SGB) and examined the importance of variables addressing storm meteorology, stand conditions, and site characteristics in predicting forest damage. Overall damage classification accuracies for a training dataset (n = 337 plots) were 72 and 81% for the single tree and SGB models, respectively, with Cohen‘s weighted linear j values of 0.71 and 0.86. For an independent validation dataset (n = 112 plots), classification accuracy dropped to 57% (j = 0.65) and 56% (j = 0.63) for the single tree and SGB models. Proportions of agreement between observed and predicted damage were significantly greater

(P < 0.05) than would be expected by chance alone for all damage classes with the training data and all but the moderate class for the validation data. Stand age was clearly the best predictor of damage for both models, with forest type, stand condition, site aspect, and distance to the nearest perennial stream also explaining much of the variation in forest damage. Measures of storm meteorology (duration and steadiness of hurricane-force winds; maximum sustained winds) were of secondary importance. The forest-wide application of our CTA model provided a realistic, spatially detailed map of predicted damage while also maintaining a relatively high degree of accuracy. The study also provides a first step toward the development of models identifying the susceptibility of forest stands to future events that could be used as an aid to incorporating the effects of large infrequent disturbances into forest management activities.

INTRODUCTION

distribution, susceptibility to subsequent disturbances, and the rate and pattern of energy flows and nutrient cycles (Tanner and others 1991; Boose and others 1994; Vandermeer and others 2001; Ostertag and others 2003; Litzgus and Mousseau 2004). Unusually large and intense hurricanes, despite their infrequency, likely play a dispropor-

Key words: large infrequent disturbance; classification tree; stochastic gradient boosting; DeSoto National Forest; predictive model; hurricane damage.

Hurricanes alter landscape-scale patterns of forest structure and composition, habitat availability and Received 16 August 2006; accepted 10 October 2007; published online 15 November 2007. *Corresponding author; e-mail: [email protected]

45

46

J. A. Kupfer and others

tionate role in restructuring ecosystem patterns and characteristics, are more likely to result in threshold-exceeding events that indefinitely alter ecosystem pattern and function, and have a wider range of environmental variation than do smaller disturbances, thereby creating a more diverse mosaic of conditions (for example, Conner and others 2005). They are nonetheless normal, integral parts of long-term system dynamics in many coastal forests in the Caribbean, Gulf and Atlantic Coast regions, which means that management plans need to recognize their effects and include the potential for such events to occur. There continues to be a particular need for research that helps land managers to better understand and predict ecosystem responses to large infrequent disturbances such as intense hurricanes (Dale and others 1998). In this study, we quantified factors associated with forest damage caused by Hurricane Katrina, one of the strongest hurricanes to hit the US coastline in the last century, and linked our findings to an empirical model of damage for a landscape in southern Mississippi, USA. At the scale of individual trees and small forest stands, the amount and type of damage caused by a hurricane (for example, uprooting, branch loss) is a function of disturbance intensity (for example, peak wind gusts) and site specific conditions, including: (1) tree height, age, health and other factors affecting susceptibility to high winds, (2) species composition, due to species-specific variations in wood strength, rooting pattern, canopy morphology, growth rate or leaf retention, (3) stand structure, condition and disturbance history, which influence the size and composition of the stand as well as the vertical wind field, and (4) soil conditions, antecedent soil moisture, geology and other factors affecting the rooting strength of individual trees (Everham and Brokaw 1996; Gresham and others 1991; Foster and Boose 1992; Veblen and others 2001; Peterson 2004; Ostertag and others 2005). At broader scales, damage patterns reflect spatial variability in tree- and stand-level factors (that is, heterogeneity in stand characteristics, environmental conditions, and disturbance history) as well as factors that emerge at the landscape- or watershed-scale to influence wind exposure or forest susceptibility, including gradients in wind and precipitation intensity and topographic setting (Boose and others 1994; Sherman and others 2001; Lindemann and Baker 2002; Platt and others 2002). Monitoring and assessing the biotic and abiotic conditions that are present in a forest ecosystem can help to explain patterns of damage that occur during an event (Brokaw and Walker 1991;

Foster and Boose 1992; Foster and others 1998) and may provide insights into post-disturbance ecosystem behavior. Hurricane Katrina provided an exceptional opportunity for addressing such questions because of its uncommon conditions and the large amount of geospatial data (for example, remotely sensed photographs and images) documenting its effects. Our specific objective was to develop and test an empirical model of forest damage resulting from Hurricane Katrina for a 153,000 ha forest management unit in southern Mississippi. Although few studies have analyzed damage at such a broad scale for hurricanes and there is a perception that such models are impractical due to an absence of fine-scale wind data, such models are feasible (Ramsey and others 2001; Lindemann and Baker 2002; Uriarte and others 2004). We used classification tree analysis to develop a model of damage severity on the basis of storm meteorology, stand conditions, and site characteristics for more than 300 locations and applied the model to an independent validation dataset. By including variables addressing wind speed, rainfall, species composition, forest structure and age, topography and floodplain conditions, we attempted to contrast the importance of different types of factors associated with landscape-scale patterns of forest damage. Hurricane damage can be both immediate and prolonged and is caused by a number of interacting factors, including high winds, heavy rainfall, the mechanical and chemical effects of storm surge, stress from hurricane-induced damage, and mortality from subsequent post-hurricane disturbances (for example, fires related to downed slash: Myers and Van Lear 1998). Here, we focus explicitly on immediate post-hurricane damage patterns outside the storm surge zone, thereby emphasizing the direct effects of high winds and precipitation.

METHODS Study Site The study was conducted in the DeSoto Ranger District of DeSoto National Forest (NF) (Figure 1). The climate is marked by mild, short winters and hot, humid summers (July mean temperatures: high = 33.3
C, low = 21.3
C; January mean temperatures: high = 15.8
C, low = 2.8
C; data for Wiggins, Mississippi, 1971–2000; Southeast Regional Climate Center: http://www.sercc.com/climateinfo/historical/historical.html). Precipitation is high (annual mean: 164 cm) and evenly distributed throughout the year.

Forest Damage by Hurricane Katrina

47

Figure 1. Location of the study area, DeSoto National Forest, including the approximate track of the eye of Hurricane Katrina.

Topography is characterized by broad and gently sloping uplands dissected by numerous streams and rivers, the largest of which have mature floodplains and noticeable topographic changes in areas transitional to upland systems. Uplands are generally dominated by pines, especially longleaf pine (Pinus palustris), loblolly pine (P. taeda), shortleaf pine (P. echinata) and slash pine (P. elliottii). Bottomlands are dominated by hardwoods, including various species of oak (Quercus spp.), sweetgum (Liquidambar styraciflua) and many others. Much of DeSoto NF has been managed for timber production, cre-

ating a diverse range of forest types, structures and ages fragmented by a large system of roads used for timber harvesting and recreation. From 1722 to 2005, 45 hurricanes made landfall on the central Gulf Coast between Houma, Louisiana and Mobile, Alabama (Graumann and others 2005), but major hurricanes such as Hurricane Katrina with winds strong enough to cause significant and widespread damage more than 100 km inland are not common (for example, Elsner and others 2006). While centered over the Gulf of Mexico on 28 August 2005, it had maximum sus-

48

J. A. Kupfer and others

tained winds of 280 km h)1 and hurricane-force winds extending more than 170 km from the eye (Graumann and others 2005). It made landfall on 29 August 2005 in southeast Louisiana as a Category 3 hurricane (on the Saffir-Simpson Hurricane Scale) with sustained wind speeds of 180– 200 km h)1 and again on the Mississippi coastline near the mouth of the Pearl River. At its closest, the hurricane‘s eye passed within approximately 12 km of the western extent of DeSoto NF, exposing the study area to the storm‘s strongest winds (Figure 1). Although the winds weakened as the storm moved inland, estimates from NHC models suggest that all of the study area experienced hurricane-force winds for at least two hours, with maximum sustained winds in DeSoto NF averaging 135–160 km h)1 (Category 1 and 2) and peak gusts of 145–225 km h)1. Rainfall for stations in and around DeSoto NF ranged from 15 to 20 cm.

Forest Damage Assessments Assessments of damage severity were conducted using aerial photographs taken by a private contractor for the U.S. Army Corps of Engineers in September 2005 and retrieved from the Mississippi Automated Resource Information System (MARIS) Technical Center website (http://www.maris.state.ms.us/HTM/DownloadData/ADS40_Imagery.html; last accessed 30 August 2007). The raw imagery used to generate the final image tiles was acquired using an ADS40 digital airborne sensor and collected simultaneously with airborne GPS and IMU data for georeferencing. The imagery was rectified to a plane of constant elevation and mosaiced to generate the final image tiles, with the end result being georeferenced, planar-rectified imagery with a 0.33 m ground resolution. To assist in our evaluation of forest damage by providing pre-storm imagery, we also acquired 1 m natural color imagery taken in 2004 as part of the USDA Farm Services Agency National Agriculture Imagery Program (NAIP). After image acquisition and rectification, we selected 380 random points for forest damage analysis using the Hawth Tools extension for ArcGIS. Field surveys at DeSoto NF indicated that damage in bottomlands differed from that in adjacent uplands. We therefore generated 400 additional random points and selected the first 35 that fell within bottomland areas to increase the sample size for those environments, giving us a total of 415 photointerpreted points. We created a 17.8 m buffer around each point, to be consistent with field assessments (see below), and damage was classified

using a four-point scale based on the estimated percentage of downed overstory trees: (1) no discernible downed trees, (2) light damage (<33% blowdown), (3) moderate damage (33–67% blowdown), and (4) heavy damage (>67% blowdown) (Figure 2). If the plot buffer included areas of obviously differing damage or land use, the plot center was shifted so that the buffer fell entirely within the dominant damage or land use class. Damage was independently classified by two different researchers with 88% agreement; where a difference of opinion occurred, the senior author made the final damage assignment. The photo-interpreted points were supplemented with data from detailed field assessments of forest damage on forty 0.1 ha circular plots established in February 2006 as part of a more comprehensive evaluation of Hurricane Katrina impacts. All standing and fallen trees greater than 10 cm in diameter at breast height (for standing trees) or base (for downed trees) were identified to species and measured. Plot locations were recorded using a WAAS-enabled Garmin eTrex Legend GPS unit with mean positional accuracy of 5 m. Based on trees that were deemed to have been overstory dominants or co-dominants, we assigned damage classes for these points using the same criterion employed for the photo-interpreted points. We did not consider trees that likely fell after the hurricane (based on their direction of fall vs. the predominant direction of blowdown) or trees that showed evidence of decomposition and clearly fell prior to the hurricane. Of these 40 plots, only 34 were used due to missing ancillary information needed for the predictive models. Because of their small number and spatial concentration in the northern third of DeSoto NF, the field-surveyed points were combined with the photo-interpreted points rather than being used as a separate validation dataset, yielding a total of 449 sample points.

Predictor Variables We used four measures of storm meteorology as indicators of storm intensity (Table 1). Estimated cumulative rainfall was based on data compiled by NOAA‘s Climate Prediction Center (Graumann and others 2005) whereas estimated maximum sustained wind speed, duration of hurricane-force winds, and steadiness of wind direction were produced by NOAA‘s Hurricane Research Division (HRD) using the hurricane wind analysis system H*Wind (Powell and others 1998). H*Wind combines observations from a range of sources and processes them to provide estimates of wind speed

Forest Damage by Hurricane Katrina

49

Figure 2. Examples of plots at DeSoto National Forest classified as having light damage (<33% canopy blowdown; top row), moderate damage (33–67% canopy blowdown; middle row) and heavy damage (>67% canopy blowdown; bottom row).

for a standardized height (10 m), surface (open terrain) and averaging period (maximum sustained 1-min wind speed). These estimates thus represent meso-scale features of the storm rather than localized wind effects. Steadiness is defined by the ratio of the vector mean wind to the scalar mean wind over the time period required for a storm to traverse a region, with low values indicating more rapid changes in wind direction during the storm‘s passage (Dunion and others 2003). The estimated hurricane wind data were created for a grid of points spaced at 0.054
intervals of latitude and longitude using 29 August 11:32 UTC landfall analysis and projecting the peak sustained winds

along the observed track of the surface circulation center at 10 min intervals using the HRD inland decay model. After downloading the data from http://www.aoml.noaa.gov/hrd/Storm_pages/katrina2005/wind_realtime.html (last accessed, 30 August 2007), we created surfaces for each of the three variables using inverse distance weighting and extracted values for each of our data points using ArcGIS v.9.2. To capture the effects of site setting and context, we extracted three topographic variables (elevation, slope angle, slope aspect) from 10 m digital elevation models (Table 1). Aspect values were grouped into 45
-wide azimuth arcs centered on

Mean: 60.5; Range: 1–>120 Frequencies, by group # to left: 1: 145 plots; 2: 46 plots; 3: 83 plots; 4: 16 plots; 5: 17 plots; 6: 13 plots; 7: 59 plots; 8: 70 plots hardwoods dominant; 8. Hardwoods: hardwoods dominant Frequencies, by group # to left: 1: 8 plots; 2: 179 plots; 3: 61 plots; 4: 138 plots; 5: 63 plots. Mean: 17.9; Range: 0–142 Mean: 35.7; Range: 0–130 Mean: 53.6; Range: 0–160

From CISC database; defined as years since stand origination From CISC database; grouped into: 1. Pine: P. palustris dominant; 2. Pine: P. taeda dominant; 3. Pine: P. elliottii dominant; 4. Pine: mixed yellow pines; 5. Pine-hardwood: non-P. taeda dominant; 6. Pine-hardwood: P. taeda dominant; 7. Hardwood-pine: From CISC database; grouped into: 1. sparse pole- and sawtimber; 2. mature sawtimber and poletimber; 3. immature poletimber; 6. immature sawtimber; 5. regeneration From CISC database From CISC database From CISC database

Mean: 44.8; Range: 4.9–97.8 Mean: 2.1; Range: 0–8.6 Frequency: North: 41 plots; Northeast: 53 plots; East: 61 plots; Southeast: 55 plots; South: 67 plots; Southwest: 57 plots; West: 46 plots Northwest: 29 plots; Flat: 40 plots Mean: 552; Range: 0–2,471 Mean: 239; Range: 0–1,575 Mean: 41.5; Range: 11.2–75.1

Extracted from 10 m digital elevation model Extracted from 10 m digital elevation model Extracted from 10 m digital elevation model; grouped into nine Classes: North: 337.5
–22.5
; Northeast: 22.5
–67.5
; East: 67.5
–112.5
; Southeast: 112.5
–157.5
; South: 157.5
–202.5
; Southwest: 202.5
–247.5
; West: 247.5
–292.5
; Northwest: 292.5
–337.5
; Flat: no slope. Based on U.S. Geological Survey hydrography data Based on U.S. Geological Survey hydrography data Based on NOAA map of hurricane track

Pine forest types: ‡70% of the basal area of trees with dominant and co-dominant crowns were softwoods; Pine-Hardwood: 51–69% softwood basal area; Hardwood-Pine: 51–69% hardwood basal area, and Hardwood: ‡70% hardwood basal area.

1

Hardwood Basal Area (ft2 acre)1) Pine Basal Area (ft2 acre)1) Total Basal Area (ft2 acre)1)

Stand condition

Distance to nearest perennial stream (m) Distance to nearest stream (m) Distance to hurricane track (km) Stand characteristics: Age (years) Forest types1

Mean: Mean: Mean: Mean:

150.4; Range: 134.2–161.6 4.37; Range: 2.15–5.63 0.716; Range: 0.510–0.799 14.0; Range: 10.2–20.5

Data values

Extracted from wind fields estimated by NOAA H*Wind model Extracted from wind fields estimated by NOAA H*Wind model Extracted from wind fields estimated by NOAA H*Wind model Data compiled by NOAA‘s Climate Prediction Center

Description/Source

Variables used to Predict Forest Damage from Hurricane Katrina at DeSoto National Forest, including Data Summaries for the 449 Sample

Storm meteorology Maximum sustained wind speed (km h)1) Duration of hurricane force winds (h) Directional steadiness (see text) Cumulative precipitation (cm) Site topography and context: Elevation (m above sea level) Slope angle (
) Aspect

Variables

Table 1. Sites

50 J. A. Kupfer and others

Forest Damage by Hurricane Katrina the cardinal and ordinal directions (north = 337.5– 22.5
; northeast = 22.5–67.5, … , northwest = 292.5–337.5
), with a ninth category for flat sites. Ideally, the potential importance of floodplain soils would have been captured directly, but finescale soils data were not available for a portion of our study area. We therefore calculated two measures of river/stream proximity: distance to the nearest perennial stream and distance to the nearest stream of any kind, perennial or intermittent (based on U.S. Geological Survey hydrography data available through MARIS). Additionally, we calculated the Euclidean distance from each plot to the nearest point of the hurricane track. Finally, data on stand characteristics were extracted from the DeSoto NF Continuous Inventory of Stand Conditions (CISC) database. This database contains information on forest attributes for individual stands approximately 5–20 ha in size with homogenous forest community characteristics. Stand data are updated periodically through field inventories using standardized data collection procedures. Because data are collected forest-wide for hundreds of individual stands, the CISC database provided an unusually comprehensive and detailed record of pre-hurricane tree cover characteristics at the scale of the entire national forest. Specific stand variables selected for our analyses are summarized in Table 1.

Data Analyses To better understand and predict forest damage severity, we used classification tree analysis (CTA), a non-parametric, probabilistic machine-learning method that recursively partitions observations with a categorical response variable based on binary splitting criteria applied to predictor variables (Breiman and others 1984). The rule that results in the greatest increase in class purity forms the first splitting rule of the tree, with the process continuing until nodes reach a defined level of homogeneity or contain a minimum number of data points. Predictor variables may be reused so hierarchical, nonlinear relationships can be derived. Classification trees have been widely used in predictive vegetation models in recent years (for example, Cairns 2001; Miller and Franklin 2002) because they provide a flexible, easy-to-interpret, nonparametric alternative to methodologies with more stringent assumptions. We began our analyses by withholding a random selection of 25% of the points to create a training dataset (n = 337 points) and a validation dataset (n = 112 points). Two primary decisions in building

51

classification trees are the criterion used to split the nodes and the method used to prune (that is, simplify) the tree. For the tree-fitting algorithm, we used the Gini splitting method, which maximizes the heterogeneity of the categories of the target variable in each child node (Breiman and others 1984). We also tried the entropy and misclassification cost methods but both resulted in substantial increases in misclassifications of the training data. Second, using an unpruned tree can result in overly detailed models that fit a training dataset well but are poor predictors for other data. To limit overfitting, we pruned the tree using an additional set of randomly withheld points (20% of the training data) such that the cross-validated error cost of the smaller tree was no more than one standard error from the minimal cross-validated error. The optimal classification tree for predicting forest damage was determined by varying two controls on tree structure, the minimum size node to split and the maximum number of tree levels. We tested minimum node sizes from 5 to 30 and maximum number of tree levels ranging from 4 to 12 and compared overall prediction accuracies for the validation dataset for all models, selecting the model with the highest predictive accuracy. Because several combinations of node size and tree levels shared the highest accuracy, we selected the most parsimonious model, that is, the tree with the fewest levels and terminal nodes. Classification tree analysis has been shown to provide an effective means for predicting class memberships but can be sensitive to outliers and unbalanced datasets. Further, early splits that might be rejected because they do not create the best initial classes might actually enable better splits at lower levels and a better overall classification (Lawrence and others 2004). Two recently developed methods for improving its predictive ability are boosting (Freund and Schapire 1996), an iterative process that develops new classification trees based on misclassifications from previous trees, and bagging (Breiman 1996), a method that uses resampling and bootstrapping to train classifiers based on random redistributions of the training dataset. In contrast to more widely used ‘single tree‘ CTA, boosting and bagging are examples of ‘voting‘ or ‘ensemble‘ methods that operate by classifying observations based on a majority or weighted majority vote of multiple trees. In addition to performing single tree CTA, we predicted forest damage class using a form of CTA termed stochastic gradient boosting (Friedman 2002) that is included as the TreeBoost algorithm in

52

J. A. Kupfer and others

the software package DTREG v. 5.0 (Sherrod 2006), which we used for all of our analyses. Stochastic gradient boosting (SGB) is a hybrid of the boosting and bagging approaches that starts by fitting an initial tree to the data. The residuals from the first tree are fed into a second tree, which attempts to reduce the error, a process that is repeated through a series of successive trees (Lawrence and others 2004). The final predicted value is formed by adding the weighted contribution of each tree. Individual trees are usually fairly small (3–6 levels deep with a limited number of terminal nodes), but the full additive series may consist of hundreds of small trees. An important benefit of the SGB approach is that it typically achieves the accuracy of other boosting methods with lower sensitivity to misclassified cases and outliers. Our final model was based on the outcome of 200 individual trees with a maximum depth of 5 for any tree in the series, a minimum node size of 3, and pruning by crossvalidation with a withheld dataset. Because it is an ensemble method, SGB does not provide a viewable decision tree like single tree CTA does. Both single tree CTA and SGB models in DTREG, however, provide a variable importance score to clarify the relationships between forest damage and the predictor variables. This score is calculated on the basis of the improvement in classification gained by each split that used the predictor, with values scaled such that the most important predictor was assigned a value of 100.0 whereas other predictors had lower scores in decreasing degrees of importance (Sherrod 2006). Predictions from the optimal single tree and SGB models were assessed with contingency matrices and accuracy measures, including: (1) overall accuracy, the percentage of plots for which the damage class was correctly predicted, (2) producer‘s accuracy, the probability that a certain damage class was predicted correctly (that is, omission errors), and (3) user‘s accuracy, the probability that a plot predicted as a certain damage class actually was that class (that is, commission errors). As some agreement between observed and predicted damage class is expected by chance, we also calculated a quadratic-weighted version of Cohen‘s kappa (j) coefficient (Cohen 1968). The weighted j takes into account the ordinal scale of our response variable (no damage < light damage < moderate damage < heavy damage) and uses a weighting function to weight degrees of disagreement between ordinal measurements. Mispredictions between ‘adjacent‘ damage classes (for example, none vs. light; moderate vs. heavy) were thus weighted less than those between more distal

damage classes (for example, light vs. heavy). As with unweighted j, values greater than 0.75 signify excellent agreement whereas values less than 0.40 indicate poor agreement. When quadratic weights are used, j is equivalent to the intraclass correlation coefficient (Fleiss 1981). Independent of j, we assessed whether predicted agreement between the observed and predicted damage was greater than that expected by chance for each of the individual damage categories. To do so, we calculated 95% confidence intervals for the observed proportion of agreement using the Wilson efficient-score method corrected for continuity (Newcombe 1998) and compared it to the proportion of agreement expected by chance. Finally, we believe our methodology could be useful for producing rapid post-disturbance damage predictions via the extrapolation of analyses from a number of field plots to a larger landscape. We therefore compared forest-wide damage patterns predicted by our single tree CTA model to damage classes mapped by the USDA Forest Service. Because of the need for an immediate posthurricane assessment, Forest Service personnel hand-mapped damage on a district-wide USGS topographic map from a helicopter shortly after the hurricane; this map was refined following a few days of ground observations. Damage classification was subjective, but approximate definitions of the three classes used were similar to those used in our analyses: light (<33% of the overstory down), moderate (33–67% of the overstory down), and heavy (>67% of the overstory down) (W. Stone, USDA-FS, personal communication). The primary reason for conducting this assessment was to prioritize areas for salvage logging operations so damage polygons were large (tens to thousands of hectares) and included considerable internal variation in damage. For example, in a small sample of mature stands classified as having suffered heavy damage (and thus expected to have lost >67% of its overstory), Meeker and others (2005) noted that severely damaged trees (those that were dead, dying or likely to die as a result of hurricane damage) constituted from 30–83% of the trees and 15–98% of the basal area per hectare.

RESULTS Classification Accuracy Overall classification accuracy of the optimal single tree model for the training data was 71.5%, with producer‘s accuracy ranging from 65 to 82% and

Forest Damage by Hurricane Katrina user‘s accuracy varying from 58 to 82% (Table 2). Most errors were made to an adjacent damage class. The weighted linear j was 0.71, and proportions of agreement were significantly greater than expected by chance (P < 0.05) for all classes. For the validation data, classification accuracy decreased, but damage class was still correctly predicted for 57% of the plots. Producer‘s and user‘s accuracy ranged from 37–75 to 39–82%, respectively, with a j of 0.65. The most common errors involved misclassification and misprediction of the light versus moderate damage classes, with the latter not predicted better than expected by chance alone (chance proportion of agreement expected: 0.15; 95% confidence interval of observed proportion of agreement: 0.13–0.38). The use of SGB increased the overall accuracy of predictions to 81% (j = 0.86) for the training data and improved producer‘s and user‘s accuracy over values for the single tree model in all cases but one (Table 2). For the validation data, however, overall accuracy (56%) and j (0.63) were essentially the same as for the single tree model. Three classes were predicted significantly better than expected by chance, but the results were again marginal for the moderate damage class (chance proportion of agreement expected: 0.15; 95% confidence interval of observed proportion of agreement: 0.15–0.41). Although SGB predicted some classes better than the single tree model for the validation data, other prediction accuracies were worse.

53

Predictor Importance and Effects Age was the best predictor of forest damage for both the single tree and SGB models (Table 3). A number of other predictors showed moderate to high importance in both models, particularly forest type, stand condition, aspect, and distance to the nearest perennial stream. The storm meteorology variables were generally of limited importance in the single tree model but moderate importance in the SGB model. The effects of specific variables on forest damage can be interpreted by examining the results of the single tree CTA (Figure 3). The initial division separates plots by age, with plots less than 31 y old classified as having no discernible damage. Older plots are then split into two groups on the basis of forest type, with the first group containing most pine and pine-hardwood plots. Plots in this group were mostly classified as having light or no damage, with specific damage levels dependent upon aspect (more damage typically on south- and westfacing slopes), age (more damage on older plots), duration of hurricane-force winds (more damage with longer duration), distance to the hurricane‘s track (more damage closer to the track), stand condition (less damage on regenerating and immature poletimber plots), species composition (more damage on plots with a larger hardwood component) and distance to a perennial stream (more damage closer to streams). The second group of older stands contained hardwood-dominated

Table 2. Confusion Matrices of Predicted Forest Damage at DeSoto National Forest based on Single Tree and Stochastic Gradient Boosting (SGB) Classification Tree Analysis Predicted None

Light

Moderate

Heavy

User‘s accuracy

Single SGB Single SGB Single SGB Single SGB Single (%) SGB (%) Training data (n = 337 plots) Observed None Light Moderate Heavy Producer‘s accuracy (%) Validation data (n = 112 plots) Observed None Light Moderate Heavy Producer‘s accuracy (%)

84 13 4 2 81.6

86 10 2 1 86.7

6 53 11 6 69.7

11 77 13 4 73.3

9 18 62 8 65.3

6 3 60 2 84.5

4 7 8 42 68.9

0 1 10 51 82.3

81.6 58.2 72.9 72.4 71.5

83.5 84.6 70.6 87.9 81.3

27 5 3 1 75.0

23 4 1 0 82.1

1 15 8 1 60.0

5 17 13 4 43.6

5 7 11 7 36.7

3 10 12 5 40.0

0 4 6 11 52.4

2 0 2 11 73.3

81.8 48.4 39.3 55.0 57.1

70.0 54.8 42.9 55.0 56.3

Bold values represent correspondence between observed and predicted damage levels.

54

J. A. Kupfer and others

Table 3. Predictor Variable Importance Scores for Single Tree and Stochastic Gradient Boosting (SGB) Classification Tree Analyses of Forest Damage Variable

Age Stand condition Forest type Aspect Total basal area Distance to perennial stream Hardwood basal area Distance to any stream Distance to hurricane track Elevation Steadiness of wind direction Duration of hurricane winds Pine basal area Cumulative precipitation Max. sustained winds Slope

Variable importance scorea

Effectb

Single tree

SGB

100.0 59.5 31.6 29.2 23.7 23.4 18.7 17.5 16.9 16.8 12.7 5.7 5.4 2.2 1.6 )

100.0 27.3 83.1 62.4 14.0 28.4 12.1 15.0 23.8 16.5 23.4 14.6 20.5 7.4 20.8 18.4

+ +1 +2 +3 + ) + ) ) ± ) + + + + ±

a The score is calculated on the basis of the improvement in classification gained by each split that used the predictor, with values scaled such that the most important predictor was assigned a value of 100.0. b Effect is the relationship between the predictor and amount of damage (+ = positive; ) = inverse): 1Greater damage on more mature stand conditions (saw timber and pole timber); 2greater damage with increasing hardwood component; 3greater damage on slopes exposed to dominant wind flow (east, southeast, south, southwest)

plots and pine/pine-hardwood stands dominated by P. taeda. These plots were mostly classified as having moderate or heavy damage. Reinforcing some of the previous findings, specific damage levels varied with stream proximity, age, distance to the hurricane‘s track and aspect.

Comparison to Damage Mapped by Forest Service To enable comparisons with the classification produced by the Forest Service, we aggregated our ‘no damage‘ and ‘light damage‘ classes, resulting in three equivalent damage categories (light, moderate, heavy). When we overlaid the sample points on the Forest Service damage map, there was only a 44% correspondence in damage class. The greatest agreement was for light damage, with 59% of the lightly damaged points located in areas defined by the Forest Service as having sustained light damage. Conversely, few of the moderate (30%) and heavy (25%) damage points were located within areas mapped as the same class, with 43% of the heavy damage plots actually occurring in areas mapped as light damage. Many of these cases were bottomland hardwood forests that suffered heavy or moderate damage but were not well-captured by the Forest Service classification due to its coarser

scale. Of the plots located in areas mapped as heavy damage, 46% sustained only light damage. Many of these misclassifications involved points that were in heavily affected areas located close to the hurricane track but that had a low susceptibility to damage (for example, young and immature stands on protected slopes). Spatially, the application of our CTA model to map damage for all of DeSoto NF resulted in much more detailed projections (Figure 4). The resulting map shows some basic similarities to the National Forest Service damage map, but it also exhibits some areas of substantial disagreement (for example, in the southern third of the study area, which was mostly classified as moderate damage by the Forest Service but light damage by our model) and finer-scale variations that were more consistent with field surveys of damage. The result is a substantially different, and more realistic, perspective on the pattern of hurricane damage (Figure 5).

DISCUSSION

AND

CONCLUSIONS

Despite fundamental differences among various types of large, infrequent disturbances such as volcanic eruptions, high intensity wildfires, exceptional floods, and extreme wind events, postdisturbance responses in all cases are shaped by the

55

Forest Damage by Hurricane Katrina

N=337

AGE ≤ 30.5

N=62 Dam=0 88.7%

AGE > 30.5 N=275

FORTYPE=1, 3, 4, 5

FORTYPE=2, 6, 7, 8 N=128

N=147

ASP≠SW

ASP=SW N=20 Dam=1 70.0%

STRM ≤ 54.1 N=82

N=127

AGE ≤ 82.5

DURAT ≤4.701

BAPINE ≤ 4

AGE > 82.5

DURAT > 4.701

COND=Mature

DISTKAT≤15.1 DISTKAT>15.1

N=57

FORTYPE=5

FORTYPE=5 N=5 Dam=1 80.0%

FORTYPE=1, 3, 4

N =4 Dam=2 100.0%

N=27

N=2 Dam=3 100.0%

N=10 Dam=0 70.0%

FORTYPE=1, 3, 4

N=11 Dam=0 45.6%

ELEV≤132.5 N=16 Dam=1 62.5%

ELEV>132.5 N=26 Dam=2 46.2%

N=16 Dam=1 68.8%

ELEV≤148.5

ASP≠ NW, N, FLAT N=21

AGE ≤ 89.5

AGE > 89.5

N=4 4

AGE ≤ 86.5

AGE > 86.5

N=18 Dam=2 44.5%

N=3 Dam=0 100.0%

N=15 Dam=3 67.7%

N=31

PEREN>307

N=42

N=3 Dam=1 100.0%

N=22 Dam=3 77.3%

N=53

PEREN≤307

N=22 Dam=0 40.0%

ELEV>148.5

N=37

ASP=NW, N, FLAT

ELEV>227.5

N=68

COND≠Mature

N=9 Dam=3 77.8%

N=71

ELEV≤227.5

N =67

N=13 Dam=1 69.3%

N=29

BAPINE > 4

N=13 Dam=2 77.0%

N=5 Dam=2 80.0%

N=42

N=11 Dam=2 81.8%

N =80

N=47

N=46

AGE > 63.5 DISTKAT≤29.5 DISTKAT>29.5

AGE ≤ 63.5

ASP≠N, E, SE

ASP=N, E, SE

STRM > 54.1

ASP=S, SW, W ASP≠ S, SW, W N=13 Dam=3 46.2%

N=18 Dam=2 72.2%

FORTYPE codes: 1 Pine: P. palustris dominant 2 Pine: P. taeda dominant 3 Pine: P. elliottii dominant 4 Pine: Mixed yellow pines dominant 5 Pine-hardwood: non-P. taeda dominant 6 Pine-hardwood: P. taeda dominant 7 Hardwood-pine: hardwoods dominant 8 Hardwood: hardwoods dominant

Figure 3. Classification tree for forest damage caused by Hurricane Katrina at DeSoto National Forest. Terminal leaves are shown as squares, with shading representing damage class (white no discernible damage or light damage; gray moderate damage; black heavy damage). Number of plots in each terminal leaf and classification accuracy are indicated. Abbreviations: AGE stand age, FORTYPE forest type; ASP aspect class, BAPINE pine basal area, DURAT duration of hurricane-force winds, DISTKAT distance to the hurricane track, PEREN distance to the nearest perennial stream, ELEV elevation, STRM distance to the nearest stream.

spatial mosaics of disturbance severity and the pattern and abundance of ecological residuals left behind (for example, surviving trees or seeds) (Turner and others 1998). Broad-scale patterns of wildfire intensity and severity, for example, are regularly mapped and studied to better understand the effects on soil conditions and biotic habitat and the implications for post-fire recovery (for example, Miller and others 2003; Lewis and others 2006), but similar studies following hurricanes or other severe wind events are less common. In this research, we found that damage patterns following Hurricane Katrina over a large, heterogeneous landscape were most strongly related to stand conditions and site characteristics. Relation-

ships between damage and independent predictors corroborated findings from other studies, for example, the associations between damage and age/stand condition, soil moisture (in this case, stream proximity and amount of hardwoods, which occur mostly in floodplain settings) and slope aspect (for example, Everham and Brokaw 1996). The use of classification tree analysis also helped to capture some of the interactions among predictors. For example, upland pine and pine-hardwood stands only suffered heavy damage within 15 km of the hurricane‘s track whereas heavy damage on less-resistant bottomland hardwood and hardwood-pine stands extended 30 km from the hurricane‘s track.

56

J. A. Kupfer and others Figure 4. Pattern of forest damage predicted for DeSoto National Forest by the single tree classification model shown in Figure 3. Inset shows damage as mapped by the USDA Forest Service.

Beyond the importance of stand conditions and site characteristics, damage was less strongly related to measures of storm meteorology. The secondary importance of wind variables does not, however, mean that hurricane intensity is unimportant as a control of forest damage; rather, it reflects the strength of the winds experienced by most of the study area even as they diminished inland from the hurricane‘s landfall along the Mississippi coast and the relationship between the spatial scales of Hurricane Katrina and DeSoto National Forest. In a study relating estimated wind gust speeds to tree damage suffered in Puerto Rico during Hurricane Hugo, Francis and Gillespie (1993) found that

uprooting and trunk snapping were rare or nonexistent below 120–130 km h)1 but accounted for 10–20 and 5–10% (respectively) of all damaged trees, above this threshold, independent of increases in gust speed. Beyond 130 km h)1, the effect of increasing gust speed on tree damage was dependent on the interplay of species characteristics, individual tree morphology, soil factors and storm duration. Maximum sustained winds from Hurricane Katrina were estimated to have averaged 135– 160 km h)1 in DeSoto NF, with peak gusts of 145– 225 km h)1. This means that baseline winds, excluding the effects of local controls and micro-

Forest Damage by Hurricane Katrina

57

Figure 5. Example of differences between damage classes as mapped at a relatively coarse scale by the USDA Forest Service (bottom left) and those predicted by classification tree analysis (bottom right) for a small portion of DeSoto National Forest.

scale wind features, exceeded the cited 120– 130 km h)1 threshold in all of our study area. The coarseness of their wind gust estimates and the importance of local contingencies caution against placing too much emphasis on the specific threshold cited by Francis and Gillespie (1993), but their general findings are consistent with observed patterns of damage in this study. Specifically, damage was most closely tied to local factors influencing stand susceptibility, with some locations more than 50 km from the hurricane‘s eye suffering heavy or moderate damage (primarily floodplain locations where greater soil moisture and unconsolidated soils lowered tree resistance to

blowdown) whereas other locations less than 20 km from the eye (mostly young stands dominated by smaller, more flexible pines) suffering only light damage. If we had studied a larger area containing a greater range of wind speeds, damage would undoubtedly have varied with storm intensity. This scale-dependent nature of damage prediction, where predictors of damage reflect differences in the range and spatial variability of abiotic and biotic conditions, also perhaps helps to explain ambiguous or even contradictory findings from previous studies of hurricane-caused forest damage (for example, the role of topographic position; Everham and Brokaw 1996).

58

J. A. Kupfer and others

We believe that our study fills an important niche in hurricane-damage research. Most assessments of hurricane effects are either plot based, with limited spatial application, or landscape based, with little accuracy at particular points. Landscapescale assessments typically appear as homogeneous brushstrokes that exclude fine-scale variability with an underlying caveat that ‘‘local conditions can result in different effects‘‘; the results are maps similar to those produced by the Forest Service (Figures 4, 5). The map produced by applying CTA in this study provided spatial detail in a broad-scale prediction of forest damage while also maintaining a relatively high degree of accuracy. In terms of plot-level accuracy, researchers agreed on damage classifications 88% of the time whereas model predictions were 72–81% (for the training data) or 56–57% (for the validation data) accurate. Model accuracies were consistently better than would be predicted by random chance alone and were higher than those associated with the broad-scale Forest Service classification while also providing more detailed and realistic damage maps. Our predictive accuracy was likely hampered by limitations in data quality and accuracy stemming from the large extent of our study area and the available data. One obvious potential source of error is human misclassification of damage from the aerial photographs, particularly in distinguishing stands in adjacent classes. This might especially have contributed to errors in predicting the Moderate Damage Class, where errors could be made on the ‘low‘ or ‘high‘ end of the class. Nonetheless, independent classifications of site damage by multiple researchers generally showed strong agreement, and a limited comparison of field- and image-based damage assessments for the same plots suggested that such error was minimal. Less easy to control were errors in the predictor variables. We classified damage at the scale of 0.1 ha plots but corresponding predictor variables extracted from the CISC database are aggregated to entire forest stands. Even though stands are defined as ‘homogeneous‘ entities, some error would occur from assigning stand-level properties to individual plots. Further, stands varied in the time since they were last inventoried, which would cause errors in nonstatic variables such as basal area. Predictive accuracy could also have been improved by the inclusion of variables that were either represented by simplified proxies (for example, floodplain soils) or were unavailable. For example, a measure of landscape openness (for example, the proximity to open areas of some minimum size) may have improved accuracy because of its rela-

tionship to surface wind speeds and turbulence. Further, hurricane damage has often been linked to not only mean sustained winds speeds but also maximum wind gust speed, but the meteorological variables that we used are meso-scale wind approximations of conditions that do not include the occurrence of hurricane-generated microbursts or tornadoes. Although it is possible to use correction factors in H*Wind to convert maximum sustained wind speeds to peak gusts (for example, Powell and others 2004), such measures still do not capture significant localized events. The results of our analyses and mapping efforts underscore that landscape-scale damage prediction using readily available or easily acquired data for a given event is feasible. Our methodology could thus serve as a complementary means of conducting rapid post-disturbance assessments of forest damage following an event through the establishment of a network of readily accessible monitoring plots that are designed to capture the range of conditions present within a management unit. Rapid, ground-based visual assessments of damage at these monitoring points following an event could then be used to quickly assess forest damage on a larger spatial scale with greater accuracy and less cost and time than could be done from assessments such as those employed by the Forest Service, particularly in cases where remotely sensed imagery may not be as readily available as it was in the case of Hurricane Katrina. Finally, although our focus was on understanding and predicting patterns of forest damage from Hurricane Katrina as a means of examining landscape-scale forest responses to a high intensity wind event, a longer-term objective is to project the susceptibility of forest stands to future events as an aid to forest management activities. In their review of the management implications of large infrequent disturbances (LIDs), Dale and others (1998) stressed that management plans need to recognize the occurrence and ecological roles of LIDs. In particular, they noted that forest systems can be managed prior to LIDs in ways that alter their vulnerability or change how they will respond to a disturbance and suggested that management actions should be tailored to particular disturbance characteristics and management goals (for example, the survival of residuals and maintenance of spatial heterogeneity that promotes a desired recovery pattern and process). Our results provide some general insights into factors that shape stand vulnerability (for example, age, composition, stream proximity, aspect), but more detailed projections of forest damage

Forest Damage by Hurricane Katrina susceptibility are beyond the scope of an empirical study such as this one where damage patterns were analyzed within the context of the characteristics of the specific event (for example, hurricane size and intensity) and the size and heterogeneity of the study area. The development of a process-based model would require linking event-specific data on tree- and stand-level damage with better or more direct estimates of storm meteorology (for example, from improved surface wind flow models), finer-scale tree- and standlevel measures (for example, tree height and diameter), and localized measures of site environment. If mechanistic, physically-based links between damage and abiotic and biotic conditions can be established, one approach to assessing susceptibility would be to: (1) construct multiple theoretical scenarios of hurricane conditions and movement based on past occurrences, (2) predict surface wind flow patterns associated with the events, and (3) model the resultant damage from each event. Susceptibility could then be defined on the basis of the summed responses (for example, the number of events producing heavy damage), allowing forest managers to identify areas of high sensitivity or areas where restoration may be more critical should an event occur.

ACKNOWLEDGMENTS We particularly appreciate the assistance of Ron Smith, Tate Thriffiley, Clint Roberts, Jeff Cotter and Wayne Stone, of the USDA Forest Service. We also thank Skeeter Dixon, Scott Franklin, Jovian Sackett and the graduate students in JAK‘s ‘Katrina Seminar‘, in which this paper was first developed, and wish to acknowledge NOAA‘s Hurricane Research Division and the U.S. Army Corps of Engineers, for developing data and products used in this study. Comments by Mike Hodgson, Ariel Lugo, and three anonymous reviewers greatly improved the quality of this manuscript. Funding was provided by the Coastal Resiliency Information Systems Initiative for the Southeast (CRISIS), Office of Research and Health Sciences, University of South Carolina.

REFERENCES Boose ER, Foster DR, Fluet M. 1994. Hurricane impacts to tropical and temperate forest landscapes. Ecol Monogr 64:369–400. Breiman L. 1996. Bagging procedures. Mach Learn 24:123–40. Breiman L, Friedman JH, Olshen RA, Stone CJ. 1984. Classification and regression trees. Belmont (CA): Wadsworth International Group.

59

Brokaw NVL, Walker LR. 1991. Summary of the effects of Caribbean hurricanes on vegetation. Biotropica 23:442–47. Cairns DM. 2001. A comparison of methods for predicting vegetation type. Plant Ecol 156:3–18. Cohen J. 1968. Weighted kappa: nominal scale agreement with provision for scaled disagreement or partial credit. Psychol Bull 70:213–20. Conner WH, Mixon WD, Wood GW. 2005. Maritime forest habitat dynamics on Bulls Island, Cape Romain National Wildlife Refuge, SC, following Hurricane Hugo. Forest Ecol Manage 212:127–34. Dale VH, Lugo AE, MacMahon J, Pickett STA. 1998. Ecosystem management in the context of large, infrequent disturbances. Ecosystems 1:546–57. Dunion JP, Landsea CW, Houston SH, Powell MD. 2003. A reanalysis of the surface winds in Hurricane Donna of 1960. Monthly Weather Rev 131:1992–2011. Elsner JB, Jagger TH, Tsonis AA. 2006. Estimated return periods for Hurricane Katrina. Geophys Res Lett 33:L08704 doi:10.1029/2005GL025452. Everham EM, Brokaw NVL. 1996. Forest damage and recovery from catastrophic wind. Bot Rev 62:113–85. Fleiss JL. 1981. Statistical methods for rates and proportions, 2nd edn. New York: Wiley. Foster DR, Boose ER. 1992. Patterns of forest damage resulting from catastrophic wind in central New England, USA. J Ecol 80:79–98. Foster DR, Knight DH, Franklin JF. 1998. Landscape patterns and legacies resulting from large, infrequent forest disturbances. Ecosystem s 1:497–510. Francis JK, Gillespie AJR. 1993. Relating gust speed to tree damage in Hurricane Hugo, 1989. J Arboricult 19:368–73. Friedman JH. 2002. Stochastic gradient boosting. Comput Stat Data Anal 38:367–78. Freund Y, Schapire RE. 1996. Experiments with a new boosting algorithm. In: Machine learning: Proceedings of the Thirteenth International Conference. San Francisco: Morgan Kaufman. pp 148–56. Graumann A, Houston T, Lawrimore J, Levinson D, Lott N, McCown S, Stephens S, Wuertz, D. 2005 (updated 2006). Hurricane Katrina, a climatological perspective, preliminary report. NCDC technical report 2005-01. NOAA‘s National Climatic Data Center, Asheville, N.C. Gresham CA, Williams TM, Lipscomb DJ. 1991. Hurricane Hugo wind damage to southeastern U.S. coastal forest tree species. Biotropica 23:420–26. Lawrence R, Bunn A, Powell S, Zambon M. 2004. Classification of remotely sensed imagery using stochastic gradient boosting as a refinement of classification tree analysis. Remote Sensing Environ 90:331–336. Lewis SA, Wu JQ, Robichaud PR. 2006. Assessing burn severity and comparing soil water repellency, Hayman Fire, Colorado. Hydrol Processes 20:1–16. Lindemann JD, Baker WL. 2002. Using GIS to analyze a severe forest blowdown in the southern Rocky Mountains. Int J Geogr Inf Sci 16:377–99. Litzgus JD, Mousseau TA. 2004. Home range and seasonal activity of southern spotted turtles (Clemmys guttata): Implications for management. Copeia 4:804–17. Meeker JR, Haley TJ, Petty S, Windham JW. 2005. Forest health evaluation of Hurricane Katrina damage on the DeSoto Na-

60

J. A. Kupfer and others

tional Forest. Jackson: USDA Forest Service, National Forests in Mississippi. Miller J, Franklin J. 2002. Modeling the distribution of four vegetation alliances using generalized linear models and classification trees with spatial dependence. Ecol Model 157:227–47. Miller JD, Nyhan JW, Yool SR. 2003. Modeling potential erosion due to the Cerro Grande Fire with a GIS-based implementation of the Revised Universal Soil Loss Equation. Int J Wildl Fire 12:85–100. Myers RK, Van Lear DH. 1998. Hurricane-fire interactions in coastal forests of the south: a review and hypothesis. Forest Ecol Manage 103:265–76. Newcombe RG. 1998. Two-sided confidence intervals for the single proportion: comparison of seven methods. Stat Med 17:857–72.

Powell MD, Houston SH, Amat LR, Morisseau-Leroy N. 1998. The HRD real-time hurricane wind analysis system. J Wind Eng Indust Aerodyn 77&78:53–64. Ramsey EW, Hodgson ME, Sapkota SK, Laine SC, Nelson GA, Chappell DK. 2001. Forest impact estimated with NOAA AVHRR and Landsat TM data related to an empirical hurricane wind-field distribution. Remote Sensing Environ 77:279–92. Sherman RE, Fahey TJ, Martinez P. 2001. Hurricane impacts on a mangrove forest in the Dominican Republic: damage patterns and early recovery. Biotropica 33:393–408. Sherrod PH. 2006. DTREG: classification and regression trees and support vector machine for predictive modeling and forecasting. Online: http://www.DTREG.com/DTREG.pdf. Tanner EVJ, Kapos V, Healey JR. 1991. Hurricane effects on forest ecosystems in the Caribbean. Biotropica 23:513–21.

Ostertag R, Scatena FN, Silver WL. 2003. Forest floor decomposition following hurricane litter inputs in several Puerto Rican forests. Ecosystems 6:261–73.

Turner MG, Baker WL, Peterson CJ, Peet RK. 1998. Factors influencing succession: lessons from large, infrequent natural disturbances. Ecosystems 1:511–23.

Ostertag R, Silver WL, Lugo AE. 2005. Factors affecting mortality and resistance to damage following hurricanes in a rehabilitated subtropical moist forest. Biotropica 37:16–24.

Uriarte M, Canham CD, Thompson J, Zimmerman JK. 2004. A neighborhood analysis of tree growth and survival in a hurricane-driven tropical forest. Ecol Monogr 74:591–614.

Peterson CJ. 2004. Within-stand variation in windthrow in southern boreal forests of Minnesota: Is it predictable?. Can J Forest Res 34:365–75.

Vandermeer JH, Boucher DH, de la Cerda IG, Perfecto I. 2001. Growth and development of the thinning canopy in a posthurricane tropical rain forest in Nicaragua. Forest Ecol Manage 148:221–42.

Platt WJ, Beckage B, Doren RF, Slater HH. 2002. Interactions of large-scale disturbances: Prior fire regimes and hurricane mortality of savanna pines. Ecology 83:1566–72. Powell MD, Bowman D, Gilhousen D, Murillo S, Carrasco N, St. Fleur R. 2004. Tropical cyclone winds at landfall: the ASOSCMAN wind exposure documentation project. Bull Am Meteorol Soc 85:845–51.

Veblen TT, Kulakowski DW, Eisenhart KS, Baker WL. 2001. Subalpine forest damage from a severe windstorm in northern Colorado. Can J Forest Res 31:2089–97.