William K. Michener and Paula F. Houhoulis

Identification and Assessment of Natural Disturbances in Forested Ecosystems: The Role of GIS and Remote Sensing


Ecologists, natural resource managers, and environmental modelers require accurate identification and assessment of the ecological impacts associated with natural disturbances in order to better understand forest ecosystem patterns and processes, to mitigate adverse impacts, and to support modeling activities. At the landscape scale, natural disturbances include fire, insect infestation, hurricanes and tropical storms, flooding, and high winds. Because many natural disturbances result in highly variable damage over large areas, accurate assessment of impacts may be difficult and time-consuming.

In this review, the role (including relevant examples) of GIS and remote sensing in identification, assessment, and monitoring of natural disturbances is discussed. Data sources, image and GIS processing techniques (data transformations, change detection algorithms), and accuracy assessment are reviewed. A case study is presented whereby four change detection approaches are evaluated for their effectiveness in discriminating vegetation changes associated with flooding in a forested ecosystem. Finally, recommendations for natural disturbance assessment are identified.


Natural disturbances play an important role in regulating forest ecosystem structure and function, as well as affecting diverse plant and animal populations and communities. Because vegetation typically exhibits abrupt changes in physiognomy and spectral characteristics in response to acute disturbances, environmental scientists are increasingly using digital images obtained by satellite remote sensors that can detect these changes over broad spatial, temporal, and spectral scales. Change detection analyses, employing Geographic Information System (GIS) coverages and satellite data obtained prior to and following a disturbance, have been used to assess vegetation responses to drought (Peters et al. 1993, Jacobberger-Jellison 1994), insect outbreaks (Muchoney and Haack 1994), dust storms (Chavez and MacKinnon 1994), high winds (Cablk et al. 1994, Johnson 1994), deforestation (Foody and Curran 1994), and other disturbances.

Various data sources and analytical approaches differing in mathematical complexity, processing and analysis intensity, and classification technique have been used to detect vegetation change. Many studies have relied upon less computationally intensive post-classification change detection techniques using images or GIS coverages from one or two dates (Aldrich 1975, Sirois and Ahern 1989, Gardner et al. 1991, Cablk et al. 1994, Dobson et al. 1995, Olsson 1995). Recently, principal components analysis (PCA), various vegetation indices, and logic rules have been implemented utilizing multitemporal satellite data (Bauer et al. 1994, Muchoney and Haack 1994, Jensen et al. 1995, Walsh and Townsend 1995).

Objectives of this paper are to examine how GIS and remote sensing have been used to assess ecological impacts of natural disturbances in forested ecosystems, and to review data sources, relevant GIS and image processing techniques, and accuracy assessment procedures. A case study is presented whereby four change detection approaches are evaluated for their ability to discriminate vegetation changes associated with flooding of a forested ecosystem. Finally, we recommend approaches that appear promising for future change detection studies and identify several research challenges.


Natural Disturbances in Forested Ecosystems

Forested ecosystems are constantly undergoing change. Many of the changes (e.g., succession, responses to climate change, etc.) are directional and occur incrementally over long periods of time. Other changes (e.g., treefall, deforestation, etc.) are more acute, ranging in size from small gaps to entire forests, and act to reset areas back to earlier successional states or entirely alter the ecosystem state altogether.

Many types of disturbances and forest change occur at broad spatial scales and are of special interest to environmental scientists and resource managers. Frequently, because of the size of the areas affected, forest disturbances and their impacts are assessed using remotely sensed data, GIS data, and appropriate change detection approaches. For example, change detection analyses employing remotely sensed data have been used to assess forest impacts of high winds (Cablk et al. 1994, Johnson 1994), fire (Lopez-Garcia and Caselles 1991), salinization (Cablk et al. 1994), climate change (Awaya et al. 1994), and deforestation and harvesting (Tucker et al. 1984, Sader 1987, Bauer et al. 1994, Foody and Curran 1994).

Insect infestations are frequently monitored using remotely sensed data since they can cause widespread forest defoliation and significantly affect commercial interests. Examples include: mountain pine beetle, Dendroctonous ponderosae Hopk. (Sirois and Ahern 1989); hemlock looper, Lambdina fiscellaria fiscellaria (Franklin 1989); spruce budworm, Choristoneura fumiferana Clemens (Buchheim et al. 1984); gypsy moth, Lymantria dispar L. (Rohde and Moore 1974, Nelson 1983, Ciesla et al. 1989, Muchoney and Haack 1994); and pear thrips, Taeniothrips inconsequens Uzel (Vogelmann and Rock 1989).

Data Sources

Sources of data for forest change detection studies vary in spectral, spatial, and temporal resolution. When available, high resolution color or color infrared aerial photographs are used to detect large-scale (local) changes or, more frequently, for assessing accuracy of small-scale (regional) changes identified from lower spatial resolution satellite data. The effectiveness of satellite data for detecting different types of forest change depends to a large extent upon the spatial resolution of the satellite sensor, which can range from 10 m (SPOT panchromatic) to 1 km (NOAA Advanced Very High Resolution Radiometer, AVHRR). For example, SPOT multispectral data (20 m resolution) have been used to identify relatively small forest stands that have been defoliated by insects (Muchoney and Haack 1994), Landsat Thematic Mapper (TM) data (30 m resolution) have been used to assess moderate forest damage resulting from the high winds and saltwater storm surge associated with Hurricane Hugo (Cablk et al. 1994), and AVHRR (1 km resolution) data have been used to assess deforestation occurring at regional to continental scales (Tucker et al. 1984). Choice of satellite data for a particular change detection study must also be based on availability, cost, and spectral resolution.

Image and GIS Processing Techniques

Various analytical approaches differing in complexity, computational intensity, and ease of interpretation have been employed in change detection studies. Although there is increasing interest in fuzzy logic (e.g., Gong 1993) and other new methods, most forest change detection studies have employed a relatively small number of techniques. Some of the most common change detection approaches include: (1) post-classification change detection differencing; (2) spectral-temporal change classification; (3) data transformations (e.g., Normalized Difference Vegetation Index, NDVI); (4) principal components analysis (PCA); (5) image differencing; and (6) change vector analysis (CVA).

Post-classification change detection differencing involves comparing classes from two or more digital data sets on a pixel or area (polygon) basis. Prior to change detection, each digital image must be independently classified by the analyst using supervised or unsupervised approaches. Examples of post-classification change detection are provided by Wickware and Howarth (1981), Estes et al. (1982), and Muchoney and Haack (1994).

Spectral-temporal change classification is based on classification of a single merged data set containing spectral data from multiple dates. When the images are obtained immediately prior to and after a disturbance or on anniversary dates (similar phenology, etc.), then only those areas that have undergone change will be significantly different and the remainder of the data will be similar. An example of spectral-temporal change classification is provided by Muchoney and Haack (1994).

Although broad-scale land use changes may often be readily detected using raw spectral data, more subtle changes such as vegetation stress may be more difficult to identify. In such cases, specific band ratios or band combinations may facilitate change detection. For example, Cablk et al. (1994) employed image differencing of Landsat TM 7/4 band ratios and NDVI data ((IR-R)/(IR+R)) to accurately identify forest stands affected by high winds and saltwater intrusion. Many of the most widely applied vegetation indices are reviewed by Perry and Lautenschlager (1984) and Nilsson (1995).

Principal components analysis (PCA) is a multivariate statistical technique that isolates inter-image change by transforming linear combinations of band data into components that account for the maximum (1st component) and successively lower proportions (2nd and higher order components) of variance among image layers. Applications of PCA techniques are reviewed by LeDrew (1987) and Jiaju (1988).

Image differencing is based on band-by-band subtraction of digital numbers (DNs) using two images (e.g., band1(yr1)-band1(yr2), etc.). Frequently, a median value is added to the differenced data set to eliminate negative values, prior to standard unsupervised classification. Examples of classifications derived by image differencing are provided by Robinove et al. (1981), Cablk et al. (1994), and Muchoney and Haack (1994).

Change vector analysis is an empirical method used to detect radiometric changes based on multidate satellite data, and is characterized by vectors representing the magnitude and direction of changes present in the data (Malila 1980, Michalek et al. 1993). Malila (1980) successfully applied CVA to brightness and greenness bands derived from Landsat TM data to detect changes in forest extent due to harvesting and regrowth. Other applications of CVA are described in articles by Michalek et al. (1993), Johnson (1994), and Lambin and Strahler (1994).

Accuracy Assessment

Classifications derived from remotely sensed images are subject to error and uncertainty. In classifying an image, the spectral response of a pixel, representing a fixed area on the ground defined by the resolution of the sensor, is used to assign it to one of a number of classes using various classification techniques. To assess classification accuracy, reference (ground truth) data are needed for a number of sample locations for each class. Accuracy is defined in terms of misclassifications, where a pixel is assigned to the wrong class. Misclassifications are usually presented in the form of a matrix which is referred to as a confusion or error matrix (Table 1). The error matrix can be used to generate various statistics that characterize the accuracy of a classification technique. For example, the overall accuracy compares the number of pixels correctly classified (those appearing on the diagonal of the matrix) to the total number of pixels sampled (see Table 1). However, this statistic can be misleading since a certain number of correctly classified pixels are expected to occur by chance alone. The Cohen's Kappa or Khat statistic allows for chance, and ranges from 0 in the case of the most confused classification to 1 in the case of the most accurate classification (Table 1). Other statistics that can be generated from the error matrix include errors of omission (producer's error) and errors of commission (user's error). These are based on individual classes, dividing the number of pixels that are incorrectly classified by either the column or row totals, respectively (Table 2). Additional discussion of accuracy assessment techniques can be found in articles by Congalton (1988, 1991), Czaplewski (1994), and Goodchild (1994).

TABLE 1. Example error matrix depicting observed classes versus actual classes of forest types.

Forest Type   Pine   Hardwood    Scrub     Row Total    

Pine          43     7           3         53           

Hardwood      7      14          9         30           

Scrub         0      8           12        20           

Column Total  50     29          24                     

Overall Accuracy = 69 / 103 = 0.699

Khat = (103(69) - 4000) / ((103)2 - 4000) = 0.470*

Khat equation

TABLE 2. Errors of omission (producer's error) and commission (user's error) using Table 1 as a reference.

Forest Type  Omission Error           Commission Error       

Pine         (7+0) / 50 * 100 = 14%   (7+3) / 53 * 100 = 19%    

Hardwood     (7+8) / 29 * 100 = 52%   (7+9) / 30 * 100 = 53%     

Scrub        (3+9) / 24 * 100 = 50%   (0+8) / 20 * 100 = 40%    

Various change detection techniques may be applied in any one study, resulting in multiple classifications. Goodchild et al. (1992) proposed a general error model, called a Probability Vector Model (PVM), for obtaining estimates of uncertainty in land cover maps. In their model, each classification scheme is treated as a "realization" and combined to form a data layer from which the uncertainty associated with a class at any point (or pixel, ij) is represented by a vector of probabilities {pij1, pij2,..., pijn} defining the probability that a pixel belongs to each class 1 through n (Goodchild et al. 1992, Goodchild 1994).


Tropical Storm Alberto presented an opportunity to examine the utility of satellite data for assessing ground cover vegetation responses to flooding in a natural forested ecosystem. Minimal wind and storm surge damage accompanied Alberto as it made landfall on the Florida panhandle. However, due to weak steering currents, the storm remained relatively stationary over southwestern Georgia and southeastern Alabama for a period of six days (July 2-7, 1994). Rainfall was especially heavy (up to 53 cm) in the Flint and Ocmulgee River basins in southwestern Georgia and flood discharges on tributaries and mainstems of the two rivers exceeded 100-year flood discharges along most stream reaches (Stamey 1995). Natural habitats in the two basins are characterized by longleaf pine trees (Pinus palustris) and wiregrass (Aristida stricta), the dominant ground cover species.

Although satellite data have been used to reconstruct regional flood history (Nagarajan et al. 1993), map water boundaries and changes in major wetland habitat types (Wickware and Howarth 1981, Walsh and Townsend 1995), and relate agricultural crop damage to severity of flooding (Yamagata and Akiyama 1988, Yamagata et al. 1988), none have related ground cover vegetation responses to flooding in natural terrestrial ecosystems. The purpose of this case study was to evaluate four different change detection approaches for their ability to discriminate vegetation responses to differential severity of flooding. Relatively sparse canopy coverage in longleaf pine stands, typical of many forest types that occur in xeric habitats, enabled satellite sensors to detect spectral characteristics of the dense ground cover vegetation. Extensive ground surveys supported evaluation of the effectiveness of different change detection approaches, and factors that affect their accuracy.

Study Area

Ichauway is a 115 km2 ecological reserve that is located in Baker County in southwest Georgia, 45 km southwest of Albany (Figure 1). The site is located along the Flint River at its confluence with Ichawaynochaway Creek. Approximately 22 km of Ichawaynochaway Creek and 19 km of the Flint River, a brownwater stream originating in the Georgia Piedmont region, are located within the reserve.

Study Area

FIGURE 1. Location of Ichauway study site showing generalized land cover and extent of flooding associated with Tropical Storm Alberto.

Remotely Sensed Data

Multispectral (XS) images of the study area were acquired from SPOT Image Corporation for September 28, 1994 and October 5, 1990. All images were processed to level 1B, were predominantly cloud-free, and had incidence angles less than 7.5 degrees. A linear regression model was used to correct for relative differences in atmospheric conditions between the two image dates (Jensen et al. 1995). Digital numbers were sampled using a 3x3 window, in 3 types of areas; dark water bodies, dense conifer forest stands, and bright bare soils. Because their appearance was consistent from year to year, the same areas in each image were used as sample sites. Mean values from the sample areas were regressed using the 1990 image as 'master' and the 1994 image as a 'slave'. Regression equations were applied to the 'slave' image using the ERDAS Imagine Modeler (Table 3).

TABLE 3. Regression Equation Used to Normalize Radiometric Characteristics of the 1994 Data with the 5 October 1990 SPOT XS Data.

Date    Band 1 (Green)  Band 2 (Red)   Band 3 (Near IR)     
Sept    y = 0.93(x)+    y = 1.02(x)+   y = 0.98(x)-     
28,     2.76,           0.29,          1.40,
1994    r2 = 0.98       r2 = 0.99      r2 = 0.99           

Twenty three ground control points (GCPs) digitized from 7.5 minute USGS topographic quadrangles (USGS quads) were used to rectify the October 5, 1990 SPOT-XS image to a Universal Transverse Mercator (UTM) map projection (RMSE = 0.29 pixels / 5.91 m). The 'slave' image was similarly rectified using GCPs obtained from the 1990 rectified image (1990 RMSE = 0.33 pixels / 6.63 m). Images were resampled to a 20 m pixel size using a nearest neighbor resampling technique (Jensen et al. 1993). To insure that the data layers used in this analysis were co-registered, the relative error between the images and the ancillary data layers was estimated by taking the difference between ten well distributed checkpoints (road intersections) whose coordinates were recorded from the rectified images and the GIS transportation layer (Wolter et al. 1995). The relative error (RMSE) was less than 5 m in each case.

Ancillary Data Layers

Ancillary data layers were used to assess classification accuracy and derive other data layers such as image masks that were used in the change detection analyses. Three ancillary data layers used in this study (landcover, groundcover, and transportation routes) were interpreted from 1:12,000 scale color infrared (CIR) aerial photographic transparencies. Data were transferred using a vertical sketchmaster to USGS quads, digitized, and attributed using ARC/INFO. Landcover classification attributes included detailed descriptions of species composition, age class, and stand density for all forested areas. Groundcover attributes included primary and secondary cover types and vegetation density. The transportation layer included linear features such as fire-breaks, state and county maintained roads, and highways. To insure that the layers were co-registered, the road network and water bodies were used as a coincident line layer, keeping these features consistent in each of the other photo-interpreted layers. Elevation spot heights and 5 ft contours were digitized from USGS quads. In a few cases, 5 ft contours were interpolated from 10 ft contours using GRASS software.

The progression of the floodwaters was monitored on site during July 1994, and maximum water levels were recorded at approximately 350 locations along Ichawaynochaway Creek and the Flint River. High water levels were surveyed with Trimble Global Positioning System (GPS) Pro XL and Basic Plus receivers and differentially corrected ( + 2 m) to a known Community Base Station. Maximum water levels were used to derive a flood boundary map by overlaying the points on topography, and extrapolating along contour lines between the points to form a polygon.

The ground cover and land cover layers were combined, and used to define a mask containing only forested areas, excluding agriculture, urban areas, roads, water, non-forested wetlands, and regenerating forest stands. This process reduced the potential confusion between flood-damaged vegetation and other land uses and changes (e.g., crop rotations).

In Situ Reference Data

One hundred and twelve sites (approximately 650 m2 per site) dominated by wiregrass groundcover were surveyed throughout the flooded area to quantify vegetation damage. Each site contained three randomly chosen plots where groundcover mortality was assessed using a 1 m quadrate divided into a 10 X 10 grid at 10 cm intervals. Presence of bare ground, detritus, and wiregrass condition (dead, live, or recovering) were recorded at all points and the data from the three plots were averaged and converted into percentages. Each site was surveyed using the GPS techniques described earlier.

Image Classification

Four change detection techniques were evaluated in this study. The first, spectral-temporal change classification (S-TCC), is based on unsupervised classification of the spectral data for the two dates. The second approach, S-PCA, was applied to a six-band merged spectral data set consisting of spectral data from both dates. Third, image differencing (S-ID) of the spectral bands from the two images was performed prior to unsupervised classification. Fourth, image differencing was based on differences in NDVI values (NDVI-ID) observed prior to and following the flood.

S-TCC was based on unsupervised classification of a single multidate data set that contained the six bands from the two dates. ISODATA, an algorithm available in Imagine 8.1, was used in the unsupervised classification. Fifty unsupervised signatures were extracted. The number of iterations was adjusted as necessary to achieve a 0.96 convergence level. The large number of classes provided relatively narrow clusters, that were visually inspected and re-classed as flooded or non-flooded based on the spatial distribution of each class in relation to the areal extent of flooding.

S-PCA was based on a merged six band data set containing all spectral bands from the two images. Examination of the eigenstructure of the transformed data indicated that the first four components accounted for almost 99% of the spectral variability among the images (Table 4). Components five and six were attributable to atmospheric and sensor variations. Therefore, only the first four components were retained for classification.

TABLE 4. Eigenstructure for Multitemporal PCs Based on Spectral Data.

Band          1        2       3       4        5       6    

1990    1    0.16    0.47    -0.20   -0.21   -0.79     0.19   

        2    0.15    0.67    -0.20   -0.41    0.55    -0.16  

        3    0.70   -0.32    -0.62    0.12    0.07    -0.02  

1994    1    0.07    0.25     0.05    0.49   -0.17    -0.81  

        2    0.05    0.40    -0.00    0.72    0.19     0.53   

        3    0.67    0.02     0.73   -0.10    0.00     0.05   

Eigenvalues 46.71   34.57    12.64    5.31    0.86     0.35   

% Variance  46.51   34.42    12.58    5.29    0.86     0.34  

% Cum. Var. 46.51   80.93    93.51   98.80   99.66   100.00 

Two variations of the image differencing approach were evaluated in this study. First, image differencing (S-ID) of the spectral bands from the two images was applied prior to unsupervised classification (as described for S-TCC). In the second approach, NDVI-ID, 1994 NDVI values were subtracted from the 1990 NDVI values to obtain a value for each pixel that represented a magnitude and direction of change. In the resulting image data set, values that are negative or close to zero indicate areas where greenness increased in 1994 or remained relatively unchanged, whereas positive values represent areas exhibiting a decrease in greenness in 1994. For this case study, positive values greater than 11 DNs were classed as flooded.

Accuracy Assessment

Data from the 112 ground survey sites were used in the accuracy assessment. GPS-derived coordinates for each of the ground survey sites were given unique identification numbers (id) and imported into ARC/INFO as a point coverage. Flood classes generated from the different change detection methods (binary masks where 1 = flood) and the flood zone layer were imported into GRID from Imagine 8.1. Site survey data (point coverage) were rasterized using a 1 m cell size and combined with each of the flood class methods and flood zones using the GRID statement gpsflood = COMBINE (gpspoints, floodzone, method(1)...method(n)). The raster value attributes (VAT) contained in the combined layer, 'gpsflood', were related and transferred to the point attribute table (PAT) of the site survey layer based on the GPS-id. The PAT was then output to an ASCII text file to be used for statistical analysis.

The ASCII file was input into the Statistical Analysis System (SAS) and merged with the groundcover mortality data. The sites were classified as live or dead based on the percentage groundcover dead ( 40% = dead). The number of dead sites classified as flooded versus those incorrectly classified as non-flooded, and the number of live sites classified as non-flooded versus those incorrectly classified as flooded, were used as a measure of how well each of the change detection methods performed. Overall accuracy and Kappa Coefficients (Khat) were calculated for each method using techniques described by Congalton et al. (1983) and Congalton (1991).

A Probability Vector Model (PVM) was used to facilitate visualization of accuracy assessment for the four change detection schemes employed. The four binary images were combined to derive probability values that ranged from 0 to 1, representing the proportion of times a pixel was classified as flooded by the different methods or realizations. Thus, a probability value equal to 0.75 indicates that the pixel was classified as flooded by three of the four techniques evaluated.


Overall accuracy and Kappa Coefficient statistics (Khat) were used to compare the different change detection techniques (Table 5). Overall accuracy was high for all techniques and ranged from 0.607 to 0.750. However, Khat values exhibited significant variation, ranging from 0.266 to 0.487. Both measures indicated that spectral-temporal change classification (also known as layered temporal change classification) was least effective in discriminating flood-affected vegetation (Table 5a; Figure 2a).

TABLE 5. Accuracy assessment of four change detection techniques used to detect vegetation responses to flooding.

           No. Dead Sites   No. Live Sites                       
Method        Det.  Und.      Det.   Und.  Accuracy   Khat   

a.  S-TCC     40     6        28     38     0.607     0.266  

b.  S-PCA      9     7        31     35     0.625     0.291  

c.  ID        34    12        45     21     0.705     0.409  

d.  NDVI-ID   33    13        51     15     0.750     0.487  

Change Detection Classifications a-d

FIGURE 2. Classification maps of Ichauway study area showing results of change detection analyses based on: (a) spectral-temporal change classification (S-TCC); (b) principal components analysis of all spectral bands (S-PCA); (c) image differencing of spectral bands (S-ID); and (d) image differencing of NDVI data using a user defined threshold (NDVI-ID).

Classification accuracy and Khat values exhibited marginal improvement when classification was based on S-PCA (Table 5b; Figure 2b). Image differencing of spectral data (S-ID) resulted in a marked improvement in classification accuracy over the first two methods (Table 5c; Figure 2c). However, image differencing based on NDVI (NDVI-ID) was the most effective technique for discriminating vegetation responses to flooding; accuracy was highest and Khat exhibited a two-fold increase over S-PCA (Table 5d; Figure 2d). Closer examination of the NDVI data indicated that digital numbers (DNs) were similar in non-flooded areas in 1990 and 1994, but were substantially lower (> -10 DNs) in the flooded area in 1994, in comparison to 1990 (Table 6). These findings indicate that ground cover vegetation exhibits a marked spectral response to flooding which is best exemplified as a decrease in NDVI in affected areas.

TABLE 6. Mean NDVI Values for the 1990 and 1994 SPOT XS Data.

Image Date   Entire Site  Non-flooded  Flooded      Standard           

10/05/90     190.2        189.7        192.5        1.9                

09/28/94     186.4        187.6        181.7        4.1                

A map based on PVM was used to visually assess sources of confusion in the classification process and to better understand the factors that affect our ability to discriminate flood-affected vegetation (Figure 3). Many of the flood-affected vegetation zones along the Flint River and Ichawaynochaway Creek that were identified by all four techniques occurred in areas that experienced highest current velocities and deepest waters as well as in localized depressions where standing water remained for several days following the flood. Within the flood boundary, 82% of the area that appeared to be affected was classified as "flooded" by at least one or more of the four techniques that are depicted in Figure 3. Furthermore, almost half (49%) of the "flooded" area was discriminated by at least two of the four techniques and 33% of the "flooded" area was detected by all four approaches. Thus, it is apparent that there is considerable agreement among the techniques in identifying vegetation changes within the flood zone. In contrast, most (20%) of the area outside the flood boundary that was misclassified as "flooded" was discriminated by only one of the four techniques. Only 6% of the area outside the flood boundary that was misclassified as "flooded" was identified by all four techniques; with many of the larger clusters (shown in red; Figure 3) frequently being associated with vegetated wetlands that were affected by the excessive precipitation.

Probability Vector Map

FIGURE 3. Map based on probability vector model showing proportion of times that a pixel was classified as flooded by four different change detection techniques.


Several factors might be expected to constrain flood impact assessments. First, the timing of the satellite data coverage relative to river conditions frequently leads to underestimates of the severity and areal extent of flood inundation (Blasco et al. 1992). Second, dense vegetation canopy and the complex relationship between hydrologic and phenologic cycles may confound vegetation spectral responses within the floodplain (Walsh and Townsend 1995). Third, accuracy assessments generally require reliable post-flood ground truth data, adequate digital elevation models, or other data that are often lacking or inadequate. Finally, vegetation may exhibit a lagged response to secondary flood-related factors (anaerobiasis, waterlogging, etc.) in areas that do not directly experience the most intense erosion and scouring.

In this study, cloud-free satellite data could not be acquired until almost three months after the flood. Relatively sparse longleaf pine overstory coupled with a dense ground cover community dominated by a single species (wiregrass) facilitated change detection analyses. Overall classification accuracy exceeded 60 per cent for the unsupervised methods used in this study, suggesting that all techniques were effective for discriminating vegetation responses to flooding. However, the approximate two-fold range in Kappa Coefficient Statistics (Khat = 0.266 - 0.487) indicated that some methods outperformed others when chance agreement was removed. Specifically, it is possible to generalize that: (1) a classification based on changes occurring in all spectral bands, S-TCC, was least effective in discriminating vegetation changes related to the flood; (2) S-PCA offered marginal improvement in classification accuracy over temporal change classification; and (3) image differencing represented the most effective method for discriminating flood-affected vegetation.


Change detection studies frequently focus on very abrupt changes related to alterations in land use (e.g., deforestation) and broad-scale natural disturbances (e.g., hurricanes, drought, insect outbreaks). The case study presented above demonstrates the feasibility for using satellite data to detect and monitor relatively subtle responses in vegetation dynamics to natural disturbances in forested ecosystems. This capability offers significant potential for increasing our understanding of ecosystem- and landscape-scale responses to natural disturbances as well as assessing changes in vegetation dynamics related to climatic variability and global climate change.

The effectiveness of change detection studies of natural disturbances may be significantly affected by temporal, spatial, and spectral resolution of the data, as well as the availability of ancillary (relevant GIS coverages) and ground truth data. When change detection analysis is based on data acquired immediately prior to and following a discrete disturbance event, spectral change may be related to ecological changes with a reasonably high degree of certainty. Otherwise, spectral changes associated with a specific disturbance may be confounded with land use change, annual phenological differences, climate, and other factors that differ between the pre- and post-disturbance imagery. Identification of changes occurring over long periods of time (e.g., succession, climate change, etc.) may require time series of images and greater automation of change detection analyses. Automation of change detection, however, requires that baseline conditions be defined prior to assessment of change. Unfortunately, very little is known about what constitutes "normal" conditions for an area. For example, can a single satellite image be considered to represent normal conditions or would the range of spectral variability present during wet and dry years (two or more images) serve more effectively as a baseline for assessing future change?

Spatial resolution may be of less concern when major disturbances (e.g., hurricanes, droughts, etc.) or land use changes (e.g., deforestation of tropical rain forests) occur over extremely large areas. More frequently, however, forest responses to disturbances are highly variable and occur in patches. Thus, the ability to discriminate local areas of change will be related to patch sizes and sensor resolution (Townshend 1981). For example, despite the 20 m sensor resolution, Sirois and Ahern (1989) found that the smallest areas affected by Mountain Pine Beetles that could be detected with SPOT XS data ranged from 1 to 2 ha in size, and contained trees with 80 to 100 per cent damaged crowns. Unfortunately, little is known about the relationships among sensor spatial and spectral resolution, minimal patch size detected, and type, variability and magnitude of damage in natural forest ecosystems.

The effectiveness of alternative change detection approaches for assessing forest disturbances is rarely evaluated within the context of a single study. Although approaches that work well within a single comparative study may not necessarily apply in other ecosystem types or for other types of disturbance, results of such studies may provide guidance or, at least, a starting point for other disturbance assessments. Muchoney and Haack (1994) evaluated four change detection approaches using multitemporal SPOT High Resolution Visible (HRV) data, ranging from standard post-classification change differencing to more analytically complex image differencing and PCA techniques, for identifying hardwood forest defoliation caused by gypsy moth infestation. In their study, overall accuracy ranged from 0.61 (post-classification, spectral-temporal) and 0.63 (PCA) to 0.69 (image differencing of spectral bands). In the case study presented in this paper, image differencing was similarly the most effective of the techniques evaluated. One interesting result, however, was the improvement in classification accuracy when image differencing was based on NDVI data as opposed to the spectral bands for the two images. Of the various techniques evaluated by Muchoney and Haack (1994) and in the case study presented in this paper, image differencing represents a relatively straightforward technique that could be easily automated for specific areas of interest.

Accuracy assessments within the context of a study and inclusion of relevant statistics (especially Kappa Coefficients) in subsequent publications can support evaluation of the effectiveness of change detection approaches for specific applications, and may facilitate future research efforts. Maps derived from probability vector modeling can be used to visually interpret classifications resulting from multiple change detection methods, and may also lead to a better understanding of factors that affect accuracy. For example, large areas that are similarly classified by all or most methods can be easily identified. Areas of disagreement (lower probability scores) may indicate mixed pixels or class uncertainty. Consistent areas of disagreement may, for example, delineate changes occurring in a particular land cover class (e.g., vegetation senescence in bottomland hardwood habitats) that are being confused with the change of interest (e.g., insect defoliation of conifer stands). Although such sources of confusion may be easily visualized, interpreted, and corrected, they may not be apparent using standard classification accuracy assessment techniques. Goodchild et al. (1992) further discuss how this modeling approach can be used to obtain standard errors associated with area estimates.

Once baseline conditions are established and change is detected in a "new" satellite data coverage, ground or aerial photography-based assessments are essential for determining whether "significant" spectral change is ecologically meaningful. In some cases (e.g., deforestation, severe defoliation associated with wind, insect outbreaks, etc.), spectral change may be easily and directly related to vegetation change, regardless of the change detection approach employed. In other cases, the ability of different change detection approaches to discriminate vegetation changes may be affected by forest stand characteristics, land cover, soil characteristics, and so forth. Logistic multiple regression represents a powerful analytical technique that may prove useful for evaluating different change detection methods and designing ground verification studies. Logistic regression has frequently been used to investigate the relationship between response probabilities of binary and ordinal response variables, and the explanatory variables (Hosmer and Lemeshow 1989). Binary response variables (e.g., defoliated, unaffected) and ordinal response variables (e.g., no effect, moderate defoliation, severe defoliation) arise in many studies of ecological disturbances. Logistic regression analysis may be effective for investigating the relationship between the spectrally defined response probability and the potential explanatory variables (e.g., degree of canopy closure, stand condition, soil moisture class, etc.). Results of such analyses could be used to reduce the number of attributes that are monitored in the field, thereby reducing sampling costs.


We thank the numerous students and technicians who participated in ground survey efforts, especially Derek Fussell; and Frank Miller, Mary Grace Chambers and Linda Garnett of the Mississippi Remote Sensing Center, and Jean Brock (JWJERC) for GIS support. This project was funded by the National Science Foundation (DEB-9520878).


Aldrich, R.C. (1975) Detecting disturbances in a forest environment, Photogrammetric Engineering & Remote Sensing 41: 39-48.

Awaya, Y., T. Hame, and N. Tanaka (1994) Vegetation change detection using Landsat data in northern Finland. Proceedings for the National Institute of Polar Research Symposium 7: 270-282.

Bauer, M.E., T.E. Burk, A.R. Ek, P.R. Coppin, S.D. Lime, T.A. Walsh, D.K. Walters, W. Befort, and D.F. Heinzen (1994) Satellite inventory of Minnesota forest resources. Photogrammetric Engineering & Remote Sensing 60(3): 287-298.

Bishop, Y., S. Fienberg, and P. Holland (1975) Discrete Multivariate Analysis-Theory and Practice. Cambridge, Massachusetts: MIT Press.

Blasco, F., M.F. Bellan, and M.U. Chaudhury (1992) Estimating the extent of floods in Bangladesh using SPOT data. Remote Sensing of Environment 39: 167-178.

Buchheim, M.P., A.L. Maclean, and T.M. Lillesand (1984) Forest cover type mapping and spruce budworm defoliation using simulated SPOT imagery. Proceedings of the 1984 SPOT Symposium, pp. 259-266.

Cablk, M.E., B. Kjerfve, W.K. Michener, and J.R. Jensen (1994) Impacts of Hurricane Hugo on a coastal forest: Assessment using Landsat TM data. Geocarto International 2: 15-24.

Chavez, P.S., and D.J. MacKinnon (1994) Automatic detection of vegetation changes in the southwestern U.S. using remotely sensed images. Photogrammetric Engineering & Remote Sensing 60(5): 571-583.

Ciesla, W.M., C.W. Dull, and R.E. Acciavatti (1989) Interpretation of SPOT-1 color composites for mapping defoliation of hardwood forests by gypsy moth. Photogrammetric Engineering and Remote Sensing 55(10): 1465-1470.

Congalton, R.G. (1988) A comparison of sampling schemes used in generating error matrices for assessing the accuracy of maps generated from remotely sensed data. Photogrammetric Engineering and Remote Sensing 54(5): 593-600.

Congalton, R.G. (1991) A review of assessing the accuracy of classifications of remotely sensed data. Remote Sensing of Environment 37: 35-46.

Congalton, R.G., R.G. Oderwald, and R.A. Mead (1983) Assessing Landsat classification accuracy using discrete multivariate analysis statistical techniques. Photogrammetric Engineering & Remote Sensing 49(12): 1671-1678.

Czaplewski, R.L. (1994) Variance Approximations for Assessments of Classification Accuracy. Research Paper RM-316, Fort Collins, Colorado: United States Department of Agriculture, Forest Service.

Dobson, E.L., J.R. Jensen, R.B. Lacy, and F.G. Smith (1995) A land cover characterization methodology for large area inventories. ACSM/ASPRS Proceedings, pp. 786-795.

Estes, J.E., D.Stow, and J.R. Jensen (1982) Monitoring land use and land cover changes. Pages 100-110 in C. Johannsen and J. Sanders, eds. Remote Sensing for Resource Management. Ankeny, Iowa: Soil Conservation Society of America.

Franklin, S.E. (1989) Classification of Hemlock Looper defoliation using SPOT HRV imagery. Canadian Journal of Remote Sensing 15(3): 178-182.

Foody, G.M., and P.J. Curran (1994) Estimation of tropical forest extent and regenerative stage using remotely sensed data. Journal of Biogeography 21: 223-244.

Gardner, L.R., W.K. Michener, B. Kjerfve, and D.A. Karinshak (1991) The geomorphic effects of Hurricane Hugo on an undeveloped coastal landscape at North Inlet, South Carolina. Journal of Coastal Research 8: 181186.

Goodchild, M.F. (1994) Integrating GIS and remote sensing for vegetation analysis and modeling: Methodological issues. Journal of Vegetation Science 5: 615-626.

Goodchild, M.F., G.Q. Sun, and S. Yang (1992) Development and test of an error model for categorical data. International Journal Geographic Information Systems 6: 87-104.

Gong, P. (1993) Change detection using principal component analysis and fuzzy set theory. Canadian Journal of Remote Sensing 19(1): 22-29.

Hosmer, D.W., and S. Lemeshow (1989) Applied Logistic Regression. New York, New York: John Wiley and Sons, Inc.

Jacobberger-Jellison, P.A. (1994) Detection of post-drought environmental conditions in the Tombouctou region. International Journal of Remote Sensing 15(16): 3138-3197.

Jensen, J.R., D. Cowen, J.D. Althausen, S. Narumalani, and O. Weatherbee (1993) An evaluation of the CoastWatch change detection protocol in South Carolina. Photogrammetric Engineering & Remote Sensing 59(6): 1039-1046.

Jensen, J.R., K. Rutchey, M.S. Koch, and S. Narumalani (1995) Inland wetland change detection in the everglades water conservation area 2A using a time series of normalized remotely sensed data. Photogrammetric Engineering & Remote Sensing 61(2): 199-209.

Jiaju, L. (1988) Development of principal component analysis applied to multitemporal Landsat TM data. International Journal of Remote Sensing 9(12): 1895-1907.

Johnson, R.D. (1994) Change vector analysis for disaster assessment: A case study of hurricane Andrew. Geocarto International 1: 41-45.

Lambin, E.F., and A.H. Strahler (1994) Indicators of land-cover change for change-vector analysis in multitemporal space at coarse spatial scales. International Journal of Remote Sensing 15(10): 2099-2119.

LeDrew, E. (1987) Application of principal components analysis to change detection. Photogrammetric Engineering and Remote Sensing 53(12): 1649-1658.

Lopez-Garcia, M.J. and V. Caselles (1991) Mapping burns and natural reforestation using Thematic Mapper data. Geocarto International 1: 31-37.

Malila, W.A. (1980) Change vector analysis: an approach for detecting forest changes with Landsat. 1980 Machine Processing of Remotely Sensed Data Symposium, pp. 326-336.

Michalek, J.L., T.W. Wagner, J.J. Luczkovich, and R.W. Stoffle (1993) Multispectral change vector analysis for monitoring coastal marine environments. Photogrammetric Engineering & Remote Sensing 59(3): 381-384.

Muchoney, D.M., and B.N. Haack (1994) Change detection for monitoring forest defoliation. Photogrammetric Engineering & Remote Sensing 60(10): 1243-1251.

Nagarajan, R., and G.T. Marathe (1993) Identification of flood prone regions of Rapti River using temporal remotely-sensed data. International Journal of Remote Sensing 14(7): 1297-1303.

Nelson, R.F. (1983) Detecting forest canopy change due to insect activity using Landsat MSS. Photogrammetric Engineering & Remote Sensing 49: 1303-1314.

Nilsson, H. (1995) Remote sensing and image analysis in plant pathology. Canadian Journal of Plant Pathology 17: 154-166.

Olsson, H. (1995) Reflectance calibration of TM data for forest change detection. International Journal of Remote Sensing 16(1): 81-96.

Perry, C.R., and L.F. Lautenschlager (1984) Functional equivalence of spectral vegetation indices. Remote Sensing of Environment 14: 169-182.

Peters, A.J., B.C. Reed, M.D. Eve, and K.M. Havstad (1993) Satellite assessment of drought impact on native plant communities of Southeastern New Mexico, U.S.A. Journal of Arid Environments 24: 305-319.

Robinove, C.J., P.S. Chavez, D. Gehring, and R. Holmgren (1981) Arid land monitoring using Landsat albedo difference images. Remote Sensing of Environment 11: 133-156.

Rohde, W.G., and H.J. Moore (1974) Forest defoliation assessment with satellite imagery. Proceedings of the Symposium on Remote Sensing of Environment 2(9): 1089-1104.

Sader, S.A. (1987) Digital image classification approach for estimating forest clearing and regrowth rates and trends. Proceedings of IGARSS '87 Symposium pp. 209-213.

Sirois, J., and F.J. Ahern (1989) An investigation of SPOT HRV data for detecting recent mountain pine beetle mortality. Canadian Journal of Remote Sensing 14(2): 104-108.

Stamey, T.C. (1995) Floods in Central and Southwestern Georgia in July 1994. Proceedings of the 1995 Georgia Water Resources Conference, pp. 313-316.

Townshend, J.R.G. (1981) The spatial resolving power of earth resources satellites. Progress in Physical Geography 5: 32-55.

Tucker, C.J., B.N. Holben, and T.E. Goff (1984) Intensive forest clearing in Rondonia, Brazil, as detected by satellite remote sensing. Remote Sensing of Environment 15: 255-261.

Vogelmann, J.E., and B.N. Rock (1989) Use of Thematic Mapper data for the detection of forest damage caused by the pear thrips. Remote Sensing of Environment 30: 217-225.

Walsh, S.J., and P.A. Townsend (1995) Comparison of change detection approaches for assessing a riverine flood hydroperiod. ACSM/ASPRS Proceedings, pp. 134-143.

Wickware, G.M., and P.J. Howarth (1981) Change detection in the Peace-Athabasca Delta using digital Landsat data. Remote Sensing of Environment 11: 9-25.

Wolter, P.T., D.J. Mladenoff, G.E. Host, and T.R. Crow (1995) Improved forest classification in the northern lake states using multi-temporal Landsat imagery. Photogrammetric Engineering & Remote Sensing 61(9): 1129-1143.

Yamagata, Y., and T. Akiyama (1988) Flood damage analysis using multitemporal Landsat TM data. International Journal of Remote Sensing 9(3): 503-514.

Yamagata, Y., C. Wiegand, T. Akiyama, and M. Shibayama (1988) Water turbidity and perpendicular vegetation indices for paddy rice flood damage analysis. Remote Sensing of Environment 26: 241-251.

William K. Michener, Ph.D.

Associate Scientist, Landscape Ecology, wmichene@jonesctr.org


Paula F. Houhoulis, M.A.

Remote Sensing Analyst, Landscape Ecology, phouhoul@jonesctr.org

Joseph W. Jones Ecological Research Center,

Route 2, Box 2324, Newton, GA 31770

ph#:(912) 734-4706, fax#:(912) 734-6650