Remote Sensing of Environment

Transcription

1 Remote Sensing of Environment 114 (2010) Contents lists available at ScienceDirect Remote Sensing of Environment journal homepage: An automated approach for reconstructing recent forest disturbance history using dense Landsat time series stacks Chengquan Huang a,, Samuel N. Goward a, Jeffrey G. Masek b, Nancy Thomas a, Zhiliang Zhu c, James E. Vogelmann d a Department of Geography, University of Maryland, College Park, MD 20742, USA b Biospheric Sciences Branch, NASA Goddard Space Flight Center, Greenbelt, MD 20771, USA c U.S. Geological Survey, Sunrise Valley Drive, Reston, VA 20771, USA d USGS Earth Resources Observation and Science (EROS) Center, Sioux Falls, SD 57198, USA article info abstract Article history: Received 17 April 2009 Received in revised form 22 August 2009 Accepted 29 August 2009 Keywords: Landsat time series stacks (LTSS) Vegetation change tracker (VCT) Forest z-score (FZ) Integrated forest z-score (IFZ) A highly automated algorithm called vegetation change tracker (VCT) has been developed for reconstructing recent forest disturbance history using Landsat time series stacks (LTSS). This algorithm is based on the spectral temporal properties of land cover and forest change processes, and requires little or no fine tuning for most forests with closed or near close canopy cover. It was found very efficient, taking 2 3 h on average to analyze an LTSS consisting of 12 or more Landsat images using an average desktop PC. This LTSS-VCT approach has been used to examine disturbance patterns with a biennial temporal interval from 1984 to 2006 for many locations across the conterminous U.S. Accuracy assessment over 6 validation sites revealed that overall accuracies of around 80% were achieved for disturbances mapped at individual year level. Average user's and producer's accuracies of the disturbance classes were around 70% and 60% in 5 of the 6 sites, respectively, suggesting that although forest disturbances were typically rare as compared with nochange classes, on average the VCT detected more than half of those disturbances with relatively low levels of false alarms. Field assessment revealed that VCT was able to detect most stand clearing disturbance events, including harvest, fire, and urban development, while some non-stand clearing events such as thinning and selective logging were also mapped in western U.S. The applicability of the LTSS-VCT approach depends on the availability of a temporally adequate supply of Landsat imagery. To ensure that forest disturbance records can be developed continuously in the future, it is necessary to plan and develop observational capabilities today that will allow continuous acquisition of frequent Landsat or Landsat-like observations Elsevier Inc. All rights reserved. 1. Introduction Forest disturbance and post-disturbance recovery are key processes in the development of forest ecosystems. The landscape pattern of forest age and structure, for example, is shaped in part by the history of disturbance and recovery processes (Peterken, 2001). Understanding these processes over space and time is also crucial to studies of terrestrial and atmospheric carbon, as they are major processes modulating carbon flux between the biosphere and the atmosphere (Hirsch et al., 2004; Law et al., 2004). While North American forests, especially those in the United States, have been proposed as a net carbon sink (Pacala et al., 2001; Liu et al., 2004), estimates of the magnitude of the sink have substantial uncertainties. Reducing such uncertainties requires improved understanding of land use history, especially forest disturbance Corresponding author. address: cqhuang@umd.edu (C. Huang). history (Schimel & Braswell, 1997; Thornton et al., 2002; Houghton, 2003). The collection of Landsat images acquired through current and previous Landsat missions (Goward et al., 2006) provide a unique data source for reconstructing forest disturbance history for many areas of the globe. With the earliest Landsat images acquired in 1972, this collection makes it possible to document forest changes that have occurred since then, while the fine spatial resolutions of Landsat images provide the spatial details necessary for characterizing many of the changes arising from both natural and anthropogenic disturbances (Townshend & Justice, 1988). Over the past 30+ years, Landsat images have been widely used in land cover and forest change analysis (Goward & Williams, 1997). While the Landsat Record has relatively dense acquisitions in many places of the world, especially in the United States (Goward et al., 2006), largely due to practical reasons, most previous studies characterized land cover change at relatively sparse temporal intervals (Singh, 1989; Lu et al., 2004). For many of the Earth's forests, reestablishment of a new forest stand following a previous disturbance /$ see front matter 2009 Elsevier Inc. All rights reserved. doi: /j.rse

2 184 C. Huang et al. / Remote Sensing of Environment 114 (2010) can occur in just a few years (e.g. Huang et al., 2009a, Fig. 1). As a result, some of those disturbances can become spectrally undetectable using observations acquired many years apart (Lunetta et al., 2004; Masek et al., 2008). Within the framework of the North American Carbon Program (NACP), the North American Forest Dynamics (NAFD) project is evaluating forest disturbance and regrowth history for the conterminous U.S. by combining Landsat observations and field measurements (Goward et al., 2008). To minimize potential omission errors that may arise from using temporally sparse observations, dense Landsat time series stacks (LTSS) are used in the NAFD study. About 30 LTSS have been assembled during the first phase of NAFD for locations selected across the U.S. (Goward et al., 2008; Huang et al., 2009a). Two major steps are involved in mapping forest disturbance using LTSS: development of imagery-ready-to-use (IRU) quality LTSS images, and forest disturbance detection. IRU quality LTSS images are defined as having minimum cloud and shadow contaminations and minimum instrument or processing related errors. Such images should also have been corrected geometrically and radiometrically such that they have the highest level of achievable geolocation accuracy and radiometric consistency. Streamlined procedures for producing such IRU quality LTSS images have been developed in a previous study (Huang et al., 2009a). Over the last few decades of remote sensing history, numerous digital change detection techniques have been developed for use with Landsat and other satellite images (see Singh, 1989; Coppin et al., 2004; Lu et al., 2004 for comprehensive reviews). These existing techniques, however, are mostly bi-temporal in nature, i.e., they can be used to analyze only one collocated image pair at a time. While each LTSS could be divided into a sequence of image pairs and a bitemporal technique could be used to analyze each image pair, such an approach would be extremely inefficient. Furthermore, bi-temporal techniques cannot take advantage of the rich temporal information contained in LTSS. As will be demonstrated in this study, temporal information is particularly useful for characterizing land cover and change processes. While algorithms capable of analyzing three or more images at a time have also been developed (e.g. Coppin & Bauer, 1996; Cohen et al., 2002; Lunetta et al., 2004), most of them suffer shortcomings similar to those of bi-temporal techniques for analyzing dense satellite observations. To improve the efficiency of land cover change analysis using LTSS, Kennedy et al. (2007) developed a trajectory-based change detection algorithm. This algorithm takes all images in an LTSS into consideration, and uses idealized temporal trajectory of spectral values to detect and characterize changes. Here we develop a different change detection algorithm called vegetation change tracker (VCT) for mapping forest change using LTSS. This algorithm is similar to the trajectory-based change detection algorithm of Kennedy et al. (2007) in that changes are detected through simultaneous analysis of all images in an LTSS. But the mechanisms based on which changes are detected are different. The VCT is based on the spectral temporal characteristics of land cover and forest change processes. It consists of two major steps. In the first step, each image in an LTSS is analyzed to create masks and to calculate spectral indices that are measures of forest likelihood. Once this step is complete for all images in that LTSS, the derived indices and masks are analyzed in a time series analysis step to map disturbances. A brief description of the VCT algorithm has been provided before (Huang et al., 2009b). Some key components of the first step have also been detailed previously (Huang et al., 2008; in press). The purpose of this paper is to provide a detailed, coherent description of all components of the VCT and to assess the disturbance products derived using this algorithm. The VCT has been used to produce disturbance products for the sites where LTSS were assembled during the first phase of the NAFD project (Goward et al., 2008; Huang et al., 2009a). Biennial disturbance products have also been produced using this approach for Mississippi (Li et al., in press) and Alabama (Li et al., 2009) as part of an effort to update a vegetation database developed through the LANDFIRE project (Rollins, 2009). Those disturbance products have been evaluated using multiple approaches, including ground-based assessment, visual assessment, designed-based accuracy assessment and assessment using field survey data collected through the Forest Inventory and Analysis (FIA) program of the USDA Forest Service. Mainly for page limit considerations, only a summary of the results derived from field based assessment, visual evaluation, and design-based accuracy assessment is provided in this paper. Full details on the design-based accuracy assessment and assessment using FIA field survey data will be provided in a follow-up paper (Thomas, in preparation). Because the VCT algorithm is designed for use with LTSS, a brief review of the LTSS concept and the procedures for developing LTSS is deemed necessary here, although details on those procedures have been provided previously (Huang et al., 2009a). Following this review Section 3 provides a detailed description of the VCT algorithm. Assessments of VCT disturbance products are provided in Section 4, followed by a summary and discussions on the properties of the VCT algorithm and its requirement on satellite observations. Fig. 1. Overall data flow and processes of the VCT algorithm.

3 C. Huang et al. / Remote Sensing of Environment 114 (2010) LTSS development A Landsat time series stack (LTSS) is defined as a temporal sequence of Landsat images acquired at a nominal temporal interval for an area defined by a path/row tile of the World Reference System (WRS). An annual LTSS consists of one image every year while a biennial LTSS consists of one image every 2 years. Due to limited data availability coupled with cloud contaminations, however, the actual temporal intervals between consecutive acquisitions can be different from the nominal interval (example acquisition dates of some LTSS can be found in Huang et al., 2009b, Table 2). Each image should have minimal or no cloud contamination and should be acquired during the peak growing season. Mainly due to data cost constraints, 1 most LTSS used in the NAFD project were biennial stacks. Each LTSS had a nominal starting year of 1984 and an ending year of 2006, consisting of Landsat Thematic Mapper (TM) and Enhanced Thematic Mapper Plus (ETM+) images acquired between 1984 and To avoid dealing with data gaps resulting from a scan-line-corrector (SLC) problem that occurred to the Landsat 7 in May 2003, no ETM+ images acquired since then (i.e., SLC-off images) were used. Most selected LTSS images were ordered from the USGS as level 1G product, which typically had substantial geometric and radiometric errors. To minimize spurious changes that often arise from misregistration errors in satellite images (Townshend et al., 1992), the level 1G images were processed using advanced correction algorithms implemented as automated modules in the Landsat Ecosystem Disturbance Adaptive Processing System (LEDAPS). Specifically, each image was registered precisely to a base image that had minimal geolocation errors. Errors caused by terrain relief were then removed through orthorectification using a digital elevation model (DEM) (Gao et al., 2009). For Landsat 5 images we undid their original calibrations, which were reportedly out of date (Markham & Barker, 1986; Chander & Markham, 2003; Chander et al., 2004; Markham et al., 2004), and re-calibrated them using the most up-to-date coefficients to calculate top-ofatmospheric (TOA) radiance and reflectance (Chander et al., 2004). Such re-calibrations were not deemed necessary for Landsat 7 images, because the ETM+ sensor has been carefully monitored since launch (Markham et al., 2004). Finally, to compensate for atmospheric scattering and absorption effects on the TOA reflectance, an atmospheric correction algorithm based on the 6S radiative transfer code was used to convert TOA reflectance to surface reflectance (Vermote et al., 1997). More details on the LTSS concept and the advanced correction algorithms can be found in Huang et al. (2009a). 3. The VCT algorithm 3.1. Algorithm overview Forest, disturbance, and post-disturbance recovery processes have a number of spectral temporal properties that can be used to distinguish them from non-forest land cover types: Due to light absorption by green vegetation and canopy shadowing, forest is one of the darkest vegetated surfaces in satellite images acquired during the leaf-on growing season in visible and some shortwave infrared bands (Colwell, 1974; Kauth & Thomas, 1976; Goward et al., 1994; Huemmrich & Goward, 1997); During the mid-growing season, undisturbed forests typically maintain relatively stable spectral signatures over many years, while most non-forest land cover types have more spectral variability, both seasonally and inter-annually; 1 When the NAFD study began in 2003, the costs for a TM and an ETM+ image from the USGS Center for EROS were $425 and $600 respectively. USGS did not open no-cost access to its Landsat images until late Most forest disturbance events result in sudden reduction or removal of forest canopy cover and woody biomass, and are often manifested by abrupt spectral changes; Depending on the nature of a disturbance, the resultant change signal in the spectral data can last several years or longer. This is because reestablishment of a new forest stand from a disturbance takes time, or no forest stand will be reestablished if that disturbance results in a conversion from forest to a non-forest land cover type. The vegetation change tracker (VCT) algorithm is developed based on these spectral temporal properties. It consists of two major steps (Fig. 1): Individual image masking and normalization: Each image in an LTSS is analyzed separately to mask water, cloud and cloud shadow, and to identify some forest samples. The identified forest samples are then used to calculate several indices as measures of forest likelihood. Time series analysis: The indices and masks derived for all images of an LTSS are used to form time series trajectories and to produce forest change products. It should be noted that because VCT uses the spectral temporal information as recorded in an LTSS to detect forest disturbance, the detected disturbances are not necessarily limited to those defined in forestry or ecology. In VCT a disturbance refers to any event that can result in significant reduction or removal of forest canopy cover and woody biomass, including harvest, selective logging, tree reduction for fuel treatment or other purposes, and damages due to fire, storm, insect or diseases, although, as will be discussed later, not all of these events can be detected reliably by the current version of the VCT. Also, throughout this paper, recovery, regrowth, and regeneration are used interchangeably. They all refer to the recovery process of a forest stand from a non-stand replacement disturbance, or the reestablishment of a new forest stand from a stand clearing disturbance Individual image masking and normalization The purpose of this step is to analyze each image individually to create initial masks for water, cloud, and shadow, and to normalize that image using known forest samples. This step has the following major processes: creation of a land water mask, identification of forest samples, calculation of forest indices, and masking of cloud and cloud shadow Land water masking While for the purpose of forest disturbance detection it is not necessary to distinguish between water and other non-forest land cover types, it is useful to separate them for many other applications. Based on known spectral properties of typical water bodies (Jensen, 1996), a pixel is flagged as a water pixel if it has low reflectance value in the shortwave infrared band (band 5) and satisfies at least one of the following two conditions: It has a trend of decreasing reflectance values from the visible to the infrared bands; or It has a low normalized difference vegetation index (NDVI) value, where NDVI is calculated using the reflectance value of the red (R red ) and near infrared (R NIR ) bands: NDVI = R NIR R red R NIR + R red Notice that turbid water or water with some surface vegetation may not satisfy the above conditions and may not be flagged as water at this step. Partial mitigation of this problem will be achieved later on ð1þ

4 186 C. Huang et al. / Remote Sensing of Environment 114 (2010) Table 1 Standard deviation values (% reflectance) used in Eqs. (2) and (3). Band 1 Band 2 Band 3 Band 4 Band 5 B These values were the average of those derived using images acquired in different years from different places. in the time series analysis step by using temporal information (see Section 3.3) Identification of forest samples Although the LTSS images have been corrected to achieve high levels of radiometric integrity, VCT uses forest samples to further normalize image radiometry and to calculate forest likelihood measures. Such forest samples are identified based on known spectral properties of forest. Specifically, dense, mature forests typically appear dark and green in a true color composite imagery, and are among the most easily distinguishable features in remote sensing imagery (Dodge & Bryant, 1976). As such, some of them can be identified reliably using histograms created from local image windows (e.g., 5 km by 5 km). Because forest pixels are typically the darkest vegetated pixels, they are generally located towards the lower end of each histogram. When a local image window has a significant portion of forest pixels, those pixels form a peak called forest peak in the histogram. In the absence of water, dark soil, and other dark non-vegetated surfaces, which are masked out using appropriately defined NDVI and brightness threshold values, forest pixels are delineated using threshold values defined by the forest peak. A detailed description of this approach has been provided by Huang et al. (2008) Calculation of forest indices The identified forest samples are used to calculate a number of indices that are indicative of the likelihood of each pixel being a forest pixel. Suppose the mean and standard deviation of the band i spectral values of forest samples within an image are b ī and SD i respectively, then for any pixel with a band i value of b i, a forest z-score (FZ i ) value for that band can be calculated as follows: FZ i = b i b i SD i ð2þ For multi-spectral satellite images, an integrated forest z-score (IFZ) value for that pixel is defined by integrating FZ i over the spectral bands as follows: sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 1 NB IFZ = NB ðfz i Þ 2 i =1 where NB is number of bands used. For Landsat TM and ETM+ images, bands 3, 5, and 7 are used to calculate the IFZ. Bands 1 and 2 are not used because they are highly correlated with band 3. The near infrared band is not included in the IFZ calculation because: 1) it is less sensitive to logging and other non-fire disturbances than the other spectral bands; and 2) spectral changes in this band do not always correlate with disturbance events. A major problem with using SD i calculated from forest samples within each individual image is that the value can vary greatly as a function of the forest type composition in that Landsat image. The SD i calculated this way will be low for images consisting of forest pixels that are spectrally similar, but can be very high for images consisting of both open canopy forests with bright backgrounds and closed canopy forests. Such a dependency of the SD i and hence the IFZ on the composition of forest types within each image makes it difficult to develop generic change detection algorithms for use over a wide range of forest biomes. To mitigate this problem, the average standard deviation values derived using images acquired in different years from different places of the U.S. are used in Eqs. (2) and (3) (Table 1). The FZ i and IFZ indices calculated using Eqs. (2) and (3) have a number of appealing properties: IFZ is an inverse measure of the likelihood of a pixel being a forest pixel. Pixels having low IFZ value near 0 are close to the spectral center of forest samples, while those having high IFZ values are likely non-forest pixels (Fig. 2). Assuming forest pixels have a normal distribution in the spectral space, FZ i could be directly related to the probability of a pixel being a forest pixel using the Standardized Normal Distribution Table (SDST) (e.g. Davis, 1986). As the root mean square of FZ i, IFZ can be interpreted similarly. Specifically, over 99% of forest pixels likely have IFZ values less than 3. Although in reality forest may not have a rigorous normal distribution, and the standard deviation values used here are not calculated from the image of interest, such an approximate probability interpretation makes it possible ð3þ Fig. 2. The IFZ values of different land cover types in eastern Virginia (a. left, WRS path 15/row 34) and Oregon (b. right, WRS path 45/row 29) show that deciduous and conifer forests have IFZ values that are generally below 3 and are different from those of other land cover types (except for some water).

5 C. Huang et al. / Remote Sensing of Environment 114 (2010) to define probability based threshold values that might be applicable to images acquired on different dates over different locations. While deciduous and coniferous forests often have different spectral characteristics, during the growing season they have similar IFZ values that are substantially more stable over time and are mostly lower than those of non-forest land cover types (Fig. 2). This observation makes it possible to detect forest changes using the IFZ index without knowing forest type, although the differences between the IFZ values of different forest types can be greater than those shown in Fig. 2 in many areas (e.g., see Section 3.3.2). In addition, the calculation of FZ i is essentially a normalization of an image using forest pixels in that image. This normalization process can substantially reduce the spatial and temporal variability of the spectral signature of forests caused by relatively homogeneous atmospheric conditions and instrument related issues (to be discussed in more details later in Section 5, Fig. 10). Fig. 3(a) shows that due to atmospheric effect, a Landsat image can have substantially higher top-of-atmospheric (TOA) reflectance values in the red band than surface reflectance (SR) values. For forest pixels, however, the IFZ values calculated using TOA and SR images are very close (Fig. 3 (b)). In addition, some of the variation caused by vegetation phenology (during the summer growing season only) and sun angle (bi-directional reflectance effects) are also normalized through the calculation of FZ i and IFZ. As will be discussed later, the selfnormalization property of these indices makes it possible for the VCT algorithm to be applicable to both TOA and SR images. In addition to the FZ i and IFZ indices, the VCT also calculates a normalized burn ratio index (NBRI) as follows: NBRI = R NIR R 7 R NIR + R 7 Where R NIR and R 7 are reflectance in the near infrared band (band 4) and band 7. This index has been correlated with field measured burn severity indices (e.g. Chen et al., 2008; Escuin et al., 2008), and is used to improve detection of fire disturbance events by VCT Cloud and shadow masking Although great effort has been taken to select images with the least cloud cover for each LTSS, due to persisting cloudy conditions in certain areas some images inevitably contained cloudy pixels (Huang et al., 2009a, Table 3). Cloudy pixels generally have high brightness values and low greenness values. If un-flagged, most likely they will ð4þ be mapped as non-forest regardless of the actual surface conditions beneath the clouds. For forest change analysis, un-flagged clouds over forest likely will be mapped as forest disturbance. Cloud shadow over forest may also be mapped as disturbance as the spectral signature of forest under shadow can be quite different from that of sunlit forest. Several cloud masking algorithms have been developed for use with images acquired by coarse resolution instruments (e.g. Simpson & Gobat, 1996; Ackerman et al., 1998; Stowe et al., 1999; Hutchison et al., 2005; Luo et al., 2008). For Landsat 7, an automated cloud cover assessment (ACCA) algorithm has also been developed for estimating the percentage of cloud cover within each image (Irish, 2000; Irish et al., 2006). The masking algorithm used in the VCT is based on observed characteristics of cloud and shadow. Clouds generally appear bright in the reflective bands and cold in the thermal band (Fig. 4(a b)). In the temperature-red band space (Fig. 4(c)), a partly cloudy image forms an elongated cluster of points extending from the upper left to the lower right. Vegetated, cloud-free observations are located near the upper left tip and cloudy pixels are located towards the lower right tail. Cloudy pixels are separated from cloud-free observations using threshold values defined by a set of linear boundaries called cloud boundaries (Fig. 4(c d)). These boundaries can be used to identify most cloud types, including high altitude cirrus clouds, which are not necessarily very bright but are often very cold (Huang et al., in press). For each cloud pixel, cloud height is calculated using its temperature and a normal lapse rate (Smithson et al., 2008). The shadow location of that cloud is then predicted according to solar illumination geometry and the calculated cloud height. For large cloud patches, however, the shadow of some cloud pixels can be covered by other cloud pixels. Therefore, only predicted shadow pixels that are also dark are flagged as actual shadows. Details on this cloud and shadow algorithm have been provided by Huang et al. (in press) Time series analysis After the masking and normalization step is complete for all images in an LTSS, temporal interpolation is used to derive interpolated values for pixels flagged as cloud, shadow, or other bad observation. The resultant masks and indices are then used to determine change and nonchange classes, and to derive a suite of attributes to characterize the mapped changes Temporal interpolation As discussed earlier, because use of measurements contaminated by cloud or shadow in land cover change analysis can result in false Fig. 3. Histograms of band 3 top-of-atmospheric (TOA) and surface reflectance (SR) (a. left) and those of the IFZ values calculated from TOA and SR (b. right) for a Landsat image (WRS path 15/row 33, acquired on July 6, 2001), showing that the IFZ values calculated from TOA and SR are very close for forest pixels (identified by the first peak in the histograms) even though the TOA and SR values for those pixels are quite different.

6 188 C. Huang et al. / Remote Sensing of Environment 114 (2010) Fig. 4. Clouds generally appear bright in the reflective bands (a) and cold in the thermal band (b). Cloudy pixels can be separated from cloud free observations using threshold values defined by a set of linear boundaries called cloud boundary in the red-temperature space (c). Cloud edges are flagged using a 2-pixel buffer from identified cloudy pixels, and shadows are identified in the direction projected from identified cloud and cloud edge pixels according to illumination geometry (d). changes, pixels flagged as cloud or cloud shadow should not be used in the time series analysis. However, ignoring such pixels and leaving them out from the analysis will result in holes in the derived change products. To avoid this problem, VCT uses temporal interpolation to derive interpolated values for such pixels. Specifically, for each pixel masked as cloud or cloud shadow in a particular year i, the temporally nearest non-cloud, non-shadow observations acquired before (p) and after (n) year i are used to calculate its value as follows: x i = x p + ði pþ x n x p n p where x is any of the indices calculated in Section 3.2. If no non-cloud, non-shadow observation can be found in the years before (or after) the current acquisition year, then the value for the current year is set to that of the temporally nearest non-cloud, non-shadow observation acquired after (or before) year i. For each flagged cloudy or shadow pixel, temporal interpolation is applied to all the indices calculated in Section 3.2. While we recognize that for pixels contaminated by cloud or shadow, there is no way to obtain their uncontaminated values, i.e., the values they would have if they were not contaminated, this temporal interpolation can provide an informed guess of those uncontaminated ð5þ values. A pixel that was contaminated by cloud or shadow in a year but was forested in the acquisition years before and after that year should be forested during that year, and Eq. (5) should give interpolated values close to those of a forest pixel. Similarly, a pixel that was contaminated by cloud or shadow in a year but was not forested in the acquisition years before and after that year should not be forested during that year, and the interpolated values from Eq. (5) likely will be different from those of a forest pixel. However, if a cloud or shadow contamination happened in the same year or the acquisition year immediately before the year when a disturbance occurred, the interpolated values can be very different from the uncontaminated values, and that cloud/shadow contamination may result in errors in the derived disturbance products (to be discussed in more details in Section 4) Determining change and no-change classes Time series analysis of forest cover and change is based mainly on the physical interpretation of the IFZ. Because the IFZ measures the likelihood of a pixel being a forest pixel, its value should change in response to forest change. Fig. 5 shows typical temporal profiles of the IFZ for major land cover and forest change processes. For persisting forest land where no major disturbance occurred during the years being monitored (throughout this paper the word persisting

7 C. Huang et al. / Remote Sensing of Environment 114 (2010) indicates that the cover type of a pixel remained the same during the entire observing period), the IFZ value stays low and is relatively stable throughout the monitoring period (Fig. 2 and 5(a)). During any year a sharp increase in the IFZ value indicates the occurrence of a disturbance in that year. A sequence of gradually decreasing IFZ values following that disturbance represents the regeneration process of a new forest stand (Fig. 5(b)). Conversion from non-forest to forest (afforestation) or regeneration of a forest stand from a disturbance that occurred before the first LTSS acquisition is documented by the gradual decrease of the IFZ from high values to the level of undisturbed forests (Fig. 5(c)). Notice that for both regeneration and afforestation processes the IFZ does not drop quickly from a high value to the level of undisturbed forest, but reduces gradually. This is because establishment of a forest stand is a gradual process; it takes at least several years for the regrowing trees to get a forest appearance in the spectral data. Finally, the IFZ is generally high throughout the time series for persisting non-forest pixels that remained non-forest during the entire observing period. While certain crops may be spectrally similar to forest and can have low IFZ values during certain seasons, their IFZ values likely will fluctuate greatly as surface conditions change from one year to another due to harvest and crop rotation (Fig. 2 and 5(d)). Based on these distinctive IFZ temporal profiles of different land cover and forest change processes, decision rules are used to identify persisting land cover types and to detect disturbances in a sequence of steps (Fig. 6). Step 1, Persisting water. Pixels identified as water by the water masks created in Section for all acquisition years are classified as persisting water. Because some water bodies may be turbid during some season and may not be masked as water during that season, any pixel that is masked as water at least half of the time and is also masked as water at least once during the first third of the observing period and once during the last third of the observing period is also classified as persisting water. Step 2, Persisting forest. Pixels not classified as persisting water in step 1 are further analyzed in this step. Persisting forest pixels are characterized by having low IFZ values throughout the entire observing period. Based on the approximate probability interpretation of the IFZ (see Section 3.2.3), most forests with closed or near close canopy cover should have IFZ values of less than a threshold value of 3. Here and in the Fig. 5. Typical IFZ temporal profiles of major forest cover change processes (a c) and non-forest (d) that are used to characterize different change processes (see Section for details). Fig. 6. Major steps and decision rules used by the VCT to determine persisting land cover types and forest disturbance classes. following discussions, this threshold value is used to separate low and high IFZ values. Step 3, Persisting non-forest. Pixels not classified as persisting forest in step 2 are further analyzed in this step. While most persisting nonforest pixels have high and often temporally variable IFZ values, some of them can be spectrally similar to certain forest pixels and can have low IFZ values during a particular season of a year. The likelihood of most non-forest pixels to have consecutive low IFZ values (CLIV), however, is low (Fig. 2). Therefore, if a pixel has a CLIV record, it likely was a forest pixel at least during the years when it had the CLIV record. Such a pixel is referred to as a once forested pixel. Obviously, the longer the CLIV record a pixel has, the more likely that pixel was a once forested pixel and less likely a persisting non-forest pixel. In VCT, a minimum number of consecutive low IFZ values (MNCLIV) is used to determine whether a pixel is a persisting non-forest pixel or was once forested. A pixel is classified as a persisting non-forest pixel if its longest CLIV record is less than the MNCLIV threshold value. For the biennial LTSS assembled through the NAFD project, a MNCLIV value of 2 has been found adequate to separate once forested pixels from persisting non-forest pixels. Step 4, Disturbance detection. A pixel not classified as any of the three persisting land cover classes in the previous steps should be once forested and should have consecutive low IFZ values during the years when that pixel remained forested. As discussed earlier, a disturbance typically results in a sharp increase in the IFZ (Fig. 5(b)). Unfortunately, an increase in the IFZ can also result from noisy observations, including un-flagged cloud, shadow, and instrument or processing related errors, because these noisy observations typically have high IFZ values. With each LTSS consisting of carefully selected Landsat images (Huang et al., 2009a), the likelihood of a pixel having un-flagged data quality problems in consecutive acquisition years should be low. Therefore, if a pixel remained forested before and after a noisy observation, that noisy observation most likely will result in a spike in the IFZ temporal profile, i.e., a high IFZ value preceded and immediately followed by low IFZ values, not consecutive high IFZ values. On the other hand, most disturbances, especially those leading to significant losses of forest canopy and live biomass, likely will result in consecutive high IFZ values (CHIV): A conversion from forest to a non-forest land cover type typically should result in non-forest signals (i.e., mostly high IFZ values) in the years following that disturbance event. For a disturbance followed by forest regeneration, including reestablishment of urban trees in areas converted from forest to an urban

8 190 C. Huang et al. / Remote Sensing of Environment 114 (2010) environment, the IFZ should remain high until the young trees grow to a stage such that they spectrally look like forest. Therefore, VCT uses the CHIV record following an IFZ hike to determine whether the increase was caused by a noisy observation or a disturbance. Only an IFZ hike followed by a CHIV record at least as long as a predefined minimum number of consecutive high IFZ values (MNCHIV) is mapped as a disturbance. For the biennial LTSS used by the NAFD, an MNCHIV value of 2 is used for most closed canopy forest ecosystems in the U.S. A higher MNCHIV value can reduce the ability to detect disturbances, especially in the southeastern U.S. where trees planted after a harvest can grow so rapidly that they can become spectrally inseparable from undisturbed forest in just 4 6years,or2 3 consecutive observations in a biennial LTSS, following that harvest (Huang et al., 2009a, Fig. 1). For open forests in the semiarid southwestern U.S., however, stress due to drought conditions can last many years, and can result in a CHIV record of more than 2. For the NAFD such a drought induced stress is not considered a disturbance. To avoid such a stress being mapped as disturbance, the MNCHIV is increased to 3 for open forest biomes. It should be noted that open canopy forests with bright backgrounds typically have IFZ values much higher than those of closed canopy forests and likely will be classified as persisting non-forest using the IFZ threshold value of 3 as definedinstep2.tominimizethisproblem,for images consisting of both closed and open canopy forest types, steps 2 to 4 are performed twice. In the first time, the initial IFZ threshold value of 3 is used to characterize forest and disturbances for areas having closed canopy forests. Pixels classified as persisting non-forest in the first iteration are reanalyzed in the second iteration, during which the IFZ threshold value in step 2 is relaxed for better characterization of forest and disturbances for areas having open canopy forests. Based on extensive examination of various sparse forests in the semiarid western U.S., the IFZ threshold value is set to 6.5 in the second iteration. We noticed that some fires, especially understory fires, did not always result in high IFZ values. If such pixels did not have other disturbances during the observing period, they likely will be mapped as persisting forest in step 2. To reduce such omission errors the VCT also checks the NBRI temporal profile for pixels mapped as persisting forest in step 2. Because fires typically result in low NBRI values (Escuin et al., 2008), they are detected by searching for significant decreases on the NBRI temporal profile. For disturbances that occurred at the beginning of the time series, there may not be a CLIV record that satisfies the MNCLIV criterion. Likewise, disturbances that occurred at the end of the time series will not have a CHIV record that satisfies the MNCHIV criterion. Therefore, use of the MNCLIV and MNCHIV threshold values as described above will not allow detection of such disturbances. To alleviate this problem, the MNCLIV is relaxed for disturbances that occurred at the beginning of the time series, and the MNCHIV is relaxed for disturbances that occurred at the end of the time series. Because of the relaxation of the MNCHIV and MNCLIV criteria, the disturbances detected at the beginning and end of the observing period of an LTSS likely will be less reliable than those detected in the middle of the observing period. If a pixel reaches step 4 but its longest CHIV record is shorter than the MNCHIV and the CHIV record is not at the end of the time series, it is classified as persisting forest Disturbance year. For a typical forest disturbance, the IFZ value increases sharply following a CLIV record (Fig. 5(b)). While the actual occurrence of the disturbance is somewhere between the acquisition dates of the two consecutive images that exhibit the sharp increase, disturbance year is defined by the acquisition year of the second image. Therefore, the disturbance year measure calculated using biennial LTSS could be up to 2 years later than the actual occurrence of a disturbance. This uncertainty will be reduced to 1 year if annual LTSS is used. When an un-flagged bad observation precedes the occurrence of a disturbance, it can add an additional error to the detected disturbance year (an example is given later in Fig. 9). This can happen even when the bad observation has been flagged up front, because the IFZ value interpolated using Eq. (5) may still be substantially higher than that of a forest observation. Regeneration from a disturbance that occurred before the first LTSS acquisition is indicated by an initially high IFZ value, which decreases gradually in later years and remains low for several consecutive years (Fig. 5(c)). This disturbance category is called pre-series disturbance (P-SD). Disturbances in this category did not necessarily occur in the first acquisition year of a given LTSS, but could have occurred years before the first acquisition year. How many years a disturbance could have occurred before the first acquisition year and still could be mapped by the VCT as a P-SD will depend on how fast the tree grows. In regions such as the southeastern U.S. where trees can grow rapidly, this disturbance category likely will only include disturbances that occurred just a few years before the first acquisition, because in such regions it only takes young trees a few years to become spectrally inseparable from undisturbed forest (e.g., Huang et al., 2009a, Fig. 1). In high latitude region or areas where trees grow more slowly, disturbances that occurred many years before the first LTSS acquisition may also be recorded in the P-SD category. Another change process that will also be recorded in the pre-series disturbance category is conversion of non-forest land to forest, or afforestation. In fact, use of information from the LTSS alone will not allow separation of afforestation from regeneration that followed a pre-series disturbance. Additional information will be needed to determine whether the pre-series non-forest condition was a nonforest land use or resulted from a pre-series disturbance Disturbance magnitude measures. Disturbance magnitude refers to the spectral change resulting from a disturbance. For the multispectral Landsat images, different disturbance magnitude measures can be calculated using different spectral bands or indices. The VCT calculates three disturbance magnitudes, with the first using the IFZ, the second using the NDVI, and the third using the NBRI. Fig. 7 shows the calculation of the IFZ disturbance magnitude. It is the difference Disturbance characterization For each disturbance detected by the VCT, a disturbance year and several disturbance magnitude measures are calculated to characterize that disturbance. In addition, two attributes are used to summarize the IFZ profile that follows the disturbance, which may be useful for characterizing the regeneration process that may follow each detected disturbance. Definitions of these attributes are illustrated in Fig. 7. Fig. 7. Disturbance year (1), disturbance magnitude (2), and slope and R2 of a linear fit (4) of the regeneration curve for a typical forest disturbance (a) and a pre-series disturbance (b). The regeneration curve is the part of the IFZ profile for the period from the disturbance year (1) to the year (5) when the IFZ value gets close to the average IFZ value of forest observations (3), which is calculated as the mean of CLIV, or to the end of the LTSS if the IFZ value does not reach the average forest IFZ value at the end of the observing period.

9 C. Huang et al. / Remote Sensing of Environment 114 (2010) between the IFZ value in the disturbance year and the mean IFZ value of consecutive forest observations, which are represented by CLIV (Fig. 7 (a)). Here use of the mean IFZ value of consecutive forest observations instead of the IFZ value of the pre-disturbance observation can minimize the impact of inter-annual variations of forest signal on the calculated disturbance magnitude value. For a pre-series disturbance, the disturbance magnitude is calculated using the IFZ value of the first LTSS acquisition and the mean IFZ value of CLIV (Fig. 7(b)). The NDVI and NBRI disturbance magnitudes are calculated the same way as the IFZ disturbance magnitude. It should be noted that other spectral indices, such as the band specific FZ i values or the tasseled cap indices (Crist & Cicone, 1984; Huang et al., 2002), can also be used to calculate disturbance magnitudes. Further studies are needed to determine which of these disturbance magnitude measures can better characterize the nature and intensity of detected disturbances Regeneration characteristics. If forest regeneration occurred after a disturbance, this regeneration process is tracked by a regeneration curve (Fig. 7). This curve refers to the portion of the IFZ profile from the disturbance year (point 1 in Fig. 7) to the point where the IFZ value gets close to the average IFZ value of CLIV (point 5 in Fig. 7), or to the end of the LTSS if the IFZ value does not reach the average forest IFZ value at the end of the observing period. To determine whether this curve tracks forest regeneration and to characterize the regeneration rate, a line is fit using this curve and the R 2 of the fit and the slope of the line are calculated. Regeneration of a new forest stand likely will yield high R 2 values, and the slope of the line may be an indicator of the growth rate of the regenerating forest. Of course, the linear fit will be statistically meaningful only when there are enough observations following a disturbance. For disturbances that occurred near the end of the LTSS, neither the slope nor the R 2 is statistically meaningful and fill values will be assigned to them. It should be noted that a spectral recovery as represented by the regeneration curve is not synonymous with ecological definitions of forest recovery. Spectral recovery tends to occur fairly quickly (within 5 15 years depending on forest type) such that even low biomass levels may give IFZ values comparable to those of mature forests. In areas like the southeastern U.S. where certain tree species grow fast enough to allow more than one forest harvest during the observing period of an LTSS, some fields may experience more than one disturbance. In such cases the VCT will detect all disturbances and will calculate the above described attributes for each detected disturbance. 4. Algorithm assessment As mentioned earlier, VCT has been tested in many places in the U.S., including Mississippi (Li et al., in press), Alabama (Li et al., 2009), and the locations where LTSS have been assembled through the NAFD project (Goward et al., 2008; Huang et al., 2009a). This highly automated algorithm was found very efficient. Except for a few parameters that needed fine tuning when both closed and open canopy forests are present within an image (see Section 3.3.2), universal or near-universal threshold values were used for all other parameters. On average it took 2 3 h to analyze an LTSS consisting of 12 or more Landsat images using an average desktop PC available in today's market. Based on our experience, use of existing bi-temporal change detection techniques to analyze such an image stack would take tens of working days, depending on the experience of an image analyst and the complexity of land cover and change processes within that image stack. Such an efficiency gap between the VCT and existing bi-temporal change detection techniques likely will not reduce as computers become faster in the future, because almost all the time required by VCT is CPU time, whereas many bi-temporal change detection techniques require significant amount of human inputs. Comprehensive validation of the entire suite of VCT products as described in Section 4 has been found extremely challenging. Linking the disturbance magnitude measures to changes in biomass or other biophysical variables requires pre- and post-disturbance measurements obtained using methods that would allow reliable retrieval of those variables. Similarly, linking the regeneration characteristics to vegetation biophysical changes associated with regeneration processes requires multi-temporal reference data sets on those biophysical variables. Such reference data sets likely will be scarce, especially for older disturbances that occurred in the 1990s and 1980s, although their availability has yet to be better understood. Therefore, we have focused on the disturbance year products in assessing the performance of the VCT algorithm, using both qualitative and quantitative methods. Qualitative assessments include limited ground validation and comprehensive visual assessment. Quantitative assessments include a design-based accuracy assessment and comparison with field measurement of forest age collected through the USDA Forest Service Forest Inventory and Analysis (FIA) program. For page limit considerations, only qualitative evaluations and a summary of the design-based accuracy assessment are provided here. The assessment using FIA data and more details on the design-based accuracy assessment will be provided in a follow-up paper (Thomas, in preparation) Ground-based assessment Ground-based assessment is often considered a preferred method for evaluating land cover classifications derived from remote sensing. Using this method to evaluate all the classes in a VCT disturbance year map, however, would require ground-based data collected in each of the acquisition years of the concerned LTSS, which likely do not exist for most disturbance classes. Nevertheless, checking the disturbances mapped by the VCT on the ground would allow a better understanding of the nature of those disturbances and the processes that occurred following those disturbances. For this purpose we conducted limited field trips in Virginia (path 15/row 34, in October 2005), Mississippi and Alabama (path 21/row 37 and path 21/row 39, in May 2007), and Oregon (path 45/row 49, in July 2007). In each field trip, we visited a list of pre-selected sites where disturbances were mapped by the VCT. Field photos were taken along with GPS coordinates during each trip. Ground evidences were easy to find for recent disturbance events that resulted in complete or near complete removal of the forest canopy, including harvest (Fig. 8(a)), fire (Fig. 8(e)), and urban development (Fig. 8(d)). Older disturbances that occurred years before the field trips and were followed by regeneration of new forest stands were often evidenced by the existence of young, even-aged forests, and the disturbance year was roughly reflected by the height of the regenerating trees (Fig. 8(b and c)). Likely due to vigorous growth of understory vegetation in Virginia and Mississippi/Alabama, evidence of less intensive disturbances such as storm damage, insect/disease defoliation, and selective logging was difficult to find during the trips. Characterized by dry environmental conditions and slow growth of both trees and understory vegetation, the eastern side of the Cascades in Oregon had many evidences of less intensive disturbances, including some that occurred many years before the field trip. There we found that some less intensive disturbances, such as selective logging and fuel treatment, were mapped successfully by the VCT (Fig. 8(f)), but many others were not Visual assessment Compared with the ground-based assessment method, visual checking of the disturbances mapped by the VCT against the input Landsat images provides a more immediate and yet reliable way to evaluate those disturbances. In general, the spectral change signals of most forest disturbances can be identified reliably by experienced

10 192 C. Huang et al. / Remote Sensing of Environment 114 (2010) Fig. 8. Field photos taken in 2007, showing the ground conditions in that year for disturbances that occurred in different years. The year in each photo caption is the disturbance year determined by the VCT. These photos can provide direct or indirect evidence on whether and when the mapped disturbances occurred, and on the nature of the disturbances (see text in Section 4.1 for details). image analysts (Huang et al., 2008; Masek et al., 2008), especially when images acquired immediately before and after the occurrence of those disturbances are available (Cohen et al., 1998). It has been shown that for each disturbance event detected by the VCT, whether the mapped location and occurrence year of each detected disturbance are correct can be determined reliably by visually examining the Landsat images acquired immediately before and after that disturbance event (Huang et al., 2009b, Fig. 3). Theoretically, all disturbance events mapped by the VCT can be evaluated this way, except those dated in the first and last acquisition year of each LTSS. While we did not have the resources to check every disturbance event mapped by the VCT against the pre- and post-disturbance Landsat images, we did browse the disturbance year maps for most locations where LTSS have been assembled, including the NAFD Phase I sites (Huang et al., 2009a) and the WRS path/rows covering Mississippi and Alabama (Li et al., 2009, Li et al., in press). The following were observed through this comprehensive visual assessment: Most disturbance year maps were found quite reasonable. Here a reasonable disturbance year map was defined as (for some

11 C. Huang et al. / Remote Sensing of Environment 114 (2010) graphic examples, see Huang et al., 2009b, Fig. 4): 1) the map had minimum speckles; 2) for human disturbance events such as harvest and logging, the mapped disturbance polygons had regular shapes or linear features that were indicative of the human origin of those disturbances; and 3) for nature disturbances such as fire and storm damages, they had irregular shapes but were often contiguous. Disturbances that were mapped by the VCT but were deemed unreasonable through this visual assessment were checked against the pre- and post-disturbance images to determine whether those disturbances were false alarms. The VCT algorithm was able to handle certain levels of bad observations in individual images. Those bad observations typically did not leave gaps or footprints in the derived disturbance year maps (Fig. 9), except the ones preceding a disturbance immediately (Fig. 9(b and e)). The VCT achieved such a level of resistance to bad observations through 1) automatic masking of cloud and cloud shadow, 2) temporally interpolating for identified bad observations, 3) requiring consecutive observations in determining forest and change, and 4) simultaneously considering the entire temporal domain of each LTSS. Of course, un-flagged, consecutive bad observations may result in false changes, and excessive bad observations may completely fool the VCT algorithm. Therefore, it is still necessary to select high quality images for use in the LTSS in order to produce satisfactory disturbance products. While it is preferable to use surface reflectance in the VCT algorithm, it can still work under circumstances when atmospheric correction cannot be performed. This is because the calculation of the FZ i and IFZ is essentially a normalization of an image using forest pixels in that image. Such a normalization approach was used in a previous study to develop a disturbance index (Healey et al., 2005). This normalization process is functionally similar to a dark object subtraction approach to atmospheric correction (Chavez, 1996), which can reduce the impact of the atmosphere when such an impact is relatively homogeneous across an image. Fig. 10 shows that the disturbance rates derived by applying the VCT algorithm to surface reflectance (SR) and top-of-atmospheric (TOA) reflectance images track each other, and that the disturbance patches derived from them are almost identical. The small differences between the disturbance rates derived using TOA and SR images (Fig. 10(a)) were probably due to non-homogeneous atmospheric effects in some of the LTSS images Design-based accuracy assessment Derivation of reference data To obtain quantitative estimates of the accuracies of the disturbance year maps produced by the VCT, we conducted design-based accuracy assessment over 6 NAFD sites according to Stehman & Czaplewski (1998). These sites were selected to represent different forest biomes and disturbance regimes in the U.S. (Table 2). For each site, validation samples were selected using a stratified random sampling method, where each sample was a 28.5 m TM pixel, and each class (see the legend of Fig. 9 for a complete list of those classes) in a draft version of the disturbance year map was used to define a stratum. Notice persisting water was included in the persisting non-forest class for the purpose of accuracy assessment. The inclusion probability of the samples in each stratum was calculated as the ratio of the number of samples selected Fig. 9. Bad observations such as cloud/shadow contamination (a, WRS path 16/row 36, acquired on August 14, 2000), missing scan lines (b, WRS path 16/row 36, acquired on September 25, 1986), or duplicate scan lines (c, WRS path 45/row 29, acquired on August 13, 1984) in individual images leave little or no signs in the disturbance year maps produced by the VCT (d, e, and f, produced with a, b, and c, respectively, as part of the inputs). When preceding a 1988 disturbance, however, the missing scan line in (b) caused the 1988 disturbance to be mapped as a 1986 disturbance (circled in (e)). The images in (a c) are shown with bands 4, 3, and 2 in red, green and blue.

12 194 C. Huang et al. / Remote Sensing of Environment 114 (2010) Fig. 10. Disturbance rates for the Virginia site (path 15/row 34) derived by applying the VCT algorithm to surface reflectance (SR) and top-of-atmospheric (TOA) reflectance images track each other, although the SR images yielded slightly higher disturbance rates in most years (a). A detailed checking of the disturbance year maps (b and c) reveals that the two maps are almost identical at the patch level. See Fig. 9 for the legend for the disturbance year maps. within that stratum over the total pixels of that stratum (Stehman et al., 2003). In order to achieve satisfactory precisions with the accuracy estimates we targeted 50 samples per stratum on average, with a minimum value of 20 for rare classes. A total of approximately 700 validation samples were selected for each validation site (Table 2). For each validation sample, all images of the target LTSS were inspected visually to determine whether it belonged to one of the persisting classes or it had disturbances. If disturbances were found at a sample location, the disturbance years were recorded. As demonstrated in Section 4.2 and in many previous studies (e.g. Cohen et al., 1998; Healey et al., 2005; Kennedy et al., 2007; Huang et al., 2009b), the disturbance year determined this way should be reliable and therefore could be used as a reference for validating the disturbance year determined by the VCT. To minimize potential difficulties in interpreting the Landsat images, for each validation sample we obtained a Digital Ortho Quarter Quad (DOQQ) from the TerraServer ( to assist the analysis. Where necessary, the high resolution imagery available from GoogleEarth was also checked. These images typically were not acquired immediately before or after the occurrence of a disturbance, but their high spatial resolutions provided a local reference for the image analysts to separate forest, non-forest, and disturbance at the Landsat level. Considering the fact that each LTSS image could have subpixel misregistration errors, the compound misregistration errors from all Landsat images in that LTSS could be over one TM pixel. To minimize the impact of such misregistration errors on accuracy estimates, any validation sample that was located on the edge of a land cover or change polygon was moved towards the inner part of that polygon such that it was at least 1 pixel away from the edge Accuracy estimates For each of the 6 validation sites, the above derived reference data sets were used to create a confusion matrix. Accuracy measures, including overall accuracy, kappa coefficient, and per class user's, and Table 2 Characteristics of the 6 validation sites where the VCT disturbance year products were evaluated using a design-based accuracy assessment method. WRS path/ row Location Land cover and forest characteristics Major disturbances Accuracy assessment sample size 12/31 Southeastern New Mostly temperate deciduous forests, agriculture, urban Urbanization, logging 697 England 15/34 Virginia Pine plantation, deciduous or mixed forests, agriculture Harvest, logging /37 Mississippi/Alabama Pine plantation, deciduous or mixed forests, agriculture Harvest, logging /27 Minnesota Temperate deciduous and mixed forests, agriculture, wetlands Wind throw, ice damage, 750 harvest 37/34 Southern Utah Semiarid, mostly shrub and grassland, pinion/juniper forests are typically Fire 645 short and sparse 45/29 Oregon Temperate evergreen forests to the west, dry grass and shrubs in the middle and to the east Fire, harvest, logging, fuel treatment 700