Classification of Boreal Using SAR Images Forest Cover Types

M a p p i n g forest cover types in the boreal ecosystem is important for understanding the processes governing the interaction of the surface with the atmosphere. In this paper, we report the results' of the land-cover classification ( f the SAR (synthetic aperture radar) data acquired during the Boreal Ecosystem Atmospheric Study's intensive field campaigns over the southern study area near Prince Albert, Canada. A Bayesian maximum a posteriori classifier was applied on the National Aeronautics and Space Administration~Jet Propulsion Laboratory airborne SAR images covering the region during the peak cf the growing season in July 1994. The approach is supervised in the sense that a combination of field data and existing land-cover maps are used to develop training areas fi)r the desired classes. The images acquired were first radiometrically and absolutely calibrated, the incidence angle effect in airborne images was corrected to an acceptable accuracy, and the images were used in a mosaic fi)rm and geoeoded and georeferenced with an existing landcover map for validation purposes. The results show that SAR images can be classified into dominant forest types' such as jack pine, black spruce, trembling aspen, clearing, open water, and three categories of mixed strands with better than 90% accuracy. The unispecies stands' such as jack pine and black spruce are separated with 98% accuracy, but the accuracy of mixed coniferous and deciduous stands suffers from co,@sing factors such as varying species composition, surface moisture, and understory effects'. To satisfy the requirements of process models', the number of cover types was reduced from eight to five general classes ~f conifer wet, conifer dry, mixed deciduous, disturbed, and open water. Reduction


INTRODUCTION
One of the major challenges of developing Earth system process models both on global and on regional scales is the accurate representation of the terrestrial vegetation. These process models work at a variety of spatial scales ranging from meters to kilometers. Depending on the application of the processes and their scales, the definition of categories of vegetation types may change. For example, for global land-atmosphere models such as Biosphere-Atmosphere Transfer Scheme (BATS), 18 general land-cover types are defined that are often inferred from maps, atlases, and national databases (Dickenson, 1994). For finer-scale process models, the availability of highresolution land-cover maps can improve the parameterization of landscape to functionally different strata. Currently, there are several approaches under investigation to statistically aggregate the high-resolution maps derived from remote-sensing techniques to a desired process model grid scale (Hall et al., 1995). These techniques are primarily focused on exploiting optical remote-sensing data such as that of the advanced very high resolution radiometer (AVHRR) and Landsat (Sellers et al., 1994;Townshend et al., 1991).
As a complement to optical remote-sensing techniques, land-cover maps derived from multipolarization, multifrequeney synthetic aperture radar (SAR) systems are an important tool for terrestrial ecologists and process modelers. Independenee of SAR data of solar irradianee and cloud cover is one significant reason for using this technique for land-cover classification, especially in northern latitude boreal forest and tropical raintbrest where the acquisition of optical data is hindered by frequent cloud cover and fire smoke. In addition, the sensi-tivity of the radar signal to moisture content and structural properties of vegetation may separate forest types, particularly when optical sensors are saturated over dense vegetation. Several studies, using a variety of classification approaches, have used SAIl images for landcover-type classification in forested regions (Saatchi et al., 1996;Ranson and Sun, 1994;Dobson et al., 1994;Cimino et al., 1986). The application of these high-resolution maps in process models has not been thoroughly explored. For example, Bonan (1993) has used the SAR-derived land-cover map over the boreal forest of interior Alaska to improve the estimation of forest assimilation.
Our classification of land cover is used to address the specific requirements of the BOREAS (Boreal Ecosystem Atmospheric Study) modeling activities. Separating functionally important land-cover types for modeling the exchange of trace gases between the land surface and the atmosphere is the ultimate goal of this study. The area has been under intensive study during the BOREAS project {br its important role in biogeochemical cycles between land and atmosphere at northern latitude (Sellers et al., 1995). The land-cover types in this region can be characterized by only a few dominant tree species. For example, the separabili~ of conifer and deciduous stands and the dry and wet conditions in this region are important for estimating the rates of photosynthesis, respiration, carbon assimilation, and nitrogen concentration. There{bre, the process of classification with the use of SAR imagery is, first, to illustrate the capability of the instrmnent to identii}" these classes, and, second, to show the spatM pattern of these classes over a region used for ecosystem processing models. In this study, we discuss the application of SAt/ data for mapping of forest types in the BOREAS area. A supervised classification approach using a maximum a posteriori Bayesian classifier is applied on the three-frequency polarimetric Jet Propulsion Laboratory airborne synthetic aperture radar (JPL AIRSAR) data to identify eight classes. Classification accuracies are computed first for the areas used {br training the classifier, then over several homogeneous sites examined during the field observation, and finally by comparing the results with a digital vegetation map assembled t'rom infrared aerial photointerpretation performed in 1984.

BOREAS EXPERIMENT
The Boreal Ecosystem Atmospheric Study is a cooperative field experiment integrating land surface climatology, tropospheric chemistry, and terrestrial ecology. In general, the experiment was designed to extend the findings of FIFE over grass prairie to boreal forests, one of the earth's largest and complex biomes, where coniferous species dominate. The biome has upland forests, extensive wetlands, some deciduous species, and many lakes, and it is a major storage of organic carbon, mostly in the soil (BOREAS Science Steering Committee, 1990). Among the primary objectives of the experiment, 1) improving understanding of the processes that govern the exchange of energy, water, heat, carbon, and trace gases between the boreal forest ecosystem and the atmosphere and 2) developing and validating remote-sensing techniques to transfer our knowledge of these processes from local to regional scales are of great importance. A relevant scientific issue is the sensitMty of the boreal forest biome to changes in physical climate and vice versa. Mapping vegetation-cover types, changes in land use, and the species composition in the region can contribute to long-term climatological research studies.

SRe Deseription
The focus of our paper is the BOREAS southern study area (SSA), which covers an area about 130 km in tile east-west direction and 90 km from north to south (Fig.  1). The southern boundary is located approximately 40 km north of the town of Prince Albert, Saskatchewan, Canada. The SSA topography, is gentle, with local elevations ranging from 550 to 730 m. Soils range from gray wooded to degraded black and are classified as brunisolic, gleysolic, chenozemic, luvisolic, and organic soil orders (Anderson and Ellis, 1976). Glacial deposits vary in thickness from 100 to 1000 m on the top of the Cretaceous bedrock. The western part of the SSA is in the Prince Albert National Park (PANP), and the eastern region falls within and around the Narrow Hills Provincial Forest.
The SSA is near the southern limit of tile boreal fi)rest and the transition to natural prairie grassland and agricultural land is 15 km to the southeast. The image data discussed in this paper are east of PANP in the area of the Narrow Hills Provincial Park. The image area also coincides with the BOREAS modeling grid (50×50 kin) used mainly for verifying remote-sensing algorithms and ecosystem modeling results. The vegetation in this area is classified as mixed boreal forest. On well-drained and san@ soil, the predominant species is jack pine (Pinus banksiana). Poorly drained sites support black spruce (Picea mariana). Mixed stands of aspen (Populus tremuloids), balsam poplar (Populus balsamifera), and white spruce (Picea glauca) are {bund on well-drained glacial deposits. In poorly drained areas throughout the study area, bogs support black spruce with tamarack (Larix laricina). The fen areas are composed mostly of sedge (Carex s'pp.) with discontinuous cover of tamarack or swamp birch (Bemla pumila). Localized logging fbr paper pulp and fence posts is common along Highways 1(16 and 120 and along Harding Road (see Fig. 1). The northeastern part of the study area encompasses a part of the Fishing Lakes burn that occurred in 1977 and 1978. Stands of small (<5 cm) jack pine regrowth now cover most of the burn areas. Major land-cover types are identified according to the needs of the BOREAS scientific applications. These land-cover types are chosen on the basis of their dominant species, canopy closure, soil organic properties, and their roles in determining the physics of the interaction of land surface and atmosphere. The land-cover categories consist of dry eonifers (e.g., jaek pine), wet eonifers (e.g., black spruce), deciduous (trembling aspen), clear cut, open water (lakes and river), brushland, treed muskeg, mixed eoniferous and deciduous trees, and regrowth (e.g., young jack pine). Between summer of 1993 and fall of 1994, forest stands of major land cover were sampled to measure tree species composition, stand geometry, biomass density, and several other forest canopy attributes. Data collections on the ground were performed for many applications and are available for all the flux tower and auxiliary sites. The flux tower sites are mainly single speeies stands.
In addition, there exists a digital vegetation map of SSA that was assembled from 1 : 12,500 scale infrared ae-rial photography and field reconnaissance notes in 1984. This vegetation map has been verified on the gronnd, but no accuracies are provided. The map consists of 40 different classes, regrouped to simplify the representation of vegetation types for dominant classes (Fig. 2). The map does not show recent changes due to tree logging, regrowth, and transformation of treed muskeg to predominantly black spruce stands.

AIRSAR DATA
The JPL airborne synthetic aperture radar (AIRSAR) was flown aboard a National Aeronautics and Space Administration DC-8 during all the intensive field campaigns (IFC) in summer of 1993, in April 1994 during the thaw period of the boreal forest, and in summer and fall of 1994. The AIRSAR operates at three frequency bands, P-band (68-cm wavelength), L-band (24-era), and C-band (5.6-era), with fully polarimetric capability. The incidence angle of the radar varied between approxi- mately 20 ° and 60 °. The radar data used for land-cover classification were acquired on 21 July 1994 and processed in synoptic mode (50-km swath). We chose this date to prevent possible errors in classification due to the partially frozen condition during the thaw period and to the leaf-off condition during the fall season. We used images from several parallel flight lines in a mosaic mode to create larger area coverage over the modeling grid. The calibration, radiometric correction, and mosaic of the images were performed in several steps.

Image Calibration
In this stu@, we made use of synoptic SAR images that were acquired with parallel flight lines in a "race track" trajectoD. The synoptic images have larger coverage (approximately 50 km) but only three polarizations. These images are often processed for the purpose of surveying the area and are not absolutely calibrated. We processed a total of' 15 synoptic images to cover all the bands and polarizations of the AIRSAR system. Calibration of images was performed by using fully polafimetric calibrated frame images processed over a part of the synoptic images. Absolute calibration constants were obtained by computing the ratios of baekscattefing coefficients from identical areas from both images and applying the calibration constants to all synoptic images. When compared with frame images, the synoptic images were absolutely calibrated with less than 0.1-dB error for all polarization Figure 3. P-band polarimetric color overlay of the AIRSAR mosaic image of the modeling subgrid within the BOREAS southern study area acquired on 21 July 1994. P-band HH, HV, and VV polarizations are in red, green, and blue, respectively. The mosaic image is coregistered with the digital vegetation map and georefereneed to universal transverse Mercator coordinates with North being parallel to the side of the image. channels. The frame images were calibrated both internally and externally by using data collected over an array of corner reflectors deployed over the Rosemond dry lake calibration site in California before and after the AIR-SAR campaign. After the absolute calibration, the images were resampled to ground range to remove the distortions in the near-range and far-range pixels.

Incidence Angle Correction
One of the disadvantages of airborne SAR data, when used for land-cover classification, is the variation of the incidence angle along the range lines across the image (20-60°). Consequently, areas with similar land-cover types produce different baekscatter signatures if they are imaged at different incidence angles; and, depending on the scene characteristics, the variation of the backseatter signature along each range line may be different. These effects can cause inaccuracies in a consistent class separation over the entire image. Correction of the image for incidence angle effects, therefore, becomes a necessary but impossible task to accomplish exactly. To have an optimal correction for incidence angle effects, several approaches have been suggested. Yueh et al. (1988) normalized the SAR data by the total power. This technique eliminated most of the incidence angle effects but at the same time changed some of the information in SAR backscatter signatures. The resulting normalized images were not able to discriminate all classes. Another method was proposed by Sader (1987), in which homogeneous areas of the same types were chosen along the range line, and the total image was calibrated such that these areas had equal baekscattering signatures. This technique will not work in areas with complex land-cover types. Rignot and Drinkwater (1993) corrected the incidence angle effects in their classification of sea-ice types by first segmenting the image along the range line, performing range-dependent clustering of the image, regrouping the clusters, and employing a supervised classification to produee self-consistent classes across the image scene. However, their technique requires that similar class types be represented in each segment of the image. Over complex land-cover types, sometimes only a limited number of classes is present over each range segment, causing difficulty in regrouping the clusters and removing the range-dependent effects. Ranson and Sun (1994) used AIRSAR images over forested land surfaces, selected a part of each image line within sapwood areas, calculated the mean and standard deviation of these pixels, and discarded all pixels falling outside of _+2 standard deviations. The remaining pixels were used to estimate the mean values at each image row, then a linear regression was used to estimate the calibration ratio for each line, and, subsequently, the entire image was calibrated by using these ratios. When employing this technique, it was found that the linear regression method did not always compensate for the inhomogeneous scene characteristic along the range line.
The synoptic images used in this study were corrected for incidence angle variations with a technique slightly different from that of Ranson and Sun (1994). We plotted the incidence angle variations for each range line, and then a nonlinear regression in conjunction with a cubic spline smoothing algorithm was used to estimate the general behavior of the incidence angle variations along each range line. The regression curve was then normalized by the mean backscattering coef~cient of the range line and then used to correct for the incidence angle effects of that range line. The entire image was then corrected line by line.

Image Mosaic
After calibration and incidence angle correction, the images from each frequency band and polarization were used in tandem to generate a mosaic image over ahnost the entire modeling subgrid. Figure 3 shows a color composite of the mosaic image at P-band (red, P-HH; green, P-HV; blue, P-W). Because the images were acquired from flight lines with the same heading, they also had an area of overlap with adjacent images. A linear feathering technique was then employed to remove the tonal inconsistencies that existed at the areas of overlap. In some areas where incidence angle effects were not optimally corrected, the feathering technique guaranteed further smoothing at the edges of images. If the overlapping regions were near the lakes where there was a dramatic change in the radar baekscatter signature, incidence angle effects could not be totally removed, and the edge effects were still obvious in the mosaic image.

CLASSIFICATION METHODOLOGY
When designing a classifier, it is important to define the mathematical basis of the classifier and, at the same time, to distinguish between the supervised and unsupervised learning procedures within the classifier. Here, we make use of a maximum-a-posteriori (MAP) Bayesian classifier developed for multifrequency polarimetric SAR data (Rignot and Chellappa, 1993). The MAP classifier models the SAR amplitudes as circular Gaussian distribution, which ineans that textural variations in radar baekscatter from tile surface are not considered to be significant enough to be incorporated into the classification scheme. In this method, the a priori distribution of image classes is modeled by using a Markov random field. From the models of the a priori distribution of classes, a model for the a posteriori distribution of the image classes is derived from the SAR image by using the Bayes' theorem. The optimal image classification of" the SAR data is defined as that which maximizes the a posteriori distribution of classes and is called the maximum a posteriori estimate of the image classes.
The MAP method is inherently different from and superior to the maximum likelihood estimation (MLE) procedure. The classifiers based on the maximum likelihood methods view the parameters (classes) as quantities whose values are fixed but unknown, and the best classification is defined to be the one that maximizes the probability of obtaining the samples actually observed (Duda and Hart, 1973). The MAP classifier views the parameters or classes as random variables with some a priori distribution. Iterative observation of the feature space converts this into an a posteriori density, thereby revising the decision about the true nature of classes. Because class distributions are analytical functions, the class charaeteristics obtained from training areas stay the same as in MLE. Air advantage of this technique, besides its mathematical rigor, is that it is general enough and, ~vith minor modification of the feature space, can be applied to both optical and SAt/ images, therefore creating the opportunity for both comparative and synergistic studies.
A number of other classifiers that have been used successfully but with limited capability fbr generalization also are available in the literature. Among them, Ranson and Stln (1994) used a combination of principle component analysis and MLE to come up with about 80% accuracy over Northern Experimental Forest near Howland, Maine. Pierce et al. (1994) introduced a knowledge-based classifier over a test site in northern Michigan. This tectmique depends on the absolute baekseatter values derived over training areas and the texture information that may be degraded ()wing to multilook averaging in polarimetric SAIl data in removing the speckle noise.
The learning procedure for the classifier is supervised in the sense that the state of the nature (class label) is known in advance, and training areas are chosen on the basis of a priori knowledge of the scene or the visual interpretation of the image. To implement the MAP classifier over the SAIl mosaic image, we first define the a priori distribution of the SAIl data for image classes by computing the average eovarianee matrix over the single training area. We eoneentrated on eight categories of training cover: 1) jack pine (jp), 2) black spruce (BS), 3) trembling aspen/mixed (TA/MX), 4) mixed jack pine and aspen (JP/TA), 5) mixed black spruce and jack pine (BS/ jP), 6) mixed strands (MX), 7) clear cut, disturbed, and nonlbrest (CC), and 8) open water (OW). For each category, we selected large homogeneous stands from the knowledge acquired during the field observation and the existing land-cover map. The average covarianee matrices are then computed over training areas for all three frequencies. Here, we used three training areas for jack pine stand, depending on the density and age, and two black spruce areas from a tower site and mature treed muskeg stand. The use of a limited number of training areas ensures realistic classification accuracy and the extrapolation of" the results to the entire image. Table 1 lists the calibrated radar backscattering coefficients, copolarized phase difference in degrees, and the coefficient of correlation in the linear domain between the complex amplitudes at HH and VV polarizations. The radar charaeteristics are obtained from the frame image, s processed over 10x10 km areas within each synoptic image. The forest stands chosen for the training areas were imaged at nearly the same incidence angle (typically, about 45 ° incidence angle)., thus the radar parameters for the image classes are assumed to be independent of the incidence angle. However, the SAIl image was classified over all of its angle variations and, although the images were corrected for incidence angle variation ahmg the range line, we can still expect some misclassiflcation, particularly near the areas of overlap. Among the training areas, we encountered some difficulty.' in identif~ng aspen stands because of their small sizes within the BOREAS modeling grid and their vicini~ to mixed stands. As a result, aspen class is labeled TA/MX to illustrate the aspen-dominated mixed stands.

RESULTS AND DISCUSSION
The map of forest types constructed fronl SAR data is shown in Figure 4. This result is obtained by using polarimetric data at P-band and the HH and [IV polarizations at L-and C-bands. Tile choice of the frequency and polarization channels for achieving the optimum classification results was made by changing the dimensionality of the classification, or, equivalently, reducing the number of elements in the covariance matrix of each pixel that are used for classification. Consequently, the optimum classification accuracy was obtained by excluding only L-and C-band VV polarizations. In this process, it was also found that the contrilmtion of P-band data was crucial in separating the classes. The reason for this combination is partly due to the calibration and radiometric inaccuracies at higher frequencies. In particular, the C-band W-polarized synoptic mosaic data suffered from banding in the image, and inaccuracies resulted from incidence angle correction. In fact, when a radar channel does not separate two image classes, it adds as a noise source to the classification and increases the classification probability" of error. The combination of polarizations and frequencies used to attain maximum separability differs from a similar technique applied on AIRSAR frame images over Alaskan boreal forest where the highest accuracy" was obtained by only L-band and C-band HV polarizations (ilignot et al., 1994). We believe the reason for this difference resides in the poor radiometric accuracy of the synoptic images 0t high frequencies in our case and in the P-band interference problem in the data used over the Alaska region.
Classification accuracy tor each class is determined by measuring the number of pixels correctly classified into the class divided by the total number of pixels in that class and is illustrated in the form of a confusion matrix. In assessing the total classification accuracy, we included open water and clear cut, though they are often separated with no difficulty within SAIl images. The con-  tribution of each frequency in the total classification was assessed qualitatively when the classifier results were examined during the dimensionality test. The results indicate that the HV polarizations contribute the most for forest-type mapping at all frequencies. As shown in Table 1, the HV channels at L-band and P-band°show the highest variability over the range of forest types because they are mainly related to the volume scattering within forest canopy and in turn sensitive to the forest biomass density. Furthermore, over forested areas, the HV backscatter is less sensitive to the incidence angle variation, and therefore the channels are less contaminated by the correction errors that may have remained over the image mosaic. The eopolarized backscatter is less variable over different stands; but, because the calibration of copolarized channels is usually better than that of cross-polarized channels, their role in separating classes is significant. For example, over low vegetation, clear cut, and open water, the HV-polarized backscatter is very low, and the copolarized backscatter signatures are the primary source for separating these classes. Table 2 shows the confusion matrix computed from the results of MLE and MAP classifiers over the training areas with 90% and 96% accuracies, respectively. For the mixed aspen and jack pine (JP/TA) class, pixels over the training area were classified with only 72% accuracy. The reason for this is the similar average copolarized backscatter values at all three bands. In general, for mixed stands, the choice of the training areas is poor compared with the monospecies homogeneous stands and, as a resuit, the mean backscatter returns for these sites are not very distinctive. Therefore, we expected poor accuracies over TA/MX sites because, over this region, most of the aspen stands are mixed with conifer trees. Jack pine and black spruce stands were classified with 100% and 99% accuracy, respectively. P-band and L-band HH polarizations are the main channels for separating these two classes. In jack pine stands, the trees are taller with less foliage and the ground surface is dry and smooth, which collectively contributes to high double-bounce return at HH polarization (Moghaddam and Saatehi, 1995). Tile black spruce stands, on the other hand, have shorter trees, more foliage, and a thick and wet moss layer and thus lower returns at HH polarization because of the absorption of the electromagnetic energy by the underlying moss layer.
To examine the ability of the classifier in separating coniferous and deciduous stands, we applied the classifier, without any changes in its current configuration, on an AIRSAR frame image acquired over the aspen tower site in Prince Albert National Park on the same date. The image covers the area south of Halkett Lake and north of dirt road Rt. 240 and is c e n t e r e d at the aspen tower site at almost 45 ° incidence angle. T h e area is covered mainly with aspen trees a n d with small scattered patches of balsam poplars (Populus balsamifera) that are similar in structure to aspen. The result of the classification is shown in Figure 5. F r o m a ~isual interpretation of the map, it appears that the classifier separates the aspen stands with no difficult. Over the tower site, the      The ground-truth data were taken from the field notes of TE-6 investigators (Sellers et al., 1995). WS is white spruce (Picea glauea), Lala is Larix laricina, Abba is Balsam Fir (Abies balsamea), Bepa is paper birch (Belula papyrifera). Numbers in ground-truth column indicate the percentage of each tree species based on the number of stems within the test plots. In vegetation cover and SAR maps, the ,mmbers indicate the percentage of image pixels of each stand classified in type of forest. classification accuracy reached 100%. This is one of the striking results of the SAR classification because, in general, the separation of coniferous and deciduous stands in boreal forests is considered one of the most challenging problems in any land-cover classification. This result also indicates that, over homogeneous stands, the structural information of the forest embedded in the SAR backseatter data becomes one of the key discriminants in the forest-type classification.
To analyze the accuracy of the SAR-derived cover map further, we compared the map with the field data and the existing land-cover map derived from infrared aerial photography. Table ;3 shows the tree species composition of 19 test sites within the modeling grid obtained from actual measurements for each site, the vegetation map, and the SAB map. The ground measurements were conducted during the intensive field campaigns in summer of 1994 and coincide with the time frame in which the SAR data were acquired. The species composition was measured on small plots within each stand and was not designed to address the species composition at the SAR pixel scale. The vegetation map is almost 10 years old and may be inaecurate because it is based on a visual interpretation of the aerial photography and does not include the changes that have occurred since then. However, we ineluded the map as an extra source for evaluating the accuracy of the SAR map. Moreover, the classifier was used to label each pixel by the dominant forest type and was never intended to estimate the species composition. Nevertheless, by performing this comparison, we are able to examine the general performance of the classifier and the capability of" SAR to identify species composition.
The SAR map was georeferenced and coregistered with the vegetation map with less than one pixel (30 m) accuracy. The center locations of the sites were identified on the images by using the GPS (ground positioning system) data. Stand compositions of 19 sites were computed over 5×5 pixels from SAR and vegetation maps. The results in Table 3 indicate that classifications of auxiliary sites and tower sites are in good agreement with tile field data and the vegetation map. Over 13 forest stands, errors in percentage of each species represented in the classification are less than 8%. The remaining six sites are mixed and contain species that are not included in the SAR classification. Over these sites, the errors in estimating species composition can increase to 20% with the exception of auxiliary site G413M, where the error exceeded 50%. These errors stem from several f:actors: 1) the spatial variability of species composition within the mixed stands is not compatible with the pixel sizes of the SAR map, 2) the location of the sites on the SAR map can be wrong owing to errors in the GPS measurements that may be larger than 100 m, and 3) the numt)er and size of plots used in the field measurements may not be adequate for the mixed stands. Furthermore, because a combination of tree geometlT, biomass, and surface conditions contributes to changes in SAR backscatter, the presence of several tree species within one SAR pixel will add to the confusion of the classifier in separating stands. These results suggest that the SAR map can be Next, we examine the accuracy of the SAR map over the entire modeling grid by computing the percentage of area covered by each forest type in the region. An area of approximately 25X35 km is taken from the middle of the SAR map, and the number of pixels of each forest type is counted and divided by the total number of pixels. In Table 4, the percentage of area covered by each type in the SAR map and in the vegetation-cover map are compared. The difference between the two maps represented by the percentage of change implies a combination of errors in both maps and changes in the land cover between the times of the two data takes. If the vegetation-cover map is considered accurate at the time of the SAR data take, then the difference can mean that 23% of the total area was classified inaccurately. Field observations during BOREAS campaigns showed that certain parts of the land cover have been altered. For example, some logged and burned areas have been forested, and some forest areas have been recently cut. Because there is no accurate information about land-cover types on a regional scale, the assessment of the accuracy of the SAR map can be difficult. Given the uncertainties in the vegetation-cover map, we expect that, on a regional scale, more than 77% of the total area can be classified accurately with SAt/ imagery.

Process Modeling Requirements
Land-cover maps can be used as one of the parametric inputs to ecosystem process models. The requirements for accuracy and spatial scale of the map depend on the ecosystem model and the application. For example, general circulation models (GCM) have incorporated 1 ° by 1 ° global land-cover classification maps (Sellers et al., 1994). Recently an AVHRR/NDVI (normalized difference vegetation index) based global land-cover map also has become available as an input to GCMs (DeFiles and Townshend, 1994). For modeling the net canopy assimi- lation in boreal or tropical forest, ecosystem models may require much finer resolution data over local or regional scales (Bonan, 1993). The BOREAS process models require five major land-cover types for the region. These are conifer wet, conifer dry, deciduous, mixed conifer and deciduous, and fen and disturbed. As an attempt to produce maps that can be readily used as input parameters to these models, we combined classes and modified the SAR and vegetation-cover maps to represent these five classes. Because pure deciduous and fen sites are rare over the modeling snbgrid mapped by SAR, we chose conifer wet, conifer dry, mixed deciduous and conifer, clearing/disturbed, and open water as typical cover types for the region. The new classes were tbrmed by grouping BS, BS/JP, and MX (mixed wet) into the conifer-wet class; JP into the conifer-dry class, and TA/MX and JP/TA into the mixed conifer-deciduous class. The clear cut and disturbed and the open water classes were not changed. The results are shown in Figure 6. By preserving the original pixel size (30 m), the new maps can be used in filture for the accurate estimation of land-use change due to environmental and anthropogenic forces. The modified SAR and vegetation-cover maps show similar patterns of land-cover types in the region. A comparison of the two images over a 25×35 km subarea is given in Table 5. Results indicate that the accuracy of the SAR image can improve when fewer classes are used. The dif-l>rence between the two maps has reduced to only 7.3% of the total area. With a reduction in the number of classes to functionally significant land-cover types, SAR data can provide maps with greater than 92% accuracy over the modeling grid.

SUMMARY AND CONCLUSIONS
This work summarizes the approach and the results of mapping ~brest types in the southern study area of the BOREAS project in the boreal forest of Canada by using SAR imagery. The images were collected by the JPL AIRSAR system and combined in a mosaic to cover the ecosystem process modeling subgrid. Eight classes were separated in the SAR image, and the classification accuracy was pertbrmed at several levels. Over 19 forest stands surveyed during the BOREAS field campaigns, the SAR map exhibited an accuracy of about 90%. The analysis showed that the map was also able to correctly predict tree species composition on the SAR pixel scale. At a larger scale, an area of 25X35 km from the SAR map was compared with a digital vegetation map based on infrared aerial photography, and more than 77% of the total area was classified aecurately. Finally, the number of classes were reduced to produce a map compatible with the requirements of the BOREAS land surface process models. The reduced map had five classes and, when compared with the vegetation map, showed similar land-cover patterns with greater than 92% accuracy over the total area. It is important to note that the classification accuracies per~brmed in this study were highly dependent on the aceuracy of the image calibration and impediments caused by errors due to incidence angle effects, aircraft motion compensation, and the image mosaic procedure. Furthermore, the results were obtained by using data from a single date. Multitemporal data can provide information about the seasonal and environmental states of the boreal forest and enhance the characteristics of the feature space for the classifier. Therefore, we believe that the accuracy obtained in this study is conservative and can be improved by incorporating multitemporal data and spaeeborne systems with better image fidelity. Some of the important results of the SAR classification were the separation of black spruce and jack pine stands and of coniferous and deciduous trees with close to 100% accuracy. These forest types are considered the dominant coniferous and deciduous stands covering large patches throughout the entire region of the boreal forest. The results also have a significant effect on modeling the canopy assimilation and biogeochemical processes for the region. Deciduous trees, because of their phenological, understory, and seasonal characteristics, represent different fimetional ~brms in ecosystem process models. Among conifers, jack pine and black sprnce trees are also treated differently in process models. Unlike the dry and sandy soils of jack pine stands, the soils of black spruce patches are often covered by a thick moss layer and are poorly drained; black spruce patches also have different characteristics due to the release of trace gases from the soil surface and canopy.