MAPPING OF FOREST TYPES IN ALASKAN BOREAL FORESTS USING SAR IMAGERY

Mapping of forest types in the Tanana river floodplain, interior Alaska, is performed using a maximum-a-posteriori Bayesian classifier applied on SAR data acquired by the NASA/JPL three-frequency polarimetric AIRSAR system on several dates. Five vegetation types are separated, dominated by 1) white spruce, 2) balsam poplar, 3) black spruce, 4) alder/willow shrubs, and 5) bog/fen/nonforest vegetation. Open water of rivers and lakes is also separated. Accuracy of forest classification is investigated as a function of frequency and polarization of the radar, as well as the forest seasonal state, which includes winter/frozen, winter/thawed, spring/flooded, spring/unflooded, and summer/dry conditions. Classifications indicate that C-band is a more useful frequency for separating forest types than L- or P-bands, and HV polarization is the most useful polarization at all frequencies. The highest classification accuracy, with 90 percent of forest pixels classified correctly, is obtained by combining L-band HV and C-band HV data acquired in spring as seasonal river flooding recedes and before deciduous tree species have leaves. In 17 forest stands for which actual percentages of each tree species are known, the same radar data are capable of predicting tree species composition with less than 10 percent error. For the same combination of observation channels, classification accuracy is 79 percent in spring on a day of intense river flooding, and 62 percent on a dry summer day with leaves on deciduous trees. In winter, using 4-look SAR data instead of 16-look, classification accuracy is 55 percent on a frozen day, and 76 percent on a thawed day. White spruce and balsam poplar stands are best separated in thawed conditions when balsam poplar trees have no leaves. From our classification, we predict that current and future spaceborne SAR systems will have limited mapping capabilities when used alone. Yet, RADARSAT combined with J-ERS-1 and ERS-1 could resolve forest types with 80 percent accuracy, separate nonforest areas resulting from commercial logging or forest wildfire, and map river edges. For comparison, a combination of green, red, and near-infrared radiance data acquired by SPOT-2 on a dry summer day yields a classification accuracy of 83 percent for the same forest stands, with limited success in distinguishing among deciduous forest types and among coniferous forest types.