NORTHERN FOREST CLASSIFICATION USING TEMPORAL MULTIFREQUENCY AND MULTIPOLARIMETRIC SAR IMAGES

Characterizing a forest ecosystem dynamics for global change studies r equires updated knowledge in terms of species composition, carbon stor age, and biophysical functioning. Often, significant areas of forest a re obscured by clouds or are under reduced solar illumination conditio ns, which limit acquisition of optical satellite data. Synthetic apert ure radar (SAR) images, however, can be acquired under these condition s. Several SAR image data sets were acquired over the Northern Experim ental Forest near Howland, Maine as part of the Forest Ecosystem Dynam ics-Multisensor Aircraft Campaign. A SAR data processing and analysis sequence, from calibration through classification, is described. The u sefulness of multifrequency temporal polarimetric SAR image data for i dentifying ecosystem classes is discussed. Our results show that with principal component analysis of temporal data sets (winter and late su mmer) SAR images can be classified into general forest categories such as softwood, hardwood, regeneration, and clearing with better than 80 % accuracy. Other nonforest classes such as bogs, wetlands, grass, and water were also accurately classified. Classifications from single da te images suffered in accuracy. The winter image had significant confu sion of softwoods and hardwoods with a strong tendency to overestimate hardwoods. Modeling results suggest that increased double-bounce scat tering of the radar beam from conifer stands because of lowered dielec tric constant of frozen needles and branches was the contributing fact or for the misclassifications.