Data fusion approach to improve forest cover classification using SAR imagery.

摘要

The interferometry coherence, backscatter, and texture information from JERS-1 synthetic aperture radar (SAR) data were investigated for discriminating general forest types (i.e., hardwood, mixed, and softwood) in the northeastern United States. The JERS-1 SAR data was then fused with Landsat 7 Enhanced Thematic Mapper Plus (ETM+) imagery to improve forest cover classification using artificial neural networks (ANNs) approaches. The study used two ANN classifiers, multi-layer perceptron network (MLP) and learning vector quantizer (LVQ), to apply the data fusion both in pixel level and decision level. Conventional statistical classifier maximum likelihood (ML) classifier and ANN classifiers were also applied to the individual imagery for comparison.; Statistical analysis showed that the combination of interferometry coherence, backscatter, and texture information of JERS-1 SAR data (called Enhanced JERS-1 (EJERS-1) imagery) was helpful for distinguishing hardwood, mixed, and softwood classes in the research area. For the individual classification of EJERS-1 and ETM+ imagery, ANN classifiers (i.e., MLP and LVQ) had higher overall accuracies than that of statistical classifier (i.e., ML) for both imagery, and the two ANN classifiers had similar overall accuracies. For the different level fusion, decision level fusion performed better than the pixel level fusion with 4% increase of overall accuracy.