MONITORING LAND-SURFACE SNOW CONDITIONS FROM SSM I DATA USING AN ARTIFICIAL NEURAL-NETWORK CLASSIFIER/

Previously developed Special Sensor Microwave/Imager (SSRM/I) snow cla ssification algorithms have limitations and do not work properly for t errain where forests overlie snow cover, In this study, we applied uns upervised cluster analysis to separate SSM/I brightness temperature (T -B) observations into groups. Six desired snow conditions were identif ied from the clusters; both sparse- and medium-vegetated region scenes were assessed. Typical SSM/I T-B signatures for each snow condition w ere determined by calculating the mean T-B value of observations for e ach channel in the corresponding cluster. A single-hidden-layer artifi cial neural network (ANN) classifier was designed to learn the SSR/I T -B signatures, An error backpropagation training algorithm was applied to train the ANN. After training, a winner-takes-all method was used to determine the snow condition. Results showed that the ANN classifie r was able to outline not only the snow extent but also the geographic al distribution of snow conditions. This study confirms the potential of using cluster means in ANN supervised learning, and suggests a nonl inear retrieval method for inferring land-surface snow conditions from SSM/I data over varied terrain.