Snow classification from SSM/I data over varied terrain using an artificial neural network classifier

The brightness temperatures (Tb's) observed by the Special Sensor Microwave/Imager (SSM/I) radiometer are sensitive to the changes in land surface snow conditions. Previously developed SSM/I snow classification algorithms have limitations and do not work properly for terrain where forests overlay snow cover. In this study, we applied unsupervised cluster analysis to define 6 snow classes in Tb observations, assessing both sparse- and medium-vegetated region classes. Typical SSM/I Tb signature, in terms of cluster means, of each snow class was determined by calculating the mean Tb's of the corresponding cluster. A single-hidden-layer backpropagation (backprop) artificial neural network (ANN) classifier was designed to learn the 6 Tb patterns. Classification performance, in terms of error rate (%), was as small as 2.4%. This study confirms the potential of using cluster means in ANN supervised learning, and suggests a nonlinear retrieval method towards making the inferences of snow classes from SSM/I data over varied terrain operational. Improvement is expected by identifying more SSM/I Tb signatures of different land surface types to train the ANN classifier.