A recurrent neural network classifier for improved retrievals of areal extent of snow cover

Accurate detection of areal extent of snow in mountainous regions is important. Areal extent of snow is a useful climatic indicator. Moreover, snow melt is a major source of water supply for many arid regions (e.g., western United States, Morocco) and affects regional ecosystems. Unfortunately, accurate satellite retrievals of areal extent of snow have been difficult to achieve. Two approaches to effectively and accurately detect clear land, cloud, and areal extent of snow in satellite data are developed. A feed-forward neural network (FFNN) is used to classify individual images, and a recurrent NN is used to classify sequences of images. The continuous outputs of the NN, combined with a linear mixing model, provide support for mixed-pixel classification. Validation with independent in situ data confirms the classification accuracy (94% for feed-forward NN, 97% for recurrent NN). The combination of rapid temporal sampling (e.g., GOES) and a recurrent NN classifier is recommended (relative to an isolated scene (e.g., AVHRR) and a feed-forward NN classifier).