Improving Mountainous Snow Cover Fraction Mapping via Artificial Neural Networks Combined With MODIS and Ancillary Topographic Data

A multilayer feedforward artificial neural network (ANN) is developed for mountainous fractional snow cover (FSC) mapping. This is trained with back propagation to learn the relationship between FSC and Moderate Resolution Imaging Spectroradiometer (MODIS) products (reflectance at seven bands, normalized difference snow index, land surface temperature (LST), and FSC) and elevation. In this paper, images from Landsat Enhanced Thematic Mapper Plus (ETM+) and MODIS products from three periods are chosen to test and validate the proposed method at the Heihe River Basin. Three binary snow cover maps derived from Landsat ETM+ images are used to calculate FSC. Two of these maps are first used to train, calibrate, and test the ANN. The other independent image is used to test the generalization ability of network. Results show that the ANN can easily incorporate auxiliary information to improve the accuracy of snow cover mapping effectively. It is also capable of mapping snow cover fraction in a complicated mountainous area with considerable generalization. For the nonindependent test set, the performance evaluation results show that the improvements of ANN-based methods are apparent compared with MODIS FSC products (higher correlation coefficient, lower root-mean-square error, and more accurate total snow cover area). For the temporal/temporal-spatial independent test set, ANN-based methods perform slightly worse than the nonindependent test set, but the accuracy of the ANN methods still shows some improvement. Elevation, LST, and FSC play more important roles in the training process of the ANN. Overall, experiment 8, which integrated all input information, is approved the best in all test sets.