An application of ANN for mountainous snow cover fraction mapping with MODIS and ancillary topographic data

Snow can strongly influence the surface radiation balance, energy exchange and hydrological processes. Accurate fractional snow cover data plays an important role in many applications. As existing fractional snow cover (FSC) products (i.e. MOD10) are not accurate enough in mountainous areas, we developed a three-layers feed-forward artificial neural networks (ANN) for mountainous FSC mapping, which is trained with back-propagation to learn the relationship between FSC and eight different schemes of input information. In the study, an image from Landsat ETM+ and corresponding MODIS data products are chosen to train, validate and test the proposed method at the upstream of Heihe River Basin. The results showed that the ANN-based methods have higher R, lower RMSE and more accurate total snow cover area. Particularly, the Exp.8 combined all input information together achieved the best performance.