A co-training, mutual learning approach towards mapping snow cover from multi-temporal high-spatial resolution satellite imagery

High-spatial and -temporal resolution snow cover maps for mountain areas are needed for hydrological applications and snow hazard monitoring. The Chinese GF-1 satellite is potential to provide such information with a spatial resolution of 8 m and a revisit of 4 days. The main challenge for the extraction of multi-temporal snow cover from high-spatial resolution images is that the observed spectral signature of snow and snow-free areas is non-stationary in both spatial and temporal domains. As a result, successful extraction requires adequate labelled samples for each image, which is difficult to be achieved. To solve this problem, a semi-supervised multi-temporal classification method for snow cover extraction (MSCE) is proposed. This method extends the co-training based algorithms from single image classification to multi-temporal ones. Multi-temporal images in MSCE are treated as different descriptions of the same land surface, and consequently, each pixel has multiple sets of features. Independent classifiers are trained on each feature set using a few labelled samples, and then, they are iteratively re-trained in a mutual learning way using a great number of unlabelled samples. The main principle behind MSCE is that the multi-temporal difference of land surface in spectral space can be the source of mutual learning inspired by the co-training paradigm, providing a new strategy to deal with multi-temporal image classification. The experimental findings of multi-temporal GF-1 images confirm the effectiveness of the proposed method.