Automatic Classification of Glaciers from Sentinel-2 Imagery Using A Novel Deep Learning Model

Abstract

The Sentinel-2 imagery provides accessible multispectral imagery, allowing better operation monitoring of glacier for climate change research, sea level rise and human life. Nevertheless, automatic glacial classification from Sentinel-2 is a challenging due to factors such as complex environment, different resolution bands and noisy or correlation in the spectral or spatial domain. In this paper, we propose an automatic glacier discrimination approach named MSSUnet to address several key research issues. First, a spatial-spectral module is used to adaptively learning the feature from different spectral band and neighboring pixels, which can better learn spatial-spectral features and reduce the impact of noise. Second, a band fusion method is applied to achieve fusion of different resolution bands in Sentinel-2 and reduce the interference of additional information. Furthermore, the proposed MSSUNet is compared with several existing neural networks on Sentinel-2 imagery to justify the advantage and improvement of the proposed approach. Experimental results show the improved performance of our proposed network over the other approaches.