Glacier classification from Sentinel-2 imagery using spatial-spectral attention convolutional model
Shuai Yan, Linlin Xu, Guojiang Yu, Longshan Yang, Wenju Yun, Dehai Zhu, Sijing Ye, Xiaochuang Yao
Abstract
Glaciers are critical components of the cryosphere and most sources of water. The Sentinel-2 imagery provides an efficient and cost-effective way for the production of inventories of glaciers. Efficient conventional glacial methods can overcome the drawbacks of manual interpretation by generating glacier maps in a fast and accurate manner. However, these methods are difficult to obtain discriminative spatial-spectral features, so that it is difficult capture the subtle differences between glaciers and surrounding material on Sentinel-2 imagery. In this paper, a novel spatial-spectral attention neural network approach, which integrates a novel spatial-spectral attention module with UNet architecture, is designed for better addressing these challenges in glacier mapping. The proposed approach builds an interdependencies between spatial-spectral domains via adaptively recalibrating the weighs of spatial-spectral features for each pixel, which is benefit to highlight the difference between glaciers and surroundings. The proposed module is effectively integrated in UNet architecture, and the optimal model overcomes the noise effect on Sentinel-2 imagery, which outperforms the original UNet approach about 1 percent. The result indicates that a consistent improvement of glacier feature extraction. Furthermore, the proposed approach is compared with several existing glacier mapping algorithm, and results demonstrate that the proposed approach has stronger generalization ability and can obtain satisfactory accuracy for glacier mapping with test accuracy about 98.61%.