Discriminative Features Based on Two Layers Sparse Learning for Glacier Area Classification Using SAR Intensity Imagery

Accurate and instant information about changes of snow and glaciers covered areas plays a vital role in hydrological and climatological research and implications. Among all the observation methods, spaceborne remote sensing has a great advantage in monitoring the glaciers located in cold high-altitude regions and inaccessible areas on a large scale. Unlike optical sensors, the synthetic aperture radar (SAR) sensor can obtain images with low limitations in terms of weather phenomena and illumination as some glaciers frequently located in cloudy regions. In this paper, we propose a multiclasses classification method for large area glacier using spaceborne single-polarimetric SAR intensity image. The proposed method takes advantage of the discrimination ability of sparse representations of features, based on which a linear classifier called supervised neighborhood embedding is constructed. Finally, we develop a gradient descent method to alternatively update the dictionary and projection matrix. Two study areas are chosen to represent the discriminative characteristics of glaciers. In the Taku glacier in Alaska, compared to the state-of-the-art methods, our proposed method achieved a suitable performance with the overall classification accuracy of 90.34%, and especially for bare ice of 91.38%. In the Baltoro glacier in Karakoram characterized by high-relief topography and thick debris cover, the overall accuracy of 72.63% and debris accuracy of 90.14% are obtained.