I. Introduction
For the improvement of earth observation system [1], remote sensing (RS) images [2], [3], especially high spatial resolution (HSR) images [4], [5], have been widely applied in various fields, such as land-use planning, target detection, and change detection. The scene classification of HSR images is an important research subject in the RS community, and it aims to automatically assign a unique label for a given image according to semantic contents [6].