I. Introduction
Remote sensed images have been widely used for observing the earth’s surface, such as change detection [1], [2], [3], [4] and land cover classification [5], [6], [7], [8], [9]. Land cover classification aims to assign a semantic label to each pixel in an image scene. Owing to the quick development of remotely sensed platforms (such as QuickBird satellites, aerial planes, and drones), high spatial resolution remotely sensed (HSRRS) images can be easily obtained [10]. An HSRRS image can capture the ground details, such as shape, boundary, and direction for ground targets, which are very helpful for land cover classification [11], [12], [13], [14], [15]. From the perspective of applications, land cover classification with HSRRS images plays an important role in urban planning [16], target recognition [17], [18], and natural disaster assessment [19], [20].