I. Introduction
With the number of remote sensing images (RSIs) increasing extensively and rapidly, deep learning semantic segmentation technologies have undergone significant advancement, especially in remote sensing surface monitoring [2]. Semantic segmentation has been a crucial technique in satellite image processing [3], aiming at pixel-level classification of various semantic targets in an image. It has proved to be highly effective in applications such as land-cover classification [4], [5], road extraction [6], building detection [7], and farmland segmentation [8] using RSIs.