Abstract: The classification and extraction of ground features based on remote sensing images is the technical basis of building a digital battlefield environment. However, it is faced with challenges such as changeable scenes, diverse categories and noise interference. Traditional deep learning model has complex structure and large amount of computation, which is difficult to meet the requirements of real-time information processing in low-performance, low-power edge computing environment. In this paper, a lightweight and complex remote sensing image fine classification method based on self knowledge distillation is proposed. The method is based on a one-step self knowledge distillation framework to improve the performance of the lightweight model by transferring knowledge from deeper portion of the networks to shallow ones. At the same time, pyramid pooling is used to fuse feature information of different scales,which significantly improves the utilization rate of global context information. This method solves the problem of low precision of remote sensing scene fine classification due to less parameters and low complexity of lightweight classification model. The proposed method achieved state-of-the-art performance and yielded a new record of pixel accuracy 80.76 % on Vaihingen and other public remote sensing datasets.