SD-Net: Boosting Few-Shot Remote Sensing Scene Classification Via Semantic Decomposing

Few-Shot Remote Sensing Scene Classification (FSRSSC) aims to develop a model capable of recognizing the novel classes with extremely limited labeled examples. Fully exploring the semantic information within images is critical for FRSSC as remote sensing images have inherent characteristics of complex backgrounds, large intra-class variations, and small inter-class differences. However, existing FSRSSC methods usually learn pooled global features or local features that do not capture elaborated semantic representations for remote sensing images. Therefore, in this paper, a novel Semantic Decomposing Network is proposed, called SD-Net. It decomposes the local feature representations extracted by a backbone network, obtaining a set of semantic feature vectors which have powerful representation ability of remote sensing scenes. Furthermore, a novel supervised training strategy for the SD-Net is proposed to make it adapt well to the novel classes. The experimental results on the WHU-RS19 and UC Merced datasets show that our proposed method obtains competitive performance comparing with current state-of-the-art FSRSSC methods.