Semisupervised Few-Shot Remote Sensing Image Classification Based on KNN Distance Entropy

Abstract

In recent years, remote sensing image processing based on deep learning has been widely applied in many scenes, but the involved deep learning technology requires large-scale labeled data, which has been a practical problem in the remote sensing field. In this study, we proposed a novel data information quality assessment method, called K-Nearest Neighbor (KNN) distance entropy, to screen the remote sensing images. The evaluation metric was used to assess unlabeled data and assign the pseudo label, which further constitutes the proposed semi-supervised few-shot classification method in this paper. The meta-task setting was adopted to verify the validity and stability of experimental results. Specifically, the KNN distance entropy metric can be used to distinguish the samples in core set or boundary set. Experimental results show that the core set samples are more suitable under the few-shot condition, for instance, the meta-task average accuracy trained by the core set samples outperforms that by boundary samples about 18 percent in the case of 45-ways and 5-shot. The proposed semi-supervised few-shot method based on KNN distance entropy achieves significant improvement under different experimental conditions. The visualization of feature distribution of screened data is shown to provide intuitive interpretation. This paper lays a meaningful foundation for screening and evaluating remote sensing images under few-shot condition, and inspires the data-efficient few-shot learning based on high-quality data in the remote sensing field.