GCT: Graph-Based Classifier Transfer Model for Few-Shot Remote Sensing Image Scene Classification

Recently few-shot remote sensing image scene classification (FSRSISC) is increasingly drawing attention for its powerful capability of acquiring better classification performance with insufficient annotated samples. One of the challenging problems of FSRSISC is how to obtain more effectual features and information from limited labeled images. To address this challenge, this paper proposes a graph-based classifier transfer (GCT) model for FSRSISC. In our proposed GCT model, a classifier transfer module based on the mechanism of analogy learning is developed to transfer the classifier parameters of the base classes into the novel classes, which makes full use of the relationship between the two types of classes. Then, a graph-based feature matching module is developed to improve the representation capability of images, which enables to effectively solve the inaccurate relational reasoning caused by the uncertainty of image representation during the analogy process. We extensively perform experiments on three common FSRSISC datasets and the results demonstrate that the proposed GCT model enables to obtain superior performance than other FSRSISC methods.