Target data guided few-shot remote sensing scene classification in reproducing Hilbert kernel space

Chunyu Du, Baodi Liu, Yanjiang Wang

Abstract: Recently, developing remote sensing scene classification (RSSC) based on few-shot has become an important issue in remote sensing image analysis. However, the inherent complexity of remote sensing images renders the embedded features to appear nonlinear. Simultaneously, limited training samples present challenges in capturing complex patterns and changes in the data, which results in a negative impact on transferring to new categories. Here, a few-shot RSSC (FSRSSC) model guided by target data within the reproduced Hilbert kernel space is proposed. Specifically, for tackling the nonlinear features of remote sensing images, the indistinguishable features in low-dimensional space are remapped into the reproduced Hilbert kernel space to make them more linearly separable. Additionally, for the problem that the model has difficulty transferring to novel categories, the information of the target data is further utilised to assist in directing the classifier’s learning, which facilitates the classifier to be better extended to unseen classes. Extensive experiments on two benchmarks reveal the superiority of the proposed approach compared with the advanced approaches.