A semi-supervised multi-source remote sensing image classification network based on adaptive pseudo-label generation

Abstract

Earth observation technology leveraging remote sensing (RS) imagery serves as a valuable non-contact detection method with broad applications in classification research. Hyperspectral (HS) image classification, while effective in various domains, faces challenges due to the unique characteristics of HS data. Fusion diverse RS data sources can mitigate redundancy and enhance classification efficiency. However, many deep learning approaches for multi-source RS classification rely heavily on abundant labeled data, which can be time-consuming and often impractical. To address the limitations in feature extraction and classification accuracy stemming from the scarcity of labeled multi-source RS image samples in complex scenes, we propose a novel semi-supervised multi-source RS image classification network based on adaptive pseudo-label generation (S2CNet-APG). This framework incorporates attention modules that effectively embed active RS features into HS features, enhancing performance through squeezing and excitation (SE) driven attention mechanisms. Our semi-supervised learning approach employs adaptive thresholds to manage the quantity of pseudo-labels derived from unlabeled samples, while maintaining the spatial consistency of the information to ensure quality. This dual strategy effectively balances the quantity and quality of pseudo-labels, enabling accurate classification with limited labeled samples and transitioning multi-source RS image classification from a supervised to a semi-supervised paradigm. We conducted extensive experiments on three real-world multi-source RS datasets, achieving superior results that validate the efficacy of the proposed method.