Label-Efficient Hyperspectral Image Change Detection via Multiscale Self-Distillation

Abstract

Methods based on deep learning have achieved remarkable success in hyperspectral image change detection (HSI-CD). However, the high dependence on massive labeled data and the computational complexity of state-of-the-art models pose significant challenges, particularly in scenarios with extreme label scarcity. To address these issues, a multiscale self-distillation change detection (MSDistillCD) framework is proposed. In the proposed framework, two core modules are designed to extract robust features and enforce consistency under limited supervision. First, a multiscale feature extraction module (MSFEM) constructs subpatches aligned at the center from a fixed maximum patch and generates inputs with explicit difference information. A shared convolutional neural network (CNN) backbone with dual global aggregation is then employed to capture embeddings that remain robust across scales. Subsequently, a self-distillation classification module (SDCM) is utilized to facilitate knowledge transfer within the network. Unlike conventional distillation settings, the smallest-scale branch is treated as the primary branch and guides the larger-scale auxiliary branches through temperature-scaled self-distillation, providing additional regularization during training while allowing efficient inference with a single branch. Experimental results on benchmark HSI datasets demonstrate that MSDistillCD achieves the best overall accuracy (OA), $\kappa \times 100$ , and $F1$ across all three datasets, while maintaining robust performance even with extremely limited training samples (e.g., 0.1%). The source code of the proposed MSDistillCD will be released at https://github.com/zhangyuan1698-sketch/MSDistillCD