An Explainable Network for Adaptive Change Detection in Unannotated Hyperspectral Images

Abstract

Multitemporal hyperspectral image change detection (HSI-CD) reveals the dynamic evolution of ground objects from high-dimensional data, serving as a core technology for surface monitoring. However, HSI-CD still faces several challenges, including data redundancy and mixed pixels induced by temporal–spectral–spatial heterogeneity, multigranular patterns arising from the diverse shapes and sizes of ground changes, and the restricted applicability of supervised methods due to a lack of manual annotations. To address these issues, we focus on physical explainability and adaptive optimization, thus proposing a novel EXplainable and Adaptive CD network (EXANet). For physical explainability, EXANet automatically captures ground changes with diverse scales, shapes, and spatial morphologies via multigranular change detection, while utilizing a subpixel feature extraction to excavate subpixel change information and alleviate the influence of mixed pixels. It constructs a low-rank latent-space representation to ensure that temporal–spectral–spatial features are extracted and redundant data are suppressed. For adaptive optimization, EXANet incorporates a self-generated weak-label guidance to reduce the dependence on manual annotations. Through reinforced fine-grained feature fusion and learning, it ensures that the change-discriminative ability of different features is adaptively evaluated and their fusion weights are dynamically adjusted. Extensive experiments on multiple real-world datasets demonstrate that EXANet outperforms state-of-the-art methods in detection accuracy and extracts multigranular change features.