Band-Selected Spectral-Spatial Transformer With Temporal Adaptive Modulation for Hyperspectral Image Change Detection

Abstract

Deep learning-based hyperspectral image (HSI) change detection (HSI-CD) methods have achieved promising accuracy. However, HSIs contain hundreds of highly correlated bands, while existing models often encode all bands with equal priority, resulting in redundant spectral computation. At the same time, bi-temporal feature fusion is commonly performed by fixed operations, such as difference or concatenation, which lack sample-adaptive fusion capability, whereas temporal cross-attention introduces considerable computational overhead and lacks an explicit change-sensitive inductive bias. These issues make it difficult to balance explicit change modeling and inference efficiency. To address these challenges, we propose B3STFormer, a band-selected spectral–spatial transformer with temporal adaptive modulation (TAM) for HSI-CD. Specifically, a spectral-aware discriminative band selection (SDBS) strategy is first designed as an offline spectral reduction step. SDBS selects a unified discriminative band subset for both temporal images by jointly considering Fisher separability of training-sample temporal differences, band-correlation structure, and diverse spectral grouping. The selected bi-temporal HSIs are then encoded by a shared spectral–spatial transformer to extract spectral–spatial features. Furthermore, a TAM module is introduced to combine temporal difference features and bi-temporal context features, and to modulate the fused representation using sample-dependent scale, shift, and gate parameters. Experiments on three public HSI-CD datasets demonstrate that B3STFormer achieves competitive CD performance with few parameters and low computational costs, yielding a favorable accuracy–efficiency tradeoff.