S2TMN: A Spectral–Spatial–Temporal Mamba Network for Hyperspectral Image Change Detection

Abstract

Hyperspectral image (HSI) change detection (CD) aims to identify land-cover changes from bitemporal hyperspectral observations by jointly exploiting spectral and spatial information. Existing methods mainly rely on convolutional neural networks or Transformer architectures. However, CNN-based methods are limited in capturing long-range dependencies due to restricted receptive fields, while Transformer-based methods suffer from high computational complexity when processing high-dimensional hyperspectral data. Moreover, insufficient temporal interaction modeling and lack of change-aware dependency learning further limit the discriminative capability of existing approaches for subtle change regions. To address these issues, this article proposes a Spectral–Spatial–Temporal Mamba Network (S2TMN) for HSI-CD. Specifically, a state space modeling framework is introduced into bitemporal hyperspectral analysis to efficiently model long-range spectral–spatial dependencies and discriminative temporal interactions with linear computational complexity. A differential-guided Spectral–Spatial Feature Mamba) module is designed to jointly capture global spatial structures and latent spectral dependencies through multidirectional spatial scanning and bidirectional spectral embedding sequence modeling. Meanwhile, differential guidance gates are incorporated to adaptively regulate selective state propagation during spectral–spatial dependency learning. Furthermore, a GRU-style Gated Spectral–Spatial Temporal module is developed to adaptively fuse bitemporal representations and enhance discriminative temporal change features for accurate change prediction. Extensive experiments conducted on three benchmark HSI-CD datasets demonstrate the effectiveness and superiority of the proposed S2TMN over several state-of-the-art methods in both quantitative and qualitative evaluations.