Hyperspectral image change detection based on gated temporal attention network

With the advancement of hyperspectral image technology towards higher resolution, change detection (CD) faces dual challenges in computational efficiency and detection accuracy. This paper proposes HyGTN, a lightweight deep learning-based CD method using gated spectral-channel temporal attention mechanisms. The core innovations include: (1) Gated Spectral-Channel Attention (GSCA) module that efficiently extracts spectral-spatial features from singletemporal hyperspectral images; (2) Gated Spectral-Channel Temporal Attention Module (GSCTAM) that integrates bitemporal attention information with self-attention mechanisms to capture discriminative CD features. The gating mechanism significantly reduces computational complexity while maintaining high accuracy, enabling edge computing deployment. Experiments on Bay Area and Farmland datasets demonstrate that HyGTN outperforms state-of-the-art methods (e.g., BIT, SST-Former) in both computational efficiency and detection accuracy (OA and Kappa coefficients), showing robust performance especially with limited training samples.