Spectral attention and visionmamba difference guided network for hyperspectral image change detection

Abstract

Transformer-based and CNN-based approaches have achieved significant success in hyperspectral image change detection (HSI-CD) in recent times. However, traditional CNN-based methods treat all extracted features equally and primarily focus on local features, which to some extent limits detection speed and accuracy. Although transformers can provide a global receptive field, they suffer from high computational complexity and pay insufficient attention to the correlation of change information across different spectral bands. In contrast, the Mamba architecture based on State Space Models (SSM) combines efficient long-sequence modeling with linear computational costs, demonstrating great potential in feature detection for low-dimensional scenarios.Building on this, this paper proposes a lightweight spectral attention and Visionmamba difference-guided network(SAVDGN) for hyperspectral image change detection. We integrate CNNs with Mamba and design the network with the aim of highlighting change information. Its goal is to extract change information from both spatial and spectral dimensions, generating highly discriminative differences. The network utilizes CNNs to hierarchically extract rich spatial features from bi-temporal images, while leveraging spectral attention and Visionmamba to generate spectral differences between the two images at different network layers. These spectral differences not only guide feature extraction in the next stage, but also support the final change decision. Compared with state-of-the-art HSI classification methods, experimental results demonstrate that SAVDGN achieves significant classification accuracy on three public HSI datasets and a superior reduction in model parameters and floating-point operations (FLOPs).