GASSM: Global attention and state space model based end-to-end hyperspectral change detection

Abstract

As an essential task to identify anomalies and monitor changes over time, change detection enables detailed earth observation in remote sensing. By combining both the rich spectral information and spatial image, hyperspectral images (HSI) have offered unique and significant advantages for change detection. However, traditional hyperspectral change detection (HCD) methods, predominantly based on convolutional neural networks (CNNs), struggle with capturing long-range spatial-spectral dependencies due to their limited receptive fields. Whilst transformers based HCD methods are capable of modeling such dependencies, they often suffer from quadratic growth of the computational complexity. Considering the unique capabilities in offering robust long-range sequence modeling yet with linear computational complexity, the emerging Mamba model has provided a promising alternative. Accordingly, we propose a novel approach that integrates the global attention (GA) and state space model (SSM) to form our GASSM network for HCD. The SSM based Mamba block has been introduced to model global spatial-spectral features, followed by a fully connected layer to perform binary classification of detected changes. To the best of our knowledge, this is the first to explore using the Mamba and SSM for HCD. Comprehensive experiments on two publicly available datasets, compared with eight state-of-the-art benchmarks, have validated the efficacy and efficiency of our GASSM model, demonstrating its superiority of high accuracy and stability in HCD.