Spectral-Spatial Subspace Detection for Improved Change Detection in Hyperspectral Images

HyperSpectral Images Change Detection (HSI-CD) plays a pivotal role in precise earth observation through remote sensing, leveraging abundant spectral and spatial information. However, prevailing methods, relying on image difference and band selection, encounter challenges in preserving spectral characteristics, neglecting nonlinear relationships among bands, and addressing high data dimensionality, thus limiting their overall efficacy. In this paper, we introduce S2HNet, a hybrid spectral-spatial network crafted to tackle end-to-end HSI-CD challenges. Our proposed S2HNet excels in efficiently extracting significant spectral-spatial information within a reduced feature space, achieved by integrating a principal component analysis (PCA) based feature subspace reduction module into a hybrid 3D-2D convolutional neural network (CNN). Consequently, the generated subspace demonstrates superior change detection performance compared to related baseline and advanced approaches. Thorough experiments were run on two publicly accessible baseline HSICD datasets: China farmland (CH) and USA irrigated farmland (USA), along with one synthesized dataset, Indian Pines (IP). These experiments rigorously validate the effectiveness and efficiency of the proposed methodology. The results demonstrate the superiority of the proposed approach, achieving the highest change detection performances, specifically 97.51% and 96.57% for CH and USA datasets, respectively, and 99.04% for our synthesized IP dataset.