Spectral-Spatial Feature Extraction for Multi-Change Detection in Bi-Temporal Hyperspectral Images

Change detection (CD) is a critical task that involves assessing alterations between bi-temporal images. Hyperspectral images (HSIs), renowned for their extensive spectral information, excel in capturing subtle changes and distinguishing between various change classes. While recent research has predominantly focused on hyperspectral binary CD, it falls short in providing detailed information about fine change classes. Existing methods, which rely on image difference and band selection, encounter challenges in preserving spectral characteristics, neglecting nonlinear relationships among bands, and grappling with the high dimensionality of data, thus limiting their overall performance. This paper proposes a novel hybrid spectral-spatial feature extraction technique for detecting multiple changes in bi-temporal HSIs. Our approach excels at extracting significant spectral-spatial information while reducing the feature space. This is achieved by integrating a principal component analysis (PCA) based feature extraction module into a well-tuned hybrid 3D-2D convolutional neural network (CNN). The resulting subspace performs superior in multiclass CD compared to baseline and advanced approaches. Comprehensive experiments were undertaken on two publicly available HSI-CD datasets: China wetland agriculture land (CH) and USA irrigated farmland. These experiments meticulously validate the effectiveness and efficiency of the proposed methodology. The outcomes unequivocally showcase the superiority of the proposed approach, attaining the highest multi-class change detection performances, notably 98.07% and 97.89%, for the USA and CH datasets, respectively.