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

Change detection (CD) is an important task, which is evaluating changes between bi-temporal images. Hyperspectral images (HSIs), which are very spectrally informative, are excellent at identifying different change classes and catching minute variations. Although hyperspectral binary CD has received most of the attention recently, specific information about fine change classes has not been adequately provided. Current approaches, which depend on band selection and image difference, struggle to deal with the enormous dimensionality of data, lose out on nonlinear interactions between bands, and preserve spectral characteristics, all of which restrict their overall performance. This paper proposes a unique hybrid spectral-spatial feature extraction method for binary and multiple CD in bi-temporal HSIs. Our method reduces the feature space while effectively extracting meaningful spectral-spatial information. This is accomplished by incorporating a feature extraction module based on principle component analysis (PCA) into an optimized hybrid 3D-2D convolutional neural network (CNN). In multiclass and binary CD, the generated subspace outperforms baseline and advanced methods. Comprehensive experiments were conducted on two publicly available HSI-CD datasets: China wetland agriculture land (CH) and USA irrigated farmland (USA). These experiments thoroughly verify the effectiveness and efficiency of the proposed methodology, demonstrating its superiority with optimal multiclass CD performances of 97.81% and 97.73%, and binary CD performances of 97.84% and 97.79%, for the CH and USA datasets, respectively, with 20% training samples.