Enhanced Hyperspectral Change Detection through Semantic, Spectral, and Spatial Analysis

Hyperspectral imaging offers a detailed and accurate representation of the Earth’s surface, enabling the detection of subtle changes in spectral signatures. Semantic segmentation is a method that aims at assigning of a label to each partition of the Hyperspectral image scene, whose pixels share similar or even identical spectral properties. This allows more understanding and characterization of specific elements of the scene, and associate them with pre-established higher-level labels. This study proposes a novel semantic spectral-spatial hyperspectral change detection approach, integrating spectral, spatial, and semantic information to identify changes in land cover, monitor natural disasters, and manage environmental resources. The incorporation of semantic knowledge reduces false alarms and enhances change detection accuracy. Spatial information is considered to detect changes in land cover and land use effectively. Hyperspectral unmixing is employed to extract end members of materials from the hypercube, and deep learning networks (Siamese Transformers or Resnet) generate classification maps for each bitemporal dataset T1 and T2. The resulting difference feature set is processed by a convolutional network to obtain the semantic-change map based on the spectral footprint. The proposed framework significantly enhances the accuracy and efficiency of hyperspectral change detection, holding promise for real-world remote sensing and environmental monitoring applications.