Deep attention driven spatial spectral framework for dimensionality reduction and land cover classification

Land cover classification is crucial in remote sensing for environmental monitoring, urban planning, and resource management. Traditional techniques often struggle with complex landscapes, leading to misclassification and lower dependability. Deep learning (DL)-based algorithms offer a more accurate and resilient solution. This paper presents an effective DL-based system for improving land cover categorization, integrating temporal and spatial branches and leveraging spectral-temporal fusion. The method achieves superior classification accuracy with minimal misclassification. Performance evaluations show the system can differentiate various land cover types with an Overall Accuracy (OA) of 99.74%, Average Accuracy (AA) of 99.58%, and a Kappa Coefficient (KC) of 0.9965 on the Pavia University (PU) dataset and 97.19% OA, 96.68% AA, and KC of 0.9687 on the Kennedy Space Center (KSC) dataset. These results establish the framework as a reliable classification system for hyperspectral imaging, paving the way for further refinements and broader applicability in real-world remote sensing tasks.