Abstract
Hyperspectral imaging has emerged as a powerful tool for remote sensing applications, offering rich spectral information across a broad electromagnetic spectrum. However, the high dimensionality of hyperspectral data poses significant challenges in analysis and interpretation. In this study, we propose a novel approach for hyperspectral image processing, focusing on dimensionality reduction, albedo recovery, and subsequent classification. Our method begins with a grouping strategy based on the electromagnetic spectrum that considers the images' physical properties, facilitating the segmentation of hyperspectral data into meaningful spectral bands. This grouping reduces the dimensionality of the data and preserves crucial spectral information. Subsequently, we integrate autoencoders to incorporate non-linear transformations in the feature extraction phase, thereby improving the model's capacity to learn intricate patterns within the data. A key goal of our methodology is to effectively embed spatial information into the representation. Albedo recovery is employed aimed at improving spatial resolution while retaining spectral fidelity. By leveraging the reduced-dimensional representation obtained through grouping and autoencoders, we reconstruct the hyperspectral image with enhanced spatial details, thereby facilitating more accurate interpretation and analysis. To assess the performance of the proposed approach, we perform experiments using three standard hyperspectral datasets: Indian Pines, University of Pavia, and Salinas.