Advancing blind hyperspectral unmixing in remote sensing: comparing deep-inspired subspace learning methods

Hyperspectral imaging has rapidly become a key technology in modern remote sensing, enabling the acquisition of hundreds of contiguous spectral bands per pixel. This rich spectral information allows for the precise identification and characterization of materials even at a sub-pixel level. However, due to limited spatial resolution, each pixel often represents a mixture of multiple materials. As a result, advanced hyperspectral unmixing (HU) techniques are essential to accurately separate and interpret these mixed spectral signatures. In this paper, we present a comprehensive overview of blind HU techniques, with an emphasis on the latest advancements in subspace learning methods. This broad category includes both Deep Nonnegative Matrix Factorization (DNMF) and approaches based on Autoencoder (AE). Our analysis covers the underlying theoretical principles, highlight key architectural differences, and assess practical implementation aspects of both approaches. Through extensive experiments on various benchmark datasets and real-world PRISMA satellite imagery, we evaluate the performance of DNMF and AE-based methods in terms of endmember extraction, abundance estimation, and computational cost. Our results offer insights into the relative strengths and trade-offs of these two deep-inspired paradigms, providing practical guidance for method selection in hyperspectral-based applications.