Unmixing in latent space: A novel unsupervised approach for geological mapping of lunar surface

Existing spectral unmixing techniques for planetary data face challenges related to noise, limited samples, and high dimensionality, which impact their effectiveness.While deep learning-based methods have achieved state-of-the-art results, they often lack explainability and generalizability. In this study, we comprehensively evaluate the state-of-the-art approaches for spectral unmixing, considering factors such as accuracy, computational expenses, explainability, generalizability, sensitivity to noise, and training sample requirements. We compare linear methods (FCLSU, EndNEt, UnDIP and CNNAEU) and a nonlinear method (PPNM) and propose a novel approach using autoencoders for hyperspectral image unmixing. Our method leverages the autoencoder's encoded representation to reconstruct input data, leading to decreased dimensionality, noise reduction, and pertinent feature extraction. Experimental results indicate its superior spectral unmixing accuracy in contrast to linear and nonlinear benchmarks.