Transitive Semantic Alignment in Class-Conditioned Latent Diffusion for Martian Hyperspectral Mineral Mapping
in
Workshop: 4th ICLR Workshop on Machine Learning for Remote Sensing
Abstract
Hyperspectral data possess rich but fragile geometric structure that is often obscured by noise, redundancy, and limited supervision. We propose \textit{Class-Conditioned Latent Diffusion Networks (CC-LDNs)}, which cast representation learning as a reverse-time diffusion process on a latent manifold. A lightweight 3D encoder maps hyperspectral patches to latent space, where classification constraints imposed on noisy latents regularize reverse diffusion dynamics. A label- and time-conditioned denoiser recovers clean representations via \textit{transitive semantic alignment}. Geometrically, CC-LDNs induce class-conditional basins of attraction that contract intra-class variability while preserving inter-class separation, yielding stable and well-conditioned latent structures. CC-LDN achieves OA of 96.51\%, 92.37\%, and 90.85\% on the HC, NF, and UP Martian datasets, respectively, outperforming eight state-of-the-art methods with fewer parameters than transformer-based models, validating latent diffusion as a geometry-grounded paradigm for hyperspectral learning under limited supervision.