Self-supervised spectral unmixing using scene-aware deep priors

High spectral resolution of hyperspectral sensors facilitates the detailed identification of material, but the coarse spatial resolution affects accurate interpretation. In this regard, spectral unmixing techniques attempt to estimate the constituent materials (endmembers) and their fractional areas (abundances) within a pixel’s instantaneous field of view (IFOV). However, many existing methods optimize denoising, endmember detection, and abundance estimation separately, which can limit robustness under noise and spectral variability. The current study proposes a self-supervised deep learning framework for joint hyperspectral unmixing and denoising. A noise-aware probabilistic encoding strategy is adopted to model a structured latent space to implicitly detect the endmembers. Further, the proposed encoding scheme and regularizations ensure spatio-spectral fidelity to improve robustness to noise. Additionally, a contrastive learning mechanism is incorporated to improve latent separability and robustness to spectral variability. The framework also integrates available partial endmember priors when present and employs a policy-based adaptive refinement mechanism to update endmember estimates under complex noise and mixing conditions. In addition, a learnable loss-balancing strategy is introduced to reduce manual tuning and stabilize optimization. Experiments on synthetic and real hyperspectral datasets demonstrate strong and balanced performance across reconstruction accuracy, spectral fidelity, and clustering quality, indicating the effectiveness of the proposed framework for practical hyperspectral unmixing.