Deep Learning-Based Spectral Unmixing of Low Resolution Sentinel-2 Data for Monitoring Small Rivers: A Case Study at Enguri River, Georgia

Spectral unmixing is crucial for analyzing mixed pixels in remote sensing, traditionally applied to hyperspectral data. While multispectral Sentinel-2 data has seen limited use in this domain, its wide availability and applicability in environmental monitoring demand effective unmixing solutions. Sentinel-2 imagery particularly suffers from spectral variability caused by environmental conditions, atmospheric residuals, and temporal changes, which are often overlooked by existing methods. In this study, we propose a novel multimodal deep generative model, the time-dependent Deep Transformer MultiSpectral Unmixing Model (tDTMSUM), to extract pure water spectra from mixed observations, particularly in narrow rivers where water pixels are often mixed with nearby land. The model integrates Sentinel-2 data with auxiliary variables capturing sources of spectral variability, using a Variational Autoencoder and channel-wise Transformer. Trained on synthetic mixtures derived from real Sentinel-2 imagery, it performs supervised endmember extraction and abundance estimation with a focus on the water endmember. Evaluation using real Sentinel-2 imagery and in-situ RoX spectrometer demonstrates that tDTMSUM outperforms state-of-the-art methods in robustness and accuracy, despite the absence of ground truth, for real-world water monitoring applications.