Development of a self-supervised deep learning framework for chlorophyll-a retrieval in data-scarce inland waters

Deep learning and remote sensing-based chlorophyll-a (Chl-a) monitoring face challenges due to the optical complexity of inland waters and the scarcity of labeled data. To address these limitations, this study develops a self-supervised learning-based deep learning (SSL-DL) framework that leverages both labeled and unlabeled data. Three SSL-DL models are developed: a predictive SSL-DL model, which learns weak labels (incomplete labels); a generative SSL-DL model, which reconstructs input reflectance to capture underlying features; and an integrated SSL-DL model, which combines both. The models are applied to Sentinel-2 imagery of Daecheong and Paldang Lakes in South Korea. Results indicate that SSL-DL models outperform baseline models, with the integrated SSL-DL model achieving the highest test NSE (improvements of 0.1–0.36 over baselines in Daecheong Lake, improvements of 0.03–0.58 in Paldang Lake). The findings highlight the significance of SSL-DL in overcoming data limitations and enhancing scalability, demonstrating the potential for broader environmental remote sensing applications.