Machine learning correction of C2-Net for chlorophyll-a retrieval in optically complex small reservoirs

Chlorophyll-a is a key indicator of phytoplankton biomass, eutrophication risk, and overall water quality in inland water bodies. Its accurate monitoring is particularly critical in small reservoirs, which serve essential functions in water supply, irrigation, and ecosystem services, yet remain largely underserved by conventional remote sensing products. Accurate retrieval of chlorophyll-a concentration in these systems remains challenging due to optical complexity, land–water adjacency effects, and scaling mismatches in atmospheric correction algorithms. This study evaluates the performance of the C2-Net atmospheric processors (C2RCC, C2X, and C2XC) applied to Sentinel-2 images of 32 small reservoirs in Extremadura, Spain, using 94 satellite scenes and in situ chlorophyll-a measurements. While C2RCC showed moderate correlation with field data, it exhibited systematic underestimation, whereas C2X and C2XC produced extreme overestimations despite preserving strong monotonic relationships. Rather than discarding these well-established processors, a machine learning-based correction framework was developed to extend their applicability to waters for which they were not originally designed. By combining rescaled processor outputs, interaction terms, and aggregate statistics, this strategy leverages the physics-based atmospheric correction already embedded in C2-Net while correcting its systematic biases through data-driven post-processing, thereby minimizing error propagation and maximizing operational applicability. The proposed approach significantly improved retrieval accuracy, reducing the RMSE by up to 35% compared to the best baseline processor. Regularized regression models provided a solid balance between accuracy, stability, and interpretability, while more complex models achieved superior peak performance. Validation results confirmed good spatial transferability across reservoirs and trophic conditions, with greater uncertainty in hypertrophic scenarios. The proposed framework offers an operational, transferable, and open-source solution for improving satellite-based chlorophyll-a monitoring in small, optically complex inland water bodies, directly applicable within existing Sentinel-2 processing workflows.