A machine learning approach for coastal water quality retrieval using satellite images: the case of the Romagna coast

Monitoring water quality in the coastal waters of the Emilia-Romagna region (Northern Adriatic Sea) represents a formidable challenge for satellite remote sensing. The Po River plume creates a dynamic Case-2 environment where phytoplankton, suspended particulate matter, and colored dissolved organic matter (CDOM) co-vary independently. These conditions frequently compromise the reliability of traditional retrieval algorithms and standard land-oriented atmospheric correction procedures. This study develops and validates a neural network (MLP) framework to retrieve Chlorophyll-a and Turbidity from Sentinel-2 MSI and Landsat-8/9 OLI imagery (2020–2023). Utilizing in-situ data from two fixed monitoring stations, a Bayesian hyperparameter optimization framework was employed to identify robust model architectures. To evaluate the methodological necessity of atmospheric correction in turbid waters, models were trained and compared using both Top-of-Atmosphere (TOA) and Surface Reflectance (SR) products. Results indicate that Sentinel-2 systematically outperforms Landsat-8/9, an advantage primarily driven by its dedicated red-edge spectral bands and finer spatial resolution. More critically, the study finds that models trained directly on TOA reflectance consistently match or exceed the performance of SR-based models. This suggests that neural networks can implicitly compensate for atmospheric effects, bypassing the systematic biases often introduced by standard correction processors in turbid nearshore zones. High-resolution spatiotemporal maps generated from the optimized models successfully captured regional hydrodynamic features, including current-driven plume confinement and seasonal biomass shifts. These findings demonstrate that TOA-based neural network modelling offers a more resilient and practical alternative for operational water quality monitoring in complex coastal environments where standard atmospheric correction remains a significant source of uncertainty.