Biochemical oxygen demand estimation using explainable ensemble learning methods

Biochemical oxygen demand over five days (BOD₅) is a cornerstone indicator of organic pollution, yet its retrieval from remote sensing is hindered by its non-optically active nature. We present an explainable ensemble-learning framework that predicts BOD₅ in Hong Kong's marine waters by fusing multi-year (2019–2023) Sentinel-2 imagery with cyclic temporal features and four physicochemical and climatic proxies—chlorophyll-a (Chl-a), salinity, suspended solids (SS) and temperature. Initially, each proxy is estimated and subsequently utilized for BOD₅ prediction using CatBoost, LightGBM, XGBoost and Random Forest. XGBoost best captures Chl-a (r = 0.81) and temperature (r = 0.99), whereas CatBoost excels for salinity (r = 0.93), SS (r = 0.85) and ultimately BOD₅ (r = 0.88). SHapley Additive exPlanations reveal the dominant predictors and spatio-temporal mapping across four representative dates shows persistently elevated Chl-a, SS and BOD₅ and depressed salinity in eutrophic Deep Bay zone. This transparent, high-accuracy framework can guide Environmental Protection Department in prioritizing field sampling and streamlining pollution mitigation.