Knowledge Distillation of PlanetScope Imagery for Metre-Scale Lake Water-Quality Mapping

Effective management of eutrophication in inland lakes requires spatially continuous information on key water-quality variables at management-relevant scales. However, metre-scale mapping of total phosphorus (reported as “Phosphorus, Total”, PPUT; µg/L) remains difficult to achieve using conventional in-situ sampling, and nearshore gradients and tributary plumes are often poorly resolved by medium-resolution satellite sensors. In this study, we exploit multi-generation PlanetScope imagery (Dove Classic, Dove-R, and SuperDove; 3–5 m, near-daily revisit) to develop a hybrid, physics-informed AI framework for PPUT retrieval in Lake Simcoe, Ontario, Canada. PlanetScope surface reflectance is combined with short-term meteorological descriptors (3–7-day aggregates of air temperature, wind speed, precipitation, and sea-level pressure) and in-situ Secchi depth (SSD) to train five ensemble-learning models (HistGradientBoosting, CatBoost, RandomForest, ExtraTrees, and GradientBoosting) across eight feature-group regimes. Inclusion of SSD yields a substantial performance gain, with mean R² increasing from ~0.67 (SSD-free) to ~0.94 (SSD-aware), confirming that vertically integrated optical clarity is the dominant constraint on phosphorus retrieval and cannot be reconstructed from surface reflectance alone. To enable scalable SSD-free monitoring, we implement a teacher–student knowledge-distillation scheme in which an SSD-aware teacher transfers its representation to a student using only satellite and meteorological inputs. The optimal student, based on a compact subset of 40 predictors, achieves R² = 0.83, RMSE = 9.82 µg/L, and MAE = 5.41 µg/L on unseen monitoring stations, and is applied to 2020–2025 PlanetScope scenes to generate metre-scale PPUT maps. A 26 July 2024 case demonstrates that >97% of the lake surface remains below 10 µg/L, while rare (<1%) but spatially coherent hotspots >20 µg/L coincide with tributary mouths and narrow channels, highlighting priority areas for management intervention. Although demonstrated here for phosphorus, the PlanetScope–KD framework is model-agnostic with respect to the target variable and can be retrained for other water-quality parameters with optical or hydro-meteorological controls, such as chlorophyll-a, dissolved oxygen, and surface water temperature. This opens a pathway toward unified, high-resolution, multi-parameter lake water-quality prediction to support adaptive monitoring and lake-basin management.