Satellite retrievals of total phosphorus in Taihu Lake using Sentinel-2 images and an optimized XGBoost model

Satellite remote sensing technology holds significant potential for water quality monitoring due to its high spatio-temporal resolution and long-term observational capabilities. However, developing satellite-retrieved models for non-optical active parameters such as total phosphorus remains challenging. In this study, we utilized measured total phosphorus data from 17 floating boat stations in Taihu Lake and Sentinel-2 satellite imagery to construct an XGBoost model based on Bayesian optimization and feature selection for the retrieval of total phosphorus levels in the lake. The model employed single bands and band ratios at various resolutions as input features, optimized model parameters through Bayesian optimization, and incrementally introduced variables with substantial contributions to evaluate and determine the optimal feature combination. Results indicated that band ratios at different resolutions constituted 79 % of the optimal feature combination, demonstrating significantly better performance than single bands. The coefficient of determination (R²) was 0.7151, and the root mean square error (RMSE) was 0.0179 mg/L. Compared to using all feature variables, the R² increased by 35 %, RMSE decreased by 11.4 %, and the accuracy rate of total phosphorus retrieval concentration reached 81.7 %. The spatial distribution of total phosphorus derived from the satellite-retrieved model aligned well with the actual conditions. The optimal band combination model exhibited high stability and generalization ability. This study provides a practical reference for constructing a satellite retrieval model to determine inactive water quality parameters, and also provides strong technical support for the construction of an air-space-ground integrated monitoring network, especially in the context of open sharing of satellite data, and offers a feasible path for realizing large-scale and low-cost water quality supervision.