Retrieval of water quality parameters based on IOA-ML models and their response to short-term hydrometeorological factors

Study region: The Honghu Lake (HHL) and Changhu Lake (CHL) in middle China. Study focus: Large-scale and high-precision estimation of water quality parameters (WQPs) is critical in explaining the spatiotemporal dynamics and clarifying their response to short-term hydrometeorological factors. Six machine learning models optimized by intelligent optimization algorithms (IOA-ML) were developed to retrieve WQPs using paired in situ measurements and near-synchronous Sentinel-2 reflectance (Rrs). Furthermore, the response of pixel-based WQPs to short-term hydrometeorological factors were explored by generalized additive model (GAM). New hydrological insights for the region: The results showed that R rs curves were significantly correlated with WQPs concentration, which provided a solid foundation for WQPs retrieval. The best IOA-ML model for total phosphorus (TP), total nitrogen (TN), and permanganate index (CODMn) was extreme gradient boosting optimized by genetic algorithm (GA-XGB), while that for dissolved oxygen (DO) and turbidity was categorical boosting regression optimized by GA (GACBR). Coefficient of determination (R2) of the best retrieval models for the test sets of TP, TN, turbidity, CODMn and DO were 0.545, 0.418, 0.794, 0.798, and 0.653, respectively. The best retrieval models were applied to two big inland lakes and revealed that TP, TN, CODMn, and turbidity in HHL increased rapidly from 2016 to 2022, especially during 2021-2022. 81.4 %91.5 % of the WQPs variations in HHL and 63.4 %-92 % in CHL can be explained by hydrometeorological factors.