Machine learning-based retrieval of total nitrogen matter from an inland highly turbid lake using Landsat 8 OLI...

Hongze Lake serves as a key regulating reservoir in the Eastern Route of the South-to-North Water Diversion Project. Its water quality directly affects the safety of water supply, as well as the ecological environment and sustainable development of the water-receiving regions. Total nitrogen (TN) is one of the major nutrients driving eutrophication in inland lakes; therefore, investigating TN concentrations in Hongze Lake and their influencing factors is essential. However, as a non-optically active substance, TN is difficult to estimate accurately using remote sensing techniques. In this study, four machine learning algorithms-random forest (RF), support vector machine (SVM), multilayer perceptron (MLP), and eXtreme gradient boosting (XGB)-were employed to automatically determine an optimal threshold that divides TN concentrations into high-value and low-value zones for separate modeling based on Landsat 8 OLI data. Among them, the RF model achieved the highest accuracy with a threshold of 1.05 mg/L; MAPE and RMSE were 14.83% and 0.22 mg/L, respectively. This algorithm was then applied to evaluate the spatiotemporal variations in TN distribution from 2013 to 2022. The main conclusions are as follows: (1) The annual variation in TN concentration ranges from 1.27 to 2.24 mg/L, with the highest and lowest values observed in 2020 and 2018, respectively. (2) The monthly variation in TN concentration ranges from 0.94 to 2.42 mg/L, with the highest and lowest values occurring in January and July. Overall, TN levels decrease with rising temperature and precipitation. (3) The average TN concentration during water diversion periods is higher than that during non-diversion periods, which is mainly attributed to nitrogen pollutants transported from upstream diversion inflow and the resuspension of large amounts of sediment-induced suspended solids.