Dissolved oxygen (DO) is crucial for the health of aquatic ecosystems, and plays an essential role in regulating biogeochemical processes in inland lakes. Traditional measurements of DO using the probe or analysis in a laboratory are time-consuming and cannot obtain data with high frequency and broad coverage. Satellites can provide daily/hourly observations within a broad scale and have been used as an important technique for aquatic environments monitoring. However, satellite-derived DO in

more »

... ers is challenging due to its non-optically active property. Here, we developed a two-step model for retrieving DO concentration in Lake Taihu from Moderate Resolution Imaging Spectroradiometer (MODIS) Aqua images. A machine learning model (eXtreme gradient boosting) was developed to estimate DO from field water temperature, water clarity, and chlorophyll-a (Chla) (root-mean-square error (RMSE) = 0.98 mg L−1, mean absolute percentage error (MAPE) = 7.9%) and subsequently was validated on MODIS-derived water temperature, water clarity, and Chla matchups with a satisfactory accuracy (RMSE = 1.28 mg L−1, MAPE = 9.9%). MODIS-derived DO in Lake Taihu from 2002 to 2021 demonstrated that DO ranged from 7.2 mg L−1 to 14.2 mg L−1, with a mean value of 9.3 mg L−1. DO in the northern region was higher than in the central and southern regions, and higher in winter than in summer. We revealed that DO in this decade (2010–2021) was considerably lower than that in the last decade (2002–2009). Meanwhile, annual mean of DO increased in 2002–2009 and decreased from 2010 to 2021. The spatial distribution of DO in Lake Taihu was related to Chla and water clarity, while seasonal and interannual variations in DO resulted from air temperature primarily. This research enhances the potential use of machine learning approaches in monitoring non-optically active constituents from satellite imagery and indicates the possibility of long-term and high-range variations in more water quality parameters in lakes.