Short-term spatial prediction of algal blooms in Lake Taihu via machine learning and GOCI observations

Abstract

Harmful algal blooms are critical issues in eutrophic lakes worldwide. However, predicting the spatial distribution of algal blooms at the pixel level is still a challenge. In this study, floating algae cover (FAC) was used to extract algal coverage via the Geostationary Ocean Color Imager (GOCI) and GOCI-II satellites. Three novel indices, the floating algae cover index (FACI), distance index (DI), and algae around index (AAI), were developed. Including these three indices and environmental factors, a total of 12 input features were utilized to predict the short-term spatial variations in algal blooms via random forest (RF), support vector regression (SVR), extreme gradient boosting (XGBoost), and deep neural network (DNN) algorithms through hour-by-hour iterations. The results indicated that the RF model exhibited better performance (R² = 0.91, RMSE = 9.08 %, N = 88,791) than the SVR model (R² = 0.79, RMSE = 13.97 %), the XGBoost model (R² = 0.84, RMSE = 12.11 %), and the DNN model (R² = 0.67, RMSE = 17.39 %). The RF model was then applied to predict the spatial distribution of FAC in Lake Taihu. The FAC values at the pixel level were predicted to have an average R² of 0.67 across the six subregions of Lake Taihu, as well as satisfactory performance (R² = 0.83, RMSE = 1.39 %, N = 68) in predicting the overall FAC of Lake Taihu after 7 h, which indicated that the model maintains a high level of accuracy at the pixel level and in overall predictions. The iterative FAC prediction model promotes the efficiency of spatial prediction of algal blooms and enables the location and intensity of bloom outbreaks to be determined hours in advance, which provides valuable technical support for the ecological management of eutrophic lakes.