Monitoring and predicting algal proliferation in drinking water sources holds significant importance for enhancing the environmental integrity of aquatic ecosystems and safeguarding human health. Utilizing multisource remote sensing dala enables ihe acquisition of high spatiotemporal resolution information regarding algal dynamics. The integration of long-term remote sensing monitoring with machine learning algorithms allows for the adaptation to the complex growth mechanisms and nonlinear characteristics associated with algal proliferation, thereby facilitating the prediction of spatiotemporal variations in algal proliferation risk. In this study, Landsal and MODIS long time-series satellite remote sensing data were employed to extract spatiotemporal variations in algal content in the Danjiangkou Reservoir from 2000 to 2020, utilizing the fraction of algal inversion and normalized difference vegetation index methods. Furthermore, the lime lag effects of meteorological factors, including temperature, atmospheric pressure, relative humidity, wind speed, and cumulative sunshine duration, on algal proliferalion were analyzed. Three machine learning algorithms, namely Support Vector Machine, Naive Bayes, and Random Forest, were deployed to predict algal proliferalion risk. The prediclive performance of ihese algorithms was evaluated and compared. The results indicated lhat ihe annual variation in algal content in the Danjiangkou Reservoir exhibited a trend of initial increase followed by a decrease, characterized by distinct seasonal periodic fluctuations, with late spring to early summer representing a period of rapid algal proliferation. Spatially, relatively higher algal content was observed in inflow tributaries and bay areas, indicating high-risk zones for algal proliferation. The algal proliferation risk prediction model for the Danjiangkou Reservoir accurately identified high-risk areas for algal proliferation and reflected short-term spatial variations. The predictions from the three algorithms exhibited overall consistency, with Support Vector Machine and Naive Bayes algorithms demonstrating higher accuracy, and a lead time of 4 to 5 days was identified as the optimal prediction window. © 2024 by Journal of Lake Sciences.
