A multi-scale framework for long-term prediction of algal bloom occurrence using satellite-derived mapping and...

Algal bloom prediction remains challenging in data-scarce lake systems where continuous in-situ bloom observations and long-term external loading records are limited. This study developed an integrated multi-scale framework for predicting algal bloom occurrence in Lake Burdur over the 1987-2024 period by combining satellite-based deep learning (DL), catchment-scale hydrological and water-quality simulation, machine learning, and explainable artificial intelligence. Long-term Landsat and Sentinel-2 imagery was used with U-Net DL models to detect algal blooms, achieving test accuracies of 94.91% and 91.53%, respectively. The Soil and Water Assessment Tool (SWAT) was implemented to simulate catchment-derived hydrological and water-quality inputs entering the lake, including point and non-point source contributions. These SWAT-derived variables, together with meteorological parameters, were used to train Random Forest, Support Vector Machine (SVM), eXtreme Gradient Boosting, and Categorical Boosting models under three temporal lag scenarios, namely short-term (t-1 to t-3), medium-term (t-1 to t-6), and long-term (t-1 to t-12). The short-term lag scenario provided the most favorable overall balance between predictive performance, stability, and generalization, with the SVM model achieving an overall accuracy of 89.74% and a balanced accuracy of 90.42%. SHapley Additive exPlanations analysis indicated that short-term bloom prediction was mainly associated with meteorological predictors, particularly Logarithmic Solar Radiation-Wind Speed Index and relative humidity, whereas SWAT-derived external Total Nitrogen to Total Phosphorus loading ratio became more influential under the medium-term scenario. These findings indicate that bloom prediction in Lake Burdur was associated with interacting recent atmospheric conditions and antecedent catchment-derived nutrient inputs. The proposed framework provides an interpretable basis for algal bloom prediction and supports future development toward early-warning applications in data-scarce lake systems.