Generalizable deep learning forecasting of harmful algal blooms using transfer learning across river systems

Harmful algal blooms (HABs) present serious challenges to water quality and resource management, with their frequency and severity escalating due to climate change and anthropogenic impacts. While deep learning (DL) models show strong predictive capabilities in HAB forecasting, challenges remain due to data scarcity and site-specific variability. This study systematically investigates transfer learning (TL) to improve the generalization and accuracy of HAB forecasts across 26 monitoring sites spanning four major rivers in South Korea. The analysis considers multiple dimensions influencing TL performance, including DL architectures, TL schemes, river system differences, and target domain sample sizes. Various DL architectures were evaluated: the Temporal Fusion Transformer (TFT), Transformer, CNN-LSTM with attention, and recurrent neural network (RNN)-based models. While RNN-based models have been commonly used in TL-based HAB forecasting, advanced architectures like TFT remain unexplored. To evaluate their effectiveness, four TL schemes with varying parameter adaptation were applied, consistently improving forecasting performance across all models and sites. Average R² increasing from 0.35 to 0.50 to 0.47–0.60, while site-specific values reaching 0.89. Improvements ranged from 12.51 % to 82.49 %, demonstrating robust generalization gains. Among TL schemes, full fine-tuning yielded the highest number of best-performing sites, especially for transformer-based models. The TFT achieved the strongest overall performance, outperforming others on most sites. Sites with higher cyanobacterial abundance generally exhibited stronger performance, while low-performing sites experienced substantial improvement. Furthermore, SHAP analysis identified key environmental drivers to enhance interpretability and management relevance. This study provides a scalable, transferable framework for advancing DL-based HAB forecasting in data-scarce environments, offering insights for early warning systems and broader environmental applications.