Optimal data pooling from multiple waterbodies to improve machine-learning predictions of cyanobacterial blooms

Accurate early warning of cyanobacterial blooms (CBs) is essential for protecting water resources and public health. However, most monitoring sites provide only small weekly datasets that limit machine-learning (ML) model generalization. Pooling data from multiple, similar waterbodies can increase sample size but also introduces heterogeneity that may reduce transferability. Here we evaluate a pragmatic pooling strategy based on cluster analysis of site-level CB patterns for 45 lakes and rivers across South Korea, and we quantify how pooling affects CB prediction using random forest (RF) and neural network (NN) models. Using hierarchical clustering, the nationwide waterbodies were grouped into eight clusters. Pooling multi-site data generally improved predictive performance even across large spatial distances, with the greatest gains when pooling was restricted to statistically similar clusters: mean improvements of approximately 14 % for RF and 17 % for NN were observed. Performance increased with dataset size and then plateaued near several hundred observations (400-500 samples). Multi-site gains were largest when pooled data originated from fewer than ten sites that shared similar dominant drivers. Feature analyses indicated that water temperature and nutrient variables were important predictors in various regions. Additionally, inter-cluster predictions suggested that rivers with weirs resembled lakes. Rainfall effects depended on antecedent timing and waterbody type. Based on these results, we recommend guidelines for selecting multi-site datasets to be pooled for improving ML-based CB predictions.