Assessment of Machine Learning Models for Chl-a Prediction Using Water Quality Data: Transferability to a Single External Site and Input Variable Configurations

Chlorophyll-a (Chl-a) is an indicator of phytoplankton biomass but is not continuously measured at all water-quality monitoring sites. This study evaluated machine learning models for estimating Chl-a from routinely monitored water-quality variables, focusing on temporal validation, evaluation at a single external site, and reduced input availability. Hourly data from 12 South Korean monitoring sites (2021–2025) were used to develop Random Forest (RF), XGBoost, LightGBM, and Extra Trees (ET) models, optimized using Bayesian optimization and five-fold cross-validation. XGBoost achieved the highest random-split internal R2 (0.9219), whereas RF had the highest external-site R2 point estimate at Yangpyeong (0.6928). However, temporal validation using 2025 data yielded substantially lower R2 values (0.3528–0.5255), indicating reduced predictive performance in a subsequent period. SHAP analysis identified dissolved oxygen and pH as the leading contributors to RF predictions. At Yangpyeong, excluding total nitrogen and total phosphorus produced an R2 of 0.6897 versus 0.6928 with all inputs, with no clear difference detected by paired block bootstrap analysis; excluding turbidity reduced performance. These findings suggest the potential for supplementary Chl-a estimation with reduced inputs at Yangpyeong and provide a basis for further transferability assessments at additional monitoring sites.