Development of data-driven algal bloom alert models with low temporal resolution data and application to Hong Kong rivers

The study focuses on the 12 rivers across Hong Kong, that have been facing varying degrees of algal bloom risks. Data-driven models are extensively employed to forecast and discern the catalysts of algal blooms, leveraging high temporal resolution data to unveil intricate numerical relationships among indicators. The efficacy of these models can be compromised due to the low temporal resolution of water quality monitoring. This study presents a data-driven algal bloom alert model employing machine learning and data discretization techniques to solve the data issue, utilizing monthly water quality data of 12 rivers and daily weather data in Hong Kong. The critical factors associated with algal blooms were then identified by feature importance analysis and false negative analysis of the optimal performance model, guiding potential algal control strategies for Hong Kong. Models that integrate data discretization outperformed those using numerical normalization, showing higher recall scores and greater stability across selected algorithms (linear regression, support vector machine, random forest, and decision tree). Permutation importance analysis identified nitrogen compounds and rising temperatures as key triggers of algal blooms, while false negative analysis highlighted total phosphorus and flow as critical factors. This study offers a potential solution for prediction and forecasting without extensive datasets and emphasizes the importance of integrating domain knowledge in both prior and posterior analyses to enhance the interpretability of data-driven models in water science. • Algal bloom alert models are developed by applying data discretization. • Key factors are identified by permutation importance and false negative analysis. • Data discretization may enhance prediction and forecasting with low-resolution data.