Improving the performance of machine learning models for early warning of harmful algal blooms using an adaptive synthetic sampling method

•The algal alert system is used worldwide for proactive management.•Alert levels of algal bloom were classified using two machine learning models.•Imbalanced class data induced biased training of the machine learning models.•Synthetic data by ADASYN increased the performance of the machine learning models. Many countries have attempted to monitor and predict harmful algal blooms to mitigate related problems and establish management practices. The current alert system-based sampling of cell density is used to intimate the bloom status and to inform rapid and adequate response from water-associated organizations. The objective of this study was to develop an early warning system for cyanobacterial blooms to allow for efficient decision making prior to the occurrence of algal blooms and to guide preemptive actions regarding management practices. In this study, two machine learning models: artificial neural network (ANN) and support vector machine (SVM), were constructed for the timely prediction of alert levels of algal bloom using eight years’ worth of meteorological, hydrodynamic, and water quality data in a reservoir where harmful cyanobacterial blooms frequently occur during summer. However, the proportion imbalance on all alert level data as the output variable leads to biased training of the data-driven model and degradation of model prediction performance. Therefore, the synthetic data generated by an adaptive synthetic (ADASYN) sampling method were used to resolve the imbalance of minority class data in the original data and to improve the prediction performance of the models. The results showed that the overall prediction performance yielded by the caution level (L1) and warning level (L2) in the models constructed using a combination of original and synthetic data was higher than the models constructed using original data only. In particular, the optimal ANN and SVM constructed using a combination of original and synthetic data during both training (including validation) and test generated distinctively improved recall and precision values of L1, which is a very critical alert level as it indicates a transition status from normalcy to bloom formation. In addition, both optimal models constructed using synthetic-added data exhibited improvement in recall and precision by more than 33.7% while predicting L-1 and L-2 during the test. Therefore, the application of synthetic data can improve detection performance of machine learning models by solving the