Optimizing radial basis function networks for harmful algal bloom prediction: a hybrid machine learning approach

Abstract

The deployment of artificial intelligence in environmental monitoring demands models balancing efficiency, interpretability, and computational cost. This study proposes a hybrid radial basis function network (RBFN) framework integrated with fuzzy c-means (FCM) clustering for predicting harmful algal blooms (HABs) using water quality parameters. Unlike conventional approaches, our model leverages localized activation functions to capture non-linear relationships while maintaining computational efficiency. Experimental results demonstrate that the RBFN-FCM hybrid achieved high accuracy (F1-score: 1.00) on test data and identified Chlorophyll-a as the strongest predictor (r = 0.94). However, real-world validation revealed critical limitations: the model failed to generalize datasets with incomplete features or distribution shifts, predicting zero HAB outbreaks in an unlabeled 11,701-record dataset. Comparative analysis with Random Forests confirmed the RBFN-FCM's advantages in training speed and interpretability but highlighted its sensitivity to input completeness. This work underscores the potential of RBFNs as lightweight, explainable tools for environmental forecasting while emphasizing the need for robustness against data variability. The framework offers a foundation for real-time decision support in ecological conservation, pending further refinement for field deployment.