Machine Learning-Based Early Warning Level Prediction for Cyanobacterial Blooms Using Environmental Variable...

Area of Science:

• Environmental Science
• Data Science
• Ecology

Background:

• Harmful algal blooms (HABs) pose significant challenges to water resource management globally.
• Predicting HABs is difficult due to infrequent occurrences and data imbalance issues in machine learning models.
• Selecting appropriate input variables for HAB prediction models is complex.

Purpose of the Study:

• To enhance the predictive performance of harmful algal bloom (HAB) models.
• To address data imbalance challenges in HAB prediction using feature selection and data resampling.
• To improve the accuracy of algal alert level predictions.

Main Methods:

• Utilized data resampling techniques, including Synthetic Minority Oversampling Technique-Edited Nearest Neighbor (SMOTE-ENN), to balance imbalanced datasets.
• Developed and compared two machine learning models (Artificial Neural Network and Random Forest) for predicting algal alert levels.
• Employed 10 years of meteorological, hydrodynamic, and water quality data for model construction and validation.

Main Results:

• Data resampling methods showed a more significant improvement in model accuracy than feature selection methods.
• Models incorporating synthetic data demonstrated superior prediction performance for caution (L-1) and warning (L-2) alert levels compared to models using original data.
• The optimal Random Forest model with SMOTE-ENN achieved 85.0% accuracy for L0, 85.7% for L1, and 100% for L2, significantly outperforming the model with original data.

Conclusions:

• Applying synthetic data generation effectively addresses data imbalance in HAB prediction.
• Improved machine learning models enhance the detection performance for early stages of algal blooms.
• Reliable HAB predictions support proactive management strategies for water resource protection.