Influence of resampling techniques on Bayesian network performance in predicting increased algal activity.

Area of Science:

• Environmental Science
• Data Science
• Ecology

Background:

• Predicting algal blooms is crucial for aquatic and human health.
• Imbalanced datasets with infrequent severe blooms hinder accurate prediction models.

Purpose of the Study:

• Investigate resampling techniques to address imbalanced data for algal activity prediction.
• Determine if resampling improves the prediction accuracy of algal blooms.

Main Methods:

• Applied Kmeans under-sampling (US_Kmeans), SMOTE, and SCUT to a Bayesian network (BN) model.
• Used the BN model to predict chlorophyll-a (chl-a) concentrations in Lake Burragorang, Australia.
• Evaluated model performance using true positive rate (TPR) and area under the curve (AUC).

Main Results:

• SMOTE generated synthetic data that resulted in the best BN graphical structure.
• All resampling techniques improved the BN's ability to detect high algal activity events.
• Resampling enhanced the prediction of increased algal activity, indicated by higher chl-a.

Conclusions:

• Resampling techniques effectively address imbalanced data in algal bloom prediction.
• SMOTE shows promise for improving the accuracy of predictive models for algal activity.
• Findings guide future data collection and model development for water quality management.