Machine Learning, Generalization, and Transfer Learning for Predicting the Exceedance of Fecal Indicator Bacteria...

Area of Science:

• Environmental science
• Public health
• Data science

Background:

• Beach water quality monitoring is crucial for public health.
• Current methods for detecting fecal indicator bacteria (FIB) are costly and time-consuming, leading to infrequent monitoring.
• Machine learning (ML) offers a potential solution for predicting FIB levels using environmental data.

Purpose of the Study:

• To evaluate the effectiveness of transfer learning (TL) in generalizing ML models for predicting FIB levels across different beach locations.
• To determine if TL can improve the accuracy of predicting FIB threshold exceedances compared to traditional ML models.

Main Methods:

• Developed ML models using environmental and FIB data from source beaches.
• Applied source-to-target generalization techniques to predict FIB levels at different beaches.
• Augmented generalization with transfer learning (TL) to improve model performance.
• Assessed model performance using metrics like specificity, sensitivity, WF1 scores, and AUC.

Main Results:

• Transfer learning (TL) combined with source-to-target generalization achieved a specificity of 0.70–0.81 and sensitivity of 0.28–0.76.
• TL significantly improved model performance over generalization alone, with 28.3% increases in WF1 scores and 5.4% in AUC.
• Model performance with TL was comparable to models trained and tested on the same beach.

Conclusions:

• Transfer learning is a viable method for generalizing ML models to predict FIB levels at new beaches.
• TL enhances the accuracy and reliability of beach water quality predictions, extending public health protection.
• Further research on selecting optimal source beaches can further refine TL applications for beach monitoring.