Development of Machine Learning Models to Predict Daily Gas Ebullition Flux in Waterways from Sediment...

Area of Science:

• Environmental Science
• Geochemistry
• Data Science

Background:

• Gas ebullition is key for sediment-to-water contaminant transport.
• Existing models for gas ebullition are limited by site-specific data and focus on water parameters, not sediment conditions.
• Current models struggle with diverse waterways and ecohydrological conditions.

Purpose of the Study:

• Develop generalized machine learning (ML) models to accurately predict gas ebullition across various waterways.
• Improve upon the predictive capabilities of existing regression models for gas ebullition.
• Identify key sediment and water parameters influencing gas ebullition.

Main Methods:

• Trained and evaluated ML models including Multivariate Linear Regression (MLR), Random Forest (RF), eXtreme Gradient Boosting (XGB), and Neural Network (NN).
• Utilized data from over 40 diverse sites across all seasons, incorporating sediment parameters, depth, and temperature.
• Compared ML model performance against ten published regression models.

Main Results:

• ML approaches significantly outperformed existing regression models, demonstrating enhanced prediction accuracy for gas flux.
• Random Forest (RF) and eXtreme Gradient Boosting (XGB) models showed superior performance (r² = 0.79 and 0.80, respectively).
• Temperature, Chemical Oxygen Demand/Total Organic Carbon (COD/TOC) ratio, and water depth were identified as the most influential parameters.

Conclusions:

• Machine learning models provide a more accurate and generalized approach to predicting gas ebullition compared to traditional methods.
• Published regression models, particularly those from lacustrine systems, have limited predictive power for broader waterway applications.
• Accurate gas ebullition prediction using common parameters offers significant potential for improved modeling, site assessment, and remediation strategies.