Identification Framework of Contaminant Spill in Rivers Using Machine Learning with Breakthrough Curve Analysis

Area of Science:

• Environmental Science
• Hydrology
• Data Science

Background:

• Minimizing damage from river contaminant accidents requires early source identification.
• Transient Storage zone Model (TSM) simulates non-Fickian Breakthrough Curves (BTCs) containing contaminant source information.

Purpose of the Study:

• To develop a framework combining ML and TSM for predicting contaminant spill location and mass.
• To identify key BTC features for accurate source prediction.

Main Methods:

• TSM was used to simulate non-Fickian BTCs.
• ML models identified 21 BTC features to predict spill location and mass.
• Recursive Feature Elimination Cross-Validation (RFECV) selected the most relevant features.

Main Results:

• The framework was applied to Gam Creek, South Korea, using data from two tracer tests.
• Six ML methods were evaluated for prediction accuracy.
• Ensemble Decision Tree models, specifically Random Forest (RF) and Xgboost (XGB), demonstrated the highest efficiency and feasibility.

Conclusions:

• The integrated ML-TSM framework effectively predicts contaminant source location and mass.
• Random Forest and Xgboost are highly suitable for real-world contaminant source identification in rivers.