Predicting ecological risks of heavy metals in watersheds based on interpretable machine learning models: under the...

Area of Science:

• Environmental Science
• Ecotoxicology
• Data Science

Background:

• Heavy metal contamination threatens watershed ecosystems, necessitating accurate ecological risk prediction.
• Limited ecotoxicological data due to high costs hinders machine learning applications.

Purpose of the Study:

• To address data scarcity in ecotoxicology by augmenting toxicity datasets using synthetic data.
• To develop and compare machine learning models for predicting heavy metal toxicity (Cr, Mn, Cu).
• To interpret key toxicity drivers and assess ecological risks in the Chaohu Lake Basin.

Main Methods:

• Employed Synthetic Minority Oversampling Technique for Regression with Gaussian Noise (SMOGN) to generate synthetic ecotoxicity data.
• Developed and compared Random Forest (RF), Extreme Gradient Boosting (XGBoost), and Support Vector Machine (SVM) regression models.
• Utilized Shapley Additive Explanations (SHAP) for model interpretability and Risk Quotient (RQ) for ecological risk assessment.

Main Results:

• Synthetic data augmentation substantially improved model performance, with RF R² increasing from 0.696 to 0.977.
• All tested models (RF, XGBoost, SVM) demonstrated higher accuracy with augmented data compared to the original dataset.
• Ecological risk assessment revealed elevated risks for Chromium (Cr) and Copper (Cu) in the Chaohu Lake Basin.

Conclusions:

• SMOGN is a feasible and effective method for overcoming data scarcity in ecotoxicology.
• Augmented datasets enhance the accuracy of machine learning models for predicting heavy metal toxicity.
• The study provides crucial data and evidence for managing heavy metal contamination in the Chaohu Lake Basin.