Improved classification of soil As contamination at continental scale: Resolving class imbalances using machine...

Area of Science:

• Environmental Science
• Soil Science
• Geochemistry

Background:

• Arsenic (As) contamination poses risks to soil health and ecosystems.
• Accurate identification of As-contaminated areas is vital for effective soil management and reclamation.
• Previous machine learning (ML) approaches often overlook the challenge of imbalanced datasets, where As-contaminated samples are rare.

Purpose of the Study:

• To develop a novel ML framework to address the imbalanced learning problem in identifying soil As contamination.
• To improve the accuracy and generalization capability of models for classifying As-contaminated soil samples using spectral data.
• To apply the optimized ML model to a global dataset for predicting high-risk As contamination areas.

Main Methods:

• Utilized visible and near-infrared (VNIR) soil spectra for As contamination identification.
• Implemented a framework combining spectral preprocessing, imbalanced dataset resampling, and model comparison.
• Employed Bayesian optimization for hyperparameter tuning of ML models.
• Selected the optimal model based on recall performance.

Main Results:

• The optimized ML model achieved a recall of 0.83, an area under the curve (AUC) of 0.88, and a balanced accuracy of 0.79 on the testing set.
• Threshold adjustment further improved the model's recall to 0.87, demonstrating excellent performance and generalization.
• The model successfully predicted areas with high soil As contamination risk on a global scale.

Conclusions:

• The developed ML framework effectively overcomes the imbalanced learning problem in soil As contamination detection.
• The study provides a robust and generalizable method for classifying As-contaminated soil samples using spectral data.
• This research offers a valuable reference for soil contamination management and reclamation strategies worldwide.