Predicting soil cadmium spatial distribution with HDXRF coupling to multi-model machine learning in the karst area

Area of Science:

• Environmental Science
• Soil Science
• Geochemistry

Background:

• Mapping soil potentially toxic metal spatial distribution is critical for safe and sustainable use and management of contaminated soil.
• Karst areas present challenges due to high soil heterogeneity and complex driving factors affecting metal distribution.
• Cadmium (Cd) contamination poses risks to ecosystems and human health, necessitating accurate spatial prediction.

Purpose of the Study:

• To develop and validate a machine learning framework for predicting soil Cd concentration in karst areas.
• To integrate high-definition X-ray fluorescence (HDXRF) spectra with soil properties and topographical features for Cd prediction.
• To identify key driving factors of Cd accumulation in both geogenically enriched and smelting-affected karst environments.

Main Methods:

• Analysis of soil samples from geogenically Cd-enriched and smelting-affected areas in southwestern China.
• Application and comparison of six different machine learning (ML) models, including Elastic Net and XGBoost.
• Utilizing SHAP (SHapley Additive exPlanations) and Geodetector methods to analyze feature importance and factor interactions.

Main Results:

• Elastic Net model achieved R²=0.879 for geogenic areas, while XGBoost achieved R²=0.893 for smelting zones.
• Key drivers of Cd accumulation differed: sand content and topography in geogenic areas; pH, silt, and organic matter in smelting zones.
• Geodetector analysis revealed synergistic interactions, such as clay-slope coupling and silt-cation exchange capacity synergy.

Conclusions:

• The integrated ML framework provides efficient and cost-effective large-scale monitoring of soil Cd levels in karst ecosystems.
• Distinct Cd accumulation mechanisms in geogenic and anthropogenic settings were elucidated.
• The study offers critical information for precise soil management strategies in vulnerable karst environments.