Transfer Learning-Based Interpretable Soil Lead Prediction in the Gejiu Mining Area, Yunnan

Area of Science:

• Environmental Science
• Soil Science
• Geochemistry

Background:

• Accurate soil heavy metal prediction is crucial for environmental monitoring and risk assessment.
• Traditional spectral modeling for soil lead (Pb) content faces limitations in small sample scenarios due to data scarcity and soil property variability.
• Developing precise predictive models for soil contaminants in data-limited regions remains a significant challenge.

Purpose of the Study:

• To propose and evaluate a novel transfer learning framework integrated with SHAP analysis for predicting soil Pb content in a small sample scenario.
• To address the limitations of traditional spectral modeling methods in predicting soil Pb content.
• To enhance the accuracy and interpretability of soil heavy metal prediction models.

Main Methods:

• A transfer learning framework was developed using a 1D-ResNet model, leveraging pH data from the European LUCAS soil database as the source domain.
• Spectral features were extracted and transferred to a target domain comprising 130 soil samples from the Gejiu mining area for Pb prediction.
• SHAP (SHapley Additive exPlanations) analysis was employed to interpret the model's predictions and understand the contribution of spectral characteristics.

Main Results:

• The transfer learning model (ResNet-pH-Pb) achieved a superior R² of 0.77, significantly outperforming direct modeling methods (PLS-Pb, SVM-Pb, ResNet-Pb).
• SHAP analysis revealed that the model effectively transferred relevant pH-related spectral features (550-750 nm, 1600-1700 nm) and optimized Pb-specific wavelengths (e.g., 919 nm, 959 nm).
• The interpretability analysis clarified the interplay between spectral characteristics and cross-component transfer learning for soil Pb prediction.

Conclusions:

• The proposed transfer learning framework with SHAP analysis offers a robust and accurate approach for soil heavy metal prediction, particularly in environments with limited sample data.
• This methodology provides a theoretical foundation for understanding the mechanisms behind spectral prediction and interpretability in soil science.
• The study demonstrates the potential of leveraging existing datasets and advanced analytical techniques to overcome data scarcity challenges in environmental monitoring.