A transfer learning-enhanced deep learning framework for efficient and interpretable soil heavy metal pollution prediction under data scarcity and spatial heterogeneity

Area of Science:

• Environmental Science
• Geospatial Analysis
• Machine Learning

Background:

• Large-scale soil heavy metal pollution assessment is hindered by data scarcity and spatial variability.
• Traditional machine learning methods face challenges with high-dimensional, heterogeneous data and limited interpretability.

Purpose of the Study:

• To develop an efficient and interpretable framework for predicting soil heavy metal pollution using multi-source data.
• To address data scarcity and improve prediction accuracy in large-scale environmental monitoring.

Main Methods:

• Proposed a transfer learning (TL)-based deep learning (DL) framework (TL-CNN) integrating convolutional neural networks.
• Utilized remote sensing (RSs), web-based (WBs), and field-sampled datasets with spatial regionalization features (SRs).
• Incorporated hierarchical feature extraction and a GradSHAP interpretability module.

Main Results:

• The TL-CNN model significantly outperformed conventional machine learning methods, achieving over 84% accuracy.
• The framework effectively handled data scarcity, reduced reliance on extensive field sampling, and minimized interpolation errors.
• GradSHAP analysis identified the crucial contribution of RSs features and spatial metrics in predictions.

Conclusions:

• The TL-CNN model demonstrates superior performance for soil heavy metal pollution prediction, especially in multi-metal scenarios.
• Leveraging multi-source heterogeneous data with TL-based DL strategies offers a promising approach for sustainable soil management.
• The framework provides both high predictive accuracy and valuable explanatory insights for environmental risk assessment.