A transfer learning-enhanced deep learning framework for efficient and interpretable soil heavy metal pollution...

Area of Science:

• Environmental Science
• Geospatial Analysis
• Machine Learning

Background:

• Large-scale soil heavy metal pollution assessment faces challenges from data scarcity and spatial heterogeneity.
• Traditional machine learning methods struggle with high-dimensional, heterogeneous data and limited interpretability.

Purpose of the Study:

• To develop an efficient and interpretable framework for estimating soil heavy metal pollution risk.
• To leverage transfer learning (TL) and deep learning (DL) with multi-source data for improved prediction accuracy.

Main Methods:

• Proposed a transfer learning-based deep learning (TL-CNN) framework integrating convolutional neural networks (CNN).
• Combined remote sensing (RSs), web-based (WBs), and field-sampled datasets (SRs) with a GradSHAP interpretability module.
• Applied the model to soil heavy metal pollution prediction in Shaoguan City (2018-2022).

Main Results:

• The TL-CNN model achieved over 84% overall accuracy, outperforming conventional machine learning methods.
• Effectively addressed data scarcity and reduced reliance on extensive field sampling.
• GradSHAP analyses identified the significant contribution of RSs features and spatial metrics.

Conclusions:

• The TL-CNN framework offers a powerful solution for large-scale soil heavy metal pollution risk estimation.
• Multi-source heterogeneous datasets combined with TL-based DL strategies are crucial for sustainable soil management.
• The approach provides both high predictive accuracy and valuable explanatory insights.