Soil Heavy Metal Content Prediction Based on a Deep Belief Network and Random Forest Model

Area of Science:

• Environmental Science
• Geochemistry
• Data Science

Background:

• Accurate prediction of soil heavy metal content is crucial for environmental monitoring and risk assessment.
• X-ray fluorescence (XRF) spectroscopy is a valuable tool for elemental analysis, but spectral data often requires sophisticated processing for quantitative analysis.
• Traditional models may struggle with the complexity and overlapping peaks in XRF spectra, limiting prediction accuracy.

Purpose of the Study:

• To develop a high-accuracy prediction model for soil heavy metal content using X-ray fluorescence (XRF) spectral data.
• To leverage the feature extraction capabilities of deep belief networks (DBN) and the regression power of tree-based models.
• To optimize model parameters for improved prediction performance.

Main Methods:

• A hybrid model integrating a deep belief network (DBN) for deep feature extraction from XRF spectra.
• Utilizing a tree-based model for regression prediction based on DBN-extracted features.
• Employing sparrow search algorithm and grid search for hyperparameter optimization.
• Applying the model to predict arsenic (As) and lead (Pb) content from overlapping spectral peaks.

Main Results:

• The hybrid DBN-tree-based model achieved high prediction accuracy for As and Pb.
• Coefficient of determination (R²) values reached 0.9884 for As and 0.9358 for Pb.
• Mean square errors were as low as 0.0011 for As and 0.0058 for Pb.
• The proposed model outperformed commonly used prediction models.

Conclusions:

• The combination of DBN and tree-based models offers a powerful approach for accurate soil heavy metal content prediction from XRF data.
• Deep feature extraction by DBN effectively addresses spectral complexity, while tree-based models provide robust regression.
• This hybrid methodology demonstrates significant potential for environmental applications requiring precise elemental analysis.