Microplastic content prediction in agricultural soils using a mixed variable selection combined with Transformer-LSTM...

Area of Science:

• Environmental Science
• Analytical Chemistry
• Agricultural Science

Background:

• Microplastic (MP) accumulation in agricultural soils threatens food safety and ecosystem stability.
• Developing accurate methods for predicting soil MP content is a critical research focus.

Purpose of the Study:

• To propose a novel predictive model for soil microplastic content.
• To enhance the accuracy and efficiency of microplastic detection in agricultural soils.

Main Methods:

• Confocal micro-Raman spectroscopy coupled with a Transformer-LSTM deep learning model.
• Integrated feature selection using synergy interval partial least squares (siPLS), successive projections algorithm (SPA), and least absolute shrinkage and selection operator (LASSO) (siPLS(SPA/LASSO)).
• Preprocessing of Raman spectra and cross-regional validation for generalization capability.

Main Results:

• The siPLS(SPA/LASSO) feature selection significantly improved prediction accuracy and spectral signal-to-noise ratio (SNR).
• The Transformer-LSTM model outperformed individual, traditional machine learning, and deep learning models.
• High prediction accuracies were achieved: PP (R²P = 0.9950), PE (R²P = 0.9974), and PS (R²P = 0.9977).
• The method demonstrated good generalization capability through cross-regional validation.

Conclusions:

• The hybrid model effectively predicts microplastic content in agricultural soil.
• This approach offers a robust solution for monitoring soil microplastic contamination.
• The study highlights the potential of advanced spectroscopic and deep learning techniques in environmental analysis.