Modelling potentially toxic elements in forest soils with vis-NIR spectra and learning algorithms

Area of Science:

• Environmental Science
• Soil Science
• Analytical Chemistry

Background:

• Forest soils are impacted by air and soil pollutants, including potentially toxic elements (PTEs).
• Conventional monitoring of PTEs is labor-intensive and costly.
• Visible-near infrared (vis-NIR) spectroscopy offers a rapid, cost-effective alternative for chemical analysis.

Purpose of the Study:

• To evaluate the capability of vis-NIR spectroscopy coupled with machine learning (ML) and deep learning (DL) for assessing PTEs (Cr, Cu, Pb, Zn, Al) in forest organic horizons.
• To compare the performance of different ML/DL algorithms (PLSR, SVMR, RF, FNN) for PTE prediction.
• To investigate the influence of organic horizon type (fragmented vs. humus) on PTE prediction accuracy.

Main Methods:

• Collected 2160 samples from 1080 forest sites across the Czech Republic, analyzing fragmented (F) and humus (H) organic layers.
• Acquired vis-NIR spectra (350-2500 nm) for all samples.
• Applied ML (PLSR, SVMR, RF) and DL (FNN) techniques to calibrate spectra with measured PTE concentrations.

Main Results:

• PTE content was higher in the H horizon compared to the F horizon.
• Sample reflectance generally decreased with increasing PTE concentration.
• Chromium (Cr) was the most accurately predicted element. Support Vector Machine Regression (SVMR) yielded the best results for the H horizon (R² = 0.88).
• Fully Connected Neural Network (FNN) provided the best predictions for Cr in combined F+H layers (R² = 0.89).
• PTEs in the F horizon were not adequately predicted by the tested methods.

Conclusions:

• Vis-NIR spectroscopy combined with ML/DL approaches can accurately estimate PTEs in forest organic horizons of the Czech Republic.
• The predictive accuracy for PTEs varies between organic layers, with better results for the H horizon.
• Large sample sizes, particularly when combined with DL methods like FNN, enhance prediction accuracy for PTEs in forest soils.