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

The surface organic horizons in forest soils have been affected by air and soil pollutants, including potentially toxic elements (PTEs). Monitoring of PTEs requires a large number of samples and adequate analysis. Visible-near infrared (vis-NIR: 350-2500 nm) spectroscopy provides an alternative method to conventional laboratory measurements, which are time-consuming and expensive. However, vis-NIR spectroscopy relies on an empirical calibration of the target attribute to the spectra. This study examined the capability of vis-NIR spectra coupled with machine learning (ML) techniques (partial least squares regression (PLSR), support vector machine regression (SVMR), and random forest (RF)) and a deep learning (DL) approach called fully connected neural network (FNN) to assess selected PTEs (Cr, Cu, Pb, Zn, and Al) in forest organic horizons. The dataset consists of 2160 samples from 1080 sites in the forests over all the Czech Republic. At each site, we collected two samples from the fragmented (F) and humus (H) organic layers. The content of all PTEs was higher in horizon H compared to F horizon. Our results indicate that the reflectance of samples tended to decrease with increased PTEs concentration. Cr was the most accurately predicted element, regardless of the algorithm used. SVMR provided the best results for assessing the H horizon (R2 = 0.88 and RMSE = 3.01 mg/kg for Cr). FNN produced the best predictions of Cr in the combined F + H layers (R2 = 0.89 and RMSE = 2.95 mg/kg) possibly due to the larger number of samples. In the F horizon, the PTEs were not predicted adequately. The study shows that PTEs in forest soils of the Czech Republic can be accurately estimated with vis-NIR spectra and ML approaches. Results hint in availability of a large sample size, FNN provides better results.