National spectral data and learning algorithms for potentially toxic elements modelling in forest soil horizons

Forest ecosystems are among the main parts of the biosphere; however, they have been endangered from the significant elevation and harmful effects of air and soil pollutants, including potentially toxic elements (PTEs). The concentration of PTEs in forest soils varies not only laterally but also vertically with depth. Forest surface organic horizons are of particular interest in forest ecosystem monitoring due to their role as stable adsorbents of the deposited atmospheric substances. Therefore, the main purpose of this study was to conduct rapid examinations of forest soils PTEs (Cr, Cu, Pb, Zn, and Al), testing the capability of VIS--NIR spectroscopy coupled with machine learning (ML) techniques (partial least square regression (PLSR), support vector machine regression (SVMR), and random forest (RF)) and fully connected neural network (FNN), a deep learning (DL) approach, in forest organic horizons. One-thousand-and-eighty forested sites across the Czech Republic at two soil layers, defining the fragmented (F) and humus (H) organic horizons, were investigated (total 2160 samples). PTEs as well as total Fe and SOC, as auxiliary data, were conventionally and spectrally determined and modelled in the combined organic horizons (F + H) and in each individual horizon using the ML and DL algorithms. Results indicated that the concentration of all PTEs was higher in the horizon H compared to the F horizon. Although the spectral reflectance of samples tended to decrease with increased PTEs concentration. Strongly significant positive correlations between all PTEs and total Fe in all horizons were obtained, which were higher in the H and F + H horizons than the F horizon. The highest correlations of PTEs with the spectra were at 460--590~nm, which is mostly linked to the presence of Fe-oxide. These results show the importance of Fe for spectral prediction of PTEs. Cr and Al were the most accurately predicted elements, regardless of the applied learning technique. SVMR provided the best results in assessing the H horizon (e.g., R\(^2\) = 0.88 and root mean square error (RMSE) = 3.01~mg/kg, and R\(^2\) = 0.82 and RMSE = 1682.25~mg/kg for Cr and Al, respectively); however, FNN predicted the combined F + H horizons the best (R\(^2\) = 0.89 and RMSE = 2.95~mg/kg, and R\(^2\) = 0.86 and RMSE = 1593.64~mg/kg for Cr and Al, respectively) due to the larger number of samples. In the F horizon, almost no parameters were predicted adequately. This study shows that given the availability of larger sample sizes, FNN can be a more promising technique compared to ML methods for assessment of Cr and Al concentration based on national spectral data in the forests of the Czech Republic.