Soil subgroup prediction via portable X-ray fluorescence and visible near-infrared spectroscopy

Lucas Benedet,Wilson Missina Faria,Sérgio Henrique Godinho Silva,Marcelo Mancini,Luiz Roberto Guimarães Guilherme,José Alexandre Melo Demattê,Nilton Curi

DOI: https://doi.org/10.1016/j.geoderma.2020.114212

IF: 6.1

2020-04-01

Geoderma

Abstract:Recently, portable X-ray fluorescence (pXRF) spectrometer and visible near-infrared (Vis-NIR) spectroscopy are increasingly being applied for soil types and attributes prediction, but a few works have used them combined in tropical regions. Thus, this work aimed at analyzing models' performance when predicting soil types at subgroup taxonomic level via pXRF and Vis-NIR separately and together. 315 soil samples were collected in both A and B horizons in three important Brazilian states. Samples undergone laboratorial analyses for soil classification and were submitted to pXRF and Vis-NIR (350–2500 nm) analyses. Vis-NIR spectral data preprocessing was evaluated utilizing Savitzky-Golay (WT) and Savitzky-Golay with Binning (WB) methods. Four classification algorithms were employed in modeling: Support Vector Machine with Linear (SVM-L) and Radial (SVM-R) kernel, C5.0, and Random Forest (RF). Predictions were made using only B horizon and using A + B horizon data. Overall accuracy and Cohen's Kappa index evaluated model quality. Both sensors displayed efficacy in soil types prediction. A + B horizons data combined using pXRF + Vis-NIR via SVM-R (WT and WB) delivered accurate predictions (89.32% overall accuracy and 0.75 Kappa index), but the best predictions were achieved using only B horizon data via pXRF with RF, pXRF + Vis-NIR (WT) with RF, pXRF + Vis-NIR (WB) with C5.0, and pXRF + Vis-NIR (WB) with RF (89.23% overall accuracy and 0.80 Kappa index). For tropical soils, soil subgroup prediction using only B horizon data obtained by pXRF in tandem with RF algorithm may be a viable alternative to assist in soil classification, especially when the acquisition of Vis-NIR is not possible.

soil science