Soil texture prediction in tropical soils: A portable X-ray fluorescence spectrometry approach

• Published in: Geoderma Vol. 362; p. 114136
• Main Authors: Silva, Sérgio Henrique Godinho, Weindorf, David C., Pinto, Leandro Campos, Faria, Wilson Missina, Acerbi Junior, Fausto Weimar, Gomide, Lucas Rezende, de Mello, José Márcio, de Pádua Junior, Alceu Linares, de Souza, Igor Alexandre, Teixeira, Anita Fernanda dos Santos, Guilherme, Luiz Roberto Guimarães, Curi, Nilton
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 15.03.2020
• Subjects: Brazilian soils; prediction; Prediction models; Proximal sensors; Soil particle size; soil texture; tropical soils; X-ray fluorescence spectroscopy; MAE; R2; CV; RF; Prediction models; Soil particle size; pXRF; GLM; RMSE; SVM; Proximal sensors; Brazilian soils
• DOI: 10.1016/j.geoderma.2019.114136
• ISSN: 0016-7061

•Soil texture of 1565 samples was determined and analyzed with pXRF in Brazil.•pXRF data was successfully predicted soil texture through SVM and RF algorithms.•Predictions were accurate independently of soil class, parent material, and horizon.•Fe and Al positively correlated to clay content due to Fe- and Al-oxides in soils.•Si positively correlated to sand, due to dominance of quartz in this fraction. Soil texture is an important feature in soil characterization, although its laboratory determination is costly and time-consuming. As an alternative, this study aimed at predicting soil texture from portable X-ray fluorescence (pXRF) spectrometry data in Brazilian soils. 1565 soil samples (503 from superficial and 1062 from subsuperficial horizons) were analyzed in the laboratory for soil texture and scanned with the pXRF. Elemental contents determined by pXRF were correlated with soil texture and used to calibrate regression models through the generalized linear model (GLM), support vector machine (SVM), and random forest (RF) algorithm. Models were created with 70% of the data using three datasets: i) only superficial horizon data; ii) only subsuperficial horizon data; and iii) data from both horizons. Validation was performed with 30% of the data. Clay content was positively correlated with Fe (0.79) and Al2O3 (0.41) reflecting the great residual concentration of Fe- and Al-oxides in this fraction. This same fraction correlated negatively with SiO2 (-0.75), while the sand fraction correlated positively with SiO2 corresponding to quartz dominance in the sand fraction of Brazilian soils. For the separated superficial and subsuperficial horizon datasets, SVM promoted the best predictions of clay (R2 0.83; RMSE = 7.04%) and sand contents (R2 0.87; RMSE = 9.11%), while RF provided the best results for silt (R2 0.60; RMSE = 6.33%). When combining both datasets, RF was better for sand prediction (R2 0.73; RMSE = 5.79%), while SVM promoted better predictions for silt (R2 0.72; RMSE = 5.77%) and clay (R2 0.84; RMSE = 7.08%). Elemental contents obtained by pXRF are capable of accurately predicting soil texture for a great variety of Brazilian soils.