Parent material distribution mapping from tropical soils data via machine learning and portable X-ray fluorescence (pXRF) spectrometry in Brazil

• Published in: Geoderma Vol. 354; p. 113885
• Main Authors: Mancini, Marcelo, Weindorf, David C., Silva, Sérgio Henrique Godinho, Chakraborty, Somsubhra, Teixeira, Anita Fernanda dos Santos, Guilherme, Luiz Roberto Guimarães, Curi, Nilton
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 15.11.2019
• Subjects: B horizons; Brazil; Digital soil mapping; discriminant analysis; fluorescence; land use; Machine learning; Parent material; Pedology; prediction; Prediction models; principal component analysis; soil sampling; support vector machines; Tropical soils; X-radiation; X-ray fluorescence spectroscopy; Brazil; Pedology; RF; Digital soil mapping; Prediction models; Machine learning; LDA; Tropical soils; Parent material; SVM; PM; PCA
• DOI: 10.1016/j.geoderma.2019.113885
• ISSN: 0016-7061

Knowledge about parent material (PM) is crucial to understand the properties of overlying soils. Assessing PM of very deep soils, however, is not easy. Previous studies have predicted PM via proximal sensors and machine learning algorithms, but within small areas and with low soil variety. This study evaluates the efficiency of using portable X-ray fluorescence (pXRF) spectrometry together with machine learning algorithms in a large area populated with a wide variety of soil classes and land uses to build accurate PM distribution maps from soil data. Samples from A and B horizons were collected from 117 sites, totaling 234 samples, along with representative PM samples of the predominant PMs in the region. Elemental contents of PM and soil samples were obtained using pXRF. Random Forest (RF), Support Vector Machine (SVM) and Linear Discriminant Analysis (LDA) models built with and without Principal Component Analysis (PCA) were used to predict PMs from A and B horizon samples separately. For validation, PM was identified in 23 different spots and compared with the predicted PM via overall accuracy and Kappa coefficient. Maps built with models excluding PCA had an overall accuracy ranging from 0.87 to 0.96 and kappa coefficient ranging from 0.74 to 0.91. Maps generated with PCA based models reached 100% overall accuracy. Prediction of PM using pXRF with samples from either A or B horizon and machine learning algorithms offers solid results even when applied in large areas with high land use and soil class variability. •Soil parent material can be accurately predicted via pXRF analysis of soils.•Maps from B horizon samples were slightly superior to A horizon samples.•Overall map predictive accuracy was strong.