Proper soil taxonomic classification makes a significant contribution toward sustainable soil management, decision making, and soil conservation. For that, a quick, environmentally-friendly, non-invasive, cost-effective and reliable method for soil class assessment is desirable. As such, this study used NixPro color and portable X-ray fluorescence (pXRF) data to characterize seven different soil orders in Brazilian tropical soils, exploring the ability of three machine learning algorithms [Support Vector Machine with Linear Kernel (SVMLK), Artificial Neural Network (ANN), and Random Forest (RF)] with and without Principal Component Analysis (PCA) pretreatment for prediction of different soils at the order and suborder taxonomic levels under both dry and moist conditions. In total, 734 soil samples were collected from surface and subsurface horizons encompassing twelve suborders. The soil profiles were morphologically described and taxonomy classified per the Brazilian Soil Classification System and the approximate correspondence was made with the US Soil Taxonomy. Soil samples were separated into modeling (70%) and validation (30%) sub-datasets, overall accuracy and Cohen's Kappa coefficient evaluated model quality. Models generated from B horizon sample with pXRF and NixPro (moist samples) data combined delivered the best accuracy for order (81.19% overall accuracy and 0.71 Kappa index) and suborder predictions (74.35% overall accuracy and 0.65 Kappa index) through RF algorithm without PCA pretreatment. Summarily, the use of these two portable sensor systems was shown effective at accurately predicting different soil orders and suborders in tropical soils. Future works should extend the results of this study to temperate regions to corroborate the conclusions presented herein.
|State||Published - 2020|