Application of proximal sensors for the prediction of soil classes and attributes in Brazil

Abstract

Knowledge of soil properties makes a significant contribution towards sustainable soil management, decision making, and soil conservation. For that, a quick, environmentally friendly, non-invasive, cost-effective, and reliable method for soil properties assessment is desirable. As such, this study used portable X-ray fluorescence (pXRF) spectrometry, visible near-infrared spectroscopy (Vis-NIR) and NixProTM color sensor data to characterize 1019 Brazilian tropical soils samples, exploring the ability of six machine learning algorithms [ordinary least squares regression (OLS), Support Vector Machine with Linear Kernel (SVMLK), Cubist Regression (CR), XGBoost (XGB), Artificial Neural Network (ANN), and Random Forest (RF)] for prediction of different soil properties. The soil samples were collected in both surface and subsurface horizons of different soil classes, under several land uses, and with varying parent materials. Numerical prediction models were built for surface, and subsurface horizons separately and combined for the following soil properties: total nitrogen (TN), cation exchange capacity (CTC), soil organic matter (SOM), and soil texture (total sand, silt, clay, coarse sand, and fine sand contents). The study also encompasses the categorical prediction of properties, such as soil taxonomic classification at order and suborder levels, and soil textural classes (complete and simplified textural triangles). The NixProTM color sensor data were scanned under both dry and moist conditions. Four preprocessing methods were applied on the raw Vis-NIR spectra: first derivative, absorbance, smoothed, and binning. Samples were randomly separated into 70% for modeling and 30% for validation. The best approach varied according to the predicted soil property. However, pXRF data were a key information for the accuracy of soil properties prediction, followed by Vis-NIR spectra, and then, NixProTM color data. The results showed the increase in accuracy via fusion of proximal sensors data for soil properties prediction: TN (R2 = 0.50), CEC (0.75), SOM (0.56), total sand (0.84), silt (0.83), clay (0.90), coarse sand (0.87), fine sand (0.82), soil order (overall accuracy = 81.19%), soil suborder (74.35%), Family particle size classes (96.55%), and USDA soil texture triangle (82.75%). The results reported in this study for the tropical soils represent alternatives for reducing costs and time needed for assessing such soil properties data, supporting agronomic and environmental strategies.