Proximal sensor data fusion and auxiliary information for tropical soil property prediction: Soil texture

• Published in: Geoderma Vol. 422; p. 115936
• Main Authors: Andrade, Renata, Mancini, Marcelo, Teixeira, Anita Fernanda dos Santos, Silva, Sérgio Henrique Godinho, Weindorf, David C., Chakraborty, Somsubhra, Guilherme, Luiz Roberto Guimarães, Curi, Nilton
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 15-09-2022
• Subjects: Machine learning; NixProTM; Prediction models; pXRF; Tropical soils; Vis-NIR; Vis-NIR; D; NixProTM; Bin; RPD; pXRF; I; RMSE; M; O; CS; SD; CV; Abs; RF; Smo; Prediction models; SH; U; Machine learning; Tropical soils; SO; PM; FD
• ISSN: 0016-7061; 1872-6259
• DOI: 10.1016/j.geoderma.2022.115936

•Does combining proximal sensor always bring synergy for predicting soil texture?•Vis-NIR, NixProTM, and pXRF data were tested for soil texture predictions.•PXRF provided key information to create accurate prediction models.•Smoothed Vis-NIR data preprocessing and dry NixProTM condition positively influenced the results.•Soil texture and textural classes were accurately predicted by at least one approach.
Soil texture is a primary variable influencing many soil chemical-physical-biological processes, providing important information for decision-making regarding sustainable soil management. The standard traditional methods for determining soil texture, however, are performed manually and are time-consuming, costly, and generate chemical wastes. As an alternative, portable X-ray fluorescence (pXRF) spectrometry and visible near-infrared spectroscopy (Vis-NIR) have been increasingly used worldwide to predict soil attributes. Other sensors (e.g., NixProTM color sensor) are also promising, but less evaluated to date. Thus, investigations towards proximal sensor data fusion for prediction of soil textural separates (clay, silt, and total, coarse, and fine sand contents) and soil textural classes (loam, loamy sand, etc) in tropical soils are rare. Therefore, this study aimed to evaluate proximal sensor data for predicting soil particle size fractions and soil textural classes (both Family particle size classes and USDA soil texture triangle) via random forest algorithm in tropical regions. A total of 464 soil samples were collected from A (n = 208) and B (n = 256) horizons in Brazil. Soil samples were submitted to laboratory analyses for soil texture and proximal sensor (pXRF, Vis-NIR, and NixProTM) scanning. Samples were randomly split into 70% for modeling and 30% for validation. The best approach varied according to the predicted attribute; however, pXRF data were key information for soil texture prediction accuracy. The best results delivered highly accurate predictions via the aforementioned proximal sensors for rapid assessment of soil texture (total sand R2 = 0.84, RMSE = 7.60%; silt 0.83, 6.11%; clay 0.90, 5.64%; coarse sand 0.87, 6.30%; fine sand 0.82, 5.27%). Categorical prediction accuracy for soil textural classes (Family particle size classes, overall accuracy = 0.97, Kappa index = 0.95; USDA soil texture triangle, 0.83, 0.73) was enhanced when the predictions were made by soil order sub-datasets. Smoothed Vis-NIR preprocessing and dry NixProTM color data positively influenced the results. The results reported here represent alternatives for reducing costs and time needed for evaluating soil texture, supporting agronomic and environmental strategies in Brazilian conditions. Further works should extend the results of this study to temperate regions to corroborate the conclusions presented herein regarding the fusion of these three proximal sensors.