Multinational prediction of soil organic carbon and texture via proximal sensors

Abstract

Novel technologies help to monitor environmental impact of human activities, but tests involving datasets from several countries, encompassing a large variability of soil properties, are still scarce. This study utilized proximal sensors to predict soil organic carbon (OC) and soil texture of samples from Brazil, France, India, Mozambique, and United States of America. In total, 1,749 samples were analyzed by portable X‐ray fluorescence spectrometry (pXRF) and visible near‐infrared diffuse reflectance spectroscopy (Vis‐NIR). Samples were randomly split into modeling (70%) and validation (30%) datasets. Sand (R2 = 0.82), silt (0.87) and clay (0.84) predictions were very accurate, despite contrasting climates, soil parent materials, and weathering degrees. Soil OC predictions were similarly successful (0.72). pXRF was the optimal sensor for soil texture predictions. Proximal soil sensing can be successfully used with a global soil database offering a clean, rapid, and accurate alternative to estimate soil texture and OC across the globe.This article is protected by copyright. All rights reserved

Low-Cost Optical Sensors for Soil Composition Monitoring

“…In both cases, accuracies were below 100%, ranging from 70% to 90% with Vis and NIR data and 97% when X pXRF-ray information was added. Additional examples can be found identifying the % of sand and clay contents using VIS and NIR data in [ 40 , 41 ], or including pXRF data [ 42 ], but no information on accuracies is presented. The used metrics include R 2 , all of them below 0.88 [ 40 ], 0.90 [ 42 ], and 0.95 [ 41 ].…”

“…Additional examples can be found identifying the % of sand and clay contents using VIS and NIR data in [ 40 , 41 ], or including pXRF data [ 42 ], but no information on accuracies is presented. The used metrics include R 2 , all of them below 0.88 [ 40 ], 0.90 [ 42 ], and 0.95 [ 41 ]. Thus, we can affirm that the proposed method poses advantages beyond the state of the art in soil texture identification; see Table 7 for a summary of the aforementioned information.…”

“…Considering the different studied samples, in all cases, the authors used natural samples [ 24 , 27 , 28 , 39 , 40 , 42 ]. In this paper, artificial samples were generated to ensure a wide variety of cases to evaluate the sensor’s performance in different scenarios, having a variation from 0 to 100% clay and sand.…”

“…In our case, the class corresponded to the percentage of sand and clay in the sample. In [ 38 , 40 , 41 , 42 ], the authors also used the percentage of soil components as a predictor.…”

“…Among different data-processing techniques in the included studies, there is a wide variety of methods, such as convolutional neural network (CNN) [ 27 ], RF [ 39 ], partial least squares regression (PLSR) [ 38 , 40 , 41 ], and multivariate image analysis (MIA) [ 24 ]. In some cases, even two different methods are combined, such as recursive feature elimination (RFE) with RF [ 42 ]. None of the included ones used DA.…”