Prediction of soil agronomic and geochemical attributes: comparison of the performance of spectral ranges from proximal sensors

Abstract

Soil characterization provides a solid support for decision-making related to geochemical mapping, environmental monitoring, and food production. For that, quick, environmentally friendly, non-invasive, cost-effective, and reliable methods for soil chemical properties assessment are desirable. As such, this dissertation used proximal sensors like portable X-ray fluorescence (pXRF) spectrometry and Nix ProTM color sensor data to accurately predict soil properties in Brazil. The objectives were to: i) predict soil fertility properties in Brazilian Coastal Plains biome; ii) predict soil organic matter content via proximal sensors (pXRF and Nix ProTM); iii) predict elementary soil contents via USEPA 3051a through elementary data delivered by pXRF, evaluating samples preparation methods (field, post-field, air-dried fine earth, macerated, and macerated and sieved) and linear and non-linear regression methods. Four regression models - simple linear regression (SLR), stepwise multiple linear regressions (SMLR), support vector machine (SVM) with Linear Kernel and random forest (RF) - were tested for prediction of different soil agronomic attributes and assessment geochemical. The soil samples were collected in both surface and subsurface horizons in profiles of different soil classes, under several land uses, management practices, sampling sites, and with varying parent materials. Prediction models were built for surface, and subsurface horizons separately and combined for the following soil agronomic properties: pH (H2O), sum of bases (SB); cation exchange capacity at pH 7.0 (CEC), and base saturation (BS) (first chapter). For soil organic matter (SOM) (second chapter) and 28 elements (third and fourth chapters) samples from surface and subsurface horizons were combined for building the prediction models. Samples were scanned with the Nix ProTM in the laboratory under both dry and moist conditions, while with pXRF only in dry condition also in laboratory. Samples were randomly separated into 70% for training and 30% for testing the prediction models. The performance of the prediction models was evaluated by the metrics: R2, root mean square error (RMSE), normalized RMSE (NRMSE), mean absolute error (MAE), and residual prediction deviation (RPD). For soil agronomic properties, the results showed that SB was predicted with high accuracy (R2 = 0.82, RMSE = 1.02 cmolc dm–3, MAE = 1.17 and RPD = 2.3) using SVM models via pXRF data. Conversely, SOM was predicted with high accuracy using combined data from pXRF and Nix Pro™ (in moist soil samples) (R2 = 0.73, RMSE = 1.09% and RPD = 2.00) via RF models. Prediction of elemental contents commonly determined by the USEPA 3051a method via pXRF data after scanning samples treated as air-dried fine earth (<2 mm) is indicated. Since, it can provide better predictions compared to other sample preparation procedures indicated above. Machine learning algorithms (SVM and RF) performed better than SLR and SMLR for the prediction of Al, Ca, Cr, Cu, Fe, Mn, Pb, Sr, Ti, V, Zn, Zr, Ba, Bi, Cd, Ce, Co, Mg and Tl in tropical soils, whose R² and RPD values ranged from 0.52 to 0.94 and 1.43 to 3.62, respectively, as well as the lowest values of RMSE and NRMSE values (0.28 to 0.70 mg kg-1). The results reported in this dissertation represent alternative methods for reducing costs and time needed for assessing such soil properties data, supporting agronomic and environmental decision making.