A comparative study of MIR and NIR spectral models using ball-milled and sieved soil for the prediction of a range soil physical and chemical parameters

Abstract

This study evaluated the influence on predicted physical and chemical parameters of soil particle sizes commonly used in the infrared spectra acquisition, < 0.100 mm (ball-milled) and < 2 mm for MIR and NIR ranges, respectively. The influences were evaluated through the accuracy (RMSEP and RPIQ) results and the chemical information extracted by multivariate classification and regression models. For this a national population of soils containing 888 samples from 225 modal soil profiles, each with the reference values of sand, silt, clay, pH_(CaCl2), pH_(Water), total carbon, organic carbon (OC), cation exchange capacity, nitrogen, aluminium and bulk density, was used. Spectra were collected in MIR and NIR ranges using samples with both particle sizes. For each soil attribute, 29 random calibration and validation sets were generated and SVM, PLS and Cubist regression models were built. This same strategy was used to classify the soil samples according to their respective horizons (1 or 2-7) using SVM, PLS-DA and random forest algorithms. Results obtained by the randomised calibration and validation set did not present positive or negative bias on the RMSEP and RPIQ values based on soil particle sizes. In general, random variations of the RMSEP and RPIQ values were observed for all soil attributes. In addition, ball-milled and < 2 mm spectral models did not present large differences in both accuracy parameters simultaneously. The median Matthews correlation coefficient values calculated by the classification models showed minor variations of 2.61% and 0.65% for samples from both particle sizes in MIR and NIR ranges, respectively. The 'Variable Importance in Projection' or VIP scores, calculated by PLS and PLS-DA models, did not show any large variation in the chemical information extracted from MIR and NIR spectra for models built using samples from both particle sizes. The results from this study show that scanning ball-milled or < 2 mm sieved soil samples will result in spectra models in MIR and NIR ranges with the same accuracy and same chemical information. This suggests there is a big potential to eliminate the ball-milling sample step in soil laboratories that use MIR and NIR vibrational spectroscopy techniques to predict soil attributes, thereby reducing the time and costs associated with soil analysis.

Keywords: Irish soil information system; Machine learning; PLS; Soil analysis; Soil particle size; Support Vector Machine.