Assessment of Soil Properties Using Spectral Signatures of Bulk Soils and Their Aggregate Size Fractions

Diffuse reflectance spectroscopy (DRS) is an emerging alternative to conventional laboratory methods of soil testing although poor estimation accuracies continue to limit its wide-spread adoption. Spectral reflectance values of different aggregate size fractions and their bulk soil have recently been shown to improve the estimation accuracy of soil textural fractions in the DRS approach. The objective of the present study is to assess this approach for estimating nine different soil parameters covering soil structure, soil chemical properties and micronutrient contents. We used three different chemometric models (PLSR: partial-least-squares regression, PLSRFS: PLSR with feature selection, and PLSRLW: locally-weighted PLSR) on three different spectral libraries (total of 1013 soil samples) to test this objective. On an average, 29% improvement in the coefficient of determination (R2) was achieved with the addition of aggregate spectra to bulk soil spectra. Improvement in the root-mean-squared error (RMSE) ranged from 5% to 48% for aggregate size distribution parameters, 7% to 40% for chemical properties, and 7% to 26% for micronutrient contents. These improvements were also observed to be a result of high correlation coefficients between estimated soil properties and the best-performing aggregate fraction spectra, which suggested that the spectral reflectance of individual aggregate size fractions holds important soil information for improving the performance of a DRS model. Segregation of soils into individual size fractions and the collection of corresponding spectra are additional burdens in the proposed approach. Nevertheless, the extent of improvements in the model performance observed in this study opens an opportunity to develop an automated system for separating soil aggregates and collecting spectra so that the information may readily be used for improving the performance of the DRS approach.