Prediction of soil texture distributions using VNIR-SWIR reflectance spectroscopy

Reflectance spectroscopy provides an alternate method to non-destructively characterize key soil properties. Differentapproaches, including chemometrics techniques or specific absorption features, have been proposed to estimate soilproperties from visible and near-infrared (VNIR, 400-1200 nm) and shortwave infrared (SWIR, 1200-2500 nm)reflectance domains. The main goal of this study was to test the performance of two distinct methods for soil textureestimation by VNIR-SWIR reflectance measurements: i) the Continuum Removal (CR) technique that was used tocorrelate specific spectral absorption features with clay, silt and sand content, and ii) the Partial Least-SquaresRegression (PLSR) method, which is a classical statistical multivariate technique, that uses the full-spectrum data. Atthis aim, the surface reflectance of 100 soil samples collected from different sites in Sicily and covering a wide rangeof textures were measured in laboratory using an ASD FieldSpec Pro spectroradiometer (350-2500 nm). The resultsof our work indicated that the PLSR technique performed better than the CR approach. Particularly, the assessment ofsoil texture accuracy performed using root mean squared error (RMSE) and coefficient of determination (R2) showedthat the CR approach allowed to obtain a moderate prediction only for the clay texture fraction. Differently, usingPLSR technique, the levels of accuracy resulted high for the clay fraction (RMSE=5.8%, R2=0.87) and satisfactoryfor the sand (RMSE=7.7%, R2=0.80) and silt fractions (RMSE=7.2%, R2=0.60). Moreover the use of PLSR techniqueallowed to establish the “key wavelengths” of the investigated spectrum range that should be considered “essential”for the prediction of soil textures, suggesting the optimal settings for airborne or satellite sensors usable in the futurefor accurate mapping of soil textures.