Estimating Atterberg limits of soils from reflectance spectroscopy and pedotransfer functions

Abstract

Atterberg limits are broadly used for engineering and geology purposes as well as in agricultural and environmental applications. Laboratory methods used for their determination are, however, laborious, destructive and tool dependent. The aim of this study was to test the feasibility of using visible near-infrared spectroscopy (vis–NIRS) as a fast and accurate alternative to the conventional measurements of Atterberg limits (LL and PL) and the PI for 229 geographically diverse soil samples originating from 24 countries. Three types of calibration techniques including Partial Least Squares (PLS) regression, Artificial Neural Networks (ANN) and Support Vector Machines (SVM) were applied to the spectral data. The performance of the best vis–NIRS models was validated using 45 independent samples and compared with two existing and one newly developed pedotransfer functions (PTF). The application of SVM yielded marginally better predictive ability than PLS and ANN for all modelled properties. The SVM models estimated LL, PL, and PI with root mean squared error (RMSE) of 7%, 5% and 7%, respectively. The newly developed PTF gave slightly better estimations than the existing ones, with RMSE values of 8%, 6%, and 6%, respectively for LL, PL, and PI. Furthermore, in terms of the sample swelling class, the SVM model correctly classified 31 of the 45 samples, compared to 34 samples for the best PTF. The results indicate a great potential of vis–NIRS for reliable estimates of Atterberg limits for soil samples of large geographical and mineralogical diversity.