Rapid and Noninvasive Assessment of Atterberg Limits Using Diffuse Reflectance Spectroscopy

In this study, Atterberg limits were estimated using pedotransfer functions (PTFs), spectral-transfer functions (STFs), and spectral PTFs (SPTFs). In the PTF approach, basic soil properties were used to calibrate PTFs using stepwise multiple linear regression and support-vector regression (SVR) approaches. In the STF approach, spectral reflectance values across the visible to near-infrared (VNIR) and mid-infrared (MIR) regions were used to estimate Atterberg limits using the SVR and partial least squares regression (PLSR) approaches along with bootstrapping. Also, two locally weighted PLSR models were developed using covariance and correlation as weighting schemes. Finally, feature selection was combined with the PLSR (PLSRFS) to identify the optimum set of spectral features to use in the estimations. In the SPTF approach, both basic properties and geometrical shape factors of spectral absorption were used for the prediction of Atterberg limits. Comparison of the coefficient of determination (R-2) and root mean squared residual values for the validation data sets across different modeling approaches suggested that the PLSRFS model along with pooled VNIR + MIR data yielded the highest R-2 of 0.77 for the liquid limit. The PLSRFS model along with VNIR data alone yielded the highest R-2 of 0.70 for the plastic limit. A maximum R-2 value of 0.65 was obtained for the plasticity index using both the PLSR and SVR models on MIR data and the PLSRFS model on pooled VNIR + MIR data. This study shows that the STF approach may be adopted as an alternate to PTFs to estimate the Atterberg limits because it can account for up to 77% of the variance in the data.