Local modeling approaches for estimating soil properties in selected Indian soils using diffuse reflectance data over visible to near-infrared region

Publication date: 1 September 2018
Source:Geoderma, Volume 325
Author(s): Abhinav Gupta, Hitesh B. Vasava, Bhabani S. Das, Aditya K. Choubey
Robust calibration algorithms are needed for the accurate assessment of soil properties in the diffuse reflectance spectroscopy (DRS) approach. Despite several studies on different calibration algorithms, the prediction accuracy of soil properties using DRS need to be improved. Specifically, the utility of local modeling approaches for small spectral libraries is less examined compared with global modeling approaches. In this study, we compared global modeling approaches such as partial-least-squares regression (PLSR), lasso, ridge regression with several locally-weighted PLSR (PLSRLW) approaches. We also examined seven different distance measures: Euclidean distance, covariance-based distance, correlation-based distance, surface difference spectrum, information-based distance, optimized principal component Mahalanobis, and locally-linear embeddings used in the PLSRLW approach for their effectiveness in modeling soil properties using DRS. A total of 954 soil samples were collected from three different states of India: West Bengal, Odisha, and Karnataka. Five soil properties such as sand content, clay content, soil organic carbon (SOC) content, extractable iron (Fe) content and extractable zinc (Zn) content were predicted using reflectance spectra over 350–2500 nm. Root-mean-squared error (RMSE) and the coefficient of determination (R²) were used as performance statistics. Among the global modeling approaches, lasso performed better than the PLSR although it is computationally more intensive than the PLSR. In general, the correlation-based PLSRLW performed significantly better than the global approaches. Specifically, the test R² values increased from 0.66 to 0.72 for prediction of sand content, from 0.59 to 0.70 for prediction of SOC content, and from 0.70 to 0.74 for prediction of Fe content by using the PLSRLW over PLSR. We also used depth of absorption peak of spectra at approximately 1400, 1900 and 2200 nm for mineralogical characterization of soil samples. The results suggested that the improvement in prediction accuracy of soil properties using the PLSRLW was achieved because calibration samples which had same mineralogy as the test sample were given higher weights. These results suggest that the prediction accuracy of soil properties may also be improved in small spectral libraries if an appropriate local modeling scheme is selected.



https://ift.tt/2GusU5j