Feasibility of proximal sensing for predicting soil loss tolerance

Abstract

Soil loss tolerance (T-value) is a vital parameter in soil conservation programs aiming to reduce erosion. Measuring the T-value is expensive, difficult, and time-consuming. No study was found that investigated the capability of the spectroscopy approach in visible (Vis) and near-infrared (NIR) regions to predict the T-value. Hence, we aimed to predict the T-value by the Vis-NIR spectroscopy. 60 soil profiles were excavated to measure the T-values according to the soil thickness method (STM), along with physico-chemical attributes and Vis-NIR spectra in the calcareous soils of southern Iran. The T-value was predicted using Vis-NIR reflectance spectra via applying different modeling approaches, including partial least square regression (PLSR), principal component regression (PCR), multiple linear and non-linear regressions-based spectrotransfer functions (MLR-STF and MNLR-STF), and support vector regression (SVR). The Vis-NIR reflectance spectroscopy can detect functional groups of organic matter and carbonate components in soil, so if the T-value is significantly correlated with these parameters, it is evidence that the Vis-NIR spectroscopy may be an effective approach in predicting the T-value. Hence, results revealed that the soil organic matter and calcium carbonate equivalent were significantly correlated (p < 0.05) with the T-value by correlation coefficients (r) of 0.77 and 0.32, respectively. Among the applied SVR algorithms to predict the T-value by Vis-NIR spectra, the Epsilon type with linear kernel algorithm (Epsilon-SVR-L) showed the best performance. The T-value was predicted with acceptable accuracy using the Vis-NIR spectroscopy and applying the PLSR, PCR, MLR-STF, MNLR-STF, and Epsilon-SVR-L models with the cross-validation R² values of 0.60, 0.57, 0.61, 0.61, and 0.64, respectively. The reflectance values at wavelengths of 420, 564, 698, 1098, 1407, 1899, 1939, 2139, 2259, 2342, and 2456 nm were recognized as the most effective and predictive bands to predict the T-value and appeared in both developed STFs. Considering accuracy, simplicity, and applicability, the developed MLR-STF is recommended to predict the T-value and recognize eroded regions to conserve the soil resources of calcareous soils.