Comparing visible-near-infrared spectroscopy with classical regression pedotransfer functions for predicting near-saturated and saturated hydraulic conductivity of calcareous soils

Abstract

Measurement of hydraulic parameters, which are very important in soil–water and environmental studies, is difficult, time-consuming, and expensive. Therefore, the present study aimed to predict near-saturated and saturated hydraulic conductivity (Kψ) of calcareous soils by: i) developing stepwise multiple linear regression (SMLR)-based pedotransfer functions (PTFs) using basic soil properties, ii) using partial least square regression (PLSR)-based method and spectral reflectance data in the visible and near infrared (Vis-NIR) range (PLSR-spectra), and iii) developing SMLR-based functions that combine (or not) basic soil properties and effective Vis-NIR spectral reflectance bands (spectro-pedotransfer functions, SPTFs, and spectrotransfer functions, STFs). Hydraulic conductivity was determined in the field from water infiltration measurements with a tension disk infiltrometer at 15, 10, 5, and 0 cm tensions at 102 locations across calcareous soils of the Fars province, Iran. The spectral reflectance of soil samples taken at the same location was in the range of Vis-NIR (400–2500 nm), while basic soil properties included particle-size distribution and its related parameters, pH, electrical conductivity, soluble sodium, potassium, calcium, and magnesium, sodium adsorption ratio, soil organic matter content (SOM), calcium carbonate equivalent (CCE), aggregates mean weight diameter, and bulk density. The developed PTFs showed relatively acceptable performance (validation determination coefficient, R2, values of 0.55, 0.52, 0.56, and 0.63 for K15, K10, K5, and K0 predictions, respectively). The PLSR-spectra performed poorer showing validation R2 values of 0.39, 0.47, 0.40, and 0.48 for predicting K15, K10, K5, and K0, respectively. Yet, the PLSR-spectra predicted CCE and SOM contents excellently (R2 ≥ 0.80 and residual prediction deviation, RPD ≥ 2.5 for both calibration and validation datasets), and clay relatively well but not pH. The effective spectral reflectance bands (mostly substitute with SOM and CCE) were very useful as input variables to predict K5 and K0 when coupled with basic soil properties resulting in SPTFs. On the other hand, to predict K15 and K10, once any soil properties were added to SMLR models, the effective spectral reflectance bands were disappeared and consequently STFs were developed. Overall, we concluded that the accuracy of the studied approaches to predict Kψ were ranked as PTF > PLSR-spectra ≈ STF for predicting K15 and K10 and PTF > SPTF > PLSR-spectra for predicting K5 and K0.