A comparison of measured and estimated saturated hydraulic conductivity of various soils in the Czech Republic

Abstract

The study aims to indirectly determine the saturated hydraulic conductivity (Ks). The applicability of recently-published pedotransfer functions (PTFs) based on a machine learning approach has been tested, and their performance has been compared with well-known hierarchical PTFs (computer software Rosetta) for 126 soil data sets in the Czech Republic. The quality of estimates has been statistically evaluated in comparison with the measured Ks values; the root mean squared error (RMSE), the mean error (ME) and the coefficient of determination (R2) were considered. The eight tested models of PTFs were ranked according to the RMSE values. The measured results reflected high Ks variability between and within the study areas, especially for those areas where preferential flow occurred. In most cases, the tested PTFs overestimated the measured Ks values, which is documented by positive ME values. The RMSE values of the Ks estimate ranged on average from 0.5 (coarse-textured soils) to 1.3 (medium to fine-textured soils) for log-transformed Ks in cm/day. Generally, the models based on Random Forest performed better than those based on Boosted Regression Trees. However, the best estimates were obtained by Neural Network analysis PTFs in Rosetta, which scored for four best rankings out of five.

Keywords

• soil parameter
• soil texture
• soil property
• prediction
• comparative assessment