ESTIMATING SOIL SOLUTION ELECTRICAL CONDUCTIVITY FROM TIME DOMAIN REFLECTOMETRY MEASUREMENTS USING NEURAL NETWORKS

Abstract

Time domain reflectometry (TDR) is a widely used method for measuring the dielectric constant (Ka) and bulk electrical conductivity (σa) in soils. The TDR measured σa and Ka can be used to calculate the soil solution electrical conductivity, σw. The σw, in turn, can be related to the concentration of an ionic tracer. Several models of the σw-σa-Ka relationship can be found in the literature. Most of these models require extensive calibration experiments in order to obtaining best-fit parameters. In this paper, we attempt to model the σw-σa-Ka relationship using neural networks (NN). We used TDR measured Ka and σa along with five different soil physical parameters (sand, silt, clay, and organic matter content and bulk density) measured in nine different soil types using three different σw levels in each soil type. In total, 2953 Ka and σa measurements were obtained. The NN estimated σw was found to have a root mean square error (RMSE) of 0.05-0.13dSm-1 for the nine different soil types whereas the RMSE of two traditional σw-σa-Ka models was 0.12-0.87dSm-1. Furthermore, the traditional models exhibited larger errors for low σa and Ka, whereas the NN estimated σw did not show any trend in the errors. A sensitivity analysis showed that the NN model was more sensitive to small changes in σa compared to Ka. Of the five soil physical parameters, the silt and clay content affected the σw-σa-Ka relationship the most. The results presented shows that using NN, the σw-σa-Ka relationship can be predicted using soil physical parameters without need for elaborate soil specific calibration experiments.