Combining Impedance Spectroscopy and Neural Networks to Improve Dielectric Water Content Measurements.

Accurate measurements of water content in soils with low electrical conductivity and minimal charged clays are straightforward.  Challenges to accuracy arise in soils with considerable electrical conductivity, either native or from irrigation water, and with large amounts of highly charged clays.  Under these conditions soil permittivity may display exceptionally high values due to interfacial polarization phenomena (Maxwell-Wagner). As a result, many dielectric sensors either fail to give data or report values well outside reason.  Conventional wisdom suggests increasing the measurement frequency and operating in a range where the above phenomena typically relax. However, progressively larger fractions of bound water become invisible at higher frequency, so there are limitations to this approach. Recent studies using Fast Fourier Transforms and neural networks have shown that broadband dielectric measurements may provide a much more accurate determination of water content compared to conventional single-frequency methods.  However, this approach is costly both monetarily and computationally.  The objective of this study is to determine whether a simplified, low-cost, multi-frequency sensor approach, including the measurement of temperature and electrical conductivity (EC), combined with neural network analysis, would improve the accuracy of water content measurements in salt affected or clayey soils.  After sensor calibration using the neural network, several soils and soilless substrates were tested and compared to conventional measurement methods.  Data showed considerable improvement of the multi-frequency measurements over conventional approaches.