Isfahan University of Technology Dryland Soil Research (DLSR) 3115-9486 2 1 2025 03 01 A data-driven approach to predict soil hydraulic conductivity: GMDH compared with ANN and multiple regression 65 79 3734 10.47176/dlsr.02.01.1047 EN Mehdi Rahmati 1 Department of Soil Science and Engineering, Faculty of Agriculture, University of Maragheh, Maragheh, Iran, 2 Forschungszentrum Jülich GmbH, Institute of Bio- and Geosciences: Agrosphere (IBG-3), Jülich, Germany 0000-0001-5547-6442 Journal Article 2025 09 27 Accurate estimation of saturated hydraulic conductivity (Ks) is essential for soil and water management, yet the reliability of the pedotransfer functions (PTFs) is often overlooked.Tthis study compares the predictive performance and the robustness of field estimates of Ks obtained by three types of PTFs: Multiple Regression (MR), Artificial Neural Networks (ANN) and the Group Method of Data Handling (GMDH) developed using 134 soil samples collected in north-western Iran, which is a region with semi-arid conditions and mixed agricultural uses whereas the dataset encompasses a wide range of structural and textural variationsto Ks prediction. In addition to traditional soil properties soil moisture deficit compared to the optimum value at the time of sampling, θd, was used as a proxy indicator of soil structural condition. Model precision was evaluated using Root Mean Square Error (RMSE) and the Nash–Sutcliffe efficiency; reliability was determined through repeated data splitting. Even though ANN provided good accuracy for the training set, its performance for the validation set was inconsistence. MR produced consistent, albeit limited performances over both the training and validation subsets. Conversely, GMDH appears to strike a good compromise between prediction accuracy, reliability, parsimony of the predictor set, and texture versus structure variables. The results point to the importance of including structural measures such as θd in PTF development and provide a basis for considering model repeatability a long with accuracy. In general, the results indicate that GMDH is a robust and feasible technique to develop accurate PTFs for Ks predictions with limited amount of data.

https://dlsr.iut.ac.ir/article_3734_9d752cb08ef466fc480fba981cfa44a1.pdf