Using soil easily measured parameters for estimating soil water capacity: Soft computing approaches

The current study examines the applicability of six different soft computing approaches, gene expression programming (GEP), neuro-fuzzy (NF), support vector machine (SVM), multivariate adaptive regression spline (MARS), random forest (RF), and model tree (MT) techniques in modeling two important soil water capacity parameters, field capacity (FC) and permanent wilting point (PWP). Geometric mean particle-size diameter (dg), soil bulk density (BD), clay and silt obtained from 192 soil samples were introduced as input variables to the applied techniques and k-fold testing procedure was used for better comparison of the soft computing models. The best accuracy was provided by the NF models followed by the GEP, while the MT approach gave the worst estimates. The performances accuracies of the soft computing models in estimation of PWP parameter were higher than those in the FC estimation. Further, the soft computing approaches were compared with the traditional multi-variable linear regression (MLR) as well as the previously developed pedotransfer functions (PTFs) and the better FC and PWP estimates which confirms the superiority of the soft computing approaches. The NF model increased the performance of the best PTF (Aina-Periaswamy) by 33% with respect to GMER in FC estimation while the SI statistics of the best PTF (Ghorbani-Homaee) was decreased by 50% using the soft computing model. The performance of the best PTF (Aina-Periaswamy) with respect to GMER was increased by 74% in PWP estimation while the SI statistics of the best PTF (Dijkerman) was decreased by 99% using the soft computing model.