Machine learning-based modeling of saturated hydraulic conductivity in soils of tropical semi-arid zone of India

Saturated hydraulic conductivity (K) is the major parameter that affects the movement of water and solutes in soil strata. Although one can estimate the K directly by using various field or laboratory methods, they turn out to be more time-consuming and painstaking while characterizing the spatial variability of K. For this reason, some recent researches employ indirect approaches such as pedotransfer functions (PTF) and surface modeling methods for estimating K of several scales. Pedotransfer functions are often developed by relating the K with readily available soil properties such as bulk density, porosity, sand content, silt content, and organic material. The present research explores the suitability of Extreme Learning Machine (ELM) in developing PTF's for K by using basic soil properties. In-situ field tests and laboratory experiments on collected samples were performed to acquire the datasets necessary for the analysis. Three competitive soft computing approaches, namely the ELM, Support Vector Machine (SVM), and Adaptive Neuro-Fuzzy Inference System (ANFIS) based on Fuzzy C-means Clustering optimized by Genetic Algorithm were exercised for developing the K models. Further, the performance of these approaches in modeling K was evaluated using various statistical mertics. The performance of ELM was found to be good in comparison to the other two models, with sufficiently good NSE values. The ELM model provided K predictions at the Murarji Peth and Punanaka sites with an NSE of 0.90 and 0.83, respectively, while at the Mulegoan site, the ANFIS model was better with R = 0.80 and NSE = 0.64.