Estimation of saturated hydraulic conductivity in spatially variable fields using various soft computing techniques

Abstract

Saturated hydraulic conductivity (Kfs) is one of the dominant and most essential soil hydrology characteristics, for understanding and duplicating various hydrological processes having environmental importance. Its estimation by (laboratory / field) method is cumbersome, time consuming and costly. In addition, the results may not be representative due to spatial variability of Kfs. This attracted researcher to address this problem by developing pedotransfer function (PTF), which estimate saturated hydraulic conductivity by using routinely measured soil properties. Objective of this study is to investigate the spatial variability of saturated hydraulic conductivity under different land use land cover by using Guelph permeameter, and to develop PTF for estimating Kfs, from soil index properties, by using soft computing techniques and thus evaluate the performance of these techniques by using statistical tests. Study is carried out at Solapur, India. Three sites (0.76ha) were identified which are having different land use land cover. The site is divided in to small grids of 10m X 10m, and observations were taken at corner of each such grid, at 15cm, 30cm and 45cm depth. In situ and laboratory tests were carried out to estimate Kfs and other basic index properties of soil. Observed data (In situ as well as laboratory) were preprocessed and then used for modeling purpose. Total dataset (Three sites, three depth with 100 sampling points; so overall 900 sample data) is segregated into six sub dataset (college, Mulegaon, Punanaka, 15cm, 30cm and 45cm). Each subset consisting of 300 observations is further split into two parts in six different ways (90% + 10%, 85% + 15%, 80% +20%, 75% + 25%, 70% + 30%, and 67% + 33%) to train the models and validate it, the combination which gives good results during training and validation is selected. For checking performance of model, various statistical parameters such as correlation coefficient (R), mean relative error(MRE), root mean square error (RMSE), normalized root mean square error (NRMSE) and Nash Sutcliffe efficiency (NSE) has been made. Scatter plots were used to evaluate the accuracies of the models. For deciding best model these checks are used, Value of R ~ 1, value and NSE ~ 1, MRE close to zero, and NRMSE is close to zero. Scatter plot point distribution should be around and close to 1:1 line. Maximum value of log saturated hydraulic conductivity was observed at Punanaka (3.842 m yr-1) and minimum value at college site (0.002myr-1). Standard deviation for Kfs was least at Punanaka (0.598m yr-1) and was maximum at college site (0.804myr- 1). Porosity has strong positive Correlation coefficient 0.9 whereas bulk density has strong negative correlation of 0.9. The performance of ELM model at all six subsets was found performing better than SVM and ANFIS model. NRMSE values of ELM model (training: testing) were found [0.02:0.06, 0.07:0.09, 0.02:0.07, 0.03:0.08, 0.07:0.10 and 0.008:0.04] at college site, Mulegaon site, Punanaka site, 15cm depth, 30cm depth and 45cm depth respectively. Saturated hydraulic conductivity was found varying spatially, land use land cover has influence on Kfs and it found declining with depth. College station has shown more variability in Kfs also variation of Kfs was found more at 45cm depth. Maximum Standard deviation was found in college site and minimum standard deviation was found at Punanaka site. Variability of porosity, bulk density and particle density was found insignificant in logarithmic scale. Soil particle size was found declining with depth. Porosity has shown strong positive correlation with Kfs, whereas bulk density has shown strong negative correlation. Performance of ELM model was found excellent in all six sub data set both during training and testing. Performance of SVM and ANFIS was not found satisfactory during testing although they are within acceptable accuracy. Saturated hydraulic conductivity has shown spatial variation; it was varying along depth as well as lateral and longitudinal direction, generally Kfs was found decreasing with depth. Kfs was found decreasing down the slope. ELM model outperformed other two models tried in this study (ANFIS and SVM). Texture of soil was founddeclining from coarse to fine with depth at majority of sampling location. Mean value of Kfs was found more at Punanaka site (15cm depth) as compared to other two sites.