Uncertainty Analysis of Artificial Neural Networks in Simulation of Saturated Hydraulic Conductivity Using Monte-Carlo Simulation

Hydraulic conductibility as a key soil property is essential for irrigation management purposes and plays an important role in understanding site-specific unsaturated water flow and transport processes. Since it cannot often be measured because of practical and/or cost-related reasons, data driven models such as artificial neural network may be applied for prediction of soil hydraulic properties. This paper aimed to assess uncertainty analysis in neural network prediction originated from different weights due to different training data sets. Here, we present a unique dataset that consists of 151 samples collected from arable land around Bojnourd City, containing of sand, silt and clay contents, saturated water content, Electrical Conductivity (EC), pH, real density, Organic matter (OM), Total Neutralizing Value (TNV) and bulk density (ρb). Bulk and real densities determined based on stepwise regression analysis as most important inputs to neural network model. Then a two layer perceptron neural network with 1000 different samples trained with some available transfer and training functions in Matlab. Results assessed with NUE as an integrated index defined as ratio of percentage of the observation coverage by 95 percent prediction uncertainty (95PPU) divided by average relative interval length of 95PPU. Uncertainty analysis results revealed that log-sigmoid and linear transfer functions with NUE values 0.57 and 0.59 performed better tan-sigmoid transfer function with NUE=0.25. Also all training functions (except of gradient descent training function) could predict saturated hydraulic conductivity with high reliability.