AN ARTIFICIAL NEURAL NETWORKS MODEL FOR ESTIMATING SOIL HYDRAULIC PROPERTIES FROM REMOTE SENSING DATA

Spatially distributed soil hydraulic properties data is required for spatially distributed hydrologic models. Limited availability of measured soil hydraulic data has prompted development of empirical models for estimating soil hydraulic properties from other available or easily measurable soil properties such as soil texture and bulk density. Even such empirically estimated soil hydraulic properties are not spatially distributed and thus spatial interpolation of measured and estimated data is required to satisfy the need of spatially distributed models. This study aims to develop, train and validate an artificial neural networks (ANN) model for estimating soil hydraulic properties in a spatially distributed format from Landsat ETM+ data for the agricultural Kharj area in arid central Saudi Arabia. Ninety sampling sites in the study area were specified by a GIS-based random spatial sampling procedure and their locations in the field were determined by high accuracy GPS. Standard procedures were used to collect the soil samples from the upper 5cm and to measure unsaturated hydraulic conductivity and water retention curve parameters. Coordinates of the sampling sites were used to identify the corresponding pixels in the rectified Landsat ETM+ image. The spectral reflectance were normalized by dividing the digital values of band 1, 3, 4, 5 and 7 by the digital value of band 2 to account for differences in illumination, surface roughness and weather conditions. The dataset were randomly split into exclusively two equal subsets for model training and validation. The ANN technique was combined with the bootstrab technique to generate uncertainty estimates. Model accuracy and reliability was evaluated using root mean square error (RMSE) between predicted and observed values of the training and validation data set, respectively.