HR: 0800h
AN: H51F-0845    [Abstracts]
TI: Incorporating Soil Hydraulic Parameter Statistics in Developing Pedo-transfer Functions
AU: Zhao, Y
EM: Yanxia.Zhao@dri.edu
AF: Desert Research Institute, 755 E Flamingo Road, Las Vegas, NV 89119, United States
AU: * Zhu, J
EM: Jianting.Zhu@dri.edu
AF: Desert Research Institute, 755 E Flamingo Road, Las Vegas, NV 89119, United States
AU: Ye, M
EM: mingye@scs.fsu.edu
AF: Florida State University, Computational Science & Department of Geological Sciences, Tallahassee, FL 32306, United States
AU: Meyer, P D
EM: Philip.Meyer@pnl.gov
AF: Pacific Northwest National LaboratoryLaboratory, P. O. Box 999, Richland, WA 99352, United States
AU: Pan, F
EM: Feng.Pan@dri.edu
AF: Desert Research Institute, 755 E Flamingo Road, Las Vegas, NV 89119, United States
AU: Hassan, A E
EM: Ahmed.Hassan@dri.edu
AF: Desert Research Institute, 755 E Flamingo Road, Las Vegas, NV 89119, United States
AB: In this study, we develop artificial neural network (ANN) based pedotransfer functions (PTFs) to predict soil hydraulic properties. The PTF approach is an efficient way of translating less costly available data, such as particle-size distributions, soil textures and other geophysical measurements, to soil hydraulic parameters required for numerical simulations and other applications. The ANN PTFs need to be trained before being used to transfer indirect measurements to soil hydraulic parameters. The traditional training process, in general, is to adjust ANN's coefficients to solely minimize the difference between the estimated and measured soil hydraulic parameters. The training process, however, did not consider the distributions of soil hydraulic parameters and the trained neural networks may yield improper distributions, which may severely affect probabilistic predictions. We incorporate the distributions of the soil hydraulic parameters into the ANN PTF development. In addition, it has been observed that PTFs can introduce unrealistic correlations between the output parameters. The unwanted artificial correlations need to be penalized during the training process, since it is well known that parameter correlations have significant effect on predictions. We achieve these two goals by adding two regularization terms to the ANN objective functions. A suite of new neural network models are developed to estimate soil hydraulic parameters. These neural network models have the same input and output variables, but different objective functions, which incorporate sequentially the site soil hydraulic parameter measurements, parameter probability distributions, and parameter correlations.
DE: 1829 Groundwater hydrology
DE: 1865 Soils (0486)
DE: 1875 Vadose zone
DE: 1876 Water budgets
SC: Hydrology [H]
MN: 2007 Fall Meeting