The unsaturated soil hydraulic properties determine the rate with which water and dissolved chemicals move through soils and rocks whose pore space is partially filled with water. Current methods to measure unsaturated hydraulic properties are generally viewed as impractical to meet the data requirements for most simulation studies. As a consequence, the empirical approach of using pedotransfer functions (PTFs) has become popular to indirectly estimate hydraulic properties or parameters from such basic soil properties as soil texture, bulk density, organic matter content, and clay mineralogy. Mathematical expressions for PTFs are typically derived by linear regression of subjectively selected basic soil properties and unsaturated hydraulic data. Relatively little attention has been paid to a systematic derivation and evaluation of pedotransfer functions because of a conspicuous lack of comprehensive databases and the myriad of soil attributes and experimental methods.

Neural network models have tremendous potential to derive PTFs to predict unsaturated hydraulic parameters in an hierarchical manner by making optimal use of soil attributes and other information that is available. These models are high-performance black-box models that have proven to be very useful for exploring input data and to assess the maximum obtainable accuracy of the predicted output. Only recently have neural network models been applied to estimate soil water retention. Calibration of the network can be done in conjunction with the bootstrap method to determine probability density functions of predicted hydraulic parameters. Errors in predicted hydraulic properties, as well as in results of flow and contaminant transport models that use these properties, can hence be quantified. Furthermore, the correlation between hydraulic output parameters can be obtained. This feature, in addition to the versatility of handling input data, makes neural network models attractive to identify relevant information in large heterogeneous sets of hydraulic and other soil data. Specific objectives of this research are (1) to develop PTFs with neural network analysis for a hierarchical input structures, and (2) to estimate the uncertainty in hydraulic properties predicted with PTFs. The anticipated results of the research would go a long way to provide scientists and practitioners with estimates of unsaturated hydraulic functions needed for computer models and geographical information systems dealing with groundwater pollution, soil remediation, climate change, remote sensing and agricultural production.

Agency
National Science Foundation (NSF)
Institute
Division of Earth Sciences (EAR)
Application #
9804902
Program Officer
L. Douglas James
Project Start
Project End
Budget Start
1998-09-01
Budget End
2000-08-31
Support Year
Fiscal Year
1998
Total Cost
$151,984
Indirect Cost
Name
University of California Riverside
Department
Type
DUNS #
City
Riverside
State
CA
Country
United States
Zip Code
92521