Development of Pedotransfer Functions with Neural Network Models

摘要

Unsaturated soil hydraulic properties determine the capacity of soils and rocks to retain and transmit water. Hydraulic properties may be needed in applications involving remediation and restoration of contaminated soils, trafficability of soils, flood control, and remotely sensed data. Current methods to measure hydraulic properties are perceived as inadequate to meet the data requirements for most (large scale) applications. Neural networks are used in our research to develop pedotransfer functions (PTFs) for the hierarchical estimation of hydraulic data from basic data such as soil texture and bulk density. Neural networks were calibrated on a database of more than 2000 soils. The predictions generally compared favorably with published PTFs. Especially noteworthy is the unsaturated hydraulic conductivity; we improved its prediction by almost half an order of magnitude compared to traditional methods. We have completed the computer program Rosetta to facilitate neural network based predictions of hydraulic parameters. The uncertainty of the estimates was shown to increase for lower water contents. We have also converted our database of soil hydraulic properties to Windows from DOS.