302-5 Artificial Neural Network Pedotransfer Functions: Model Development Issues.



Wednesday, October 19, 2011
Henry Gonzalez Convention Center, Hall C, Street Level

Mahesh R. Gautam and Jianting Zhu, Division of Hydrologic Sciences, Desert Research Institute, Las Vegas, NV
Ability of artificial neural network (ANN) in capturing  potential nonlinear relationship between soil hydraulic properties (SHP) and their predictors, no requirement of a priori form of models in ANN modeling, and better predictive power have made ANN a favorable choice for soil hydraulic properties pedotransfer function (PTF) development. However, as in other alternative approaches, there are problems associated with ANN PTF development. This study looks into some of the major problems in ANN PTF development and provides guidance for input selection, selection of ANN learning approach under data-limited condition, and dealing with bias caused by scale problem or limited sampling.

The inputs for PTF range from inexpensive soil texture related limited data (e.g. silt, sand, clay contents) to detailed particle size distribution, as well as other expensive and inexpensive data. An understanding of relative importance of inputs for SHP prediction contributes towards better ANN PTF model development. One noted practice in ANN PTF literature for identifying the importance of inputs is the use of stepwise model development. We reviewed other alternative approaches in ANN literature that can be potentially used for identifying the importance of inputs. In addition, we compare a few popular neural network learning methods in Matlab environment (e.g. Levenberg Marquardt “trainlm”, bayesian regularized “trainbr”, etc.) under data limited condition. ANN developed with bayesian regularization provided better solution compared to others under such situation. Finally, we introduced an alternative approach for bias correction that can be applied within ANN modeling framework.

See more from this Division: ASA Section: Biometry and Statistical Computing
See more from this Session: General Biometry & Statistical Computing: II