|Nemes, Attila - UNIV OF MD,COLLEGE PK,MD|
|Quebedeaux, Bruno - UNIV OF MD,COLLEGE PK,MD|
Submitted to: Biological Systems Simulation Group Proceedings
Publication Type: Abstract Only
Publication Acceptance Date: March 11, 2008
Publication Date: N/A
Technical Abstract: Environmental and crop simulation models use a wide range of inputs that include soil hydraulic properties. For many applications, use of laboratory determined soil water retention and hydraulic conductivity data is not feasible; therefore those need to be estimated. The current version of the Agricultural Policy Environmental Extender (APEX) model – that is being used to evaluate on-site benefits of conservation practices on cultivated croplands is among the users of the Rawls et al. (1982) pedotransfer functions (PTFs) to estimate soil water retention. The Rawls et al. linear regression PTFs have widely been considered valid and applicable for US soil conditions. Most studies, that verify PTFs, use three generic measures to test the performance of such equations, namely RMSR, ME, R-square, but these measures have limited capabilities in unveiling unexpected data patterns. We used additional measures and a non-parametric k-Nearest Neighbor (k-NN) pattern recognition technique as an alternative to linear regression equations to identify inconsistencies in the estimation of soil hydraulic properties and test the adequacy of the Rawls et al. (1982) PTFs. Data from the original data collection used by Rawls et al. as well as a second independent data set were used. We found that the Rawls et al. PTF delivers sub-optimal and biased estimates for US conditions, for which these equations were widely considered to be valid. Matching the data used for the above PTFs with the original sources revealed that this result is mainly due to incorrect conversion of organic matter and/or organic carbon data into common units, misrepresentation of organic matter data when those were in fact missing, and uneven representation of soil regions in the data collection. Correcting those data significantly reduced estimation bias for an independent data set. Use of the k-NN pattern recognition technique further reduced bias in the estimates compared to those obtained using simple linear regression. The alternative estimation technique and an improved data set are suggested for use.