Tuesday, November 6, 2007
183-2

Using Support Vector Machines to Develop Pedotransfer Functions for Water Retention of Soils in Poland.

Krzysztof Lamorski1, Yakov Pachepsky2, Cezary Slawinski1, and Ryszard Walczak1. (1) Institute of Agrophysics, Polish Academy of Sciences, Doswiadczalna 4, Lublin, 20290, Poland, (2) USDA-ARS-BA-ANRI-EMSL, 173 Powder Mill Road, Beltsville, MD 20705

Pedotransfer functions (PTF), which estimate soil hydraulic parameters from better known soil properties, are the important data source for hydrologic modeling. Recently artificial neural networks (ANNs) became the tool of choice in PTF development. Training of ANN consists of finding of minimum of the mean-squared error as dependent on the neuron weights. None of training algorithms can guarantee that the global rather the local minimum will be found. Recent developments in machine learning methods include the growing research and application of the alternative data driven method called Support Vector Machines (SVMs). SVMs have gained popularity in many traditionally ANNs dominated fields. Using the SVM eliminates the local minimum issue - the minimum found is always the global one. The objective of this work was to see whether using the SVM to develop PTFs may have some advantages compared with the ANN. We have used the Soil Profiles Bank of Polish Mineral Soils that includes hydraulic properties for about 1000 soil samples taken from 290 soil profiles. This database was repeatedly randomly split into training and testing datasets, and both SVMs and ANNs were trained and tested for each split. The PTF performance was evaluated by the determination coefficient and the root-mean-squared error obtained with the test data sets. The accuracy of estimates from bulk density, sand and clay was slightly better than estimates from bulk density, sand and silt. There was no statistically significant difference (P<0.05) between the average ANN and SVM performance parameters for most of the 11 matric potential measurement levels. The SVM performed slightly better where the significant difference was found but this difference should not be important for many practical purposes. Overall, the ANN did not demonstrate the tendency to generate worse predictions after being stuck in local minima for the database of this work.