Multimodel prediction of water flow in a field soil using pedotransfer functions

Authors: Guber, Andrey; Pachepsky, Yakov; JACQUES, DIEDERIK - SCK-CEN, MOL, BELGIUM; Van Genuchten, Martinus; NEMES, ATTILA - U. OF CA, RIVERSIDE, CA; SIMUNEK, JIRI - U. OF CA, RIVERSIDE, CA; NICHOLSON, THOMAS - US NRC, ONRR; CADY, RALPH - US NRC, ONRR

Submitted to: American Geophysical Union
Publication Type: Abstract Only
Publication Acceptance Date: 3/20/2007
Publication Date: 5/22/2007
Citation: Guber, A.K., Pachepsky, Y.A., Jacques, D., Van Genuchten, M.T., Nemes, A., Simunek, J., Nicholson, T.J., Cady, R.E. 2007. Multimodel prediction of water flow in a field soil using pedotransfer functions. American Geophysical Union, May 22-25, 2007, Acapulco, Mexico.

Interpretive Summary:

Technical Abstract: Calibration of variably saturated flow models with field monitoring data is complicated by the strong nonlinearity of the dependency of the unsaturated flow parameters on the water content. Pedotransfer functions (PTFs) are routinely utilized to relate these parameters to readily available data on soil and sediment properties. Pedotransfer functions are empirical regression-type relationships, and their accuracy outside of the development region is generally unknown. Combining predictions using various independent models, often called multimodel prediction, has become a very popular technique in climate prediction and is now increasingly being used also in subsurface hydrology. The objectives of this work were (a) to compare different methods of multimodel prediction of the field soil water regime using pedotransfer functions, and (b) to see whether the calibration of a flow model with field data can be replaced by multimodel predictions. The multimodel prediction in this work consisted of running the Richards model with outputs of individual PTFs and then combining the obtained outputs into a single prediction. We compared weighing predictions from individual models by (1) using only the best model, (2) assigning equal weights, (3) using the unconstrained superensemble (i. e. regressing measured values to outputs of individual models), (4) using singular value decomposition in the regression, (5) using Bayesian model averaging, and (6) applying weights derived from the Kullback-Leibler information for each model. We evaluated the weighing methods in terms of their accuracy (i. e. errors in reproducing the training, or hindcast, datasets), and reliability (i.e., errors in reproducing the test datasets). The procedure was applied to a large database of measured soil water contents and pressure heads at 5 depths along a 6-m transect in a layered loamy soil collected over 384 consecutive days. Soil textural fractions, organic matter contents and bulk densities were averaged along the transect and used as input in the ensemble of 19 published PTFs developed from large datasets in different regions of the world. The HYDRUS 1D software was used to simulate water content time series for each of the PTFs at each of the five depths. The training (i.e., determining the weights) was done with daily water contents using moving windows that were from one to five months long; all data outside the windows were used to test weights. Probability distributions of RMSE for each window size were obtained and used to evaluate both accuracy and reliability. The unconstrained superensemble was found to be the worst weighing method characterized by low reliability but high accuracy because of multicollinearity in the regression input. Assigning equal weights led to accuracy and reliability about two times worse than other weighing methods. The other methods showed a weak trend of decreasing accuracy and increasing reliability with increasing length of training period. Selecting the best model was as good as other weighing methods for the 35 cm depth data, but its accuracy and reliability were two times worse than that of the Bayesian model averaging and regression with singular value decomposition at larger depths. The two latter methods had comparable accuracy and reliability. Weighing based on Kullback-Leibler information gave the same accuracy and reliability as the best model method. The two best weighing methods (Bayesian model averaging and regression with singular value decomposition) had average accuracy and reliability RMSE values of about 0.01 cm3cm-3 at 35 cm depth, and of about 0.005 cm3cm-3 at larger depths for one month monitoring and 13 months of testing. Calibrating the Richards model resulted in RMSE values of 0.009 cm3cm-3 at 35 cm depth and from 0.004 to 0.006 cm3cm-3 at larger depths. This indicates that monitoring of the s