Multimodeling – An Emerging Approach to Improving Process-Based Modeling of Soil Systems

Authors: Pachepsky, Yakov; Guber, Andrey; VAN GENUCHTEN, MARITINUS - Federal University - Brazil

Submitted to: World Congress of Soil Science
Publication Type: Proceedings
Publication Acceptance Date: 2/3/2010
Publication Date: 7/30/2010
Citation: Pachepsky, Y.A., Guber, A.K., Van Genuchten, M.T. 2010. Multimodeling – An Emerging Approach to Improving Process-Based Modeling of Soil Systems. World Congress of Soil Science. p. 32-35, published on DVD.

Interpretive Summary: Multiple models have been developed to simulate the same soil processes. Because each model has been evaluated based on its performance with unique data, it is difficult to predict how the models will perform when soil processes or initial conditions vary from those used for initial validation. This problem was first encountered in the field of atmospheric science, where the concept of “multimodeling” was developed to improve predictive ability by combining the results from several models. We applied the “mutimodeling” concept to soil processes. By researching and evaluating the combined predictions from several models, we were able to devise a method that gave excellent accuracy and reliability for predictions of soil water flow. These results are important for soil hydrological observations and forecasting, in that the work has identified the opportunity to improve modeling results by applying the “multimodeling” approach.

Technical Abstract: Environmental systems usually are approximated in mathematical terms by making simplifying assumptions that lead to multiple model structures which may produce results that are equally consistent with available observations. An increasing number of papers are now being published on various applications of multimodeling in which predictions from various independent models are combined, rather than attempting to find the best model. Multimodeling consists of assigning weights to the simulation results from the various models, and then combining these results into a single prediction. We constructed a multimodel using 14 independent Richards equation-based individual models by employing different pedotransfer functions. The individual models were not calibrated. Soil water contents were monitored for 300 days with multisensory capacitance probes at eight depths in four locations. Simulations using seven different methods to assign weights to individual models were compared with observed soil water time series. The multimodel was by far more accurate and reliable than the individual models. The concurrent use of several models, and mutimodeling in particular, presents an opportunity to better understand and forecast soil processes.