Assimilation of Soil Moisture using Support Vector Machines & Ensemble Kalman Filter
Location
Space Dynamics Laboratory
Event Website
http://water.usu.edu/
Start Date
3-26-2004 2:45 PM
End Date
3-26-2004 3:00 PM
Description
We are presenting a hybrid data assimilation methodology that combines two state-of-art techniques: Support Vector Machines (SVMs) and Ensemble Kalman filter (EnKF). SVMs methodology, based on statistical learning theory, provides statistically sound and robust approach to solve the inverse problem and thus to build statistical models. The traditional use of SVMs is for solving classification, regression, and ranking problems. The inclusion of kernel transformation of input space into the feature space allows the approach to deal with nonlinearities. The second component, EnKF, is an extension of Kalman Filter (KF), a well-known tool in prediction update, which is based on the Bayesian theory that the a posteriori probability distribution of the state of model can be improved by using a set of observations. EnKF extends KF to deal with the non-linearities in data assimilation and ensemble generation problems, and is mostly applied in weather, ocean and hydrologic prediction modeling.
In this paper, our problem is soil moisture prediction. Soil moisture distribution is helpful in predicting and understanding of various hydrologic applications, including weather changes, energy and moisture fluxes, drought monitoring, irrigation scheduling and rainfall-runoff generation processes. In the present research, ground measurements at various points were used to build a SVM. Subsequent observations were assimilated to update predictions from SVM model by coupling it with EnKF. Using soil moisture and meteorological data at a given time, soil moisture values were predicted forward in time by a Monte Carlo-based SVM. EnKF used newly acquired ground data to update the predicted values. In this way both model predictions and ground data are assimilated minimizing the prediction error, and making the predictions and observations statistically consistent. Our results are quite encouraging and suggest that this approach can be utilized in predicting other hydrologic processes.
Assimilation of Soil Moisture using Support Vector Machines & Ensemble Kalman Filter
Space Dynamics Laboratory
We are presenting a hybrid data assimilation methodology that combines two state-of-art techniques: Support Vector Machines (SVMs) and Ensemble Kalman filter (EnKF). SVMs methodology, based on statistical learning theory, provides statistically sound and robust approach to solve the inverse problem and thus to build statistical models. The traditional use of SVMs is for solving classification, regression, and ranking problems. The inclusion of kernel transformation of input space into the feature space allows the approach to deal with nonlinearities. The second component, EnKF, is an extension of Kalman Filter (KF), a well-known tool in prediction update, which is based on the Bayesian theory that the a posteriori probability distribution of the state of model can be improved by using a set of observations. EnKF extends KF to deal with the non-linearities in data assimilation and ensemble generation problems, and is mostly applied in weather, ocean and hydrologic prediction modeling.
In this paper, our problem is soil moisture prediction. Soil moisture distribution is helpful in predicting and understanding of various hydrologic applications, including weather changes, energy and moisture fluxes, drought monitoring, irrigation scheduling and rainfall-runoff generation processes. In the present research, ground measurements at various points were used to build a SVM. Subsequent observations were assimilated to update predictions from SVM model by coupling it with EnKF. Using soil moisture and meteorological data at a given time, soil moisture values were predicted forward in time by a Monte Carlo-based SVM. EnKF used newly acquired ground data to update the predicted values. In this way both model predictions and ground data are assimilated minimizing the prediction error, and making the predictions and observations statistically consistent. Our results are quite encouraging and suggest that this approach can be utilized in predicting other hydrologic processes.
https://digitalcommons.usu.edu/runoff/2004/AllAbstracts/13