A multi-layer soil moisture data assimilation using support vector machines and ensemble particle filter

Hybrid data assimilation (DA) is greatly used in recent hydrology and water resources research. In this study, one newly introduced technique, the ensemble particle filter (EnPF), formed by coupling ensemble Kalman filter (EnKF) with particle filter (PF), is applied for a multi-layer soil moisture prediction in the Meilin watershed based on the support vector machines (SVMs). The data used in this paper includes six-layer soil moisture: 0-5 cm, 30 cm, 50 cm, 100 cm, 200 cm and 300 cm and five meteorological parameters: soil temperature at 5 cm and 20 cm, air temperature, relative humidity and solar radiation in the study area. In order to investigate this EnPF approach, another two filters, EnKF and PF are applied as another two data assimilation methods to conduct a comparison. In addition, the SVM model simulated data without updating with data assimilation technique is discussed as well to evaluate the data assimilation technique. Two experimental cases are explored here, one with 200 initial training ensemble members in the SVM training phase while the other with 1000 initial training ensemble members. Three main findings are obtained in this study: (1) the SVMs machine is a statistically sound and robust model for soil moisture prediction in both the surface and root zone layers, and the larger the initial training data ensemble, the more effective the operator derived; (2) data assimilation technique does improve the performance of SVM modeling; (3) EnPF outweighs the performance of other two filters as well as the SVM model; Moreover, the ability of EnPF and PF is not positively related to the resampling ensemble size, when the resampling size exceeds a certain amount, the performance of EnPF and PF would be degraded. Because the EnPF still performs well than EnKF, it can be used as a powerful data assimilation tool in the soil moisture prediction. (C) 2012 Elsevier B.V. All rights reserved.