|Van Den Berg, Martinus -|
Submitted to: Water Resources Research
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: June 30, 2010
Publication Date: November 15, 2010
Repository URL: http://handle.nal.usda.gov/10113/58519
Citation: Crow, W.T., Van Den Berg, M.J. 2010. An improved approach for estimating observation and model error parameters for soil moisture data assimilation. Water Resources Research. 46:W12519-1 - W12519-12. Interpretive Summary: Accurate information regarding the availability of soil water is valuable for a wide range of agricultural applications (including irrigation scheduling, fertilizer application optimization, crop yield forecasting and water quality modeling). The current state-of-the-art for estimating large-scale surface soil moisture variations over agricultural landscapes are data assimilation systems which optimally merge soil moisture predictions from water balance models with remotely-sensed surface soil moisture retrievals obtained from satellite observations. This paper addresses a key technical challenge for such systems. Namely, how do you appropriately weight information in the satellite observations relative to (potentially conflicting) information provided by the land surface model? Currently, we have no reliable methodology for obtaining such weighting over most agricultural areas. This paper presents the first feasible technique for obtaining relative vital weighting information over poorly-instrumented portions of the globe. When applied within a land data assimilation system, our new technique is demonstrated to provide more accurate surface soil moisture predictions within three separate USDA ARS experimental watersheds.
Technical Abstract: The accurate specification of observing and/or modeling errors presents a remaining challenge to the successful implementation of many land data assimilation systems. Recent work has developed adaptive filtering approaches which address this issue. However, such approaches possess a number of known weaknesses including a required assumption of serially uncorrelated error in assimilated observations. Recent validation results for remotely-sensed surface soil moisture retrievals call this assumption into question. Here we propose and test an alternative system for tuning a surface soil moisture data assimilation system which is more robust to the presence of auto-correlated observing errors. The approach is based on the application of a triple collocation approach to estimate the error variance of remotely-sensed soil moisture retrievals. Using this estimate, the variance of assumed modeling perturbations is tuned until normalized filtering innovations have a temporal variance of one. Real data results over three highly instrumented watershed sites in the United States demonstrate that this approach is superior to a classical tuning strategy based on removing the serial auto-correlation in Kalman filtering innovations and nearly as accurate as a calibrated colored Kalman filter in which auto-correlated observing errors are treated optimally.