Comparison of efficient algorithms for assimilating satellite partial column retrievals with WRF-Chem/DART

Abstract for oral presentation at the Third AMS Symposium on the Joint Center for Satellite Data Assimilation (JCSDA) at the 95th AMS Annual Meeting, January 4-8, 2015, Phoenix, AZ

Comparison of efficient algorithms for assimilating satellite partial column retrievals with WRF-Chem/DART

Arthur P. Mizzi*, Ave F. Arellano+, David P. Edwards*, and Jeffery Anderson#

*Atmospheric Chemistry Division National Center for Atmospheric Research Boulder, CO 80307

+Department of Atmospheric Science University of Arizona Tucson, AZ 85721

#Institute for Mathematics Applied to Geosciences National Center for Atmospheric Research Boulder, CO 80307

mizzi@ucar.edu 303-497-8987

We have successfully incorporated WRF-Chem (the WRF regional forecast model with online chemistry) into DART (the Data Assimilation Research Testbed) with forward operators to assimilate MOPITT (Measurement of Pollution in the Troposphere) and IASI (Infrared Atmospheric Sounding Interferometer) satellite instrument carbon monoxide (CO) trace gas retrieval and a DA algorithm that provides greater forecast skill (when compared to not assimilation CO retrievals) for meteorological (u, v, and T) and chemistry (CO) state variables (DA Algorithm). Due to the nature of satellite observing platforms and trace gas retrieval algorithms, retrieval data sets contain large volumes of data with low information content per observation requiring large data storage capabilities and high assimilation computational costs. In addition the error covariance contains cross-correlations that complicate the use of sequential ensemble Kalman filter algorithms. Additionally the retrievals contain contributions from the retrieval prior the effect of which should be minimized (i.e., we want to assimilate the observation information and not the first guess). To address those issues other researchers have proposed assimilating a truncated or filtered retrieval that is obtained by projecting the retrieval onto the leading left singular vectors (or eigenvectors) of a function of the averaging kernel (Joiner and Da Silva, 1998; Segers et al., 2004). Migliorini et al. (2008) proposed assimilating the “quasi-optimal” retrieval formed by subtracting the prior profile from retrieval (effectively the product of the averaging kernel and the true state). This paper continues that work by assimilating forms of a truncated retrieval that address the: (i) low information content, (ii) influence of the prior, and (iii) cross-correlation of the error covariance. We examine two methods. In the first method we rotate the “quasi optimal” retrieval equation by the leading left singular vectors of the averaging kernel and scale the result by the inverse square root of the rotated retrieval error covariance. In the second method (instead of scaling) we perform a second rotation by the leading left singular vectors of the rotated error covariance. Both methods are efficient in that they provide for data compression, remove the contribution of the retrieval prior, and diagonalize the error covariance. We compare the results from those algorithms with our existing DA Algorithm.