14.3 Effective Merging of Satellite and Ground Aerosol Measurements Using an Ensemble Kalman Filter Based Approach

Thursday, 16 January 2020: 2:00 PM
259A (Boston Convention and Exhibition Center)
Jing Li, Peking Univ., Beijing, China; and X. Li, J. Wei, B. E. Carlson, and A. A. Lacis

In the recent two decades, many satellite sensors and algorithms have been developed to retrieval aerosol properties from space, leading to the production of multiple global aerosol datasets. However, satellite aerosol optical depth retrieval still bears relatively large uncertainty due to instrumental design, retrieval assumptions, etc. Ground observations are more accurate, but its spatial coverage is limited. Therefore, there is always the question of how to effectively combine these two types of measurements to yield a globally covered aerosol dataset with improved accuracy. In our previous research, we developed an ensemble based method to objectively estimate the spatial representativeness of surface aerosol observation sites. With this information in hand, the next step would be to extend the information obtained at one ground station to the larger area that it can represent, or in other words, to improve the accuracy of satellite retrievals in the larger area using measurements made at this site. For this purpose, we further develop an ensemble Kalman Filter (EnKF) based approach to merge satellite and ground observations. The two key aspects of applying the EnKF are the construction of an ensemble that can properly represent the variance of the background field (in our case, satellite data), and the correct estimate of the observation (in our case, surface observation) error which include both the absolute error and the representation error. To address the first question, we use all monthly observations with more than 2/3 global coverage from 8 platforms (AATSR, MISR, MODIS-Aqua, MODIS-Terra, VIIRS, SeaWiFS, AVHRR, PARASOL) to generate a 474 member ensemble, as well as a monthly ensemble set with ~40 members for each month. For the second questions, the representation errors of ground observations are estimated as the standard deviation of MODIS 3km products within a 1 by 1 degree grid. Moreover, we also enforce a localization scheme with 3000km radius to avoid spurious correlation in the background covariance matrix due to insufficient sampling. In our initial investigation, we merged MODIS, MISR and AERONET data. Results show that the merged dataset have greatly reduced the uncertainty of satellite data at the location of ground observation sites, which is not surprising. The more important feature is that satellite data at other locations beyond the ground observation site is also improved by as much as 20%, as shown by regional 3-fold and leave-one-out cross validation. This technique paves a promising way for combining the information of all difference types of measurements to yield a better estimate of aerosol properties and its space-time variability.

Figure caption: a. Percentage changes in the multi-year annual mean between the merged field and the original satellite field; b. Percentage reduction of the root mean square error (RMSE) at the AERONET locations after assimilation ; c. Percentage reduction of RMSE at the stations where observation is not assimilated into the satellite field calculated from cross validation.

- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner