«Tsuyoshi Thomas Sekiyama 関山 剛 Department of Geophysics, Graduate School of Science, Tohoku University 東北大学大学院理学研究科 ...»
Data assimilation of satellite-borne lidar aerosol observations
and its validation with Asian Dust
Tsuyoshi Thomas Sekiyama
Department of Geophysics, Graduate School of Science,
教授 早坂 忠裕 教授 山崎 剛 准教授 佘 偉明 准教授 2012 平成24年 Abstract We have developed a four-dimensional ensemble-based data assimilation system for satellite-borne lidar aerosol observations. The system consists of a local ensemble transform Kalman filter (LETKF), a global aerosol model, and an observational operator that emulates a lidar instrument. Firstly, the system was assessed with simulated satellite-borne lidar aerosol observations by using observing system simulation experiments (OSSEs), in which the Cloud-Aerosol Lidar and Infrared Pathfinder Satellite Observations (CALIPSO) Level 1B data were emulated. The Level 1B data contain attenuated backscattering coefficients at 532 and 1064 nm and depolarization ratios at 532 nm, which are not subjected to retrieval processes. The performance of the system was validated by the Method for Object-based Diagnostic Evaluation (MODE) tool using sulfate and dust aerosol optical thickness (AOT) distributions in East Asia as analysis objects. The system worked successfully under the OSSE conditions, and it was found that, compared with a free-running simulation, the data assimilation system had the ability to produce a better analysis of sulfate and dust plumes. For example, the three-month mean centroid distance of aerosol plumes (the difference between the true and estimated values) was improved from 600 km to 400 km for sulfate aerosols and from 750 km to 330 km for dust aerosols; the three-month mean area ratio of aerosol plumes (the ratio between the true and estimated values) was improved from 0.49 to 0.76 for sulfate aerosols and from 0.51 to 0.72 for dust aerosols. The data assimilation successfully improved not only the aerosol plume analysis but also dust emission estimation in the OSSEs. The OSSE results strongly suggest that this data assimilation system would work appropriately even if the real CALIPSO data were used.
Furthermore, by conducting OSSE sensitivity tests, we explored the optimal parameter settings of inflation and localization for the LETKF algorithm, and assessed the robustness of the system against decreases in ensemble members and observational data amount. This OSSE study was the first in which lidar observations were emulated and assimilated successfully, to the best of the author’s knowledge.
Secondly, we applied the same data assimilation system to real CALIPSO aerosol observations in spring 2007. The results were validated via two independent observations: the extinction coefficient profile of dust and spherical aerosols, which were measured at a ground-based lidar observatory managed by the National Institute for Environmental Studies of Japan, and the visual weather reports of aeolian dust events in Japan. Consequently, aerosol particle types were successfully discriminated by the data assimilation without retrieval processes, and detailed three-dimensional structures of dust aerosol outflows from source regions over oceans and continents were well reproduced. For example, the threat score ranging i between 0 (worst) and 1 (best), which is often used for weather forecast verification, was remarkably improved from 0.42 to 0.74 for the heavy dust plume over Japan on May 28, 2007.
These results suggest tremendous promise for the beneficial use of the satellite-borne lidar that provides horizontally sparse but vertically and temporally dense information. This system will be able to supply the initial conditions for aerosol forecasting even in remote areas and across clouds. The CALIPSO aerosol observations were successfully assimilated for the first time, to the best of the authors’ knowledge.
Finally, we optimized the emission flux estimation of Asian Dust for spring 2007 by using the same data assimilation system and CALIPSO aerosol observations. This estimation was based on the assumption that the dust emission error is caused by the uncertainties of snow cover and soil moisture in the model. In this estimation, we installed a model bias correction process and one-way variable localization into the data assimilation system. The one-way variable localization utilizes only the covariance between wind direction and aerosol concentration to analyze aerosol plumes when horizontal winds of the reanalysis are simultaneously assimilated with aerosol observations. This bias correction and variable localization prevented filter divergence, and stabilized the data assimilation system. The estimation results were evaluated through a comparison with independent ground-based lidar observations. The data assimilation procedure resulted in an increase in the dust emission from the Gobi region during the dust event from March 25 to April 3, 2007, and it improved the analysis of dust concentrations in the leeward region. Without data assimilation, the dust emission was underestimated owing to the wet surface conditions of the dust source region in the model. It was found that the spatial pattern of the dust emission estimation result in the Gobi region was in good agreement with the results from an independent inverse analysis. These results indicated that the data assimilation was a powerful tool for rebuilding the upstream information of aerosol plumes from the satellite-borne lidar observations.
This dissertation shows the potential benefits of combining a state-of-the-art data assimilation scheme with satellite observations in the exploration of aerosol science. Although we focus on mainly the analysis of Asian Dust in this study, the promise of our global data assimilation system is not regionally limited and it is able to be applied to aerosols other than dust. The products of this study will contribute to wide-ranging research areas such as climate change, weather prediction, air pollution, marine biology, agriculture, and public health.
List of Figures
List of Tables
1.3. General Methodology
1.4. Outline of Thesis
2. Development of the Aerosol Data Assimilation System
2.2. Data Assimilation System
2.2.1. Data Assimilation Scheme
188.8.131.52. Ensemble Kalman Filter
184.108.40.206. 4-Dimensional Expansion
220.127.116.11. State Vector Augmentation
2.2.2. Global Aerosol Model
2.2.3. Lidar Observation Operator
2.3. Concluding Remarks
3. Simulation Study of EnKF Data Assimilation of Satellite- Borne Lidar Aerosol Observations
3.2. Experimental EnKF Conditions
3.3. OSSE Design
3.3.2. Making a Nature Run
3.3.3. Making the Simulated Observation Data
3.4. Evaluation Tools
3.4.1. Traditional Methods
3.4.2. Method for Object-based Diagnostic Evaluation
3.5. Results and Discussion
iii 3.5.1. Standard Test
18.104.22.168. MODE Scores
22.214.171.124. Flux Verification
3.5.2. Sensitivity Tests
126.96.36.199. Ensemble Size Test
188.8.131.52. Covariance Inflation and Localization Scale Test
184.108.40.206. Observation Error Test
220.127.116.11. Data Density Test
4. Data Assimilation of Real Satellite-Borne Lidar Aerosol Observations.............. 72 4.1. Introduction
4.2. Description of the Data Assimilation System
4.2.1. Observational Data
4.2.2. Data Assimilation Scheme
4.2.3. Experimental Design
4.3. Results and Discussion
4.3.1. Comparison with CALIPSO data
4.3.2. Comparison with Ground-Based Lidar Observations
4.3.3. Comparison with Weather Reports
5. Dust Emission Estimation by Data Assimilation
5.2. Effects of Snow Cover and Soil Moisture on Asian Dust
5.4. Results and discussion
5.4.1. Case Studies
5.4.2. Comparison with Another Inversion Analysis Result
5.4.3. Comparison with Independent Lidar Observations
6.2. Future Directions
Figure 1-1. Geometric relationship between the truth χt, analysis χa, observation χø, and first guess χƒ in the linear minimum variance estimation approach.
Figure 1-2. The schematic diagram of EnKF when we have an initial mean analysis with its ensemble members and an observation with its error information. The location of each circle symbolizes the system state. Each meandering arrow indicates the temporal progress of each ensemble member driven by a forecast model. χa indicates the analysis estimated optimally.
Figure 2-1. The schematic diagram of the 4D-LETKF data assimilation scheme. In this case, each time window is 48 hours long. The analysis is obtained at the intermediate time of this assimilation time window at 24-hour intervals. Red stars indicate the analysis. Red arrows illustrate a temporal transition of the state. Blue circles indicate observations, and the size of each blue circle symbolizes the localization weight that depends on the distance from the analysis grid to the observation grid. Each observation is weighted and used twice. The same observation has the same number in this illustration.
Figure 2-2. The correlation between surface dust concentrations and the dust emission within the yellow square area centered at (43ºN, 112.5ºE). This spatial distribution indicates the response pattern of the dust concentrations to an increase in the dust emission within the yellow square area. This correlation distribution was derived from the 32-member ensemble forecast from 00 UTC to 24 UTC on 30 March 2007. Green arrows indicate the mean surface winds on this day.
Figure 3-1. An example plot of the one-day CALIPSO orbit tracks (1 March, 2007)........ 56 Figure 3-2. Snapshots of sulfate aerosol optical thickness (AOT) at 00UTC on 30 March 2007 derived from (a) the Nature Run MASINGAR result and (b) the default MASINGAR result.
Figure 3-3. The same as Fig. 3-2, but snapshots of dust aerosol optical thickness (AOT).. 58 Figure 3-4. (Upper) The daily-mean cloud cover ratio (CCR) of the Nature Run model result on 30 May 2007, which is the sum (varying from 0 to 16) of the 16-layer CCRs below 150 hPa. Each layer’s CCR varies from 0 to 1, in which a zero ratio means a perfectly clear sky. (Lower) The real cloud snapshot taken by the infrared channels of the geostationary meteorological satellite MTSAT-1R located at 140ºE on the same day. The green circle in the upper panel indicates the approximate location of the snapshot of the lower panel.
v Figure 3-5. A distribution of lidar aerosol observations on 30 May 2007 (a) derived from the real CALIPSO/CALIOP after screened by the cloud-aerosol discrimination (CAD) scores, and (b) derived from the OSSE virtual lidar with the standard settings after screened by the cloud-cover ratio and the aerosol signal threshold. Because the CALIOP vertical resolution is higher than the Nature Run resolution, the total data amount of (a) is larger than that of (b).
Figure 3-6. Frequency distributions of the selected observations to be assimilated in the global troposphere: (a) the real CALIPSO/CALIOP observations (Sekiyama et al., 2010), (b) the OSSE simulated observations with the perfectly clear sky condition, (c) the OSSE simulated observations with the standard conditions. The aerosol signal threshold was set to a 2×10-5 m-1 extinction coefficient at 532 nm for (b) and (c).
These observations were previously screened by the cloud cover ratio and the aerosol signal threshold, but not yet averaged to the model resolution. Thus, the total observation number of (a) tends to be larger than that of (b) or (c). Blue line and squares indicate 532nm parallel attenuated backscattering coefficients; red line and squares indicate 532nm perpendicular attenuated backscattering coefficients; gray line and squares indicate 1064nm total attenuated backscattering coefficients. The X-axis shows the number of observations. The Y-axis shows the intensity of attenuated backscatter, and is expressed logarithmically.
Figure 3-7. Cross sections of lidar observations (total attenuated backscattering coefficients at 532 nm) at approximately 17 UTC on 27 May 2007 over East Asia along the orbit path shown in (a), (b) derived from the real CALIPSO/CALIOP data, and (c) derived from the Nature Run. The simulated observations (c) have vertically discrete layers depending on the model resolution, distributing them into strips. The real CALIOP observations (b) are not yet averaged to either the CAD score resolution or the model resolution.
Figure 3-8. A schematic example of various observation and analysis combinations. (a)–(c) These all yield the same RMSE, whereas (d) has the best RMSE. However, (a) would probably be evaluated as the best subjectively.