Bret Schichtel (Washington U.) and Paul Wishinski , (VT DEC),
7/01/96 Summary Draft
A central issue of the Ozone Transport Assessment Group (OTAG) is to address the role of transported ozone in the non-attainment regions of the Eastern US. Trajectory analysis has been identified by the OTAG Air Quality Analysis Workgroup as one activity to aid in this endeavor. Poirot and Wishinski (1996) have already begun some trajectory analysis for high elevation sites in the Eastern US (see Air Trajectory Residence Time Analysis Investigation of Ozone Transport Pathways: 1989-95 )
Trajectories identify a likely pathway of an airmass in route to a receptor (back trajectory) or from a source (forward trajectory). Back trajectories identifies the locations where source emissions could influence individual receptor concentrations, while forward trajectories identify potentially where source emissions will go. Trajectories are usually calculated from trajectory models, such as ARL-ATAD (Heffter, 1980), HY-SPLIT (Draxler, 1992) and CAPITA Monte Carlo (Schichtel and Husar, 1995). These models require input of gridded, three or four dimensional meteorological data. The models also employ alternative assumptions and parameterizations of atmospheric physics, for example, trajectories can be calculate at fixed (or mixed) elevations, pressures or isentropes in the atmosphere, and processes such as atmospheric mixing may or may not be employed.
An inherent limitation of trajectories is their uncertainty. Generally, the longer an airmass is track forwards or backwards in time, the more uncertain is its position (Kuo et al., 1985; Rolph and Draxler, 1990; Kahl and Samson, 1986). The uncertainties arise due to errors in the input meteorological data, and limitations of the parameterizations and initial conditions used in the models. These uncertainties have raised concerns that performing similar trajectory analyses using different models, meteorological drivers, initial conditions, etc. could lead to different conclusions.
To begin to address these concerns, a comparative study was undertaken between back trajectories calculated from the HY-SPLIT and CAPITA Monte Carlo models. The same meteorological wind fields and initial conditions were used for each model. Hence, any differences in the results were due to the different methodologies and parameterizations used in each model to calculate the back trajectories.
The comparison was conducted by calculating back trajectories from Whiteface Mt., New York, and Boston Mt., Arkansas during June - August 1995. Trajectories were generated for four receptor times per day, at 3 AM, 9 AM, 3 PM and 9 PM, EST with a maximum duration of 4.5 days. The trajectories were compared individually and as an ensemble via Residence Time Analysis (Ashbaugh et al., 1985; Poirot and Wishinski, 1985; and Poirot and Wishinski, 1986).
The two models used in the comparison were the HY-SPLIT and CAPITA Monte Carlo models. Both models can be used to calculate forward and backward three dimensional trajectories, using various input meteorological drivers. The HY-SPLIT model has continually evolved over the past fifteen years and continues to be developed today. (Draxler and Taylor, 1982; Draxler, 1987; Rolph et al., 1992). This model has been used in a number of organizations for studies ranging from the characterization of transport to the simulation of atmospheric SO2 and SO4 (Wishinski and Poirot, 1996; Rolph et al., 1992; Draxler 1994).
The CAPITA Monte Carlo model was first developed by Husar and Patterson (1981) in the early 1980's, and has been re-implemented as an IBM-PC tool by Schichtel and Husar (1995). The model has been used for trajectory analysis (NESCAUM, 1993), simulations of SO2 and SO4 ambient concentrations and wet deposition fields, as well as the retrieval of emission fields from ambient monitoring data (Schichtel, 1996). A web based server is currently being developed to allow for the calculation of North American trajectories from 1991 - 1995.
The primary difference between the models is that the Monte Carlo Model employs mixing in the calculation of the back trajectories while the HY-SPLIT model does not. The mixing is employed using a random process such that no two trajectories calculated for the same receptor and receptor time are the same (Table 1). This allows for each receptor airmass to be simulated by multiple trajectories, where as HY-SPLIT uses one trajectory. A second difference is that each model uses different techniques for integrating the wind fields when calculating the trajectories. In this study, back trajectories were computed for each model using the 4-D meteorological data from the National Meteorological Center's Nested Grid Model (NGM) (Rolph, 1992).
Table 1. Features of the HY-SPLIT and CAPITA Monte Carlo Model.
|# Trajectories per Receptor Time|
Individual back trajectories were qualitatively compared by overlying the single trajectory from the HY-SPLIT model over the 10 trajectories from the Monte Carlo model for each receptor and receptor time. This was done in both the spatial and height domain (Figures 1 and 2). The mixing processes employed by the Monte Carlo model are evident in these Figures. As the 10 trajectories from the Monte Carlo model move further back in time they are spread throughout the lower 1-2 km of the atmosphere due to intense vertical mixing. Horizontally, the trajectories slowly fan out, and occasionally bifurcate into two separate airmasses.
Examining these comparison for the summer of 1995 it was found that the trajectories often predicted very similar receptor airmass pathways with the airmasses was coming from different spatial and vertical directions. This is shown in Figure 1 where in Figure 1C the trajectories from Boston AR for both models show the airmass traveling along the surface coming from the gulf of Mexico moving in a clockwise direction, while in Figure 1B a sinking airmass from the west impacts the Boston Mt. Receptor. A consistent pattern that arose was that the two models compared well when the 10 Monte Carlo trajectories remained grouped together in space. In these situations, the employed mixing processes did not significantly vary the horizontal positions of the trajectories, and all of the trajectories were transported in the same general direction as predicted by the HY-SPLIT model.
The models predicted different airmass pathways about as often as they predicted similar pathways. Three examples where they differed considerably are presented in Figure 2. In Figure 2C, the trajectories from the Monte Carlo model show that the receptor airmass was composed of two different airmasses; one high elevation airmass coming from the northwest and another surface airmass approaching from the south. These airmass converged about one day prior to arriving at the receptor. The trajectories from the HY-SPLIT model closely follows the surface airmass from the Monte Carlo model. A similar pattern is also seen in Figure 1A. In Figure 1B, both models predict the receptor airmass moving along the surface, but the Monte Carlo trajectories display substantial spreading and come more from the North than predicted by the HY-SPLIT model.
In order to identify how the trajectories from the two models differ in the aggregate, ensembles of trajectories were compared via the residence time analysis. The particular technique for residence time analysis used in this study is described by Poirot and Wishinski (1985) and Poirot and Wishinski (1986). The general approach is to grid the trajectory domain, and keep track of the "residence time hours" of each trajectory over each grid square in its path. The grid used in this analysis is presented in Figures 1 and 2. The resulting residence time probability plots identify the most likely regions that the receptor airmass previously resided, and the likely airmass pathways in route to the receptor .
Trajectories from any location and time can be included in the residence time analysis. Usually the trajectories are selected based upon characteristics of a particular receptor and time or characteristics of the trajectories themselves. This allows for the generation of transport climatologies for specific scenarios of interest. In this comparison we were interested in seeing how the two models performed under high ozone conditions. This was accomplished by using only those trajectories associated with the 50% largest ozone concentrations at the Whiteface Mt. and Boston Mt. receptor sites.
The results for Whiteface Mt. are presented in Figure 3. The contours have been selected for plotting so that they bound the smallest numbers of grid squares accounting for 25% , 50%, and 75% of the total residence time hours. The area shaded pink, gray and tan each include 25% of the total residence time hours for this scenario, and represent areas of decreasing residence time probabilities away from the receptor. The least probable 25% of scenario residence time hours are in the unshaded, white area of the map. As shown, the residence time plots from both models are quite similar. The residence time plots show two lobes of high residence with one lobe along the Atlantic coast and the other along the US and Canadian boarder. While the residence time patterns are quite similar, the 10-particle Monte Carlo trajectories provide denser spatial coverage, resulting in "smoother" residence time contours. The more detailed texture of the HY-SPLIT residence time plots is an artifact of the less dense spatial coverage of the single-particle HY-SPLIT trajectories over the relatively short, 3-month comparison period. The residence time plot from the Monte Carlo model has less texture and is more coherent than from the HY-SPLIT model. This is most like due to the fact that 10 times more trajectories were included in the Monte Carlo residence time plots.
Inherent to the residence time analysis is that air parcels are more likely to reside in nearby grid cells than distance grid cells. This creates sharply decreasing gradients from the receptor site, as seen in Figure 3. This property of residence time plots washes out the residence times for the more distance grid cells, reducing the ability to identify airmass pathways. Poirot and Wishinski (1986) called this bias central tendency. Based upon geometric reasons, they determined that weighing each transfer matrix element by the distance from the source to the receptor removes the central tendency. With the central tendency removed it is possible to identify regions of high residence for other reasons then their proximity to the receptor. Figure 4 presents the distance weighted residence time plots for the Monte Carlo and HY-SPLIT models. As shown, for both models, the two lobes of high residence are accentuated showing that the airmasses are likely to resided along a belt extending from southern Illinois over Lake Erie and from Virginia along the Atlantic coast to the Whiteface Mt. receptor. Both plots also show a region of low residence over West Virginia flanked by the two lobes of high residence.
The distance weighted residence time plots for Boston Mt. are presented in Figure 5. The same conclusions concerning regions of high residence can be drawn from both the HY-SPLIT and Monte Carlo results. That is, the highest likelihood of residence extends in a belt east of the Boston Mt. receptor through North Carolina and to southern Ohio. The belt also extends south of the receptor to the Gulf of Mexico. In the HY-SPLIT results, high distance weighted residence is also seen in northern Wisconsin. This is not seen to the same extent in the Monte Carlo results, and is most likely "noise" in the HY-SPLIT results due to the smaller number of trajectories used in the analysis.
This study was conducted in response to concerns raise by members of the OTAG Air Quality Analysis workgroup over the inherent uncertainties of trajectories, and the affect of these uncertainties upon the results of trajectory analyses. To address these concerns, back trajectories from the HY-SPLIT and CAPITA Monte Carlo model were compared individually and as an ensemble for June - August 1995. Both models used the same NGM meteorological data as input, but the Monte Carlo model employed atmospheric mixing while the HY-SPLIT model did not.
It was found that individual trajectories at times compared very well and at other times predicted substantially different airmass pathways. The large differences occurred when the mixing processes in the Monte Carlo model caused large scale dispersion of the 10 trajectories simulating each receptor airmass. When the mixing processes where unimportant and the Monte Carlo trajectories remained grouped together they generally followed similar paths as the HY-SPLIT trajectories.
In the residence time analysis it was found that there were no critical differences between the residence time plots from the HY-SPLIT model and the Monte Carlo model. This is an indication that there are no systematic differences between the back trajectories from the two models, and the differences between the individual trajectory tend to average out when aggregated over "long" time periods (in this case, a period as short as 90 days appears to be sufficient).
Ashbaugh L.L, Malm, W.C., Sadeh, W.Z. 1985. A residence time probability analysis of sulfur concentrations at Grand Canyon National park. Atmos. Environ. 19, 1263 - 1270.
Draxler R.R. and Taylor A.D. 1982. Horizontal dispersion parameters for long-range transport modeling. J. appl. Met. 21, 367-372
Draxler R.R. 1987. Sensitivity of a trajectory model to the spatial and temporal resolution of the meteorological data during CAPTEX. J. Clim. appl. Met. 26, 1577-1588.
Draxler R.R. 1992. Hybrid single-particle Lagrangian integrated trajectories (HY-SPLIT): Version 3.0 - user's guide and model description. NOAA Technical Memorandum ERL ARL-195. Air Resources Laboratory, Silver Spring Maryland.
Draxler, R.R., J.T. McQueen, and B.J.B. Stunder. 1994. An evaluation of air pollutant exposures due to the 1991 Kuwait oil fires using a Lagrangian model. Atmospheric Environment 28:2197-2210.
Heffter J.L. 1980. Air Resources Laboratories atmospheric transport dispersion model (ARL-ATAD). Technical Memo ERL ARL-81, NOAA, Rockville.
Husar, R.B. and D.E. Patterson. 1981. Monte Carlo simulation of daily regional sulfur distribution: Comparison with SURE sulfate data and visual range observations during August 1977. J. of Applied Meteorology 20, 404-420.
Kahl, J.D. and Samson, P.J. 1986. Uncertainty in trajectory calculations due to low resolution meteorological data, J. Clim. Appl. Meteorol., 25, 1816-1831.
Kuo, Y, M.S. Skumanich, P.L. Haagenson and J.S. Chang. 1985. The accuracy of trajectory models as revealed by the observing system simulation experiments. Mon. Wea. Rev., 11, 1852-1867.
NESCAUM. 1993. 1992 Regional Ozone Concentrations in the North Eastern United States, North Eastern States for Coordinated Air Use Management (NESCAUM), Ambient Monitroing and Assessment and DataManagement Committees report.
Poirot, R.L. and P.R. Wishinski. 1985. Regional apportionment of ambient sulfate contributions to a remote site in northern Vermont. in Transactions APCA Int. Spec. Conf. on "Receptor Methods for Source Apportionment: Real World Issues and Applications" T. G. Pace, Ed., Williamsburg, VA.
Poirot, Richard L and Paul R. Wishinski. 1986. Visibility, sulfate and air mass history associated with the summertime aerosol in Northern Vermont. Atmos. Environ. 20, 1457-1469.
Rolph, G.D. and Draxler, R.R. 1990. Sensitivity of three-dimensional trajectories to the spatial and temporal densities of the wind field. J. of Appl. Met. 29, 1043-1054.
Rolph G.D. 1992. NGM Archive. NCDC Report TD-6140, July, National Climatic Data Center.
Schichtel, B.A. and Husar, R.B. 1995. Regional simulation of atmospheric pollutants with the CAPITA Monte Carlo model. J. Air & Waste Manage. Assoc. Accepted for publication.
Schichtel, B.A. 1996. The Retrieval of Pollutant Emission Fields from Ambient Concentration and Precipitation Chemistry Data.. Doctoral Dissertation presented at Washington University, St. Louis, MO.
Wishinski and Poirot. 1996 Air Trajectory Residence Time Analysis Investigation of Ozone Transport Pathways: 1989-95. Located at: "http://capita.wustl.edu/otag/Reports/Restime/Restime.html"
Figure 1. Comparisons of back trajectories from the HY-SPLIT and CAPITA Monte Carlo models for A) Whiteface Mt. NY at 07/09/95 15 (EST) B) Boston Mt. AR at 07/18/95 15 (EST) C) Boston Mt. AR at 07/17/95 15 (EST). The large dots are the HY-SPLIT trajectory segment endpoints, while the smaller multi-colored dots are the CAPITA Monte Carlo trajectory segment endpoints for the 10 trajectories. Each trajectory segment endpoint represents the estimated position of the airmass every 3 hours back in time from the receptor.
Figure 1 Continued
Figure 2. Comparisons of back trajectories from the HY-SPLIT and CAPITA Monte Carlo models for A) Boston Mt. AR at 07/11/95 15 (EST) B) Whiteface Mt. NY at 07/08/95 15 (EST) C) Boston Mt. AR at 07/08/95 15 (EST). The large dots are the HY-SPLIT trajectory segment endpoints, while the smaller multi-colored dots are the CAPITA Monte Carlo trajectory segment endpoints for the 10 trajectories. Each trajectory segment endpoint represents the estimated position of the airmass every 3 hours back in time from the receptor.
Figure 2 Continued
Figure 3. Residence time plots for Whiteface Mt., NY calculated from back trajectories for the 50% of the highest ozone concentrations during June - August 1995 using trajectories generated from the A) CAPITA Monte Carlo model and B) HY-SPLIT model.
Figure 4. Distance weighted residence time plots for Whiteface Mt., NY calculated from back trajectories for the 50% of the highest ozone concentrations during June - August 1995 using trajectories generated from the A) CAPITA Monte Carlo model and B) HY-SPLIT model.
Figure 5. Distance weighted residence time plots for Boston Mt.,
AR calculated from back trajectories for the 50% of the highest
ozone concentrations during June - August 1995 using trajectories
generated from the A) CAPITA Monte Carlo model and B) HY-SPLIT