1. OBJECTIVES AND BACKGROUND
The Air Force Combat Climatology Center is producing high-resolution gridded climatologies by using a method based on a deterministic mesoscale model (Zack et al. 1997). The purpose of the model is to produce high-resolution gridded climatologies over regions with complex terrain or limited observed data. Traditional climatologies generated by summarizing observations made at points are of limited use in mountainous or data sparse regions. This paper examines the initial results of this method and compares the modeled climatologies with the traditional techniques.
The Air Force Combat Climatology Center (AFCCC) provides climatological support to Department of Defense agencies. A typical project generates some type of climatological information based on hourly weather data that have been archived from airport weather observations. This is useful information for that site, but in mountainous or littoral regions, the information may not be representative of nearby locations. Another problem is availability of data. AFCCC has data for over 16,000 sites around the world. Many of these sites, however, are located in industrial countries which leaves large regions of the world with few surface weather observations. Even when there is a site that has taken observations, the period of record may be incomplete. Also, the quality of these observations may be subject to the local observing practices and equipment.
There are numerous locations worldwide that have no observing sites. In these areas the local climate is often determined by an interpolation (Inverse Distance Weighting, kriging, etc.) of the nearest point climate information. However, if there are insufficient points or a lack of spatial continuity, the results can be unrealistic since they do not explicitly account for the effects of topography and other mesoscale features.
There are two strategies to help minimize these problems. One can either build a probabilistic or a deterministic model. Probabilistic models have demonstrated an ability to create accurate, high-resolution precipitation and temperature gridded climatologies in complex terrain (Carrega, 1995; Daly et al., 1994). Models of this type use statistical relationships between a particular meteorological element and other information that is available (elevation, aspect, etc.).
If there is sufficient knowledge of the phenomena being estimated, it is possible to use a deterministic model to estimate information where we have no observed data (Isaaks and Srivastava, 1989). Advances in atmospheric numerical modeling systems (ANMS) and computer hardware have now made the deterministic approach possible (Cox, 1998). For AFCCC's purposes, there are several advantages to using an ANMS coupled with statistical methods. First, ANMS output many other elements in addition to temperature and precipitation such as relative humidity, pressure, wind speed, wind direction, clouds, etc. Second, the information is also output at many levels in the vertical. Additionally, ANMS can be run in parts of the world that have few observational sites. Finally, ANMS explicitly consider the effects of complex terrain and its influence on the mesoscale environment. This ability to have climatologies for many different elements at various levels throughout the atmosphere (surface, 1000 ft., 2000 ft., etc.) for any location in the world is the primary reason AFCCC chose to use a mesoscale model to produce its high resolution gridded climatologies. The process of producing these climatologies is moving from a prototype stage to a production stage.
2.
RESEARCH
ACCOMPLISHED
2.1. Methodology
Development
MESO Inc. has developed the method called CLImate statistics by a dynamical MODel (CLIMOD) (Zack et al. 1997) based on a deterministic mesoscale model called the Mesoscale Atmospheric Simulation System (MASS) (Manobianco et al.,1996). The methodology involves conducting research for several different climate regimes around the world at a model resolution of 10 km and 40 km. The basic methodology involves initialization of the model with large-scale gridded data fields. The model then uses the large-scale gridded data fields and lateral boundary conditions and periodically ingests the available observed data through one of several possible data assimilation techniques (Newtonian relaxation periodic reanalysis, etc.). Next the model dynamically fuses the available observations with its knowledge of the surface characteristics of the earth and the basic principles of physics to generate estimates of local climate statistics at locations for which no observational data is available.
The method was initially applied and evaluated on a 10-year period in the Eastern Great Lakes Region for the months of January and July (Zack et al. 1997). More recently, the method has been extended to produce data for all 12 months of a 10-year period (1973-1982) over a region in Far East Asia and the Middle East. Model output is created hourly as a 3-D grid on 20 sigma-p levels at both 40 km and 10 km resolutions. The data can then be interpolated to any level or point of interest. The data are then summarized to produce a set of grids and descriptive statistics typically used in climate analysis.
The quality of the CLIMOD method is determined by comparing the model derived climatologies with the climatologies at existing observing locations. NCDC’s Global Historical Climatology Network (GHCN) (Vose, et al. 1992) provides extensive coverage of temperature and precipitation monthly means for nearly 500 sites around Far East Asia (Figure 1). At each location, the mean temperature and precipitation for the month of January is compared with the model derived mean temperature and precipitation for the same period of record. The modeled values are determined by an inverse distance weighted interpolation from the nearest model gridpoints to the station location.
Figure 1. Location of the model domain at 40 Km
resolution. Locations of GHCN temperature observing sites are indicated with
terrain elevation in meters.
2.2. Temperature Results
Typically, when performing a regional climate analysis for our customers, we will take the mean temperatures at existing observing sites, and construct a contoured analysis for visual interpretation (Figure 2). However, this kind of analysis has several severe limitations.
Figure 2. Contoured analysis of mean January
temperature from the GHCN. Hatched regions indicate areas beyond 250-km range
to the nearest observing site.
First, while observation sites may be fairly dense in some locations, there are large regions over mountainous terrain and especially the oceans where no surface observation sites exist. A second limitation is that a simple interpolation scheme like IDW is not representative between sites where strong mesoscale effects are present (e.g. land/ocean boundaries and mountainous terrain). Another limitation is that climatologies from the GHCN database only include temperature, precipitation, and pressure. While climatologies may be obtained from other sources of surface data, these sources are even more sparsely populated in this region than those in the GHCN.
The use of the CLIMOD method to “fill in the gaps” of the above-mentioned deficiencies is showing some promise. Figure 3 shows the model-derived mean temperature over the entire model domain (40 km resolution). This figure demonstrates that the model is producing a physically consistent climatology that reflects mesoscale climate influences. Most notable are the colder mean temperatures over the mountainous terrain on a mesoscale level.
Figure 3. Mean January temperature from the 40km model
domain.
Significant differences between the observed and modeled results are shown in Figure 4. Overall, the model tends to have a warm bias of 1.6 °C. This bias, however, is largely a result of the analysis over the oceans. Interpolating over water between land-based observation sites fails to account for the moderating effect the oceans have on air temperature. The model does, however, account for this effect and correctly models warmer temperatures over water. Strictly over the land areas, Figure 4 shows that a majority of the area is within ±1 °C. Another region of significant warm bias is the mountainous terrain in extreme northern China/Russia. Here the model failed to capture the extremely cold mean temperatures seen in this area. There are two possible reasons for this. (1) The vertical resolution of the model near the ground, while higher than in the free atmosphere above the PBL, is still not sufficient to capture shallow inversions which often occur in this region in the winter and contribute significantly to a very low mean temperature. (2) The coarse resolution of the NMC Reanalysis data used for lateral boundary conditions fails to properly represent these cold shallow inversions. The combination of these two factors results in climatologies that look like the Reanalysis data near the model boundaries.
Many regions where the differences are greatest are located in between the surface observation sites where mesoscale effects such as mountainous terrain or ocean influences are significant. In these areas it is inappropriate to simply interpolate between observations and it is believed the model makes more physical and climatological sense than the observed.
Figure 4. Difference between modeled and observed mean
January temperature. Solids indicate regions where the model is warmer than
observed.
The CLIMOD performance can be measured by analyzing the differences between the modeled climate and the observed climate at the points where surface observations are recorded. The GHCN provides over 500 observation sites in the region with which to compare the modeled mean temperatures.
An examination of the two distributions reveals several similarities between them (Figure 5). Both modeled and observed show peaks occur near +4, -4, and -18 °C. The model is skewed slightly more toward the warmer temperatures and fails to accurately identify the tail of the distribution at extreme cold temperatures. With the exception of the extreme minimum temperatures, the modeled distribution is remarkably similar to the observed.
Figure 5. Distributions of observed mean temperature
(C) (top) and modeled mean temperature (bottom) at the GHCN locations.
A scatter plot of modeled vs. observed shows a very strong linear correlation with an R2 of 0.96 (Figure 6). Only a handful of modeled temperatures depart significantly from observed. A cursory examination of several of these outliers indicates the modeled resolution of 40 km was insufficient to resolve mesoscale features such as terrain and ocean at these locations.
Further examination of the distribution of errors shows that the model has a mean error of +0.8 °C, a mode of +0.3 °C, and is skewed slightly toward a warm bias (Figure 7). 44% of the locations are modeled within ±1 °C of observed and 75% are within ±2 °C.
These are remarkable results when considering that the extremely limited amount of direct observational data that was ingested in the model. Another significant consideration that needs to be made when viewing these results is that this project is only a first attempt at using this method and there is considerable room for optimizing the methodology and model configuration.
Figure 6. Comparison of modeled
vs. observed mean January temperature at GHCN observing locations in Far East
Asia.
Figure 7. Distribution of the errors in modeled mean January temperature.
2.3. Precipitation Results
The modeled precipitation climatology for the region is shown in Figure 8. A brief analysis reveals the following characteristics:
Figure 8. Modeled mean January
precipitation (inches).
1. Precipitation drops off to zero at the model boundaries.*
2. The model correctly identifies mesoscale features such as up-slope precipitation maximums and rainshadow minimums.
3. The rainfall maximum over the ocean southeast of Japan coincides with the observed mean position of the polar frontal zone.
4. The maximum over the northeast region of Taiwan is a result up-slope conditions from the Northeasterly trade winds.
5. The very dry conditions in continental Asia are consistent with observed.
A comparison of the modeled and observed distributions reveals several similarities (Figure 9). The mean precipitation at all sites is slightly higher from the model than the observed but only by about 10%. The model precipitation amounts are somewhat more spread out but this is largely a reflection of the large difference in the extreme maximums. The differences between the quartiles were all less than 0.10 inches.
In general, the model appears to produce
a distribution of precipitation climatology consistent with expected
conditions. However, when the modeled precipitation is compared with the
observed mean precipitation at the GHCN sites, significant differences are
observed.
A scatter plot analysis of modeled vs. observed shows precipitation modeling is much less skillful than temperature modeling. However, it needs to be noted that a 40-km model resolution, which works well for temperature, is not adequate to capture fine-scale precipitation patterns that will likely dominate the climatology in complex terrain. Higher model resolution combined with the more complete microphysics will likely have a significant positive impact on precipitation.
Figure 9. Frequency distributions of observed (top)
and modeled (bottom) mean January precipitation at 383 GHCN sites around Far
East Asia.
The traditional linear correlation is compared with the ranked correlation (Figure 10) where the lowest rank corresponds to the driest station. In this analysis, there is a stronger correlation between the driest observed and the driest modeled sites. The locations with the most precipitation are also the same locations where the model is the wettest. The lower traditional R2 value of 0.53 compared to a much higher rank R2 of 0.74 indicate that a few deviant pairs are contributing significantly to a low correlation. These locations are identified as A, B, and C in the figure.
Figure 11 shows the error between modeled and observed mean precipitation. As noted earlier, the model appears to be too dry along its boundaries, and since this appears to be a boundary value problem, these sites are not included in this analysis. The regions that appear to contribute significantly to the correlation in Figure 10 are identified as:
A – Taipei, Taiwan
B – Continental Southeast China
C – Four sites in Japan.
Figure 10. Modeled vs. Observed mean January
precipitation at 383 GHCN stations. Bottom is ranked values where the lowest
ranks are the driest stations. The solid lines show the expected 1-1
correlation.
Taipei’s large discrepancy is a result excessive low-level convergence due to the moisture rich tropical Northeasterly trade winds moving into up-slope conditions along the windward side of the island mountains. The island’s close proximity to the model boundary is heavily influenced by the coarse Reanalysis data used for lateral boundary conditions. The Reanalysis data is not fine enough to properly resolve Taipei and Taiwan as a whole but the CLIMOD 40 km simulations do. Thus, a conflict is created in the model between the Reanalysis data and the solution from the model in the vicinity of the Island. This conflict creates added convergence due to the wind differences between the model and the Reanalysis and this leads to extra precipitation. The results for Taipei would likely be significantly better if it were further into the interior of the 40 km grid.
The locations at C present a similar problem. However, all are located within the leeward side of a significant mountain barrier. Although the model does demonstrate the “rainshadow” effects on the leeward side of the mountains, it is apparently not precipitating enough moisture out as systems move across the mountains.
In general, it appears that the model is over-precipitating at many locations where there is mountainous terrain and an ample source of moisture (see Korea and Japan).
Figure 11. Model error at GHCN
locations. Positive values (circles) indicate regions where the model is wetter
than observed.
A contrary problem is seen in the locations over continental Southeast China (B) which are much drier than observed. The cause for this appears to be a result of boundary conditions. The model does identify this zone as a region of locally higher precipitation (Figure 8). However, the precipitation amounts quickly diminish to zero inland from the oceans (toward the model boundary).
Ideally a normal distribution of errors strongly peaked near zero is expected (Figure 12). This distribution peaks at 0.04 inches and has a mean of only 0.12 inches. Its skewness is largely influenced by the extreme case at Taipei, Taiwan whose modeled error is in excess of 12 inches (off chart).
Figure 12. Distribution of errors
between modeled and observed mean precipitation at GHCN locations.
Figure 13 shows the cumulative frequency distribution of root mean square errors (RMSE) between modeled and observed mean precipitation. While the mean error seems rather excessive at 0.60 inches, 50% of the errors were less than 0.25 inches and 80% of the errors were less than 1 inch.
Figure 13. Cumulative frequency
distribution of RMS Errors between modeled and observed mean precipitation.
3.
Conclusions
The use of a mesoscale atmospheric model to calculate climate descriptive statistics is producing successful results. The model is very skilled at producing a temperature climatology with a correlation of 0.96 and a mean error of only 0.8 °C. Although less skillful, precipitation modeling still produced encouraging results. The model demonstrates the ability to identify mesoscale precipitation patterns influenced by the oceans and terrain. With a mean error of 0.12 inches, over 80% of observing sites had an error of 1 inch or less.
The CLIMOD modeling method can be extended to also provide descriptive climatological summaries for dozens of additional DOD impact parameters (e.g. clouds, visibility, IR tranmissivty, winds, thunderstorms, turbulence, icing, soil moisture, frozen ground, etc.). Additional research will be required to determine the model’s skill for each of these variables. This project has been a significant first step in developing a methodology that will provide useful climate and historical weather information for the many data sparse regions around the globe. Support for additional development will turn this methodology into a robust production capability.
4.
Acknowledgements
The authors wish to thank Capt John Thompson for preparing the GHCN data for analysis. Pam Price, and Mary Bousquet of MESO Inc. were invaluable in providing the data needed to perform this analysis.
5.
References
Carrega, P, 1995: A method for the Reconstruction of Mountain Air Temperatures with Automatic Cartographic Applications. Theor. and Appl. Clima., 52, 69-84.
Cox, R., B. Bauer, T. Smith, 1998: A Mesoscale Model Intercomparison. Bull. Amer. Meteor. Soc., 79, 265-283.
Daly, C., R. Neilson, D. Phillips, 1994: A Statistical-Topographical Model for Mapping Climatological Precipitation over Mountainous Terrain. J. Appl. Meteor., 33, 140-158.
Isaaks, E., R. Srivastava, 1989: Applied Geostatistics. Oxford University Press, 561 pp.
Manobianco, J., J. Zack, G. Taylor, 1996: Workstation-based Real-time Mesoscale Modeling Designed for Weather Support to Operations at the Kennedy Space Center and the Cape Canaveral Air Station. Bull. Amer. Meteor. Soc., 77, 653-672.
Vose, R. S., Richard L. Schmoyer, Peter M. Steurer, Thomas C. Peterson, Richard Heim, Thomas R. Karl, and J. Eischeid, 1992: The Global Historical Climatology Network: long-term monthly temperature, precipitation, sea level pressure, and station pressure data. ORNL/CDIAC-53, NDP-041. Carbon Dioxide Information Analysis Center, Oak Ridge National Laboratory, Oak Ridge, Tennessee.
Zack, J. W., K.. T. Waight III, M. D. Bousquet, C. E. Graves, S. Yalda and G. E. Van Knowe, 1996: An evaluation of local climate statistics generated from the output of a 3-d mesoscale atmospheric model.. Preprints, 11th Conference on Numerical Weather Prediction, Norfolk, VX, Amer. Meteor. Soc., 379-381.