VISIBILITY AND FOG FORECASTING BASED ON
DECISION TREE METHOD.
Ferenc Wantuch
Hungarian Meteorological Service,
greatbarrierreefKitaibel Pál u. 1 H-1525 Budapest, Hungary; E-mail: wantuch.f@met.hu
Abstract-The paper describes a visibility and fog forecasting model developed and ud at the Hungarian Meteorological Service (HMS) for last 3 years. The investigated model is a perfect prognostical model (PP). Characteristics of the model, such as input data, statistical approach, decision trees and threshold numbers, are described in this paper. The model was tested for both measured sounding and predicted data. The verification of the model led to very good results, so it was applied to aeronautical forecasting as well as to nowcasting. Information and short review about different type of other visibility models are also given.
nice的中文意思Key-words: NWP parameters, perfect prognosis(PP), model output statistics (MOS), FOGSI-index, decision tree.
1. Introduction
f开头的英文名Visibility forecast is very important for transportation, especially for air traffic where its accuracy is prominent. The WMO/ICAO requirements are very rigorous in aviation meteorology (ICAO-1998). Verifications regarding to the aeronautical forecasts show, that the reason of poor Terminal Weather Forecasts – in about 70 per cent of all cas - are the weak or not suitable visibility predictions. At HMS earlier it was not available any special numerical method, which could be an aid for forecasters in the prediction of the visibility, so they could u only traditional tools.
斯奎拉奇The European forecasters u different methods in the practice. One possibility is the diagnosis of fog from satellite images (Kerenyi et al., 1995). Some EUMETSAT, Central Institute for Meteorology and Geodynamics Austria (ZAMG), Swedish Meteorological Institute (SMHI), and Météo France u NOAA AVHRR, and Meteosat images in order to analy fog and low cloud from satellite data. Another possibility is the improvement of 1-D-models applied in UK, Sweden, Portugal, Belgium (Stesl and Ottoy, 1999)and also in France. Some ca studies have been validated with promising results.
skycityThe third way is the u of statistical methods and decision support systems for fog and low cloud fo
recasting. In the frame of it different methods, like decision trees, linear regression, Kalman-filter (Kilpinen and Juha, 1992) and neural network (Pasini at al., 1999) were considered for probability forecasts. In general, the results of all the methods were promising, so we considered the problem from statistical point of view.
Let us denote by y the estimated parameter, that is the predictand and let detected meteorological elements which are the predictors. In this ca we have to construct a function:
p x x x ,.....,2,1ε+=),.....,(2,1p x x x f y
behind是什么意思(1) where ε is the error of the method. One can u this function in the following estimated form: )~,.....~,~(~2,1p x x x f y = (2)
where p x x x ~,.....~,~
2,1 are known from NWP. This is the basic concept of the perfect prognosis method. Suppo that (2) is constructed directly from p x x x ~,.....~,~
2,1. ε+=)~,.....~,~(2,1p x x x f y (3)
In this way we get a model output statistical method. Bad on this idea an automatic visibility forecast method can be constructed. The input data of the visibility prediction is in the given ca the ALADIN mesoscale model output. This is a hydrostatic, spectral limited area numerical weather prediction which was developed by collaboration between Météo-France and some Central-and Eastern-European hydrometeorological rvices including HMS. The main dynamical characteristic of the model, like the preparation of initial, lateral boundary conditions, the physics and post-processing was discusd by Horányi at al. (1996). In order to find a connection with visibility at first we made a comprehensive statistical rearch of direct measurements and derived physical quantities. The best correlation was received by the fog stability index. The index was calculated according to the following formula:
FOGSI = 2 | T sfc - T 850 | + 2 ( T sfc - T d sfc ) + 2 W 850
(4)
where
T sfc temperature near the surface,
T d sfc dew point near the surface,
T 850 temperature on 850 hPa level, W 850 wind speed on 850 hPa level.
FOGSI index takes into account the temperature gradient, (that is the measure of the stability), the impact of moisture near the surface and the mixing by wind.
2.The result of statistical rearch
The FOGSI index is highly correlated with the obrved visibility especially in autumn-winter time when fog and mist frequently occurs. Becau of the strong relationship, we could u a regression connection bad on two years long datat as follows:
Visibility FOGSI =−+•133
045.. (5) Fig. 1. shows connection between the FOGSI index and the obrved visibility (measured in kms) in October 1996. Bad on this figure we can make the following considerations. There is a critical interval of FOGSI, above the upper limit which the calculation of visibility by regression is adequate to u. On the other hand if the FOGSI number is smaller than the lower limit of this domain, one can predict fog in all cas. The variance of the visibility values inside the critical interval is very high, conquently the statistical method is uncertain. It means that in this interval on
e can not decide about the visibility bad on FOGSI, e.g. if FOGSI is equal to 25 it might reprent whether fog, mist or good visibility in the same time. If data are taken into account only from the critical interval, then a statistical connection between FOGSI and visibility becomes very poor (Fig. 2.). Its physical reason is, that veral other effects, which play a great role in development of visibility, were neglected in FOGSI definition. Therefore one has to introduce some new weather predictors and methods. Such kind of parameters is reasonable to lect, which can be computed from TEMP data and NWP model output as well. After thorough investigation the mean relative humidity of lower air layers (925 hPa -surface) and upper layers ( 850-700 hPa), the near surface wind speed and relative humidity were chon to be included into the decision process.
Having examined a large number of cas, it was found that in winter period the cold air can be accumulated near the surface, mainly in the valleys and basins. Sometimes the surface temperature is colder 2-5 degrees with, in extreme cas this difference reaches 10 0C as compared with teh temperature of the 850 hPa level. This inversion stratification called “the cold air pad” (Tóth,1984; Bóna, 1986), which reprents a very stable state of the atmosphere. We came to the conclusion, that it is necessary to treat cold air pad situations parately and for the days other threshold numbers have to be determined . In order to
specify different visibility categories inside the critical interval, one had to work out a new procedure. The steps of the process are summarid in Fig. 3.
3. Decision tree
In this chapter the principles and steps of the decision-making procedure will be discusd and demonstrated in Fig. 4. The main characteristic of this tree is that each physical condition of the air column reprents only one category. The threshold numbers bad on two years of surface and radiosounding measurements of Budapest-Lőrinc.
Suppo that the lowest layer of the air (between surface and 925 hPa) is dry or moderately wet and windy, then the visibility depends on the water content near the surface. In this ca the process us simply the regression line for determining the visibility. If the lowest layer is dry and the wind is weak, then the visibility depends on the water content of the air near the surface. If this layer is wet, then we choo mist otherwi a good visibility category is lected. If the lowest layer is medium wet, then four subclass are constructed. In the subclass the radiative cooling effects of the atmosphere above a ground level point are reprented. This influence was modelled as the difference of mean relative humidity between the upper and the lower layers as follows:
1) if the air near the surface is dry and we do not include radiative cooling effect near the
surface, then can be calculated the visibility by the help of the regression line.
2.) if the air near the surface is wet and we have not radiative cooling influence in the air
column, then the decision is the misty weather.
3.) if the air near the ground level is wet and we have radiative cooling effect, then is fog
formation we expected.
4.) if the air near the surface is medium wet and we include radiative cooling, in this ca is
misty weather predicted.
斟酌5.) Finally, supposing that the air near the ground level is very wet, then in ca of radiative cooling we expect foggy otherwi misty category.
For cold air pad situations similar decision tree was constructed. The main differences are in the values of the threshold numbers. If very high relative humidity and weak winds occur together, it will
2006年时代周刊风云人物
be foggy weather.
4.Test results
The test of diagnostic method prented in this paper led to the following results. Fig. 5 illustrates the visibility at 00 UTC for each day of January in 1997, where JANAKTL means the measured, JANMODSZ the computed values. With regard to reliable visibility forecast good estimation of the small values is especially important. As it is shown, under 5 km both lines give similar range of sight, even if the dotted line sometimes a little bit underestimates the real data, so it makes the for. aviation more safety. For larger visibility values the difference is not so important.
The next two figures illustrate the correlations between measured and computed visibilities with the u of decision tree procedure (JANMODSZ) (Fig. 6.) and without (JANSSI) (Fig. 7.). High correlation (0.83) was reached with the more developed method, as oppod to the low correlation received 0.40 apply only the simple FOGSI index (4). According to our experiences the described decision tree procedure improves the results in all cas.
An even more illustrative picture is prented in Fig. 8., where one can follow the hour by hour (continuos line) changes of real visibility compared to the 48 hour forecasts (columns). The run of o
brved and predicted visibility values is comparable to each other, although some hours long shift might be detected. Regression coefficients were calculated bad on the radiosounding data of Budapest and ud for Szeged. It can be concluded, that equation (5) adequate for most of the places of Hungary. A possible explanation is, that it is due to the relatively smooth surface of the country.
5. Conclusion
The described method is mainly ud in aeronautical meteorology. After the test period this method was installed at HMS Weather Forecasting and Aviation Meteorology Department. According to 3 years long experience, efficiency of the method strongly depends on the quality of the ALADIN model output near the ground level.
Another application area is nowcasting, where the application of the above outlined procedure for fog formation and dissipation, as well as the horizontal visibility are led to significant improvements.
什么是abc
References
mazeHorányi,A.., Ihász, I. and Radnóti, G., 1996: ARPEGE/ALADIN: A numerical weather prediction mod
el for Central-Europe with the participation of the Hungarian Meteorological Service. Időjárás 100, 277-301.
ANNEX 3, International Standards and Recommended Practices, Meteorological Service for International Air Navigation To the Convention on International Civil Aviation, Thirteenth edition – July 1998 (ICAO).
Bóna, M., 1986: Hideg-légpárnák aeroszinoptikai vizsgálata. Meteorológiai Tanulmányok.
54. szám .
Stesl, J.-P. and Ottoy, H., 1999: Den Fog Forecasting with an Interactive Expert. Cost-78 Project II.3, final report, Fog and Low Clouds: Statistical Methods and Decision Support Systems for Fog and Low Cloud Forecasting.
Kerényi, J.,.G. Szenyán, I., Putsay, M.. and.Wantuch, F., 1995.: Cloud detection on treshold technique for NOAA-AVHRR images for the Carpathian Basin, Proceedings of 1995
Meteorological Satellite Data Urs` conference, Winchester, United Kingdom, 4-8
Sept 1995, pp: 565-569.
Kilpinen and Juha, 1992. The Application of Kalman Filter in Statistical Interpretation of Numerical Weather Forecasts. 12th Conference on Probability and Statistics in the Atmospheric Sciences, June 22-26, 1992, Toronto, Ont, Canada, 11-16.
Pasini, A., Pelino, V. and Potestà, S.,1999. A Neural Network model for visibility nowcasting from surface obrvations: results and nsitivity to physical input variables, submitted to Journal of Geophysical Rearch, D.
Tóth, P., 1984.: Parametrizáció bevezeté a hideg légpárnák keletkezé és feloszlásának analizi céljából. Meteorológiai Tanulmányok. 51. szám
