Experimental study on severe convection forecast in Ningxia

Experimental study on severe convection forecast in Ningxia by non-linear retrieval and variational assimilation using satellite data HU Wen-dong 1, 2 SHEN Tong-li 3 DING Jian-jun 1 YANG You-lin 1 LIU Jian-jun 1 CHEN Xiao-juan 2 WANG Chenwei 3 (1 Key Laboratory of Meteorological Disaster Preventing and Reducing in Ningxia, 2 Ningxia Meteorological Observatory, Yinchuan China, 750002, 3 Nanjing Institute of Meteorology, Nanjing China, 210044)

1. Introduction • The application of satellite data is mainly confined to the synoptic, satellite meteorological conception model, and subjective, qualitative explanation in provincial meteorological agency. • The numerical weather prediction (NWP) models are getting more and more perfected, their products have become one of the indispensable parts for modern meteorology and. But the problem of initial field remains unsolved, thus the meso-scale NWP cannot work efficiently.

• The characteristics of meteorological satellite data make themselves play a very important role on data assimilation. • The hardware performance, technical staff in provincial meteorological services are not good enough to support the operational implement of complicated retrieval and assimilation system. • Simplicity and easy-going should be the basic features in provincial meteorological branch.

2. Satellite data and derived elements • The Geostationary Meteorological Satellite infrared data of 1999, 2000, and 2001 from May 1 st to Sept. 30 each year were used，the total samples is 6839. • The gray-level data at every meteorological station in Ningxia were read correspondingly. • The gray-level temporal change in a certain position was calculated as it did in the reference [1]. • Vk, i, j=Ck, i, j- Ck-1, i, j (1) • Here C is gray-level, k is time (hour), and i, j means the coordinate of the station, or geographic position. This element indicates the change range of gray-level in a certain temporal interval, and reflects the stability of weather system. • In the same data file, the spatial gradient was calculated as: • Gi, j= (2)

• This element reflects the consistency degree of the cloud. A small gradient means more similar near the taken pixel, and the weather tendency of cloud there is not obvious, vice versa. • The bright temperature was calculated and both the change rate and gradient were derived. • The hourly precipitations of 21 meteorological stations during that period were collected. • The fields of relative humidity on surface and at different levels of atmosphere were calculated for the study using the surface observation and sonde data.

3. Retrieval of meteorological satellite data • 3. 1 Assimilation and satellite data retrieved humidity • Chou[2], Charney[3] conducted experiment using satellite data. Kou[4] indicated that the precipitation forecast is very sensitive to the initial humidity. Because of the wide change of humidity both in space and time, and the sparsely location of meteorological observation, it is hard to describe the detail of moisture distribution and it is impossible to meet the need of NWP on initial condition. Shen[5] studied the assimilation with limited area model using satellite data and improvement of precipitation was achieved. Cui[6] adjusted the humidity at 850～ 400 h. Pa and found the ability of rainfall prediction was elevated. Lin[7], Zhu[8] researched the torrent rain NWP with satellite data retrieved humidity and obtained satisfactory results. • In the light of the studies above, humidity retrieval using satellite was carried out to provide good initial field for NWP in Ningxia.

• 3. 2 Non-linear retrieval and optimal fitting • Because of the non linear essential of atmosphere movement, the linear process is only an approximate approach under certain conditions. • It is hard to acquire ideal effect based on this relation. Retrieval with satellite data is a complex issue and a nonlinear model is more reasonable. • Regression: mature technology and easy to be applied but linear only. • If the very relation is nonlinear, the nonlinear regression can be done with special process. Here below are the cases of nonlinear regression.

• 1. fit with polynomial. Take x, x 2, x 3…xn as linear independent factors. • 2. Transforming with some functions to linear regression, such as Gagin’s[9], Bocchieri’s work [10]. • 3. Normalize the variable as Bocchieri[11] did with mathematic model and then to set up regression model. • The method in this paper is, try to find out more precise relations between the independent and dependent variables with suitable mathematic model developed by fitting on the samples. The linear relations can be established for regression after transform, and the further study can be conducted.

• In this case, the key to the problem is to find out the non linear relation. • In fact, because of the widely scattering of the sample, it may not be a strict linear relation after the transform, but the linear degree will be improved and it’s beneficial for the quality of regression. • Xie[12], Chu[13] adopted the fitting method on meteorological use, however it is quite rare.

• The works [9 -13] mentioned above assumed the fixed function and try to get the parameters only. It is limited and the function chosen may not an optimal one. • In this paper, we try to find out a cluster of mathematic functions to meet the physics demands, and get the parameters. So the optimal function can be found out in a certain extent. • Considering the shape of variables distribution, 48 functions were designed to fit the samples. • The functions chosen were sorted by the fitting effect, or residual error. The best 2 functions were listed in table 1.

Table 1. The fitted functions Items (number of passed functions) A. Relative humidity and gray-level (33) 1 the best 0. 3343302 -0. 43019449 -1. 6622662 0. 48286566 a= b= c= d= e= 0. 32236575 0. 27838399 -0. 55497431 0. 93163654 -0. 45718758 B. Relative humidity And temporal change of gray-level (30) y=a+b. cos(cx+d) a= 0. 49123164 b= 0. 097777643 c= 9. 1731609 d= -4. 1514746 a= b= c= d= 0. 15391716 0. 44475178 -2. 5936814 3. 7911048 … … … F. Relative humidity And special gradient of bright temperature (12) a= b= c= d= 2 the second 0. 65758276 4373. 3971 8078. 6531 8998. 4372 a= b= c= 1. 5260107 0. 70485075 0. 15384818

• 3. 3 Over fitting • The coefficients were improved with the best functions. All the absolute coefficients of others were elevated more than 0. 1. • But overfitting phenomenon appeared for gray-level. Because the function pursued the fitness too much, indulged special details and the particular non typical samples, the fit function is too sensitive and caused violent vibration. (Runge phenomenon [14, 15]. ) • In order to keep the stability and avoid the overfitting, the 2 nd function was taken for gray-level. • The linearity improvement of bright temperature is the most remarked, the absolute correlation increased 0. 135, and the others increased more than 0. 1 except temporal change of bright temperature which increased 0. 084. • Generally speaking, the correlation improved to a better level. The lowest is 0. 148 and the highest is 0. 367.

• 3. 4 Nonlinear regression • The regression can be conducted after the transform to set up the relation between humidity and the IR data of geostationary meteorological satellite. For concision, let: • • V 1=gray-level V 2=temporal change of gray-level V 3=spatial gradient of gray-level V 4=bright temperature V 5=temporal change of bright temperature V 6=spatial gradient of bright temperature V 7=relative humidity • Then • V 7＝a 1Ψ 1(V 1)+a 2Ψ 2(V 2)+a 3Ψ 3(V 3)+a 4Ψ 4(V 4)+a 5Ψ 5(V 5)+a 6Ψ 6(V 6)+b (7) • Here ai, b are coefficient of regression and Ψi() are the nonlinear functions selected above and i=1，2，3，… 6.

Table 2. effect of regression before and after transform. Total deviati on Correlati on Order of the factors Standard error Residual F Before 19. 864 . 312 V 6, V 5, V 1, V 3, V 2, V 4 . 1928 184. 192 89. 027 After 46. 662 . 478 V 6, V 4, V 2, V 1, V 3, V 5 . 1783 157. 394 244. 733

• The regression with independent fitting transform is obviously better than the original one. The correlation enhanced 53. 2% from 0. 312 to 0. 478, the standard error decreased from 0. 1928 to 0. 1783 and the residual decreased 27. 189, from 184. 192 to 157. 394. The f was employed to examine the significance of the regression： (8) Here n: number of samples and p: number of factors The bigger the correlation, the bigger the f, and the better the regression. The f of transformed regression is 2. 75 times of that of original one. Take degree of freedom as 6 and the number of samples ∞, α＝ 0. 05, then F 0＝ 2. 22。 Both fs of the 2 regressions developed are bigger than F 0, so the regression models are acceptable or significant, and the nonlinear regression is better.

Table 3. Regression parameters, A: original, B: fitting transformed. Non-normalized coefficients Standard error normalized coefficients A B A Intercept -5. 025 -. 347 1. 067 V 1 11. 918 . 135 1. 986 . 028 1. 759 V 2 -. 912 1. 614 . 227 . 258 V 3 . 588 -. 289 . 407 V 4 2. 691 . 825 V 5 -. 572 V 6 -1. 306 B t A B -4. 709 -5. 203 . 106 6. 000 4. 786 -. 341 . 290 -4. 011 6. 256 . 054 . 059 -. 165 1. 444 -5. 348 . 540 . 051 1. 460 . 326 4. 983 16. 025 -. 431 . 233 . 318 -. 209 -. 064 -2. 460 -1. 356 -0. 036 . 290 . 061 -. 189 -. 018 -4. 510 -. 589

• BP Network to retrieve • Generally 4 layers are enough to form a difficult network, and we think that the forecast is a difficult problem and a 4 -layer network is needed. • Experiment was conducted to find the optimal structure, and the best is that has 6 nodes in each hide layer, and the network to retrieve was built by training No. Nodes in 2 nd hide layer 6 error 1 Nodes in 1 st hide layer 6 2 5 4 0. 051472 3 4 4 0. 051505 4 3 3 0. 051599 5 2 4 0. 051649 6 3 4 0. 051656 7 7 2 0. 051665 8 7 6 0. 051668 … … 49 1 6 0. 054079 0. 051363

• 3. 5 Retrieval precipitation of satellite data for heavy • Geographically, Ningxia located in the northwest China, far inside the mainland with an arid climate. Take Yinhcuan city, the capital as an example; the annual precipitation is only about 200 mm. Heavy rainfall must go with severe convection. Without abundant moisture, even though the convection happens, lack of enough energy supply of latent heat of condensation provided by moisture, the convection will hardly develop, and the heavy rainfall will not occur. • After all, heavy rainfall in Ningxia must accompany with severe convection and, the severe convection must be supported by plentiful moisture. • For the torrent rainfall, moisture is the key condition. 990 samples were selected taking relative humidity as threshold and the criterion was set as 75% experimentally. With the nonlinear approach employed, the relative humidity retrieval models at different levels of atmosphere were developed and the accuracy was further improved.

4. Brief summary of experiment • Using the models developed above, relative humidity retrieval can be processed easily with satellite data, and it can satisfy the requirement of provincial meteorological department. • After the retrieval using the approach in this paper and neutral network, variational assimilation were conducted as reference [16] did. • Relative humidity quality control. In space, no introduction here. • Then the retrieved humidity was put into MM 5 V 3 with other initial fields of T 213 output, for a sudden occurred heavy rainfall in northern Ningxia on July 21 2003, which caused a flood. • The results show that the precipitation region forecast with retrieved humidity is highly consistent with observation, and the precipitation products are helpful to the operational forecast, and spin up was improved 25%. While the forecast without retrieved data failed in this experiment.

• • Satellite data: July 21 2003, 08: 00 am, FY－2 IR The grid is 60 km west and 80 km south to the actual position in the images. The information was read directly from satellite data after geographic calibration, in stead of from the images above.

• • Retrieved relative humidity after quality control and originals at 400 and 500 h. Pa. Dot line(blue): humidity of T 213, solid line(red): retrieved after quality control. It is clear that the retrieved relative humidity after quality control is consistent with the original in macroscale, so it is beneficial for the model to run stable. • On the other hand, the mesoscale systems retrieved are much more obvious compared with the T 213 s.

• A case of heavy rainfall • a convective torrential rain was observed at 14: 00 -15: 00, July 21 2003 in north part of Ningxia. Maximum hourly precipitation reported 22. 3 mm, with 3. 2, 1. 6, 1. 8 and 0. 1 mm in some other stations and the hour next. • Spatial and temporal distribution of precipitation

• MM 5 V 3 precipitation without satellite data. The convective precipitation only, with a very tiny value 0. 2128 mm at the south end of Ningxia (hundreds km away) the maximum forecasted 4 h later than the observation (spin-up) • The MM 5 can not indicate this convective rainfall at all without the assimilation.

• The satellite data at different levels were retrieved and quality controlled (The relative humidity fields above), then they were put into the MM 5 v 3. • 14: 00, 15: 00, 16: 00 July 21, 2003

• 17: 00 18: 00 precipitation (mm) • The precipitation got enhanced greatly, the convective rainfall was forecasted in north part of Ningxia since 14: 00, the exact area of observed, and middle part also, maybe false, without observation evidence. 15: 00 north: stronger 0. 2 mm, middle: disappear. North only. 16: 00: 0. 8. 17: 00 1. 1 mm, 18: 00 0. 5 mm and the area shrinking. 19: 00 stopped.

5. Conclusion and discussion • With optimal fitting, the mathematical relations between the relative humidity and meteorological satellite data were found out and, all of the relations are nonlinear. • After the transform with these nonlinear functions, nonlinear regression was achieved, and the effect of regression was enhanced. • For the heavy rainfall, the capability of retrieval improved further. Comparing with the traditional regression, it is not only much more reasonable, but achieved a rather high precision also. • The primary experiment indicated the NWP forecast made a distinct advance with retrieved data for a sudden occurred severe convection with an intensive precipitation and caused a flood, meanwhile the NWP without assimilation failed to forecast.

Reference: • 1. HU Wen-dong, Chen Xiao-guang, LIU Jian-jun. Analysis on the relation between the summer hourly precipitation and geostationary meteorological satellite data in Ningxia [J], Ningxia engineering and technology, (in Chinese) 2003，2(4): 310 -314 • 2. CHOU Ji-fan. Application of historical data in NWP[J], China Science B, (in Chinese) 1974，89: 875 -883 • 3. Charney J G, Halem M, Jastrow R. Use of Incomplete Historical Data to Infer the Present State of Atmosphere[J]. J Atmos Sci 1969 26: 1160 -1163 • 4. Kou, Y. H. Assimilation of Precipitable Water Measurements into a Mesoscale Numerical Model [J]. Mon. Wea. Rev, 1993, 121 (5): 1215 -1238. • 5. SHEN Tong-l, MIN Jin-zhong, WU Cheng-ou. Experiment and design of the variational analysis system for the satellite cold data in the limited area [J]. plateau meteorology, (in Chinese) 1996，15 1: 58 -67 • 6. CUI Cun-guang. Moisture initial condition and sensitive experiment on heavy precipitation with meso-scale model [J], Meteorology, (in Chinese) 26（11）： 3－7 • 7. Lin Lei, ZHU Min, YU Fan et al. Meso-scale NWP application and case experiment on combination of moisture retrieval using satellite data [J]， meteorology science，(in Chinese) 2001，21（3）： 279 -290

• 8. ZHU Min, YU Fan, ZHENG Wei-zhong et al. Moisture retrieval using satellite data and primarily analysis on its application in torrent rainforecast [J]，Journal of meteorology，(in Chinese) 2000，58（4）： 470 -478 • 9. Gagin. A, Rosefield. D. The Relationship between Height and Precipitation Characteristics of Summertime Convective Cells in South Florida[J]. J. Atmos. Sci. 1985, 42(1): 84 -94 • 10. Bocchieri. J, R. Use of the Logit Model to Transform Predictors for Precipitation Forecasting[C]. Preprint 6 th conf. On Probability and Statistics in Atmos. Sci. Amer. Meteor. Soc. 1979: 49 -54 • 11. Bocchieri. J, R. A New Operational System for Forecasting Precipitation Type[J]. Mon. Wea. Rev. 1979, 107: 637 -649 • 12. XIE Ding-sheng, LIANG Feng-yi, YI Ai-min. nonlinear forecast on heavy rain in Guandong [J], catastrophe，(in Chinese) 1999，14（2）： 18 -22 • 13. CHU Rong-zhong, JIA Wei. Precipitation estimation with biline polarized radarⅠ．Ordering partner approach [J], plateau meteorology，(in Chinese) 1999， 18（1）： 97 -108 • 14. FENG Kan. Numeric calculation [M]. Beijing: defense industry press house, (in Chinese) 1978. 1— 40. • 15. LI Yue-sheng. Numeric approach[M]. Beijing: people education press house, (in Chinese) 1978. 65— 68, 76— 132. • 16. ZHENG Zuo-fang, SHEN Tong-li, ZHANG Xiu-li. Quality control on assimilation with satellite data and its application on heavy rain forecast [J], catastrophe, (in Chinese) 2003，18（1）： 1－8

thanks for your attention 谢谢 merci 胡文东 宁夏气象防灾减害重点实验室，宁夏银川，750002, HU Wendong Key Laboratory of Meteorological Disaster Preventing and Reducing in Ningxia, Yinchuan China, 750002, hu. wendong@163. com, 86 -951 -5043015