SEARCH

SEARCH BY CITATION

Keywords:

  • storm characteristics;
  • hypothesis testing;
  • support vector machine

Abstract

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. The new storm identification method
  5. 3. Storm discrimination and the warning technique
  6. 4. Summary and future work
  7. Acknowledgements
  8. Appendix A. The SVM Classifier
  9. References

A new storm identification and warning technique is proposed which exclusively uses radar data as input. The new identification method assembles contiguous storm points to constitute 2D storm components and improve the vertical association of storm components to construct 3D storms, which can overcome the deficiencies existing in traditional identification methods. Based on the evolution properties and characteristic distributions, strong storms and general storms are specified to train support vector machines (SVMs) which then can be used to discriminate storms. The performance of the SVM shows that it can indicate the intensity and development of a storm, providing an important aid in severe weather warning. Copyright © 2011 Royal Meteorological Society


1. Introduction

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. The new storm identification method
  5. 3. Storm discrimination and the warning technique
  6. 4. Summary and future work
  7. Acknowledgements
  8. Appendix A. The SVM Classifier
  9. References

Convective weather systems such as hailstorms, rainstorms and tornadoes may bring serious damage to daily production and life. Therefore, it is meaningful to promote the prediction level of convective weather and give a practical warning alarm to reduce casualties and property losses. With high temporal and spatial resolution, Doppler weather radar can provide information such as reflectivity, radial velocity and spectral width, of importance in nowcasting compared to other traditional procedures. Storm identification and monitoring through radar systems and computer technology become an important technique for severe weather warning.

Storm identification can be roughly divided into two categories: cross-correlation methods and centroid methods. The cross-correlation method takes the 2D reflectivity data to calculate the motion vectors of subregions that can be used to forecast the storm movement (Rinehart and Garvey, 1978; Li et al., 1995). This method cannot identify and track individual storms, which constrains obtaining detailed storm parameters. In contrast, the centroid method (Crane, 1979; Browning et al., 1982; Rosenfeld, 1987) identifies storm entities within single radar volume scan data and tracks storms across consecutive scans followed by forecasting storm positions based on the movement of the identified storm centroids. Because the centroid method can provide characteristic parameters of storms it draws more attention in severe weather warning. Typical centroid methods include TITAN (Thunderstorm Identification, Tracking, Analysis and Nowcasting) (Dixon and Wiener, 1993) and SCIT (Storm Cell Identification and Tracking) (Johnson et al., 1998), which are widely used throughout the world. They generally take three steps to perform storm identification using a reflectivity threshold (e.g. 35 dBZ in TITAN) or several thresholds (e.g. 30, 35, 40, 45, 50, 55 and 60 dBZ in SCIT). The procedure is: (1) contiguous sequences of points are searched and recorded along radials to create 1D storm segments if the reflectivity exceeds the threshold; (2) storm segments are combined to constitute 2D storm components if a continuity requirement is satisfied by neighbouring storm segments on the same PPI scan, and, (3) storm components are associated with each other to construct 3D storms on different PPI scans. The defect of this procedure lies in the fact that it needs firstly to search storm segments in radials and secondly compose storm components in azimuths, possibly causing identification errors and track failures especially for storms with irregular shapes or variable wind speeds (Han et al., 2009). Occasionally, several storm components may actually be detected as one when they are clustered closely, or a single storm component may be detected as several (Johnson et al., 1998).

Nowcasting has become an increasingly important and specialized subject when forecasters need to make critical operational decisions in the face of rapidly developing weather situations in weather services (Liljas, 1998). Radar-based nowcasting techniques are superior to NWP models for periods less than a few hours when the model performance has not yet stabilized (Li and Lai, 2004). Traditional schemes are mainly for location forecasting which has been based primarily on the extrapolation of radar reflectivity echoes (Wilson et al., 1998), or longevity forecasting which has been based on storm characteristics and clear-air signatures (Henry, 1993; Wilson and Megenhardt, 1997; MacKeen et al., 1999). Fewer studies have focused on storm properties indicating the convective stage or intensity which can provide real-time warning information. Hand and Conway (1995) used a simple conceptual model to depict and forecast the life cycle of a convective storm as a sequence of six stages. However, the initial stage is only determined by the corresponding reflectivities whose values were suggested by experience and by the beam height which varies with the distance from the radar. Besides, little use was made of conventional surface observations while assumptions and approximations were often applied in ascertaining the storm stage that many of these would require investigation and refinement. Furthermore, the geostationary satellite data and mesoscale NWP model were involved in the scheme, which constrains the temporal and spatial resolution. Hu et al. (2007) computed the convection index for storm intensity which can serve as a warning tool according to the statistical distribution of storm characteristics. However, the contributions of different characteristics are considered to be the same (equal weights are adopted in the convection index). Additionally, choices of strong storms and membership functions lack uniform and precise standards.

In this paper, a new storm identification and warning scheme is proposed. Taking account of defects of commonly used identification methods, the new storm identification algorithm is designed to improve the identification pattern. Storms appearing in Jiangsu Province of China during convective weather processes from 2007 to 2009 are identified with the new algorithm from the data of a new generation of Doppler weather radar in Nanjing. According to the evolution properties and characteristic distributions of storms, strong storms and general storms are specified to train the support vector machine (SVM) which then can be used to discriminate storms. In this way, effective warning alarms can be provided when a strong storm emerges.

2. The new storm identification method

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. The new storm identification method
  5. 3. Storm discrimination and the warning technique
  6. 4. Summary and future work
  7. Acknowledgements
  8. Appendix A. The SVM Classifier
  9. References

2.1. The method of assembling contiguous points for 2D storm component

Avoiding defects of traditional storm identification algorithms used in TITAN and SCIT, a new algorithm is designed to identify 2D storm components simultaneously in both directions. The basic idea of the algorithm is very simple. Points with reflectivity exceeding the threshold (one threshold as in TITAN or several thresholds as in SCIT) determined in advance are searched to serve as storm points. Contiguous storm points constituting a 2D storm component are then assembled group by group until all storm components have been identified. Contiguous points are defined as the eight points around a reference point on a 2D plane.

In the first step of assembling, a storm point serves as the original reference to assemble its contiguous points which serve as the references in the second step. The initial reference point is excluded from the second step to avoid repeated assembling. The storm points contiguous to the reference points assembled in the second step serve as the references in the third step and each reference point and its contiguous points which have been already assembled are excluded from the assembling for the next reference point to avoid repeated assembling. This procedure is repeated one after another until no storm points contiguous to the reference points can be found in a step. Consequently, the storm points that have been assembled constitute a 2D storm component. Other storm components are identified in the same way until all storm points have been assembled. Figure 1 is a schematic diagram of the algorithm for assembling contiguous points. After all storm components are identified, those whose areas are less than a threshold are eliminated for the sake of noise filtering (e.g. storm component 1 in Figure 1).

thumbnail image

Figure 1. A schematic diagram of the algorithm for assembling contiguous points. R represents a storm point whose reflectivity exceeds the threshold. The number at the top left corner represents the serial number of the storm component to which the storm point belongs, while the number at the top right corner represents the serial number of the step in which the storm point is assembled

Download figure to PowerPoint

Because the new algorithm identifies 2D storm components simultaneously in both directions, it can overcome identification deficiencies existing in traditional algorithms. Figure 2 illustrates some instances in which the new algorithm gives reasonable identification results while TITAN and SCIT not. For storm component 1, SCIT and TITAN separate it into two parts (segments in radials B–D, and segments in radials E–G) because of omitting the point Dh due to the length limitation of the storm segment in SCIT and TITAN (the segment is required to be not shorter than two range gates), which is a case of false split. For storm components 2 and 3, SCIT identifies them as one because some contiguous points whose reflectivity values are not more than the DROPOUT REF DIFF below the reflectivity threshold are included (DROPOUT REF DIFF is the difference in effective reflectivity of points below the threshold that may still be included in a radial segment identified with the reflectivity threshold (the default value is 5 dBZ) (Johnson et al., 1998)), which is a case of false merger. This kind of false split or merger may usually occur when SCIT implements higher reflectivity thresholds. As for storm component 4, TITAN and SCIT give no identification since they need firstly to identify 1D storm segments in radials satisfying the length requirement.

thumbnail image

Figure 2. A schematic diagram of identification results of 2D storm components. Radials and range circles are represented by uppercase letters from A to R and lowercase letters from a to m, respectively. An Arabic numeral (1, 2, 3 or 4) represents a point whose reflectivity exceeds the threshold. A letter ‘r’ represents a point whose reflectivity is not more than DROPOUT REF DIFF below the reflectivity threshold as in SCIT. A group of points marked with the same Arabic numeral represents a 2D storm component identified by the new algorithm

Download figure to PowerPoint

Figure 3 gives two real cases to display some identification problems in TITAN and SCIT and the improvement achieved by the new algorithm. In Figure 3(a), TITAN isolates a smaller southwest part from the neighbouring storm segments which have already been combined with the segments in the northwest, while, by contrast, the new algorithm can identify the storm component as a whole. In Figure 3(b), because including some points whose reflectivity values are not more than the DROPOUT REF DIFF below the reflectivity threshold in a storm segment, SCIT improperly identifies a storm component 3 which is unreasonable under the threshold of 50 dBZ. In addition, storm components 1 and 4 are not identified by SCIT under the area threshold since many storm points are omitted along several radials due to the length requirement of 1D storm segments. By contrast, the new algorithm identifies storm components 1, 2, 4 and 5, which is more reasonable.

thumbnail image

Figure 3. Identification results of two real cases by the new algorithm, TITAN and SCIT under an area threshold of 20 km2. The reflectivity data is collected by Nanjing Doppler radar at an elevation angle of 0.5° at (a) 0642 UTC on 30 July 2007 and (b) 0009 UTC on 21 March 2009. In (a), both the new algorithm and TITAN use a threshold of 35 dBZ. The new algorithm identifies one storm component but TITAN identifies two including a separating part on the southwest marked by the circle. In (b), both the new algorithm and SCIT use thresholds of 30, 35, 40, 45, 50, 55 and 60 dBZ. The new algorithm identifies storm components 1, 2, 4 and 5 while SCIT identifies 2, 3 and 5

Download figure to PowerPoint

2.2. The configuration of 3D storms

Generally, the raw data of a weather radar are of the volume scan and have multiple elevation angles. Therefore, a vertical association of 2D storm components is attempted for 3D storms on different PPI conical scans. In SCIT, components at consecutive elevations whose centroids are horizontally within a threshold (e.g. 5, 7.5 and 10 km) are associated, with only one-to-one association being allowed. In fact, the horizontal distance between storm components on consecutive elevations may exceed the threshold due to storm slanting (Figure 4(a)), leading to a failure association for SCIT. Additionally, when there are several components at a higher elevation and one storm component at a lower elevation (Figure 4(b)), one-to-one association can cause identification error.

thumbnail image

Figure 4. Examples that 2D storm components at consecutive layers cannot be properly associated for 3D storms by SCIT because of (a) a slanting storm and (b) several components at a higher layer. In (a), C1 and C are centroids of storm components at the higher layer and at the lower layer respectively, and C1′ is the projection of C1 at the lower layer. The horizontal distance CC1′ is larger than the given threshold, and thus they can not be associated by SCIT. In (b), C1 and C2 are centroids of the two storm components at the higher layer, C is the centroid of the storm component at the lower layer, and C1′ and C2′ are projections of C1 and C2 at the lower layer. Either of the components at the higher layer may be associated with the one at the lower layer, but not both, since SCIT uses one-to-one association

Download figure to PowerPoint

To deal with these problems in SCIT, the new method uses the following strategies to perform the vertical association: (1) components at consecutive elevations whose centroids are horizontally within a range threshold (e.g. 10 km) are associated. If several components at a higher elevation can be associated with a component at a lower elevation, all of them are associated (i.e. many-to-one association is introduced) (Figure 4(b)); (2) when non-associated components remain after (1), the component at an elevation is associated with the component at the consecutive elevation whose horizontal boundary involves the horizontal position of the centroid of the former component (Figure 4(a)).

For each identified 3D storm, parameters such as the centroid position, top height, base height, volume, mean reflectivity, maximum reflectivity and storm-based vertically integrated liquid water can be calculated for analysis.

3. Storm discrimination and the warning technique

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. The new storm identification method
  5. 3. Storm discrimination and the warning technique
  6. 4. Summary and future work
  7. Acknowledgements
  8. Appendix A. The SVM Classifier
  9. References

3.1. Data and the identified storms

A threshold of 35 dBZ is commonly considered as an indicator of a convective echo in many studies (Dixon and Wiener, 1993; Henry, 1993; Roberts and Rutledge, 2003; Mecikalski and Bedka, 2006). Therefore, it is adopted in this research as the reflectivity threshold for storm identification. Given that noise and clutter appear generally near the ground, the area threshold for storm components is set to 20 km2 at the lowest elevation and 10 km2 at other elevations.

According to the weather phenomena observed by the automatic weather stations (AWSs) in Jiangsu Province of China (Figure 5), convective weather processes including hailfall and 1 h heavy rainfall (1 h precipitation of more than 20 mm which is defined as the short-time heavy rainfall in China, hereinafter inclusive) are chosen from 2007 to 2009. From the volume scan data collected by the new generation Doppler weather radar in Nanjing during these processes (some technical parameters of the radar are listed in Table I), storms are identified using the new storm identification method described in Section 2.

thumbnail image

Figure 5. The Doppler weather radar in Nanjing and the automatic weather stations (AWSs) in Jiangsu Province of China. The circle and points respectively represent the location of the radar and the AWSs, respectively

Download figure to PowerPoint

Table I. Some technical parameters of the new generation Doppler weather radar in Nanjing
Technical parametersValues
Operating frequency2860 MHZ
Transmitter peak power698.9–730.9 KW
Receiver noise figure3.06 dB
Pulse width1.57 µs, 4.50 µs
Reflectivity calibration error0.53 dBZ
Elevation governing error0.07°
Azimuth governing error0.03°
Volume coverage patternVCP21 (precipitation mode with 14 elevations)
Maximum detection range460 km
Time resolution6 min

To track storms and deal with splitting and merging events accurately, manual analysis is adopted for storm tracking using radar images. Forty tracked storms suitable for analysis were selected based on the following criterions: (1) the distance to the radar is within 250 km and, (2) the evolution process of the life cycle is accomplished in Jiangsu Province. The 40 tracked storms distribute at 876 time intervals altogether, including 15 tracked storms producing hailfall or 1 h heavy rainfall in their life cycles.

3.2. The specification of strong storms and general storms

It is necessary to choose proper storm characteristics to evaluate the storm intensity for the specification of strong storms and general storms. First of all, it is evident that the maximum reflectivity (Rmax) can indicate the storm intensity. In addition, storm echo top height (ET) and storm-based vertically integrated liquid water (VIL) are another two important characteristics. A high ET value means that the storm develops to a high altitude and has a strong updraft so that the size of water droplets or solid particles can be large. A high VIL value means that the storm has plenty of water content and a significant convection so that the probability of appearing heavy rainfall or hailfall is high (Greene and Clark, 1972; Steven and Wolf, 1997). Besides, storm centroid height (Z) is a characteristic that can give an indication of the evolution stage of a storm (Baeck and Smith, 1998). When the storm is initiated, the water vapour is taken from the low altitude to the high, causing relatively small Z at this moment. When the storm reaches the mature stage, the updraft is intensive with plenty of water vapour condensed at the high altitude and the storm has a larger Z. When the storm begins to dissipate, the downdraft is dominant and the water vapour falls to the ground in the form of liquid or solid precipitation, with Z decreasing.

Through the analysis of the 15 tracked storms producing hailfall or 1 h heavy rainfall, Z, ET and VIL have obvious sudden increase and decrease before and after the hailfall or the 1 h heavy rainfall. The sudden increase or decrease time is determined from the following standards which are suggested by experience to be appropriate for storms in Jiangsu Province of China: (1) 1.5 h before and 0.5 h after the hailfall time, or 0.5 h before and 0.5 h after the 1 h heavy rainfall time; (2) closest to the hailfall or 1 h heavy rainfall time and, (3) remarkable increase or decrease compared with the mean variation of storm characteristics. Analysis shows that the sudden increase or decrease time determined by Z, VIL and ET is approximately the same. The sudden increase indicates that the strong updraft makes the water vapour accumulate quickly and condense to big drops or ice particles leading to rainfall or 1 h heavy hailfall. The storm weakens rapidly due to the drag effect when the hailfall or 1 h heavy rainfall occurs, presenting a sudden decrease of the characteristics. Z, VIL, ET and Rmax values of the 15 tracked storms at time intervals from the sudden increase time to the sudden decrease time (106 time intervals altogether) are remarkably higher than for the storms at other time intervals. Therefore, these 106 storms (storms are at 106 time intervals altogether) can be specified as strong storms and the corresponding characteristics can be regarded as characteristics of strong storms.

However, it should be noted that some tracked storms experienced similar evolution processes to the 15 tracked storms and had similar characteristic values at some time intervals to those 106 specified strong storms, but there was no convective weather such as hailfall or 1 h heavy rainfall associated with them. The reason may be attributed to: (1) the convective weather did occur but it was not observed by the AWSs (e.g. a small storm did not move above any AWS when it generated convective weather), or (2) the storm would have generated convective weather if it had a regular evolution, but it encountered merging or splitting events in its life cycle. Obviously, these storms at some time intervals should be also treated as strong storms instead of general storms to give warning alarms, so general storms need to be specified through a more precise method, whether or not the convective weather is observed by the AWSs.

The characteristic probability distributions of the 106 specified strong storms and all 876 storms (the 40 tracked storms are at 876 time intervals altogether) are illustrated in Figure 6, and the corresponding mathematical expectations µ and standard deviations σ are calculated in Tables II and III. It can be seen that characteristics of the strong storms and all storms approximately obey the normal distributing, and µ for strong storms is larger than for all storms, while σ is similar. This means that the larger the characteristic value, the higher the probability that the storm belongs to strong storms but, at the same time, the storm still has some probability that it belongs to general storms even if the characteristic value is very high. Characteristic distributions of general storms are not given, but its rough form can be deduced from those of strong storms and all storms (Figure 7). Because there is an overlapping area between distributions of general storms and strong storms, the characteristics of a general storm may conform to the characteristic distributions of strong storms. Similarly, the characteristics of a strong storm may conform to the characteristic distributions of general storms. It is assumed that X0 is the value of a characteristic. If the characteristic value of a general storm is less than X0, the probability that it is classified as a strong storm by mistake is the shaded area A. If the characteristic value of a strong storm is more than X0, the probability that it is classified as a general storm by mistake is the shaded area B. B can not be acquired because distribution parameters for general storms are unknown while A can be achieved based on the characteristic distributions of the 106 specified strong storms. Specifying general storms from all storms requires A to be as small as possible, so a hypothesis testing procedure can be designed as follows. H0: a general storm is classified as a strong storm. If the probability of this error does not exceed some error threshold α0, H0 is rejected and the storm is specified as a general storm. A corresponding characteristic value threshold for general storms can be calculated through α0, and if the characteristic value of a storm is less than the threshold, the storm is specified as a general storm. Because all characteristics of a strong storm should obey corresponding characteristic distributions of strong storms, a storm is specified as a general storm if the value of any of the four characteristics is less than the corresponding threshold. Table IV gives the characteristic thresholds and the number of general storms specified under different error thresholds, α0.

thumbnail image

Figure 6. Characteristic probability distributions of the 106 specified strong storms and all 876 storms. (a) Z; (b) Rmax; (c) VIL; (d) ET. equation image all storms, equation image strong storms

Download figure to PowerPoint

thumbnail image

Figure 7. Probability distributions of storm characteristics abstracted from Figure 6. The solid, dotted and dashed curves represent characteristic probability distributions of strong storms, all storms and general storms, respectively. X0 is a characteristic value. The shaded area A denotes the probability that a general storm is classified as a strong storm by mistake if its characteristic value is less than X0. The shaded area B denotes the probability that a strong storm is classified as a general storm by mistake if its characteristic value is more than X0. equation image general storms, equation image all storms, equation image strong storms

Download figure to PowerPoint

Table II. Mathematical expectations and standard deviations of characteristics of the 106 specified strong storms
Strong stormsZ (km)Rmax (dBZ)VIL (kg·m−2)ET (km)
µ3.7160.6210.199.79
σ1.044.353.712.02
Table III. Mathematical expectations and standard deviations of characteristics of all 876 storms
All stormsZ (km)Rmax (dBZ)VIL (kg·m−2)ET (km)
µ3.0258.196.687.34
σ1.015.693.892.74
Table IV. Characteristic thresholds and the number of general storms specified under different error thresholds α0
α0Z (km)Rmax (dBZ)VIL (kg·m−2)ET (km)Ngeneral
0.12.3855.055.447.2495
0.052.0153.464.096.47382
0.011.350.51.565.09194

3.3. Storm discrimination using the SVM

It is necessary to discriminate whether a storm is a strong storm or not at a time interval to give a real-time warning. With the SVM classifier which is trained by the strong storms and general storms specified in Section 3.2, a storm can be classified as a strong or general storm by the SVM. A brief introduction and description of the SVM classifier used for the present study is given in Appendix A.

The set of specified strong and general storms is randomly divided in two parts. One half is used to train the SVM, and the other is used to verify the recognition result of the SVM. For the purpose of facilitating comparison, the four characteristics are normalized. Table V gives the SVM training and verification results by storms specified under different error thresholds α0 and penalization parameters C. w*1, w*2, w*3, and w*4 are the coefficients (weights) corresponding to Z, Rmax, VIL and ET, respectively. The accuracy rate is the proportion that the SVM gives correct recognition, the missing rate is the proportion that the SVM specifies strong storms as general storms by mistake and the vacancy rate is the proportion that the SVM recognizes specified general storms as strong storms by mistake.

Table V. Training and verification results of SVMs by storms specified under different error thresholds α0 and penalization parameters C
α0Cw*1w*2w*3w*4b*Accurate rateMissing rateVacancy rate
0.010.20.73581.05291.74091.6705− 2.34110.94670.05330
 0.50.97261.40382.38652.4503− 3.03820.95330.04670
 0.80.85311.62832.70552.8298− 3.3680.960.040
0.050.21.02280.63971.74361.8553− 2.86460.91390.08610
 0.51.29790.7452.59232.7818− 3.69390.93030.06970
 0.81.74110.9093.20893.4117− 4.4260.93850.04920.0123
0.10.20.95690.80491.91861.5985− 3.10630.90330.09670
 0.51.13310.94932.71371.9935− 3.74850.920.080
 0.81.45311.16883.31042.33− 4.37190.930.06670.0033

It can be seen from Table V that w* and accuracy rates increase with C, while missing rates decrease with C. These phenomena are comprehensible because: (1) the convex hulls of the two groups of training points become closer to each other as C increases and, therefore, the hyper plane can correctly classify more storms, yielding higher accurate rates and lower missing rates, and, (2) the convex hull for general storms expands more than that for strong storms so that the hyper plane moves toward training points of strong storms since the distribution space of general storms is larger than that of strong storms, characterized by w* increasing. Among the four coefficients w*, w*3 and w*4, respectively corresponding to VIL and ET, are larger, while w*1 and w*2, corresponding to Z and Rmax. are smaller, with the former two coefficients twice or three times as much as the latter two. This means that VIL and ET contribute more to the storm intensity and their contributions are twice or three times as much as Z and Rmax.

In Table V, SVMs give encouraging results in which all accuracy rates are above 90% and the best results are achieved by the SVMs when α0 = 0.01. However, the specified general storms used by these SVMs are relatively few (see Table IV) and thus the results lack persuasiveness. For this reason, SVMs trained by storms specified under different α0 are verified by storms specified under other α0. Corresponding verification results are given in Table VI and it is optimal for the SVM when α0 = 0.05 and C = 0.8. Consequently, this SVM is used for practical storm discrimination and severe weather warning. When the SVM indicates the emergence of a strong storm, the warning alarm is provided.

Table VI. Verification results of SVMs trained by storms specified under different α0 and verified by storms specified under other α0
α0CStorms when α0 = 0.01Storms when α0 = 0.05Storms when α0 = 0.1
  Accuracy rateMissing rateVacancy rateAccuracy rateMissing rateVacancy rateAccuracy rateMissing rateVacancy rate
0.010.20.96930.02050.01020.92680.01660.0566
 0.50.95080.01640.03280.86360.01330.1231
 0.80.94880.01230.03980.85520.010.1348
0.050.20.87330.126700.93180.06320.005
 0.50.920.0800.94840.03990.0116
 0.80.94330.056700.94680.02830.025
0.10.20.810.1900.88320.11680
 0.50.850.1500.90780.09220
 0.80.880.1200.92620.07380

3.4. The performance of the warning technique

The warning technique is applied for a tracked hailstorm on 26 July 2007 (Figure 8). This storm was initiated in the southeast of Jiangsu Province and then moved towards the northeast. It had a remarkable strengthening at 0844 UTC when Z, VIL and ET increased by 1 km, 2.9 kg m−2 and 3.3 km respectively, and the SVM indicates that it strengthened to a strong storm at this time. There was a more intense strengthening at 0850 UTC when Z, Rmax, VIL and ET increased by 1.3 km, 6 dBZ, 7.7 kg m−2 and 3.9 km respectively. The storm remained strong until hailfall appeared at 1015 UTC. After that, Z decreased gradually from 5.0 km at 1015 UTC, 4.1 km at 1021 UTC and 3.3 km at 1027 UTC until 2.6 km at 1033 UTC. A remarkable weakening appeared at 1033 UTC when Rmax, VIL and ET respectively decreased by 9 dBZ, 5.7 kg m−2 and 4.2 km, and the SVM indicates that it weakened to a general storm at this time. The storm then dissipated as soon as it moved eastward over the sea. From 0844 UTC to 1033 UTC, strong storms lasted more than 100 min, which seems a long time for a single cell storm. The tracked storm merged with an intensive storm to its southeast at 0926 UTC, which could affect its structure and inner flow although the characteristic values did not change obviously. This merger event might prolong its life cycle and put back the hailfall time.

thumbnail image

Figure 8. The evolution process of the hailstorm on 26 July 2007. The four curves represent variations of normalized Z, Rmax, VIL and ET, respectively. The squares and circles on the time coordinate axis respectively represent general storms and strong storms discriminated by the SVM. The discrimination results coincide with variations of the four characteristics. equation imageZ, equation imageRmax, equation imageVIL, equation imageET

Download figure to PowerPoint

4. Summary and future work

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. The new storm identification method
  5. 3. Storm discrimination and the warning technique
  6. 4. Summary and future work
  7. Acknowledgements
  8. Appendix A. The SVM Classifier
  9. References

A new storm identification and warning scheme has been proposed. Compared with traditional methods such as TITAN (Thunderstorm Identification, Tracking, Analysis and Nowcasting) and SCIT (Storm Cell Identification and Tracking), the new storm identification method assembles contiguous storm points simultaneously in both directions on a radar PPI scan to identify 2D storm components. In addition, the vertical association of 2D storm components between consecutive PPI scans for 3D storm is improved by introducing the position relationship between centroids and boundaries of storm components and by allowing many-to-one association. Applications show that the new method can give reasonable identification in contrast to problems encountered by TITAN and SCIT.

Storms identified with the new method were analysed using data collected by the Doppler weather radar in Nanjing of Jiangsu Province of China for convective weather processes from 2007 to 2009. Based on the storm evolution properties and probability distributions of characteristics Z, Rmax, VIL and ET, strong storms and general storms are specified to train support vector machines (SVMs). Experiments verify that SVMs can give encouraging recognition of strong and general storms. In addition, SVMs show that VIL and ET make primary contributions to the storm intensity, with twice or three times as much as the secondary contributions made by Z and Rmax. The optimal SVM is applied to a hailstorm case and the results show that it can indicate the storm intensity and development. When the SVM indicates the emergence of a strong storm, the warning alarm is provided and a closer watch on the storm evolution is required.

Although some results are achieved, many problems are still waiting to be studied. These include the influence of merging and splitting on storms, detailed analysis on outer environments and inner flow fields of storms, evolution differences between hailstorms and rainstorms. As a result, future work will analyse storm structures using multiple data collected by Doppler radars, satellites and radiosondes, and to provide more assistance to nowcasting and severe weather warning.

Acknowledgements

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. The new storm identification method
  5. 3. Storm discrimination and the warning technique
  6. 4. Summary and future work
  7. Acknowledgements
  8. Appendix A. The SVM Classifier
  9. References

The authors are grateful to the anonymous referees for their constructive criticism and suggestions regarding an earlier version of this paper. The authors also thank the Jiangsu Meteorological Observatory for supplying weather radar data and weather observation data.This study is supported by Natural Science Foundation of Jiangsu Province (BK2009415), Research Fund for the Doctoral Program of Higher Education of China (20093228110002), Key University Science Research Project of Jiangsu Province (10KJA170030), National 863 Project (2007AA061901), Project of State Key Laboratory of Severe Weather of Chinese Academy of Meteorological Sciences (2010LASW-A01), Nanjing Weather Radar Open Laboratory Foundation (BJG201001) and projects CX09B_227Z, HKQX09-03 and 2009Y0006.

Appendix A. The SVM Classifier

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. The new storm identification method
  5. 3. Storm discrimination and the warning technique
  6. 4. Summary and future work
  7. Acknowledgements
  8. Appendix A. The SVM Classifier
  9. References

The Support Vector Machine (SVM) is a new popular machine learning technique (Vapnik, 1998). Generally, it is an algorithm that predicts a data classification given relevant features. In the SVM approach, the optimization criterion is the maximization of the margin that is the minimum distance from any training point to the separating hyper plane. The SVM produces a model that predicts the class label by setting parameter values of an optimization problem based on its input data. With the breakthrough progress achieved in both theoretical studies and practical implementation in recent years, SVMs are widely used in various fields (Shawe-Taylor and Cristianini, 2000).

For the present study of storm classification, the four characteristics are taken as the input index vector, i.e. x = [Z, Rmax, VIL, ET], and two-class categories are taken as the output index given that y is set to 1 for strong storms and − 1 for general storms. As illustrated in Figure 6, the larger the characteristic value is, the higher probability that the storm belongs to strong storms. Therefore, quasi-linear SVMs can be employed, which is referred to Deng and Tian (2004):

  • 1.
    The training set T = {(x1, y1), …, (xl, yl)} ∈ (X × Y)l is established, where l is the number of storms, xiX = [Zi, Rmaxi, VILi, ETi], yiY = {1, − 1} and i = 1, …, l;
  • 2.
    Penalization parameters C (C > 0) are chosen for quasi-linear separable problems. C can be considered as the reduction factor of convex hulls of the two point groups in 4D space according to the geometrical bisection method. The following optimization problem is designed and solved:
    • equation image
    s.t. equation image,
    • equation image(A.1)
    and then the optimization solution equation image is obtained;
  • 3.
    equation image is calculated. A positive component equation image less than C is selected from equation image, based on which equation image is calculated, and,
  • 4
    the hyper plane (wx)+ b* = 0 is constructed, and the decision function f(x) = sgn((wx)+ b*) is achieved.

References

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. The new storm identification method
  5. 3. Storm discrimination and the warning technique
  6. 4. Summary and future work
  7. Acknowledgements
  8. Appendix A. The SVM Classifier
  9. References
  • Baeck ML, Smith JA. 1998. Rainfall estimation by the WSR-88D for heavy rainfall events. Weather and Forecasting 13: 416436.
  • Browning KA, Collier CG, Larke PR, Menmuir P, Monk GA, Owens RG. 1982. On the forecasting of frontal rain using a weather radar network. Monthly Weather Review 110: 34552.
  • Crane RK. 1979. Automatic cell detection and tracking. IEEE Transactions on Geoscience Electronics 17: 250262.
  • Deng N, Tian Y. 2004. A New Method in Data Mining: Support Vector Machine. Science Press: Beijing, China.
  • Dixon M, Wiener G. 1993. TITAN: thunderstorm identification, tracking, analysis, and nowcasting—a radar-based methodology. Journal of Atmospheric and Oceanic Technology 10: 785797.
  • Greene DR, Clark RA. 1972. Vertically integrated liquid water—a new analysis tool. Monthly Weather Review 100: 548552.
  • Han L, Yu X, Zheng Y, Chen M, Wang H, Lin Y. 2009. Statistic characteristics of severe convective storm during Warm-Season in the Beijing-Tianjin region and its vicinity. Chinese Science Bulletin 54: 24932498.
  • Hand WH, Conway BJ. 1995. An object-oriented approach to nowcasting shower. Weather and Forecasting 10: 327341.
  • Henry SG. 1993. Analysis of thunderstorm lifetime as a function of size and intensity. Proceedings of 26th International Conference on Radar Meteorology, Norman, OK, 24-28 May 1993, American Meteorological Society: Boston, MA; pp 138140.
  • Hu S, Gu S, Zhuang X, Luo H. 2007. Automatic identification of Storm cells using Doppler radars. Acta Meteorological Sinica 21: 353365.
  • Johnson JT, MacKeen PL, Witt A, Mitchell ED, Stumpf GJ, Eilts MD, Thomas KW. 1998. The storm cell identification and tracking algorithm: an enhanced WSR-88d algorithm. Weather and Forecasting 13: 263276.
  • Li PW, Lai EST. 2004. Applications of radar-based nowcasting techniques for mesoscale weather forecasting in HongKong. Meteorological Applications 11: 253264.
  • Li L, Schmid W, Joss J. 1995. Nowcasting of motion and growth of precipitation with radar over a complex orography. Journal of Applied Meteorology 34: 12861300.
  • Liljas E. 1998. COST-78 co-operations on SAFs. Proceedings of SAF Training Workshop—Nowcasting and Very Short Range Forecasting, Madrid, Spain, 9–11 December 1998, European Organisation for the Exploitation of Meteorological Satellites: Darmstadt, Germany; pp 2333.
  • MacKeen PL, Brooks HE, Elmore KL. 1999. Radar reflectivity-derived thunderstorm parameters applied to storm longevity forecasting. Weather and Forecasting 14: 289295.
  • Mecikalski JR, Bedka KM. 2006. Forecasting convective initiation by monitoring the evolution of moving cumulus in daytime GOES imagery. Monthly Weather Review 134: 4978.
  • Rinehart RE, Garvey ET. 1978. Three-dimensional storm motion detected by convective weather radar. Nature 273: 287289.
  • Roberts RD, Rutledge S. 2003. Nowcasting storm initiation and growth using GOES-8 and WSR-88D data. Weather and Forecasting 18: 562584.
  • Rosenfeld D. 1987. Objective method for analysis and tracking of convective cells as seen by radar. Journal of Atmospheric and Oceanic Technology 4: 422434.
  • Shawe-Taylor J, Cristianini N. 2000. Support Vector Machines and other Kernel-based Learning Methods. Cambridge University Press: London, United Kingdom.
  • Steven AA, Wolf PL. 1997. VIL density as a hail indicator. Weather and Forecasting 12: 473478.
  • Vapnik V. 1998. Statistical Learning Theory. John Wiley & Sons: New York, NY.
  • Wilson JW, Crook NA, Mueller CK, Sun J, Dixon M. 1998. Nowcasting thunderstorms: a status report. Bulletin of the American Meteorological Society 79: 20792099.
  • Wilson JW, Megenhardt DL. 1997. Thunderstorm initiation, organization, and lifetime associated with Florida boundary layer convergence lines. Monthly Weather Review 125: 15071525.