A Technique to Censor Biological Echoes in Radar Reflectivity Data

Valliappa Lakshmanan Cooperative Institute for Mesoscale Meteorological Studies, University of Oklahoma, and National Oceanic and Atmospheric Administration/National Severe Storms Laboratory, Norman, Oklahoma

Search for other papers by Valliappa Lakshmanan in
Current site
Google Scholar
PubMed
Close
,
Jian Zhang National Oceanic and Atmospheric Administration/National Severe Storms Laboratory, Norman, Oklahoma

Search for other papers by Jian Zhang in
Current site
Google Scholar
PubMed
Close
, and
Kenneth Howard National Oceanic and Atmospheric Administration/National Severe Storms Laboratory, Norman, Oklahoma

Search for other papers by Kenneth Howard in
Current site
Google Scholar
PubMed
Close
Full access

Abstract

Existing techniques of quality control of radar reflectivity data rely on local texture and vertical profiles to discriminate between precipitating echoes and nonprecipitating echoes. Nonprecipitating echoes may be due to artifacts such as anomalous propagation, ground clutter, electronic interference, sun strobe, and biological contaminants (i.e., birds, bats, and insects). The local texture of reflectivity fields suffices to remove most artifacts, except for biological echoes. Biological echoes, also called “bloom” echoes because of their circular shape and expanding size during the nighttime, have proven difficult to remove, especially in peak migration seasons of various biological species, because they can have local and vertical characteristics that are similar to those of stratiform rain or snow. In this paper, a technique is described that identifies candidate bloom echoes based on the range variance of reflectivity in areas of bloom and uses the global, rather than local, characteristic of the echo to discriminate between bloom and rain. Every range gate is assigned a probability that it corresponds to bloom using morphological (shape based) operations, and a neural network is trained using this probability as one of the input features. It is demonstrated that this technique is capable of identifying and removing echoes due to biological targets and other types of artifacts while retaining echoes that correspond to precipitation.

Corresponding author address: V. Lakshmanan, 120 David L. Boren Blvd., Norman, OK 73072. Email: lakshman@ou.edu

Abstract

Existing techniques of quality control of radar reflectivity data rely on local texture and vertical profiles to discriminate between precipitating echoes and nonprecipitating echoes. Nonprecipitating echoes may be due to artifacts such as anomalous propagation, ground clutter, electronic interference, sun strobe, and biological contaminants (i.e., birds, bats, and insects). The local texture of reflectivity fields suffices to remove most artifacts, except for biological echoes. Biological echoes, also called “bloom” echoes because of their circular shape and expanding size during the nighttime, have proven difficult to remove, especially in peak migration seasons of various biological species, because they can have local and vertical characteristics that are similar to those of stratiform rain or snow. In this paper, a technique is described that identifies candidate bloom echoes based on the range variance of reflectivity in areas of bloom and uses the global, rather than local, characteristic of the echo to discriminate between bloom and rain. Every range gate is assigned a probability that it corresponds to bloom using morphological (shape based) operations, and a neural network is trained using this probability as one of the input features. It is demonstrated that this technique is capable of identifying and removing echoes due to biological targets and other types of artifacts while retaining echoes that correspond to precipitation.

Corresponding author address: V. Lakshmanan, 120 David L. Boren Blvd., Norman, OK 73072. Email: lakshman@ou.edu

1. Introduction

Weather radar data are used operationally to warn of impending severe weather (Kitzmiller et al. 1995) and to create high-resolution precipitation estimates (Fulton et al. 1998). Radar data are routinely assimilated into numerical weather models and used for the prediction of convective systems (Sun and Wilson 2003). Simmons and Sutter (2005) demonstrated that expected fatalities due to tornadoes after Doppler radar installation in the United States were 45% lower and expected injuries were 40% lower.

All of these uses of weather radar require that radar echoes correspond, broadly, to precipitation. By removing ground-clutter contamination, rainfall from the radar data using the National Weather Service Weather Surveillance Radar-1988 Doppler (WSR-88D) can be improved (Fulton et al. 1998; Krajewski and Vignal 2001). A large number of false positives for the mesocyclone detection algorithm (Stumpf et al. 1998) are caused in regions of clear-air return (McGrath et al. 2002; Mazur et al. 2004). A hierarchical motion-estimation technique segments and forecasts poorly in regions of ground clutter (Lakshmanan et al. 2003). Hence, a completely automated algorithm that can remove regions of nonprecipitating echo such as ground clutter, anomalous propagation, radar artifacts, and clear-air returns from the radar reflectivity field would be very useful in improving the performance of other automated weather radar algorithms.

Steiner and Smith (2002) describe the causes, effects, and characteristics of such contamination in weather radar data. Kessinger et al. (2003) and Lakshmanan et al. (2007a) determined individual features, and combinations of features, that can be used to remove range gates of radar reflectivity data that correspond to “bad” echoes. Local neighborhoods in the vicinity of every range gate in the three WSR-88D radar moments (reflectivity, velocity, and spectrum width) were examined and used for automated removal of nonprecipitating echoes. Steiner and Smith (2002) used a decision tree to classify range gates into two categories—precipitation and nonprecipitation, and Kessinger et al. (2003) used a fuzzy-rule base with features that included some that were introduced by Steiner and Smith (2002). Lakshmanan et al. (2007a) used a neural network to classify radar range gates into precipitation or nonprecipitation and followed the pixelwise classification with clustering. A cluster was censored if the majority of its pixels were determined to be nonprecipitating echo.

a. Biological echoes

The methods of Steiner and Smith (2002) and Kessinger et al. (2003) worked well for anamalous propagation (AP) because AP echoes are characterized by high reflectivities, high local variance (“texture”) in the reflectivity field, and low velocities. When followed with the clustering-based postprocessing of Lakshmanan et al. (2007a), the quality of the resulting fields met the high threshold necessary for fully automated quality control (QC) of radar data. However, biological contaminants cannot easily be removed by means of such local texture or vertical profile features. Although the technique of Kessinger et al. (2003) includes fuzzy rules to identify biological scatters, its performance is not robust enough to use as an automated mask.

Biological echoes are difficult to discriminate from true precipitation because they share several characteristics of precipitating echoes. Biological contaminants (such as birds, bats, or insects) are moving and therefore have nonzero Doppler velocities (Gauthreaux and Belser 1998)—the magnitude and texture of these scatters is very similar to that of widespread rain. Biological echoes, especially during peak migration seasons of several biological species, have radar reflectivity values in the horizontal dimensions that are similar to those of snow or rain. In the vertical dimension also, as illustrated in Fig. 1, it is difficult to distinguish between biological artifacts and rain from local characteristics alone. The three top panels and the middle-left panel of Fig. 1 show a case of biological echoes from the La Crosse, Wisconsin, radar (KARX) at 0404 UTC 25 May 2008, and the middle-center, middle-right, and two bottom panels of Fig. 1 show a case of winter precipitation from the Gaylord, Michigan, radar (KAPX) at 1923 UTC 17 January 2009. Note that when looking at just the local neighborhood of a pixel in the horizontal plane (for local texture features) or in the vertical plane (for the vertical profile) there is little to distinguish the two cases.1 The global view shown in the top-left and middle-center panels of Fig. 1 is required to discriminate between biological echoes and light snow.

To discriminate between biological echoes and light rain/snow, it is necessary to consider the characteristics of the entire echo and not just the vertical and horizontal neighborhood of a single pixel. Biological echoes tend to be circularly symmetric and centered around the radar (see Fig. 2, left two panels). The reflectivity intensity tends to reach maximum at a certain distance from the radar and then drops with range from the radar. This is probably because the migrating biological population peaks at a certain height. Another reason for the drop in power as the distance from the radar increases could be that the biological target fills less and less of the radar’s sampling volume. These are tendencies and are not universally valid—storm cells may be circular, may pass right over the radar, and may exhibit a very similar reflectivity profile. In addition, storm echoes may be embedded in an area of biological contamination, as shown in the center-right panel of Fig. 2. Light snow passing over the radar can have some of the characteristics of bloom, as shown in the rightmost panel of Fig. 2.

2. Method

Because biological echoes have a global profile that can be used to distinguish them from precipitating echoes whereas other artifacts need to be discriminated based on local characteristics, we followed the strategy of adding a feature to the local texture-based neural network that would be a probability that the pixel in question belongs to a biological echo.

To evaluate this probability, we computed several features and trained a neural network with one input. Then, a feature field was created from this probability by assigning to a pixel the bloom probability if it met certain morphological (value, shape, and contiguity) criteria. The block diagram of the technique is shown in Fig. 3.

a. Bloom radius

Extensive analyses of radar data as they relate to bird movements in the atmosphere along with independent bird observations have found that birds migrating at night frequently depart 30–45 min after local sunset (Gauthreaux and Belser 1998). As birds leave their diurnal stopover sites and climb to typical altitudes (see diagrams at http://virtual.clemson.edu/groups/birdrad/COM4A.HTM) of migration, they enter the radar beam and appear as rapidly expanding circular (or nearly circular) patterns in a base reflectivity image. So do insects except that the insects fly at a lower altitude and slower speed (Markkula 2008). The radius of the circle impacted by the birds depends on the maximum height at which birds could fly. This maximum range is called bloom radius in the current study.

The QC technique tries to identify and censor bloom echoes when the surface temperature at the radar site is at least 4°C.2 Simply clustering echoes based on contiguity will result in precipitation embedded within the bloom (such as in the center-right panel of Fig. 2) also being considered to be part of the bloom and being potentially censored. To identify echo over radar as being bloom, the following steps are carried out:

  1. Only range gates with an elevation of less than 4 km above ground level are considered, following studies carried out by Gauthreaux and Belser (1998) that indicated that this was where biological echoes are concentrated.

  2. The values of reflectivity factor Z at constant range in the “hybrid scan” (lowest unblocked reflectivity at every range/azimuth gate) are averaged.

  3. The values of averaged Z as it varies in range are fitted to line segments.

  4. The longest line segment (Pearson correlation coefficient of 0.9 or better) the slope of which is negative is considered to be the candidate bloom’s radius.

This process of computing average Z at a certain range and finding the longest negative-slope line is illustrated in Fig. 4, center panel. Following the procedure outlined above takes into account the expected drop-off in returned intensity by range of biological echoes while stopping the bloom detection when high reflectivities are encountered.

b. Bloom probability

If no “long enough” (at least 10 km in length) line segment was identified, then it is assumed that no bloom is present in the radar image. If a line segment of longer than 10 km is identified, then the bloom radius is set to the endpoint of the line segment and several statistics are computed on the radar echoes within the bloom radius: 1) mean reflectivity, 2) variance of reflectivity, 3) symmetry of the mean of octants of the bloom, 4) variance between the mean of the octants, 5) fraction of the bloom that is filled with echo, and 6) bloom radius. These features are used as inputs to a neural network that was trained to output the probability that the echo in question corresponds to bloom.

The training of the neural network was carried out on a dataset consisting of 34 examples of good data around the radar and 54 examples of biological artifacts. This dataset was divided 60/40 into a training dataset and a validation dataset. The good data points (which are scarce because we needed to find examples of storms with the appropriate reflectivity values directly over the radar) were repeated based on random selection so that the two classes had equal a priori probability in both the training and the validation datasets. The architecture of the neural network—one hidden layer consisting of a single node—was set arbitrarily, and the validation dataset was used to carry out early stopping based on cross entropy (Bishop 1995). To ensure that the output of the neural network is a probability, the transfer function at the output node was chosen to be a sigmoid and the error measure to be minimized was chosen to be the cross entropy.

c. Identifying bloom pixels

The output of the neural network is the probability that within the bloom radius biological echoes are present. While a candidate radius has been identified and a bloom probability has been calculated, not all pixels within that radius will correspond to bloom and not all echoes beyond this radius will be nonbiological. This is because the bloom radius was estimated from the variation in the average Z across all azimuths at a certain range. There could be storm echoes embedded inside the biological echo (see Fig. 5, top-left panel). It is also possible that the bloom may be nonsymmetric, extending beyond the bloom radius in one direction while the decrease of Z with range may have been stopped by the occurrence of a large enough storm in another, as in Fig. 5, bottom-left panel.

Because of nonsymmetry and embedded storm echoes, the extent of bloom echoes varies from radial to radial. The extent of bloom is assumed to be the nearest distance at which a storm echo is seen or when reflectivity values fall below a threshold. To ensure that these checks are tolerant of noise, a local 3-km neighborhood in the radial direction around every range gate is examined. To ensure that storm echoes at all tilts are taken into consideration, the reflectivity composite is used.

The algorithm assumes that a storm echo has been seen if all the gates in the 3-km neighborhood contain values above 35 dBZ. Then, all pixels connected to this 35-dBZ pixel that have values above 25 dBZ are also marked as corresponding to storm echoes. The algorithm assumes that the bloom echo has been seen completely when all of the gates in the 3-km neighborhood contain values below 10 dBZ. The extent of the bloom is the nearest distance at which either of these conditions—a storm echo or below threshold—happens. All pixels in this radial until that range are given the probability of bloom that was output from the neural network. All pixels beyond that range are assigned a bloom probability of zero.

Because the morphological (shape based) operations to extract embedded storm echoes are launched only if a 35-dBZ value is seen, weaker precipitation echoes embedded in bloom will either be identified all as bloom (leading to loss of precipitating echo) or identified all as precipitating, leading to precipitation estimates where there is no precipitation.

d. Second-stage neural network

The bloom probability result from the first neural network is assigned to every pixel in the image using morphological operations, thus creating a local feature field. This feature is provided as one of the inputs to a second, local-feature-based neural network. The second neural network had 21 of the inputs chosen through feature selection as described in Lakshmanan et al. (2007a) and a 22nd determined by following image morphological operations on the result of the first-stage neural network. Thus this second network had, as input features, 1) Doppler velocity, 2) mean of Doppler velocity, 3) standard deviation of Doppler velocity, 4) minimum standard deviation of Doppler velocity in neighborhood, 5) spectrum width, 6) reflectivity at lowest tilt, 7) neighborhood mean of reflectivity, 8) standard deviation of reflectivity, 9) minimum standard deviation of reflectivity in neighborhood, 10) spatial reflectivity of the reflectivity field, or “spin” (Steiner and Smith 2002), 11) inflections (Kessinger et al. 2003), 12) reflectivity at second tilt, 13) mean reflectivity at second tilt, 14) difference between reflectivity value and mean, 15) minimum standard deviation, 16) maximum value in the vertical, 17) vertically integrated liquid (Greene and Clark 1972), 18) difference between the two lowest tilts, 19) echo top of 0 dBZ, 20) echo top of 20 dBZ, 21) height of maximum, 22) fraction of neighborhood filled, and 23) probability that this pixel is part of a biological echo.

e. Clustering

The second-stage neural network was trained as in Lakshmanan et al. (2007a) and was followed by the same cluster-based postprocessing as in that paper. The only change is that the clustering is now on two attributes—the reflectivity maximum and the bloom probability—so that pixels have to be connected in both the reflectivity field and the bloom probability field to be considered to be a cluster. The output of the second-stage neural network is averaged within these clusters, and if the cluster average probability of being precipitating echo is less than 0.5 then the entire cluster is censored.

As was explained in Lakshmanan et al. (2007a), using a cluster of pixels in this manner greatly increases the expected accuracy of this neural network, since the neural network would have to be wrong on more than one-half of the pixels of a cluster to wrongly classify a cluster. So, even an average classifier will have extraordinary performance once its results are subject to a statistical averaging operation. The output of the cluster-based postprocessing formed the final mask used to censor the reflectivity field.

3. Results and conclusions

The training of the first-stage neural network (to perform the discrimination between biological echoes and storm echoes within a computed bloom radius) was carried out on a dataset of 88 cases split 60/40 into a training dataset and a validation dataset. Skill scores [critical success index (Donaldson et al. 1975), Heidke skill score (Heidke 1926), probability of detection, and rate of false alarm (Wilks 1995)] of the trained network on the validation set are shown in Fig. 6, top-left panel. In this context, the probability of detection refers to the probability of retaining precipitation and false alarm refers to nonprecipitating echoes that have not been censored while misses are precipitating echoes that have been wrongly censored.

The probabilities from the first neural network formed the 22nd input to the second-stage neural network that operated on a pixel-by-pixel basis. This neural network was trained with many more data—nearly 1.5 million training patterns without velocity and more than 5 million patterns with velocity data (each pattern corresponds to a nontrivial pixel in the radar data that needs to be classified; see Lakshmanan et al. 2007a for details). Skill scores of the trained second-stage network on the validation set are shown in Fig. 6, top-right and bottom-left panels.

a. Real-time performance

The technique described in this paper has been implemented and is being run in real time to censor biological echoes in radar data. Some examples of the technique’s performance on independent cases are shown in Fig. 7. The top row of Fig. 7 shows extensive bloom echoes over Texas at 0430 UTC 26 March 2009. These bloom echoes were all removed by the QC algorithm, even the deep, nonsymmetric bloom around Laughlin Air Force Base, Texas (KDFX). The middle row of Fig. 7 shows some bloom and some light rain over the southern plains at 0600 UTC 27 March 2009. All of the bloom echoes were removed by the QC algorithm while all of the precipitation, even the light precipitation such as over Lubbock, Texas (KLBB), was retained. The bottom row of Fig. 7 shows some snow echoes over New Mexico and Colorado and bloom echoes elsewhere at 0600 UTC 27 March 2009. All of the bloom echoes were censored by the algorithm [even the relatively deep bloom around Dallas/Fort Worth, Texas (KFWS)], and all of the snow echoes were retained.

b. Discussion

This paper has concentrated on biological targets centered on the radar because that is the category that the local texture or vertical profile–based methods of quality control such as those of Steiner and Smith 2002, Kessinger et al. 2003, and Lakshmanan et al. 2007a fail to handle properly. If the biological echoes are some distance away from the radar, they typically have weak reflectivities and do not affect higher-elevation scans. Therefore, such biological echoes can typically be removed very well by techniques that rely on local texture and vertical profiles. As shown in Fig. 8, such echoes may not be identified as bloom, but they are censored by the second-stage neural network.

The National Weather Service has been upgrading the WSR-88D network with dual-polarization capabilities. Dual-polarization radars provide more information about scatters than do single-polarization radars (Bringi and Chandrasekar 2001). Zrnic and Ryzhkov (1998) showed promising results in identifying biological targets such as birds and insects with a research dual-polarization radar in Oklahoma. However, the effectiveness of dual polarization to address blooms over different geographical regions will not be fully known until the dual-polarization upgrades are near completion for the conterminous United States. The technique discussed in this paper, in combination with techniques that employ the additional moments from dual-polarization radar (e.g., Zrnic et al. 2001), may be useful in accurately identifying biological scatters. The WSR-88D will be upgraded with dual polarization in the coming 3 yr (2009–12), but gap-filling radars, Terminal Doppler Weather Radars, and commercial radars, as well as Mexican and Canadian radar networks, do not have dual-polarization capability and can potentially benefit from techniques that are similar to those presented in the paper.

c. Summary

Based on the performance of the algorithm on the validation set, and its performance in real time, we can conclude that a technique that identifies candidate bloom echoes based on the range variance of reflectivity in areas of bloom and that uses the global, rather than local, characteristic of the echo is capable of discriminating between bloom and rain. It is possible to compute a probability that every range gate corresponds to bloom using a neural network trained on historical cases. This probability can be used as an additional feature to a traditional radar QC algorithm that is based on local texture. Such a two-stage intelligent machine algorithm is capable of identifying and removing echoes due to biological targets and other types of artifacts while retaining echoes that correspond to precipitation.

Acknowledgments

Funding for this research was provided under NOAA–OU Cooperative Agreement NA17RJ1227.

REFERENCES

  • Bishop, C., 1995: Neural Networks for Pattern Recognition. Oxford, 504 pp.

  • Bringi, V., and V. Chandrasekar, 2001: Polarimetric Doppler Weather Radar: Principles and Applications. Cambridge University Press, 636 pp.

    • Search Google Scholar
    • Export Citation
  • Donaldson, R., R. Dyer, and M. Kraus, 1975: An objective evaluator of techniques for predicting severe weather events. Preprints, Ninth Conf. on Severe Local Storms, Norman, OK, Amer. Meteor. Soc., 321–326.

    • Search Google Scholar
    • Export Citation
  • Fulton, R. A., J. P. Breidenbach, D-J. Seo, D. A. Miller, and T. O’Bannon, 1998: The WSR-88D rainfall algorithm. Wea. Forecasting, 13 , 377395.

    • Search Google Scholar
    • Export Citation
  • Gauthreaux Jr., S. A., and C. G. Belser, 1998: Displays of bird movements on the WSR-88D: Patterns and quantification. Wea. Forecasting, 13 , 453464.

    • Search Google Scholar
    • Export Citation
  • Greene, D. R., and R. A. Clark, 1972: Vertically integrated liquid water—A new analysis tool. Mon. Wea. Rev., 100 , 548552.

  • Heidke, P., 1926: Berechnung des Erfolges und der Güte der Windstärkvorhersagen im Sturmwarnungsdienst (Calculation of the success and of the quality of the wind strength forecasts in the Storm Warning Service). Geogr. Ann., 8 , 301349.

    • Search Google Scholar
    • Export Citation
  • Kessinger, C., S. Ellis, and J. Van Andel, 2003: The radar echo classifier: A fuzzy logic algorithm for the WSR-88D. Preprints, Third Conf. on Artificial Applications to the Environmental Sciences, Long Beach, CA, Amer. Meteor. Soc., P1.6. [Available online at http://ams.confex.com/ams/pdfpapers/54946.pdf].

    • Search Google Scholar
    • Export Citation
  • Kitzmiller, D. H., W. E. McGovern, and R. F. Saffle, 1995: The WSR-88D severe weather potential algorithm. Wea. Forecasting, 10 , 141159.

    • Search Google Scholar
    • Export Citation
  • Krajewski, W. F., and B. Vignal, 2001: Evaluation of anomalous propagation echo detection in WSR-88D data: A large sample case study. J. Atmos. Oceanic Technol., 18 , 807814.

    • Search Google Scholar
    • Export Citation
  • Lakshmanan, V., R. Rabin, and V. DeBrunner, 2003: Multiscale storm identification and forecast. J. Atmos. Res., 67 , 367380.

  • Lakshmanan, V., A. Fritz, T. Smith, K. Hondl, and G. Stumpf, 2007a: An automated technique to quality control radar reflectivity data. J. Appl. Meteor., 46 , 288305.

    • Search Google Scholar
    • Export Citation
  • Lakshmanan, V., T. Smith, G. Stumpf, and K. Hondl, 2007b: The warning decision support system—Integrated information. Wea. Forecasting, 22 , 596612.

    • Search Google Scholar
    • Export Citation
  • Markkula, I., 2008: Insect migration case study by polarimetric radar. Preprints, Fifth European Conf. on Radar in Meteorology and Hydrology, Helsinki, Finland, Finnish Meteorological Institute, 5.2.

    • Search Google Scholar
    • Export Citation
  • Mazur, R. J., V. Lakshmanan, and G. J. Stumpf, 2004: Quality control of radar data to improve mesocyclone detection. Preprints, 20th Int. Conf. on Interactive Information and Processing Systems (IIPS) for Meteorology, Oceanography, and Hydrology, Seattle, WA, Amer. Meteor. Soc., P1.2a. [Available online at http://ams.confex.com/ams/pdfpapers/71377.pdf].

    • Search Google Scholar
    • Export Citation
  • McGrath, K. M., T. A. Jones, and J. T. Snow, 2002: Increasing the usefulness of a mesocyclone climatology. Preprints, 21st Conf. on Severe Local Storms, San Antonio, TX, Amer. Meteor. Soc., 5.4. [Available online at http://ams.confex.com/ams/pdfpapers/44897.pdf].

    • Search Google Scholar
    • Export Citation
  • Simmons, K. M., and D. Sutter, 2005: WSR-88D radar, tornado warnings, and tornado casualties. Wea. Forecasting, 20 , 301310.

  • Steiner, M., and J. A. Smith, 2002: Use of three-dimensional reflectivity structure for automated detection and removal of nonprecipitating echoes in radar data. J. Atmos. Oceanic Technol., 19 , 673686.

    • Search Google Scholar
    • Export Citation
  • Stumpf, G. J., A. Witt, E. D. Mitchell, P. L. Spencer, J. T. Johnson, M. D. Eilts, K. W. Thomas, and D. W. Burgess, 1998: The National Severe Storms Laboratory mesocyclone detection algorithm for the WSR-88D. Wea. Forecasting, 13 , 304326.

    • Search Google Scholar
    • Export Citation
  • Sun, J., and J. W. Wilson, 2003: The assimilation of radar data for weather prediction. Radar and Atmospheric Science: A Collection of Essays in Honor of David Atlas, Meteor. Monogr., No. 52, Amer. Meteor. Soc., 175–198.

    • Search Google Scholar
    • Export Citation
  • Wilks, D. S., 1995: Statistical Methods in Atmospheric Sciences: An Introduction. Academic Press, 467 pp.

  • Zrnic, D., and A. Ryzhkov, 1998: Observations of insects and birds with a polarimetric radar. IEEE Trans. Geosci. Remote Sens., 36 , 661668.

    • Search Google Scholar
    • Export Citation
  • Zrnic, D., A. Ryzhkov, J. Straka, Y. Liu, and J. Vivekanandan, 2001: Testing a procedure for automatic classification of hydrometeor types. J. Atmos. Oceanic Technol., 18 , 892913.

    • Search Google Scholar
    • Export Citation

Fig. 1.
Fig. 1.

Biological echoes and light snow are indistinguishable based on just local texture (the top-center and top-right panels vs the middle-right and bottom-left panels) and vertical profile (the middle-left panel vs the bottom-center panel). The “global” view shown in the top-left and middle-center panels is required to discriminate between biological echoes and light snow.

Citation: Journal of Applied Meteorology and Climatology 49, 3; 10.1175/2009JAMC2255.1

Fig. 2.
Fig. 2.

(left), (center-left) Examples of biological echoes: note that the characteristics of an entire echo may help to identify regions in which biological echoes are more likely. (center-right) One difficulty is storms embedded in biological echo. (right) Another difficulty is light snow moving over the radar. From left to right, the four panels represent data from 0431 UTC 4 Aug 2008 at Memphis, TN (KNQA), 0318 UTC 5 Aug 2008 at Omaha, NE (KOAX), 0607 UTC 5 Aug 2008 at Des Moines, IA (KDMX), and 0558 UTC 26 Mar 2009 at Glasgow, MT (KGGW).

Citation: Journal of Applied Meteorology and Climatology 49, 3; 10.1175/2009JAMC2255.1

Fig. 3.
Fig. 3.

Block diagram illustrating the stages of the technique described in this paper. (a) Finding the bloom radius is described in section 2a, (b) the bloom neural network is described in section 2b, (c) the morphological operations to find the bloom pixels are described in section 2c, and (d) the second-stage neural network is described in section 2d.

Citation: Journal of Applied Meteorology and Climatology 49, 3; 10.1175/2009JAMC2255.1

Fig. 4.
Fig. 4.

(left) Reflectivity from KARX at 0407 UTC 25 May 2008. (center) Finding the extent of bloom using average Z at constant range. (right) Reflectivity after bloom echoes have been censored.

Citation: Journal of Applied Meteorology and Climatology 49, 3; 10.1175/2009JAMC2255.1

Fig. 5.
Fig. 5.

Why morphological operations are required to clean up the reflectivity field once bloom has been identified: (top-left) Instance of storm echoes in one direction and bloom in the other [data are from Kansas City, MO (KEAX), at 0414 UTC 26 May 2008]. (top-right) The biological echoes have been censored by the technique of this paper while the storm echoes are retained. (bottom-left) Instance of highly nonsymmetric bloom with storms embedded in bloom [data are from Brownsville, TX (KBRO) at 0414 UTC 20 May 2008]. (bottom-right) Most of the biological echoes have been removed, and all the storm echoes have been retained.

Citation: Journal of Applied Meteorology and Climatology 49, 3; 10.1175/2009JAMC2255.1

Fig. 6.
Fig. 6.

Measures of skill of the three neural networks that compose the complete automated algorithm on their respective validation sets as the threshold on the output of the neural network is varied. The censoring is done cluster by cluster, with a cluster being censored if the mean second-stage neural-network output on the cluster’s pixels is less than 0.5.

Citation: Journal of Applied Meteorology and Climatology 49, 3; 10.1175/2009JAMC2255.1

Fig. 7.
Fig. 7.

Performance of algorithm in real time on 26 and 27 Mar 2009. (left) The circled areas of bloom in the reflectivity composite field have (right) been correctly removed by the algorithm described in this paper while areas of snow (marked by a rectangle in bottom-left panel) have been correctly retained.

Citation: Journal of Applied Meteorology and Climatology 49, 3; 10.1175/2009JAMC2255.1

Fig. 8.
Fig. 8.

Biological echoes that are not centered around the radar do not need special handling. (left) Reflectivity composite from Mobile, AL (KMOB), at 1139 UTC 20 Jul 2005 shows birds leaving from an overnight roost. The circular patterns here are not centered on the radar. (right) Quality-controlled field: even though they were not identified as bloom, the biological echoes have been removed because they have weak reflectivities and do not show vertical continuity.

Citation: Journal of Applied Meteorology and Climatology 49, 3; 10.1175/2009JAMC2255.1

1

The “table readout” and cross sections in these images were created with the help of the display software described in Lakshmanan et al. (2007b).

2

We found, after analysis of radar data in 2005–08, that this was the lowest surface temperature at which bloom could be found. Radar echoes around KAPX at 0400 UTC 17 Apr 2009 illustrate the existence of biological echoes at temperatures as low as 4°C.

Save
  • Bishop, C., 1995: Neural Networks for Pattern Recognition. Oxford, 504 pp.

  • Bringi, V., and V. Chandrasekar, 2001: Polarimetric Doppler Weather Radar: Principles and Applications. Cambridge University Press, 636 pp.

    • Search Google Scholar
    • Export Citation
  • Donaldson, R., R. Dyer, and M. Kraus, 1975: An objective evaluator of techniques for predicting severe weather events. Preprints, Ninth Conf. on Severe Local Storms, Norman, OK, Amer. Meteor. Soc., 321–326.

    • Search Google Scholar
    • Export Citation
  • Fulton, R. A., J. P. Breidenbach, D-J. Seo, D. A. Miller, and T. O’Bannon, 1998: The WSR-88D rainfall algorithm. Wea. Forecasting, 13 , 377395.

    • Search Google Scholar
    • Export Citation
  • Gauthreaux Jr., S. A., and C. G. Belser, 1998: Displays of bird movements on the WSR-88D: Patterns and quantification. Wea. Forecasting, 13 , 453464.

    • Search Google Scholar
    • Export Citation
  • Greene, D. R., and R. A. Clark, 1972: Vertically integrated liquid water—A new analysis tool. Mon. Wea. Rev., 100 , 548552.

  • Heidke, P., 1926: Berechnung des Erfolges und der Güte der Windstärkvorhersagen im Sturmwarnungsdienst (Calculation of the success and of the quality of the wind strength forecasts in the Storm Warning Service). Geogr. Ann., 8 , 301349.

    • Search Google Scholar
    • Export Citation
  • Kessinger, C., S. Ellis, and J. Van Andel, 2003: The radar echo classifier: A fuzzy logic algorithm for the WSR-88D. Preprints, Third Conf. on Artificial Applications to the Environmental Sciences, Long Beach, CA, Amer. Meteor. Soc., P1.6. [Available online at http://ams.confex.com/ams/pdfpapers/54946.pdf].

    • Search Google Scholar
    • Export Citation
  • Kitzmiller, D. H., W. E. McGovern, and R. F. Saffle, 1995: The WSR-88D severe weather potential algorithm. Wea. Forecasting, 10 , 141159.

    • Search Google Scholar
    • Export Citation
  • Krajewski, W. F., and B. Vignal, 2001: Evaluation of anomalous propagation echo detection in WSR-88D data: A large sample case study. J. Atmos. Oceanic Technol., 18 , 807814.

    • Search Google Scholar
    • Export Citation
  • Lakshmanan, V., R. Rabin, and V. DeBrunner, 2003: Multiscale storm identification and forecast. J. Atmos. Res., 67 , 367380.

  • Lakshmanan, V., A. Fritz, T. Smith, K. Hondl, and G. Stumpf, 2007a: An automated technique to quality control radar reflectivity data. J. Appl. Meteor., 46 , 288305.

    • Search Google Scholar
    • Export Citation
  • Lakshmanan, V., T. Smith, G. Stumpf, and K. Hondl, 2007b: The warning decision support system—Integrated information. Wea. Forecasting, 22 , 596612.

    • Search Google Scholar
    • Export Citation
  • Markkula, I., 2008: Insect migration case study by polarimetric radar. Preprints, Fifth European Conf. on Radar in Meteorology and Hydrology, Helsinki, Finland, Finnish Meteorological Institute, 5.2.

    • Search Google Scholar
    • Export Citation
  • Mazur, R. J., V. Lakshmanan, and G. J. Stumpf, 2004: Quality control of radar data to improve mesocyclone detection. Preprints, 20th Int. Conf. on Interactive Information and Processing Systems (IIPS) for Meteorology, Oceanography, and Hydrology, Seattle, WA, Amer. Meteor. Soc., P1.2a. [Available online at http://ams.confex.com/ams/pdfpapers/71377.pdf].

    • Search Google Scholar
    • Export Citation
  • McGrath, K. M., T. A. Jones, and J. T. Snow, 2002: Increasing the usefulness of a mesocyclone climatology. Preprints, 21st Conf. on Severe Local Storms, San Antonio, TX, Amer. Meteor. Soc., 5.4. [Available online at http://ams.confex.com/ams/pdfpapers/44897.pdf].

    • Search Google Scholar
    • Export Citation
  • Simmons, K. M., and D. Sutter, 2005: WSR-88D radar, tornado warnings, and tornado casualties. Wea. Forecasting, 20 , 301310.

  • Steiner, M., and J. A. Smith, 2002: Use of three-dimensional reflectivity structure for automated detection and removal of nonprecipitating echoes in radar data. J. Atmos. Oceanic Technol., 19 , 673686.

    • Search Google Scholar
    • Export Citation
  • Stumpf, G. J., A. Witt, E. D. Mitchell, P. L. Spencer, J. T. Johnson, M. D. Eilts, K. W. Thomas, and D. W. Burgess, 1998: The National Severe Storms Laboratory mesocyclone detection algorithm for the WSR-88D. Wea. Forecasting, 13 , 304326.

    • Search Google Scholar
    • Export Citation
  • Sun, J., and J. W. Wilson, 2003: The assimilation of radar data for weather prediction. Radar and Atmospheric Science: A Collection of Essays in Honor of David Atlas, Meteor. Monogr., No. 52, Amer. Meteor. Soc., 175–198.

    • Search Google Scholar
    • Export Citation
  • Wilks, D. S., 1995: Statistical Methods in Atmospheric Sciences: An Introduction. Academic Press, 467 pp.

  • Zrnic, D., and A. Ryzhkov, 1998: Observations of insects and birds with a polarimetric radar. IEEE Trans. Geosci. Remote Sens., 36 , 661668.

    • Search Google Scholar
    • Export Citation
  • Zrnic, D., A. Ryzhkov, J. Straka, Y. Liu, and J. Vivekanandan, 2001: Testing a procedure for automatic classification of hydrometeor types. J. Atmos. Oceanic Technol., 18 , 892913.

    • Search Google Scholar
    • Export Citation
  • Fig. 1.

    Biological echoes and light snow are indistinguishable based on just local texture (the top-center and top-right panels vs the middle-right and bottom-left panels) and vertical profile (the middle-left panel vs the bottom-center panel). The “global” view shown in the top-left and middle-center panels is required to discriminate between biological echoes and light snow.

  • Fig. 2.

    (left), (center-left) Examples of biological echoes: note that the characteristics of an entire echo may help to identify regions in which biological echoes are more likely. (center-right) One difficulty is storms embedded in biological echo. (right) Another difficulty is light snow moving over the radar. From left to right, the four panels represent data from 0431 UTC 4 Aug 2008 at Memphis, TN (KNQA), 0318 UTC 5 Aug 2008 at Omaha, NE (KOAX), 0607 UTC 5 Aug 2008 at Des Moines, IA (KDMX), and 0558 UTC 26 Mar 2009 at Glasgow, MT (KGGW).

  • Fig. 3.

    Block diagram illustrating the stages of the technique described in this paper. (a) Finding the bloom radius is described in section 2a, (b) the bloom neural network is described in section 2b, (c) the morphological operations to find the bloom pixels are described in section 2c, and (d) the second-stage neural network is described in section 2d.

  • Fig. 4.

    (left) Reflectivity from KARX at 0407 UTC 25 May 2008. (center) Finding the extent of bloom using average Z at constant range. (right) Reflectivity after bloom echoes have been censored.

  • Fig. 5.

    Why morphological operations are required to clean up the reflectivity field once bloom has been identified: (top-left) Instance of storm echoes in one direction and bloom in the other [data are from Kansas City, MO (KEAX), at 0414 UTC 26 May 2008]. (top-right) The biological echoes have been censored by the technique of this paper while the storm echoes are retained. (bottom-left) Instance of highly nonsymmetric bloom with storms embedded in bloom [data are from Brownsville, TX (KBRO) at 0414 UTC 20 May 2008]. (bottom-right) Most of the biological echoes have been removed, and all the storm echoes have been retained.

  • Fig. 6.

    Measures of skill of the three neural networks that compose the complete automated algorithm on their respective validation sets as the threshold on the output of the neural network is varied. The censoring is done cluster by cluster, with a cluster being censored if the mean second-stage neural-network output on the cluster’s pixels is less than 0.5.

  • Fig. 7.

    Performance of algorithm in real time on 26 and 27 Mar 2009. (left) The circled areas of bloom in the reflectivity composite field have (right) been correctly removed by the algorithm described in this paper while areas of snow (marked by a rectangle in bottom-left panel) have been correctly retained.

  • Fig. 8.

    Biological echoes that are not centered around the radar do not need special handling. (left) Reflectivity composite from Mobile, AL (KMOB), at 1139 UTC 20 Jul 2005 shows birds leaving from an overnight roost. The circular patterns here are not centered on the radar. (right) Quality-controlled field: even though they were not identified as bloom, the biological echoes have been removed because they have weak reflectivities and do not show vertical continuity.

All Time Past Year Past 30 Days
Abstract Views 0 0 0
Full Text Views 365 69 8
PDF Downloads 333 83 8