This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING
1
A Quality Control Method of Ground-Based Weather Radar Data Based on Statistics Nan Li , Zhenhui Wang , Kangyuan Sun, Zhigang Chu, Liang Leng, and Xingchao Lv
Abstract— The detection of meteorological targets using ground-based weather radars usually suffers from ground clutter and beam blockage. These nonmeteorological or weakened signals should be identified so quality control should be implemented before weather radar data can be used. Conventional quality control methods aim at differentiating between echo structures of ground clutter and meteorological targets, and use terrain information to calculate beam blockage regions based upon standard atmospheric refraction. However, it is difficult to achieve the goal for long-term large data sets by conventional methods due to the complexity and diversity of weather radar echoes. In this paper, regions of ground clutter and beam blockage are first identified through the statistics on spatial distribution of reflectivity and fuzzy logic classification, and then they are used as masks to remove data from the scan. The new method is applied to data of the Nanjing weather radar in China. By the aid of a proposed evaluation scheme and the visual recognition, quality control results of the new method are compared with those of the conventional methods. It is found that the new method can provide better identification of ground clutter or beam blockage and thus better quality control results. The new scheme has a good prospect in operational service for its principle advantages, easy applicable conditions, and better performance compared with conventional methods. Index Terms— Beam blockage, fuzzy logic classification, ground clutter, statistics, weather radar.
I. I NTRODUCTION
W
EATHER radar is a kind of active remote sensing instrument that uses electromagnetic waves to determine the location and the intensity of precipitation. Modern weather radars can provide high temporal and spatial resolution data, and they are very helpful to meteorological service. Manuscript received April 5, 2017; revised August 6, 2017, October 11, 2017, and November 15, 2017; accepted November 18, 2017. The work was supported in part by the China Commonweal Industry Research Project under Grant GYHY201306078 and in part by the National Natural Science Foundation of China under Grant 41305031. (Corresponding author: Nan Li.) N. Li is with the CMA Key Laboratory for Aerosol-Cloud-Precipitation, School of Atmospheric Physics, Nanjing University of Information Science and Technology, Nanjing 210044, China, and also with the Laboratory of Straits Meteorology, Xiamen 361012, China, and also with the Hubei Key Laboratory for Heavy Rain Monitoring and Warning Research, Institute of Heavy Rain, CMA, Wuhan 430205, China (e-mail:
[email protected]). Z. Wang, Z. Chu, and X. Lv are with the CMA Key Laboratory for AerosolCloud-Precipitation, School of Atmospheric Physics, Nanjing University of Information Science and Technology, Nanjing 210044, China. K. Sun is with the Jiangsu Institute of Meteorological Sciences, Nanjing 210009, China. L. Leng is with the Hubei Key Laboratory for Heavy Rain Monitoring and Warning Research, Institute of Heavy Rain, CMA, Wuhan 430205, China. Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/TGRS.2017.2776562
However, the hydrometeor detection capability of a groundbased weather radar is usually constrained by the surrounding terrain especially at low elevation angles, which seriously affects the application of radar data both in quantity and quality. On the one hand, not all of echoes detected by a weather radar are produced by meteorological objects, and many of them result from ground clutter. A number of quality control methods on identifying ground clutter have been proposed based on differences in echo features between ground clutter and meteorological targets. Some progress has been achieved in recent years. Steiner and Smith [1] used a decision tree to remove nonprecipitation echoes, which made use of the 3-D reflectivity structure including vertical extent, spatial variability, and vertical gradient. The proposed technique worked well for ground clutter of anomalous propagation either separated from or embedded within precipitation. In their algorithm, clear-air echoes and gust fronts were removed as nonprecipitation echoes. Clear-air echoes can be retained in the method described in this paper. Berenguer et al. [2] adopted a fuzzy logic algorithm that used echo features of ground clutter such as shallow vertical extent, high spatial variability, and low radial velocities to calculate the possibility of ground clutter, and removed those echoes that exceeded a certain threshold. The proposed fuzzy logic algorithm gave better results than the clutter mask technique with the mean clutter map. Lakshmanan et al. [3] adopted a neural network technique to discriminate between precipitating and nonprecipitating areas on a region-by-region basis and it could be applied in real time. The algorithm contained a number of horizontal and vertical echo features that were described in previous works. Compared with the radar echo classifier implemented in the Weather Surveillance Radar-1988 Doppler (WSR-88D), the proposed method showed a better performance on the test cases. Sadouki and Haddad [4] proposed a combination of textural approach and fuzzy approach to classify echo types and identify ground clutter, and applied the technique to two radars in Algeria and France. With two of the textural parameters as input for the fuzzy logic system, the method seemed to have satisfying results by choosing proper parameters for different situations. Doppler analysis provides another means for discriminating between clutter and precipitation signals in radars that are coherent. Clutter signals would notch a band of frequencies at either side of zero Doppler frequency because they usually have close-to-zero mean Doppler velocity with a narrow spectrum width compared with precipitation signals. Torres and Zrnic [5] proposed a ground clutter elimination
0196-2892 © 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. 2
method by using regression techniques which could be easily designed with a set of orthogonal polynomials and could be implemented to nonuniformly sampled signals. Comparisons of the regression filters and the fifth-order elliptic filters implemented in WSR-88D on actual signals indicated that the regression filters gave better performance. Hubbert et al. [6], [7] introduced the clutter phase alignment (CPA) as the discriminator of ground clutter, and used a fuzzy logic algorithm to distinguish between clutter echoes and precipitation echoes. The CPA technique aimed at the signal phase rather than the amplitude. The phase of the received signal should be constant for stationary targets, and thus, the sum of received time series was high for clutter targets and low for moving targets. CPA was affected by both mean velocity and spectrum width, but was a more robust indicator of ground clutter since it took the primary characteristic of stationary targets. The comparison with quality controlled data of WSR-88D at Denver indicated that the improvement of the proposed technique was obvious, especially along the zero-velocity isodop. On the other hand, the surrounding terrain could cause radar beam blockage, and thus, make echoes weaker than the actual situation and sometimes even undetectable. Therefore, the identification of radar beam blockage is also important besides the identification of ground clutter. Previous methods used the digital elevation model (DEM) to simulate the shielding effect of topography on the radar beam. Delrieu et al. [8] used a digitized terrain model to calculate the partial and total beam blockage for a ground-based X-band weather radar in France. They adopted Gaussian power gain pattern for the angular and range weighting functions of the radar measurement to give weighted illuminated areas for various sizes of the radar resolution volume. The results showed that 87% of the measured mountain return variance could be explained by the geometrical estimates when the resolution volume with the 15-dB beamwidth was considered. Bech et al. [9] evaluated the performance of beam blockage corrections based on a simple interception function between the radar beam and the terrain with three years of radiosonde data. They chose three different targets surrounding the radar to evaluate the simulated radar beam shielding under different propagation conditions. Results showed that the beam blockage correction was generally robust for standard propagation conditions, but it would yield inaccurate results by the departure from the standard propagation. Accordingly, they suggested that the monitoring of propagation conditions as a requirement. Kucera et al. [10] and Krajewski et al. [11] explored the use of geographic information system (GIS) in computing the beam blockage rate for the WSR-88D. They found that GIS information provided reasonable beam blockage areas and also the physical interpretation of the beam blockage effect on rainfall estimation. It was noted that the DEM resolution played an important role in resolving blocked patterns, and higher DEM resolution provided better results. Bech et al. [12] studied the variability of beam propagation conditions and assessed their effects on beam blockage corrections for rainfall estimation of the Nordic Weather Radar Network. They used a beam propagation model to simulate
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING
the interaction between the radar beam and the terrain and derived the correction factors under the assumption of standard propagation conditions. The correction improvement depended on the distance from the radar and on the blockage degree related to the vertical profile of refraction. Shakti et al. [13] proposed a modified DEM method to evaluate the beam blockage of an X-band radar in Hakone mountainous areas of Kanto, Japan. In their method, they attempted to include factors such as power loss due to ground clutter filtering and the calibration error. It was found that the simple DEM method alone was insufficient to correct beam blockage, but the modified DEM method gave results in good agreement with the rain gauge measurements. In addition, they found that a higher resolution of the DEM made a limited contribution to improve the correction. Although studies in recent years improved the identification of ground clutter and beam blockage, the traditional methods leave difficult problems to be solved. The frequently used methods of ground clutter identification focus on the echo features such as large reflectivity gradient and small radial velocity. However, some precipitation echoes also have large reflectivity gradient such as convective rainfall or have small radial velocity when azimuth angles of the radar are nearly perpendicular to the wind orientation. These situations would cause some precipitation to be wrongly identified as ground clutter. Meanwhile, some ground clutter would be missed when mixed with precipitation. Additionally, hardware performance of the specific radar would affect the identification results of the methods dealing with signal processing. The frequently used methods of beam blockage identification use the DEM and the assumption of standard atmospheric refraction to calculate the interaction between the radar beam and the terrain. However, standard refraction would not be satisfied under local weather and climate conditions, and the actual spatiotemporal distribution of the refractive index is difficult to be acquired in real time with limited radiosonde data. In addition, the accuracy of topography information could be constrained by vegetation growth and building development. Moreover, some factors related to radar hardware issues would also affect the calculation accuracy based on the DEM, including power gain pattern, power loss due to ground clutter filtering, and calibration errors. A few scholars mentioned statistics in the analysis of conventional quality control methods. Kucera et al. [10] and Krajewski et al. [11] calculated the ratio of the number of reflectivity samples exceeding a certain threshold to the total number of reflectivity samples. They merely used this ratio as a reference to qualitatively analyze the partial radar beam occultation and to assess the results given by the DEM. However, in their studies, the statistical results under the threshold of 10 dBZ may be affected by the clear-air echo. Similarly, Gou [14] proposed an idea of analyzing partial beam blockage based on probabilistic characteristics of echoes. In fact, statistical techniques have the potential to identify ground clutter and beam blockage through in-depth investigation. It may be difficult to completely distinguish different categories of echoes at one time interval, but for the long-term measurement of a specific radar, statistics can
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. LI et al.: QUALITY CONTROL METHOD OF GROUND-BASED WEATHER RADAR DATA
reflect significant differences between categories of echoes within the period. Based on this idea, a new ground clutter and beam blockage identification and quality control scheme is proposed in this paper. The proposed scheme and a specific weather radar are introduced in Section II. The quality control results of the new and the conventional methods are analyzed in Section III. A summary and prospect is given in Section IV. II. P ROPOSED M ETHOD BASED ON S TATISTICS A. Basic Principle The spatial distribution of ground clutter is almost fixed, and the echo intensity of ground clutter is stronger than that of precipitation, clear air or beam blockage. Although some ground clutter may have similar echo intensity to precipitation, the probability of large reflectivity appearing in regions of ground clutter would be significantly higher than in other regions because ground clutter appears no matter whether rainfall occurs or not. Therefore, this difference of probability can be used in the identification of ground clutter based on statistics. Given a large reflectivity threshold, regions on which reflectivity samples have a significantly larger proportion above the threshold than other regions would be the places where ground clutter appears. The appearance of beam blockage (partial or total beam blockage) is accompanied by echoes of precipitation or clear air, and its echo intensity is significantly weaker than the echoes of precipitation or clear air along the radial direction on both sides. The location of beam blockage is fixed along radial directions, and the coverage is related to the atmospheric refraction, but changes little. The probability of small reflectivity (including no echo) appearing in regions of beam blockage would be significantly higher than in other regions. Therefore, this difference of probability can be used in the identification of beam blockage based on statistics. It is worth noting that another case with a high probability of small reflectivity is the region where there are few echoes within the statistical period. This situation can be avoided by narrowing the analysis range or extending the statistical period so that precipitation or clear-air echoes can be uniformly distributed within the range. Given a small reflectivity threshold and a chosen analysis range that at least covers the farthest occlusion that causes beam blockage and has uniformly distributed echoes within the statistical period, regions on which reflectivity samples have a significant larger proportion below the threshold than other regions would be the places where beam blockage appears. After the azimuth angles of beam blockage regions are determined, the identified beam blockage regions within a small range can be extended for the entire radar detection range. For the volume scan mode of a specific ground-based weather radar, the farthest distance that ground objects appear at low elevations can be found since the radar beam height increases with the elevation angle and the radial range. Both ground clutter and beam blockage at an elevation are expected to be identified simultaneously, given an analysis range which covers the farthest ground object intercepting the radar beam
3
and a statistical period when echoes can be uniformly distributed within the analysis range. Consequently, quality control can be implemented using the identified regions of ground clutter and beam blockage as masks to remove data from the scan. B. Identification Algorithm of Ground Clutter and Beam Blockage Plane position indicator (PPI) reflectivity of an elevation at which ground clutter and beam blockage appear is used to accomplish statistical analysis and identification. All of the PPI reflectivity (including no echo) is resampled on a fixed grid polar coordinate. Afterward, a reflectivity threshold Z t is set, and whether the reflectivity Z of the i th sample exceeds the given threshold Z t on a grid is judged, that is, 1 Zi > Zt Pθ,ϕ,r,i = (1) 0 Zi ≤ Zt where θ , ϕ, and r represent azimuth, elevation, and range coordinates, respectively. The ratio of the number of times that Z exceeds Z t on the grid is then calculated N Pθ,ϕ,r,i (2) Pθ,ϕ,r = i=1 N where N is the number of times that all valid reflectivity samples appear on the grid. This calculation is implemented for all grids, and the spatial distribution of Pθ,ϕ,r within the analysis range can be obtained, based on which categories of grids can be discriminated as described below. When a smaller Z t is set, Pθ,ϕ,r is smaller on the grids where beam blockage appears compared with Pθ,ϕ,r on the grids where clear-air echo, precipitation, or ground clutter appears, which is larger. When a larger Z t is set, Pθ,ϕ,r is larger on the grids where ground clutter appears compared with Pθ,ϕ,r on the grids where beam blockage, precipitation, or clear-air echo appears, which is smaller. The fuzzy c-means (FCM) algorithm [15], [16] is adopted as the cluster analysis technique to classify grids into two categories to identify regions of ground clutter or beam blockage with a larger or a smaller Pθ,ϕ,r for a given reflectivity threshold. The objective function and the constraints of the FCM clustering algorithm are J (u, v) =
c n
2 um i j di j
i=1 j =1
⎧ c ⎨ u = 1, 1 ≤ j ≤ n, ij and i=1 ⎩ u i j ∈ [0, 1], 1 ≤ j ≤ n, 1 ≤ i ≤ c.
(3)
where c is the number of the clusters (c = 2), n is the number of the grids, u i j is the degree of membership of the j th grid P j to the i th clustering center v i , di j = ||P j − v i || is the distance between P j of the j th grid and the i th clustering center, and m (m > 1) is the fuzzy partition matrix exponent for controlling the degree of fuzzy overlap between clusters, with larger values indicating a greater degree of overlap. In this paper, a suggested value of 2 is adopted for m according to
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. 4
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING
the previous references. The clustering algorithm performs the following steps: 1) randomly initialize the cluster membership values u i j ; 2) calculate the cluster centers n m j =1 u i j P j v i = n m j =1 u i j
TABLE I S OME BASIC T ECHNICAL PARAMETERS OF THE CINRAD AT N ANJING IN C HINA
3) update u i j according to ui j =
c k=1
1 P j −v i P j −v k
2 m−1
4) calculate the objective function J (u, v); 5) repeat steps 2–4 until J (u, v) improves by less than a specified minimum threshold (0.00001) or until after a specified maximum number of iterations (100). Consequently, the minimum value of the objective function is yielded and the corresponding optimal classification results are obtained. The FCM algorithm is implemented twice for the two different reflectivity thresholds. Grids of ground clutter with a larger Pθ,ϕ,r would be identified from others for the larger threshold and grids of beam blockage with a smaller Pθ,ϕ,r would be identified from others for the smaller threshold. The appropriate larger and the appropriate smaller reflectivity thresholds can be selected according to the difference between the two clustering centers under different reflectivity thresholds. Based on the basic principle and the identification algorithm, there should be independence of echo features between categories, and thus, the optimal clustering should achieve the maximum difference between the centers of the two clusters. C. Identification Results for a Specific Weather Radar In this paper, the reflectivity data of the China New Generation Radar (CINRAD) in Nanjing of Jiangsu province is used to implement the identification of ground clutter and beam blockage with the new scheme presented above. The CINRAD is a single-polarization Doppler weather radar and is used for the operational service in China. The Nanjing radar is an S-band CINRAD and has technical parameters which are almost the same as WSR-88D. Some important technical parameters are listed in Table I. The measuring range of the reflectivity for the Nanjing radar is from −10 to 70 dBZ, and the reflectivity resolution is 0.5 dBZ. The Nanjing radar is located in East China (118.70°E, latitude 32.19°N, and 138.2 m above the sea level). It is about 200–300 km away from the eastern coast where the weather is dominated by the East Asian monsoon. Here, it is dry and rainless from November to April and wet and rainy from May to October, i.e., the flooding season. Enough samples can be obtained for both the clear-air echoes and precipitation echoes within the flooding season period. In addition, the reflectivity data were expected to have acceptable accuracy and consistency during this period of time because routine annual calibration was done through the internal hardware
before the flooding season. Reflectivity data at 0.5° elevation of the Nanjing radar from May to October, 2012 were collected for testing the algorithm. This data set provides 33 864 volume scans. Affected by the surrounding terrain, the Nanjing radar detects both ground clutter and beam blockage at the lowest elevation of 0.5°. The ground clutter appears within 80 km from the radar and the beam blockage appears at around azimuth 135° and 225° (the azimuth is counted clockwise from the north). To avoid echoes from the ocean, the analysis range is set to 200 km within which all echoes are from echoes over the land. Considering that the radar beamwidth is about 1° and the reflectivity gate length is 1 km, the grid of the polar coordinate for resampling is set to 1° × 1 km. All of the PPI reflectivity at 0.5° elevation is resampled on the uniform polar coordinate grids. The number of reflectivity samples on each grid is counted for every 0.5-dBZ reflectivity interval from −10 to 70 dBZ as well as for those of no echo, smaller than −10 dBZ and larger than 70 dBZ. Different reflectivity thresholds Z t are set to calculate the ratio of the number of times that Z exceeds Z t on grids, as shown in Fig. 1. It can be seen from Fig. 1 that spatial distributions of Pθ,ϕ,r are centrosymmetric; therefore, the spatial distribution of echoes is uniform within the analysis range. When larger thresholds Z t are used [Fig. 1(f)–(h)], ground clutter near the radar site is considerable due to their larger proportions above the thresholds compared with the other categories of echoes. For a larger threshold Z t under which the ground clutter has the maximum difference from the others, the fuzzy clustering technique is used to divide the grids into two categories, and the category with a larger Pθ,ϕ,r corresponds to the regions of ground clutter. When smaller thresholds Z t are used [Fig. 1(a)–(e)], beam blockage can be seen around azimuth 135° and 225°, and the smaller the Z t is, the more the obvious of beam blockage. The radar beam is shaded in these areas and thus is not able to detect echoes or only partially detect echoes, leading to a very small Pθ,ϕ,r . It is noted that Pθ,ϕ,r is also small in distant areas from the radar, and this is because the radar resolution volume is difficult to be completely filled due to beam broadening with distance. For a smaller threshold Z t under which the beam
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. LI et al.: QUALITY CONTROL METHOD OF GROUND-BASED WEATHER RADAR DATA
5
Fig. 2. Statistical and identification results using the 0.5° PPI reflectivity of the Nanjing radar from May to October, 2012. (a) Difference of clustering centers between the two clusters under different reflectivity thresholds. (b) Identified regions of ground clutter (red). (c) Identified regions of beam blockage (red).
Fig. 1. Spatial distribution of the ratio of the number of times that reflectivity Z exceeds a given threshold Z t for 0.5° PPI of the Nanjing radar from May to October, 2012. Thresholds are chosen as (a) −10, (b) −5, (c) 0, (d) 5, (e) 10, (f) 15, (g) 20, and (h) 25 dBZ.
blockage has the maximum difference from the others, the fuzzy clustering technique is used to divide the grids into two categories and the category with a smaller Pθ,ϕ,r corresponds to the regions of beam blockage or distant areas where radar resolution volumes cannot be usually completely filled. Fig. 2(a) shows the difference of clustering centers between the two categories for different reflectivity thresholds. It can be seen that there is a smooth variation of the difference until 27 dBZ. When the threshold becomes very large, Pθ,ϕ,r may become very small for all grids so that there would be not two distinct categories to classify. In this situation, the centers of the two divided clusters would be close to zero and thus their difference would be nearly zero. The first maximum difference is at −10.5 dBZ, and then the difference begins to decrease because the beam blockage category is gradually blended with clear-air echo or precipitation echo with the increasing threshold. Therefore, −10.5 dBZ is used as an appropriate smaller threshold to yield the beam blockage category. A minimum difference is at 6.5 dBZ and then the
difference begins to increase. Around this minimum difference and the threshold, the statistical properties of clear-air echo and ground clutter are mixed in grids and thus the clustering cannot give desired results. The second maximum difference is at 17 dBZ, and then the difference begins to decrease because the category with a smaller Pθ,ϕ,r is gradually blended with ground clutter with the increasing threshold. Therefore, 17 dBZ is used as an appropriate larger threshold to yield the ground clutter category [Fig. 2(b)]. Although radar resolution volumes that cannot be completely filled appear frequently in distant areas for long-term statistics, they may be completely filled for a certain time. The elimination of echoes in these areas can ensure the reflectivity accuracy, but can also reduce the number of available samples. Therefore, only echoes in beam blockage regions are to be eliminated after the identification of beam blockage. Since the beam blockage regions expand along radial directions [Fig. 1(a)], the summation of Pθ,ϕ,r on grids in each radial, Pθ,ϕ , is calculated. The fuzzy clustering technique is used to divide all radials into two categories and the category with a smaller Pθ,ϕ corresponds to the radials in which the beam blockage regions lie. Consequently, the regions of beam blockage are determined by the intersection of the grids with a smaller Pθ,ϕ,r and the radials with a smaller Pθ,ϕ [Fig. 2(c)]. For the quality control of individual radar scans, the reflectivity from the scan is resampled into the uniform grids. The identified regions of ground clutter and beam blockage are then used as masks to remove data. If actual echoes are resampled into grids of ground clutter or beam blockage, they are identified as ground clutter or beam blockage and are removed in order to accomplish the quality control.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. 6
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING
TABLE II
III. Q UALITY C ONTROL R ESULTS OF THE N EW AND C ONVENTIONAL M ETHODS
S TATISTICAL R ESULTS OF Q UALITY C ONTROL ON G ROUND C LUTTER BY THE C ONVENTIONAL AND THE N EW M ETHODS FOR 20 T IMES ON AUGUST 15, 2012
A. Testing Approach of Quantitative Evaluation In order to assess performance of the new method proposed in this paper, quantitative evaluation is needed. Taking point-by-point comparison between the actual value and the predicted value, contingency tables were originally used to evaluate forecast models [17]. In this paper, a similar approach is proposed based on the comparison between the visual recognition which is considered true and the automatic identification by the new method or the conventional method. For a reflectivity sample at a point, if both of the visual and the automatic methods identify it as ground clutter or beam blockage, a successful identification is considered. If the visual recognition identifies the sample as ground clutter or beam blockage while the automatic method does not, a failure is considered. If the visual recognition does not identify the sample as ground clutter or beam blockage while the automatic method does, a false identification is considered. Consequently, the probability of detection (POD), the false alarm ratio (FAR), and the critical success index (CSI) can be calculated as Ns (4) POD = Ns + N f Ni FAR = Ns + Ni Ns CSI = Ns + Ni + N f
(5) (6)
where Ns, Nf, and Ni represent the number of times for the successful identification, the failure identification, and the false identification, respectively. Values of POD, FAR, and CSI are between 0 and 1. A larger value of POD indicates a higher probability of correct identification, a smaller value of FAR indicates a lower possibility of false identification, and a larger value of CSI indicates higher identification accuracy. B. Comparison of Identification Results Between the New and the Conventional Methods An actual rainfall process on August 15, 2012 is used to evaluate the performance of the new method as well as the two representative conventional methods. On the one hand, the ground clutter is identified by the new method and a fuzzy logic method used for the CINRAD in operational service [18], which is similar to the National Center for Atmospheric Research Radar Echo Classifier [19]. In this specific conventional single-polarization method, five kinds of echo features are adopted to distinguish between ground clutter and meteorological echoes, including the horizontal texture of reflectivity, the vertical gradient of reflectivity, SPINchange [1], mean radial speed, and mean spectral width. The fuzzy membership function is used to calculate the identification index and the sample is identified as ground clutter if the identification index exceeds a given threshold of 0.6. On the other hand, the beam blockage is identified by the new method and a DEM method based on GTOPO30 (a global DEM with a horizontal grid spacing of approximately 1 km) under the
assumption of standard atmospheric refraction. Meanwhile, the ground clutter and the beam blockage are visually recognized point-by-point, which are considered true. Afterward, the identification results are used to calculate POD, FAR, and CSI for testing the new and the conventional methods. Table II gives the identification results of ground clutter using the conventional and the new methods for the 20 consecutive times from 12:40 to 14:53 on August 15, 2012. It can be seen that the POD of the new method is higher than the conventional method, and the FAR of the conventional method is much higher than the new method. It means that the new method correctly recognizes more ground clutter samples than the conventional method, and the conventional method falsely recognizes much more nonground clutter samples as ground clutter samples than the new method. Therefore, the new method has a larger CSI value compared with the conventional method. Fig. 3 gives the original and the quality controlled reflectivity at 0.5° elevation of the radar at 13:22 on August 15, 2012, focusing on the ground clutter close to the radar site. In Fig. 3(b), the conventional method wrongly recognizes the echoes in the larger white circle as ground clutter and modifies their reflectivity values. Meanwhile, it misses the ground clutter echoes in the other smaller white circles. The conventional method makes these erroneous identifications, because sometimes there is no obvious difference between ground clutter echoes and meteorological echoes. Some precipitation echoes have features similar to ground clutter echoes such as large reflectivity gradients, and some ground clutter echoes have features similar to precipitation echoes such as small reflectivity gradients. In addition, it may be difficult to use horizontal and vertical gradients in distinguishing between clutter and precipitation for echoes far away from the radar because the radar resolution volume and the vertical
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. LI et al.: QUALITY CONTROL METHOD OF GROUND-BASED WEATHER RADAR DATA
Fig. 3. PPI reflectivity (in dBZ) at 0.5° elevation of the Nanjing radar at 13:22 on August 15, 2012 before and after the quality control using the frequently used conventional method (a fuzzy logic technique for echo structure difference between ground clutter and precipitation) and the new method. (a) Reflectivity before the quality control. (b) Modified reflectivity for ground clutter regions identified by the conventional method. (c) Reflectivity after eliminating ground clutter regions identified by the new method. See text for more details. TABLE III S TATISTICAL R ESULTS OF Q UALITY C ONTROL ON B EAM B LOCKAGE BY THE C ONVENTIONAL AND THE N EW M ETHODS FOR 20 T IMES ON AUGUST 15, 2012
interval between adjacent elevations increase with the distance from the radar. By contrast, the new method holds primary spatiotemporal distribution characteristics of echoes in the statistical period and thus avoids the overlapping features of echo structure between precipitation and ground clutter in an individual radar scan. In Fig. 3(c), the new method gives the more satisfactory result that all of the ground clutter echoes are eliminated with little false identification. Table III gives identification results of beam blockage by the conventional and the new methods for the 20 consecutive
7
Fig. 4. PPI reflectivity (in dBZ) at 0.5° elevation of the Nanjing radar at 13:22 on August 15, 2012 before and after the quality control using the frequently used conventional method (DEM technique under standard atmospheric refraction) and the new method. (a) Reflectivity before the quality control. (b) Reflectivity after eliminating beam blockage regions by the conventional method. (c) Reflectivity after eliminating beam blockage regions by the new method. See text for more details.
times. The POD of the new method is higher than the conventional method, while the FAR of the new method is lower than the conventional method. Therefore, the new method has a higher CSI compared with the conventional method. Fig. 4 gives the original and the quality controlled reflectivity at 0.5° elevation of the radar at 13:22 on August 15, 2012, focusing on the beam blockage. In Fig. 4(b), the conventional method identifies three beam blockage regions labeled as 1, 2, and 3. In region 1, there is an area (in the white circle) which is not recognized as beam blockage, but it should be judged as partly beam blockage because the reflectivity along the azimuth angles of this area is significantly weaker compared with adjoining areas. Compared with the original reflectivity in Fig. 4(a), region 2 seems obviously larger than the size that it should be. The beam blockage in region 3 is not obvious along radial directions so it should be ignored. These unreasonable identification results may be due to the fact that actual atmospheric refraction did not meet the standard atmospheric refraction when rainfall occurred in these areas during this time period. In addition, the accuracy of topography information may be limited by the variation of vegetation and buildings. By contrast, in Fig. 4(c), the new method gives more satisfactory results, which show a good agreement with the obvious weaker echoes along the radial directions and indicate the beam blockage regions around azimuth 135° and 225° in long-term operational service. IV. C ONCLUSION Ground-based weather radars are often affected by surrounding terrains, and thus, radar data are usually contaminated by ground clutter and beam blockage. Therefore, it is necessary to remove these nonmeteorological signals or weakened signals from meteorological signals. However, it is hard to ensure the performance of the conventional quality control
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. 8
methods due to the complexity and diversity of radar echoes in an individual radar scan. By comparison, due to the fixed locations of terrain, the differences between ground clutter, beam blockage, and meteorological echo can be accumulated through statistics of echo intensity over a period and thus become obvious. Accordingly, a new quality control method is proposed in this paper based on the statistics of echo intensity on fixed grids. In the new scheme, certain statistical conditions of the number of samples and the spatiotemporal range are needed for the identification of ground clutter or beam blockage. Ground clutter has fixed positions and appears no matter whether rainfall occurs or not. The frequency of a large reflectivity appearing at regions of ground clutter is significantly larger than other regions if enough reflectivity samples (e.g., more than 30) are provided. In comparison, the appearance of beam blockage needs to be distinguished from the appearance of precipitation or clear-air echoes since the primary feature of beam blockage is the much weaker echoes along radial directions than those on both sides. Therefore, the identification of beam blockage through statistics on the frequency of a small reflectivity requires enough samples so that the uniform spatial distribution of precipitation or clear-air echoes can be achieved except the beam blockage regions. Another requirement is to narrow the coverage from the radar which only needs to ensure that the farthest terrain causing beam blockage is covered. In this way, the uniform spatial distribution of precipitation echo or clear-air echo would be more easily satisfied, and statistical results for a smaller coverage can be extended along the radial directions to the entire radar detection range. The new method can be implemented by the following steps. 1) PPI or volume scan data of a weather radar that satisfies the statistical conditions for the identification of ground clutter or beam blockage are collected. 2) The PPI reflectivity at the elevation where ground clutter or beam blockage appears is resampled onto fixed grids in polar coordinates. 3) The ratio of the number of times that the reflectivity exceeds different reflectivity thresholds on each grid is calculated. Under each reflectivity threshold, the FCM clustering algorithm is used to classify all grids with these ratios into two clusters, and the difference between the centers of the two clusters is calculated. Corresponding to the two reasonable maximum ones among all difference values, a suitable smaller and larger thresholds can be selected for the identification of beam blockage and ground clutter, respectively. 4) The clustering result under the selected smaller threshold is considered as the optimal clustering for the identification of beam blockage, and one cluster with a smaller ratio is defined as beam blockage regions in the uniform grids. The clustering results under the selected larger threshold is considered as the optimal clustering for the identification of ground clutter, and one cluster with a larger ratio is defined as ground clutter regions in the uniform grids. 5) For quality control of individual radar scans, the reflectivity data from the scan are resampled into the uniform
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING
grids, and the identified ground clutter and beam blockage regions are then used as masks to remove data from the scan. The new method is implemented on the reflectivity at the lowest elevation of Nanjing weather radar from May to October, 2012. In order to quantitatively evaluate the quality control results, an evaluation scheme with POD, FAR, and CSI indices is introduced. With the aid of visual recognition, evaluation scores of the results by the new and the two representative conventional methods are computed for an actual precipitation process of 20 consecutive times. The comparison and analysis suggest that the new method can effectively improve the identification results of ground clutter and beam blockage compared with the conventional methods for single-polarization weather radars. The successful implementation of the proposed new method provides the prospect of further application. The new method is applicable to other operational radar systems whenever large samples of reflectivity that meet the necessary statistical conditions can be acquired, regardless of hardware performance of different radars. However, it might be difficult for the new method to identify clutter resulting from anomalous propagation of the electromagnetic wave because the anomalous propagation is likely not to be stationary within the statistical period. This problem is required to be addressed in the future studies, and inspiration may be acquired through comparisons with other skilled methods. ACKNOWLEDGMENT The authors would like to thank the anonymous reviewers for their valuable suggestions regarding earlier versions of this paper. R EFERENCES [1] M. Steiner and J. A. Smith, “Use of three-dimensional reflectivity structure for automated detection and removal of nonprecipitating echoes in radar data,” J. Atmos. Ocean. Technol., vol. 19, no. 5, pp. 673–686, May 2002. [2] M. Berenguer, D. Sempere-Torres, C. Corral, and R. Sánchez-Diezma, “A fuzzy logic technique for identifying nonprecipitating echoes in radar scans,” J. Atmos. Ocean. Technol., vol. 23, no. 9, pp. 1157–1180, Sep. 2006. [3] V. Lakshmanan, A. Fritz, T. Smith, K. Hondl, and G. Stumpf, “An automated technique to quality control radar reflectivity data,” J. Appl. Meteorol. Climatol., vol. 46, no. 3, pp. 288–305, Mar. 2007. [4] L. Sadouki and B. Haddad, “Classification of radar echoes with a textural–fuzzy approach: An application for the removal of ground clutter observed in Sétif (Algeria) and Bordeaux (France) sites,” Int. J. Remote Sens., vol. 34, no. 21, pp. 7447–7463, Nov. 2013. [5] S. M. Torres and D. S. Zrnic, “Ground clutter canceling with a regression filter,” J. Atmos. Ocean. Technol., vol. 16, no. 10, pp. 1364–1372, Oct. 1999. [6] J. C. Hubbert, M. Dixon, S. M. Ellis, and G. Meymaris, “Weather radar ground clutter. Part I: Identification, modeling, and simulation,” J. Atmos. Ocean. Technol., vol. 26, no. 7, pp. 1165–1180, Jul. 2009. [7] J. C. Hubbert, M. Dixon, and S. M. Ellis, “Weather radar ground clutter. Part II: Real-time identification and filtering,” J. Atmos. Ocean. Technol., vol. 26, no. 7, pp. 1181–1197, Jul. 2009. [8] G. Delrieu, J. D. Creutin, and H. Andrieu, “Simulation of radar mountain returns using a digitized terrain model,” J. Atmos. Ocean. Technol., vol. 12, no. 5, pp. 1038–1049, Oct. 1995. [9] J. Bech, B. Codina, J. Lorente, and D. Bebbington, “The sensitivity of single polarization weather radar beam blockage correction to variability in the vertical refractivity gradient,” J. Atmos. Ocean. Technol., vol. 20, no. 6, pp. 845–855, Jun. 2003.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. LI et al.: QUALITY CONTROL METHOD OF GROUND-BASED WEATHER RADAR DATA
[10] P. A. Kucera, W. F. Krajewski, and C. B. Young, “Radar beam occultation studies using GIS and DEM technology: An example study of Guam,” J. Atmos. Ocean. Technol., vol. 21, no. 7, pp. 995–1006, Jul. 2004. [11] W. F. Krajewski, A. A. Ntelekos, and R. Goska, “A GIS-based methodology for the assessment of weather radar beam blockage in mountainous regions: Two examples from the US NEXRAD network,” Comput. Geosci., vol. 32, no. 3, pp. 283–302, Apr. 2006. [12] J. Bech, U. Gjertsen, and G. Haase, “Modelling weather radar beam propagation and topographical blockage at northern high latitudes,” Quart. J. Roy. Meteorol. Soc., vol. 133, no. 626, pp. 1191–1204, Jul. 2007. [13] P. C. Shakti et al., “Correction of reflectivity in the presence of partial beam blockage over a mountainous region using X-band dual polarization radar,” J. Hydrometeorol., vol. 14, no. 3, pp. 744–764, Jun. 2013. [14] Y. Gou, “The optimization and evaluation of quantitive precipitation estimation based on multi-radar mosaic,” (in Chinese), Ph.D. dissertation, Chinese Acad. Meteorol. Sci., Beijing, China, 2014. [15] J. C. Bezdek, Pattern Recognition With Fuzzy Objective Function Algorithms. New York, NY, USA: Plenum, 1981. [16] J. C. Bezdek, R. Ehrlich, and W. Full, “FCM: The fuzzy c-means clustering algorithm,” Comput. Geosci., vol. 10, nos. 2–3, pp. 191–203, Dec. 1984. [17] I. T. Jolliffe, and D. B. Stephenson, Forecast Verification: A Practitioner’s Guide in Atmospheric Science. Hoboken, NJ, USA: Wiley, 2012. [18] T. Wu, Y. Wan, W. Wo, and L. Leng, “Design and application of radar reflectivity quality control algorithm in SWAN,” (in Chinese), Meteorol. Sci. Technol., vol. 41, no. 5, pp. 809–817, Oct. 2013. [19] C. Kessinger, S. Ellis, and J. Van Andel, “The radar echo classifier: A the WSR-88D,” in Proc. 3rd Conf. Artif. Intell. Appl. Environ. Sci. (AMS), Long Beach, CA, USA, Feb. 2003, pp. 1–11. Nan Li received the B.S. and M.S. degrees in atmospheric sciences from Nanjing University, Nanjing, China, in 2004 and 2007, respectively, and the Ph.D. degree in atmospheric remote sensing and atmospheric sounding from the Nanjing University of Information Science and Technology, Nanjing, in 2011. He is currently an Associate Professor with the School of Atmospheric Physics, Nanjing University of Information Science and Technology. His research interests include radar meteorology, atmospheric dynamics, and astronomical climatology. Zhenhui Wang received the degree from the Department of Atmospheric Sounding, Nanjing University of Information Science and Technology (formerly, Nanjing Institute of Meteorology), Nanjing, China, in 1978. He is currently a Professor with the School of Atmospheric Physics, Nanjing University of Information Science and Technology. He has been involved in atmospheric sounding and remote sensing since 1978. His research interests include atmospheric parameter inversions from meteorological satellites, ground-based radiometers, and radar systems.
9
Kangyuan Sun received the B.S. and M.S. degrees from the Nanjing University of Information Science and Technology, Nanjing, China, in 2010 and 2013, respectively. He is currently pursuing the Ph.D. degree with Nanjing University, Nanjing. His research interests include the data processing of wind profiler radars and Doppler weather radars.
Zhigang Chu received the M.S. degree in atmospheric sounding and the Ph.D. degree in atmospheric physics from the Nanjing University of Information Science and Technology, Nanjing, China, in 2009 and 2013, respectively. He is currently a Lecturer with the School of Atmospheric Physics, Nanjing University of Information Science and Technology. His research interests include remote sensing for understanding and quantifying weather and cloud/precipitation microphysics.
Liang Leng received the B.S. and M.S. degrees in atmospheric sounding from the Nanjing University of Science Information and Technology, Nanjing, China, in 2008 and 2011, respectively. He is currently an Engineer with the Institute of Heavy Rain, Wuhan, China. His research interests include the quality control and the automatic recognition algorithm of weather radar data.
Xingchao Lv received the B.S. degree from the Nanjing University of Information Science and Technology, Nanjing, China, in 2017, where he is currently pursuing the M.S. degree. His research interests include the quality control and data processing of ground-based and spaceborne weather radars.