Quality control of rainfall measurements in Cyprus

Transcription

1 Meteorol. Appl. 13, (2006) Quality control of rainfall measurements in Cyprus Claudia Golz 1, Thomas Einfalt 1 & Silas Chr. Michaelides 2 1 einfalt&hydrotec GbR, Breite Str. 6-8, D Luebeck, Germany 2 Meteorological Service, Nicosia, Cyprus c_golz@einfalt.de; thomas@einfalt.de; silas@ucy.ac.cy doi: /s x The basic condition for using precipitation data from raingauges and radars is data quality control. This aspect is important for comparing and using rainfall data, for example in models. In the scope of the EU-project VOLTAIRE (Validationofmultisensors precipitation fields and numerical modelling in Mediterranean test sites) rain data from Cyprus have been analysed. Different quality control methods have been applied to the rainfall data of 158 raingauges and the data of 11 events (in 2002 and 2003) of the C-Band radar in Kykkos. The first results of the use of ground clutter algorithms for radar data in Cyprus are presented in the paper. Keywords: Data quality control, radar data, Cyprus, VOLTAIRE. Received April 2005, revised December Introduction Precipitation data could be used more intensively in many fields of activity (meteorology, hydrology, etc.), but this is dependent on the improvement of data quality (Einfalt et al. 2004) delivered by the measuring devices and of the quality enhancement procedures (Golz et al. 2005a) available for the measured data. Data quality control (QC) can be divided into quality check and data correction (Michelson et al. 2004). The data quality check is the process by which data are analysed in order to categorise them (either in a discrete or a continuous scheme). It includes the identification of measurement errors. Data correction is the process by which data that have been labelled as suspicious or error loaded in the quality check are modified in such a way that either they will pass the quality check after correction or their quality is improved. Not all errors can be corrected. Only quality checked or corrected data (depending on the purpose of use) should be accepted for further processing steps. Table 1 provides an overview of selected known data problems for one single data source (ground or space radar, continuous or daily raingauge), which should be minimised or eliminated by means of check and correction algorithms. If the data are to be used fully and competently, it is essential to document the QC methods applied to them. In VOLTAIRE a metadata model compatible with international data description conventions has been developed, implemented and used for this purpose (Golz et al. 2005b). 2. What is VOLTAIRE? VOLTAIRE (Validation of multisensors precipitation fields and numerical modelling in Mediterranean test sites) is an EU-funded project. One of the objectives of VOLTAIRE is to compare data quality control schemes for radar (ground and space) and raingauges. The gaugeadjusted and radar-derived precipitation fields are to be used as ground validation for the TRMM radar in the Mediterranean area (around Cyprus). This is a basic condition for preparing for European participation in the future Global Precipitation Measurement (GPM) satellite mission, which will offer the possibility of measuring precipitation using an onboard precipitation radar at higher latitudes than is presently possible (35 N) using the TRMM radar. Another aim of the VOLTAIRE project is to improve the data quality of ground radar and raingauge data by reducing or correcting the measurement errors, for example in mountainous terrain. This is an important requirement in comparing precipitation fields, as represented by numerical models, by adjusted ground-radar and by space-borne radar. More detailed information on the project can be obtained from the official VOLTAIRE website at org. 3. Ground clutter and speckle correction methods Ground clutter is difficult to avoid in radar measurements with a low elevation angle, since it occurs when fixed objects such as buildings or mountains reflect the 197

2 Claudia Golz, Thomas Einfalt & Silas Chr. Michaelides Table 1. Overview of project related relevant rain data quality items. Data source Ground radar Continuous raingauge Daily raingauge Space radar Error types 1. Attenuation 1. Data gaps 1. Data gaps 1. Attenuation 2. Ground clutter 2. Plugging 2. Sums of several days 2. Convective/stratiform 3. Z-R Relation, convective/ stratiform 3. Plausibility errors 3. Plausibility errors 3. Homogeneous beam filling 4. Bright band, vertical profile 4. Zero rainfall 5. Radial anomalies 5. Timing errors 6. Speckle 7. Anomalous propagation 8. Homogeneous beam filling radar pulses and produce non-meteorological echoes. These echoes are generally recognised as stationary and as pixels with high reflectivity values. For non- Doppler radars in particular, ground-clutter analysis plays an important role. Doppler radars can detect clutter more easily through the comparison of potential clutter pixels with the corresponding movement speed, since clutter does not move. The term speckles means radar reflections over a very small area, usually variable in time, e.g. reflections from airplanes. Within the scope of VOLTAIRE, three different clutter algorithms were chosen to minimise or eliminate ground clutter in 2-D radar data: a cluttermap, the texturebased algorithm and the segment size algorithm Cluttermap A fixed cluttermap is a commonly used and well-cited method of removing permanent ground clutter (see, for example, Harrison et al. 2000). A cluttermap contains the image pixels, which are constant or, in most cases, influenced by ground clutter, e.g. high buildings or mountains. Such a map is not variable, which can lead to problems if, for example, signals of water vapour from power plants are detected by the radar. Pixels flagged by the cluttermap are removed and have to be interpolated using the surrounding pixels Texture based algorithm The texture-based algorithm by Gabella & Notarpietro (2002) detects small areas such as speckles, which have a high gradient compared to their neighbourhood. It also helps to remove artefacts on radar images, e.g. thin lines. The texture-based method can be applied to polar and Cartesian data in combination with any other method of clutter removal (e.g. cluttermap). This algorithm is composed of two parts: a spatial-proximity filter and a test of compactness. The first part eliminates pixels that are weakly spatially correlated to the surrounding ones. Therefore, Gabella & Notarpietro chose a 5 5 pixel window around the pixel in question and if the differences 198 between the pixel in the middle and the surrounding ones are below a given threshold, they are accepted as representing a meteorological echo. Otherwise, the pixel is replaced. The second part of the algorithm tests the number of pixels that belong to the same group (i.e. set of connected pixels). The ratio of the total number of pixels in one group to the number of pixels defining its boundary should be greater than a given threshold. Pixel values flagged by the texture-based algorithm are removed and have to be interpolated using the surrounding pixels Segment size algorithm The algorithm segment size is a speckle filter. It computes the number of connected image pixels with values greater than zero constituting a segment. With a defined segment size threshold, also depending on the pixel size, it is possible to eliminate speckle with few pixels on a radar image. In contrast to the texture-based algorithm, this method works only on segments (i.e. pixels with values greater than zero) which are surrounded by zero-value pixels. The procedure sets the pixels in such segments to zero. 4. Evaluation methods An objective evaluation of the efficiency of the algorithms is quite difficult. While it is easy to prove when the data have been modified, it is difficult to decide when the modification is really an improvement to the data. Therefore, the evaluation here is a combination of subjective judgement and use of a spatial variance Spatial variance The spatial variance used for evaluation is based on the standard deviations in dbz for 3 3-pixel fields. All standard deviations of the 3 3-pixel fields are averaged over a radar image to calculate a value describing the degree of uniformity. An image with a lot of singular points (e.g. ground clutter), which is very non-uniform, would have a high spatial variance whereas an image with pixels of one value, which is very uniform, would have a very low spatial variance. In the case of ground

3 Quality control of rainfall measurements in Cyprus clutter correction in particular, it is desirable to reduce the spatial variance. 5. Results of the Cyprus offline tests The process of checking the QC filter on a past rainfall event is called an offline test. This is the basis of evaluating which filter is reasonable for later permanent real-time ( online ) use. The analysed volume data of the C-Band radar at Kykkos in Cyprus are polar data ( ) with a pixel size of 500 m 1. The number of elevations varied for the inspected events from two to eight, and the most commonly used measurement interval was 15 minutes. In general, the events comprise radar data of two or three days, during which a precipitation field passed Cyprus. In preparation for the offline tests, the radar data were manually checked for different measurement errors (ground/sea clutter, attenuation, radial anomalies and bright band). The results of the manual check give an overview of the observations of the 11 events and indicate which data can be used for the offline tests. One hour on 4 February 2003 (02:15 03:00) was skipped because the values were not related to rain (Figure 1). Table 2 contains the percentages of all usable images for the four measurement errors by event: ground/sea clutter, attenuation, radial anomalies and bright band signature. The last column, called extreme ground/sea reflection, illustrates the days ( 100%), for which the radar was set to measure very low elevations ( 2,0 / 1,8 / 1,5 / 1,3 ) for hardware test reasons. This results in partly unreliable data, because of strong ground and sea reflection (Figure 2). Table 2 provides an overview of the events that are deemed suitable for a particular quality control algorithm. From the table, it is obvious, that all images ( 100%) are contaminated with ground/sea clutter. A Figure 1. Polar radar images of 0 elevation, 4 February 2003, 0200 (upper image) and 0215 (lower image). first step is to look at the different clutter correction algorithms. The results of the three ground clutter algorithms (described in Section 3), namely cluttermap, segment size algorithm, texture- based algorithm, and a combination of the texture-based and cluttermap algorithms have been compared. The cluttermap used for the Cyprus data was produced by Politecnico di Torino on the basis of 1660 precipitation-free images defined visually. It was prepared for two different thresholds: 10 dbz and 25 dbz. This means that if a pixel exceeded this threshold on one image it was counted for the cluttermap with a frequency of 1. Since the number of clutter pixels fluctuates from image to image and from day to day, a threshold and a temporal percentage of the presence of each potential clutter pixel was computed. For the Table 2. Percentage of images with observed measurement errors. Events Ground/sea clutter (%) Attenuation (%) Radial anomalies (%) Bright band (%) Extreme ground/sea reflection (%) Feb May Jan Feb Feb Mar Mar Mar Dec Dec Dec

4 Claudia Golz, Thomas Einfalt & Silas Chr. Michaelides averaged spatial variance [dbz] original segment size texture based cluttermap texture based & cluttermap correction methods Figure 3. Averaged spatial variances with standard deviations for original and clutter corrected radar data of the event of May Figure 2. Polar radar images of 0 (upper image) and 2,0 (lower image) elevation, 22 January 2003, offline tests the cluttermap with the threshold of 10 dbz was chosen and a pixel was flagged as a clutter pixel if its value was larger than 415 (25% of the 1660 images). This signifies that in more than 25% of all the investigated clear-sky images, the pixel was a pixel with a value larger than 10 dbz. The chosen clutter algorithms were applied to the eleven events for all elevations. The results of the ground clutter corrections were evaluated visually on the basis of single images and statistically with the calculation of spatial variances (Section 4.1). Table 3 contains the spatial variances (SV) averaged over the selected events for the 0 elevation. The mean values of the SV are given with the corresponding standard deviations. As described in Section 4.1, a reduction of the SV is interpreted as an indication of successful reduction of ground clutter. The image then becomes smoother. The results show that the algorithm segment size has the smallest reduction of the tested methods (around 5%). In contrast, the cluttermap algorithm and the texture based algorithm both resulted in reductions of SV up to 63%. The combination of the texture based algorithm and the cluttermap shows the most significant decrease, e.g. 79% for the event of May Figure 3 presents the development of the averaged spatial variances for this event. It can be seen that the cluttermap and the texture based filter have the best results according to the SV criterion. In addition to the calculation of the spatial variances, the corrected images were checked visually. Figure 4 shows one example with a single image for 18 March It contains the original image and the corrected images, after applying the different methods. It is obvious that the dark clutter area of the mountains (black circle) can only be successfully removed by applying a cluttermap filter. Also, the small points (black rectangle) which can be associated with sea clutter or with other non-precipitating echoes can be successfully removed using the texture-based algorithm. The segment size algorithm is not cleaning the image as visibly as the texture based one, because it corrects only small clutter areas (black triangle). Therefore, the combination of cluttermap application and texture-based algorithm Table 3. Averaged spatial variances [dbz] and standard deviations for original and clutter corrected radar data. Events Original Segment size Texture based Cluttermap Texture based & Cluttermap Feb ± ± ± ± ± May ± ± ± ± ± Jan ± ± ± ± ± Feb ± ± ± ± ± Feb ± ± ± ± ± Mar ± ± ± ± ± Mar ± ± ± ± ± Mar ± ± ± ± ± Dec ± ± ± ± ± Dec ± ± ± ± ± Dec ± ± ± ± ±

5 Quality control of rainfall measurements in Cyprus Figure 4. Cartesian radar images of 0 elevation, 18 March 2003, produces the best result. This result was obtained for all analysed Cyprus events. 6. Conclusions Data quality control is essential for all rain data. This paper presents the first results of offline tests using the Cyprus radar data. Three different ground-clutter correction methods namely cluttermap, texture based and segment size algorithms were applied to 11 events in 2002 and The visual check of the radar images and the evaluation using calculated spatial variances showed that the combination of a cluttermap and the texture based algorithm produces the best result. This combination satisfactorily eliminated the ground clutter/speckle. Acknowledgements A large part of this work has been funded and implemented within the scope of the EU-project VOLTAIRE (EVK CT-00155). References Einfalt, T., Arnbjerg-Nielsen, K., Golz, C., Jensen, N. E., Quirmbach, M., Vaes, G. & Vieux, B. (2004) Towards a roadmap for use of radar rainfall data use in urban drainage. J. Hydrol. 299: Gabella, M. & Notarpietro, R. (2002) Ground clutter characterization and elimination in mountainous terrain. Proc. ERAD (2002), Copernicus GmbH, ISBN , pp Golz, C., Einfalt, T., Gabella, M. & Germann, U. (2005a) Quality control algorithms for rainfall measurements. Atmos. Res. 77: Golz, C., Einfalt, T., Gabella, M. & Galli, G. (2005b) Metadata definition. Public VOLTAIRE document VOLTAIRE_HDF5.pdf. Available from voltaireproject.org/ Harrison, D. L., Driscoll, S. J.& Kitchen, M. (2000) Improving precipitation estimates from weather radar using quality control and correction techniques. Meteorol. Appl. 6: Michelson, D., Einfalt, T., Holleman, I., Gjertsen, U., Friedrich, K., Haase, G., Lindskog, M. & Jurczyk, A. (2004) Weather radar data quality in Europe: quality control and characterization. Available from COST 717 website: 201