morel 2003

Upload: george-tsavd

Post on 08-Jan-2016

212 views

Category:

Documents


0 download

DESCRIPTION

Morel 2003

TRANSCRIPT

  • COMPARING LIGHTNING DETECTION AND SATELLITE-BASED DETECTION FOR CONVECTIVE SYSTEM IDENTIFICATION : THE CASE OF IMPACT TECHNOLOGY VERSUS METEOSAT RAPID SCAN DATA

    Christophe MOREL 1, Fatiha EL MAHDAOUI 2, Stphane SENESI 1 and Ernest NDri KOFFI 3

    1 Nowcasting R&D. DPrevi. Mto-France. 42, avenue G. Coriolis. 31057 Toulouse Cedex 1. France

    2 Service Radar et Applications Nouvelles. DMN Casablanca. BP 81 06 Oasis. Morocco

    3 Department of Geosciences.University of Fribourg. Perolles CH-1700 Fribourg. Switzerland

    ABSTRACT

    The Rapid Developing Thunderstorms (RDT) satellite-based discrimination method has been tuned using METEOSAT Rapid Scan infrared images. Results of this tuning shows a weaker quality of discrimination compared to previous results obtained from GOES infrared images. Possible explanations are listed. Performances of this satellite-based discrimination of convective systems are also presented, in terms of detection precocity, as compared to the first occurrence of cloud-to-ground (CG) lighting flash detected by an IMPACT technology network (the western European maritime lightning location system). It shows that although the RDT software allows a very early tracking of convective systems, time of first discrimination often occurs after the detection of the first CG lightning flash. However, there are some convective systems for which the satellite-based discrimination happens before first occurrence of CG lightning flash and it is observed that these convective systems have a tendency to be more electrically active and to have longer durations than the whole sample.

    1. INTRODUCTION The RDT (Rapid Developing Thunderstorms) product of the EUMETSAT SAF Nowcasting aims at automatically detect, track and characterize convective systems from infrared MSG (METEOSAT Second Generation) data. One of the challenges of the RDT software is to design an efficient method to automatically discriminate convective systems, as soon as possible after cloud formation, among all the tracked cloud systems from their satellite characteristics only. The current discrimination method is presented in section 2. Koffi et al (2001) showed that the RDT algorithm is able to detect most of the convective systems before they trigger their first lightning flash, provided that they are not obscured by an upper cloud shield (for instance, using GOES images at a 15 minutes time-resolution, more than 80% of these convective systems are detected by the RDT software before their first flash occurrence). They also briefly investigated the use of the cloud top cooling rate in order to discriminate convective systems before the occurrence of their first detected flash. Finally, these authors evidenced the great sensitivity of these results to the frequency of the infrared images used (METEOSAT, GOES at 30, 15 and 7 minutes time-resolution).

  • A study has been conducted (see section 3) to i) quantify the quality of the RDT discrimination method (in terms of percentage of good detections and false-alarms rate) from Meteosat Rapid Scan Service images (10 minutes time-resolution), and ii) assess the performance of this satellite-based discrimination of convective systems, in terms of detection precocity, as compared to the first occurrence of cloud-to-ground flash. Cloud-to-ground lightning data used in this study comes from the western European maritime lightning location system (see figure 1A). This network is based on the IMPACT technology (Cummins et al, 1998).

    Figure 1A Figure 1B

    Figure 1: A): Sensor locations for the western European maritime lightning location system. B): location accuracy of this network in summer 2002 (the red line shows the limit of a location

    accuracy better than 2 km for 50% of the detected lightning flashes)

    2. OVERVIEW OF THE RDT ALGORITHM 2.1 The detection of cloud systems The detection algorithm allows to define the object which will represent the cloud systems as observed in the infrared 10-12 m channel of geostationnary satellites. Once these objects are defined, a number of morphological (area, aspect ratio) and radiative features (average and minimum brightness temperature,) are computed in order to characterize the corresponding cloud systems. More precisely, the objects detected in the RDT software are connected zones (8-connectivity) of pixels having a brightness temperature lower than a given temperature threshold. Moreover, the RDT detection method is based upon an adaptative temperature thresholding of infrared images (Crane, 1979) which allows to select for each cloud system the temperature threshold corresponding to the warmest brightness temperature which keeps the cloud system isolated from other clouds (see Morel et al, 2002 and Tzanos et al, 2001 for a full presentation of this method). Figure 2A shows a brightness temperature vertical cross-section of idealized cloud systems (seen as local minima of brightness temperature) and for each of these cloud systems, the temperature threshold chosen by the detection algorithm to define their associated object (red line). The RDT detection algorithm has 5 parameters:

    - coldT which is the coldest possible temperature threshold which could be used to define a cloud system (usually, its value is 55C).

    - warmT which is the warmest possible temperature threshold which could be used to define a cloud system (usually, its value is between 10C and 5C).

    - T which is the brightness temperature step (usually, its value is 1C). Possible temperature thresholds are then : warmT , TTwarm , TTwarm 2 ,... , coldT .

    - towerT which is the minimum vertical extension that a cloud system has to exceed in order to be retained (usually, its value is 3C).

    - minA which is the minimum area threshold that a cloud system has to exceed in order to be retained (usually ,its value is one infrared pixel).

  • Figure 2A Figure 2B

    Figure 2: Diagrams illustrating the principle of the RDT detection algorithm. A): current method. B): version used to tune the satellite-based discrimination method from GOES data

    2.2 The tracking of cloud systems The goal of the tracking algorithm is to link objects detected in two consecutive infrared images and corresponding to the same cloud system. Once the tracking is done, it is then possible to derive the trajectory (i.e. the time series of objects) of cloud systems. The RDT tracking algorithm is fully presented in Morel et al (2002). It is based on the computation of the geographical overlapping of objects. It handles merges and splits of cloud systems and has been specifically designed to be efficient for the tracking of very small (i.e less than five pixels) cloud systems:

    - Use of an estimated speed of cloud systems to compute the overlapping (Morel and Snsi, 2002). - Initialisation of the speed of cloud systems from cross-correlation technique (Morel et al, 2000). - Introduction of a spatial tolerance to ease the existence of geographical overlapping between very

    small cloud systems (Tzanos et al 2002). 2.3 The discrimination of convective systems among all tracked cloud systems The last step of the RDT algorithm is the so-called discrimination of convective systems. The goal of this discrimination step is to identify automatically among all the tracked cloud systems, those corresponding to the convective ones. A quality indicator (percentage of confidence) is attached to this automated discrimination. Two discrimination methods are currently implemented in the RDT software:

    - one based on lightning data (optional input of the software). - and another based on infrared characteristics of the cloud systems only. This satellite-based

    discrimination method of convective systems has been developed in order to propose a possible discrimination of convective objects over regions where lightning detection data are not available or reliable and in an attempt to identify convective systems before lightning occurs.

    Each time an infrared image is processed and once the detection and tracking of cloud systems is performed, these two discrimination methods are applied to each detected cloud systems as follows. Concerning the lightning-base discrimination method, if at least one lightning flash is detected below an object of the trajectory of a given cloud system then this cloud system is diagnosed as convective (with a percentage of confidence set to 100%). Concerning the satellite-base discrimination method, the user of the RDT software has to choose the value of the relative cost FA/NDC of false-alarms (i.e. cloud systems ill-classified as convective) against misses (i.e. convective systems not classified as convective by the discrimination method). The range of FA/NDC values lies between 0.1 and 50 and a value of FA/NDC less than 1 leads to lower misses number (but increases false-alarms). The principle of the satellite-based discrimination is then the following: First, each cloud system is assigned to a given class of trajectories (A , T, S) where A is an interval of age, T is an interval of vertical extension and S is an interval of horizontal extension (area). Possible classes (A , T, S) come from a previous tuning of the satellite-base discrimination method. During this tuning, a best discrimination parameter (see below) and a tuned discrimination threshold have been assigned to each of the classes for several FA/NDC values.

  • For a given cloud system, the age, the vertical and horizontal extensions are defined as follows:

    - The age is the tracking duration of this cloud system up until this processed image (i.e. the time difference between time of first detection of this cloud and time of the current processed image).

    - The vertical extension is defined as the minimum value of the minimum brightness temperature of all the objects corresponding to this cloud system since its first detection.

    - The horizontal extension is defined as the maximum value of the area of all the objects corresponding to this cloud system and having minimum brightness temperatures inside the interval of its class of vertical extension.

    Once the class (A , T, S) of a given cloud system has been found, then the RDT software checks (according to the results of the previous tuning of the satellite-based discrimination method) which discrimination parameter and threshold have to be used for this class and for the specified FA/NDC value. Then, this discrimination parameter is computed for this cloud system and if its value is larger than the discrimination threshold then the cloud system is classified has convective with a percentage of confidence derived from this value and from the values encountered in this class during the tuning of the satellite-based discrimination method There are two kinds of possible discrimination parameters in the satellite-based discrimination method:

    - Those based on the peripheral gradient of IR10.8 brightness temperature of the detected objects (i.e. gradient computed only along the edge of the object). Their usefulness comes from the fact that strong values of peripheral temperature gradients mainly occur for convective clouds because of the sharp edge of their cold anvil observed during their development. The parameters which were used during the tuning of the satellite-based discrimination method are the mean and 95 percentile of this peripheral gradient (Morel and Snsi, 2002).

    - Those based on the cooling rate of cloud systems (infrared brightness temperature difference between two consecutive objects of a trajectory). The usefulness of these parameters is physically based on the fact that a strong cooling is mainly observed for convective systems because of their rapid vertical extension during their development phase (Koffi et al, 2001). The parameters which were used during the tuning of the satellite-based discrimination method are the cooling rates based on the mean and minimum brightness temperature.

    The tuning of the discrimination parameters and thresholds has been performed on one month of GOES IR10.8 infrared images of summer 1999 (06-16/06/99, 03-10/07/99, 21-29/07/99, 13-17/09/99) at full temporal resolution (15 minutes). The RDT software version used during this tuning was a former one:

    - the detection method was not the same (see Tzanos et 2001 and figure 2B). - the tracking algorithm was not as efficient as the current one concerning the tracking of small clouds

    (the introduction of spatial tolerance was not effective). The tuning has been done for several values of FA/NDC (0.1, 0.3, 0.5, 0.7, 0.8, 0.9, 1, 2, 3, 5, 10 and 50) on the following classes (A , T, S):

    - classes of age (in minutes) [;[ 150 , [;[ 3015 , [;[ 6030 , [;[ 9060 , [;[ 12090 , [;[ 180120 , [;[ 180 . - classes of vertical extension (in C): ];] 50 , ];] 4050 , ];] 3040 , ];] 2030 , ];] 1020 . - classes of horizontal extension (in km2): [;[ 1000 , [;[ 200100 , [;[ 500200 , [;[ 1000500 , [;[ 1000 .

    Figure 3 shows the quality of the discrimination using GOES data, in terms of false alarm rate (FAR) and percentage of good detections (POD), for several classes of age and. POD and FAR are defined as follows: Let Ncv be the number of convective systems. Let Ngood be the number of convective systems that are correctly discriminated as convective by the satellite-based discrimination method. Let Nfalse be the number of non-convective systems that are wrongly discriminated as convective by the satellite-based discrimination method.

    Then, falsegood

    false

    NNN100FAR

    and cv

    good

    NN

    100POD .

    This figure first shows that, for the youngest cloud systems, which have been detected only once (i.e. with an age lower than 15 minutes), the discrimination is rather difficult: the detection rate hardly reaches 40% while the false alarm rate already reaches 25%. However, as soon as a system has been detected twice (red

  • curve), the discrimination efficiency is much better, and allows compromises like 50 to 5 or 80 to 40. The 90% detection rate can be reached with around 15% of false alarms after one hour of tracking. Finally, the discrimination is almost perfect after 3 hours of tracking.

    Figure 3: Quality of the discrimination for several age intervals. Tuning from GOES data.

    3. DISCRIMINATION QUALITY FROM METEOSAT RAPID SCAN INFRARED DATA In this section, results of the tuning of the satellite-base discrimination method from METEOSAT Rapid Scan infrared data and the current version of the RDT detection and tracking algorithms are presented. This tuning has been performed from METEOSAT Rapid Scan infrared images of summer 2002 (17/0608/07, 11/0719/07, 22/0705/08) over a geographical domain covering western Europe and Mediterranean Sea (figure 1B). These images are available at a 10 minute frequency and a spatial resolution of around 6 km over Europe. Corresponding lightning data from the western European maritime lightning location network were also used as ground truth for convection diagnosis. Parameters of the detection algorithm (see section 2) used during this tuning were:

    coldT = 55C / warmT = 5C / T = 1C / towerT = 3C / minA = 1 km2.

    Running the detection and tracking algorithms on these images has led to a convective sample of 1898 convective systems (i.e. with electrical activity below their RDT objects) and 226 157 non-convective systems (i.e. without electrical activity below their RDT objects). Only cloud system trajectories starting inside the domain shown by the red line in figure 1B were considered. The tuning has been done for roughly the same values of FA/NDC (0.1, 0.3, 0.5, 0.7, 1, 2, 3, 5, 10 and 50) on the following classes (A , T, S):

    - classes of age (in minutes) [0;10[ , [20;10[ , [30;20[ , [40;30[ , [50;40[ , [60;50[ , [70;60[ , [80;70[ , [90;80[ , [100;90[ , [110;100[ , [120;110[ , [130;120[ , [140;130[ , [150;140[ , [160;150[ , [170;160[ , [180;170[ ,

    [190;180[ , [200;190[ , [210;200[ , [240;210[ , [270;240[ , [300;270[ , [330;300[ , [360;330[ , [390;360[ , [420;390[ , [450;420[ , [;450[ .

    - classes of vertical extension (in C): ]50;] , ]40;50] , ]30;40] , ]20;30] , ]10;20] , ]0;10] , ]5;0] .

    - classes of horizontal extension (in km2): [100;0[ , [200;100[ , [500;200[ , [1000;500[ , [;1000[ . Figure 4A displays the discrimination quality of this tuning, in terms of POD and FAR, for several classes of age. It shows that the satellite-based discrimination from METEOSAT Rapid Scan data and the up-to-date detection and tracking RDT algorithms is not as efficient as when using GOES data and previous detection and tracking RDT algorithms (figure 3). Indeed, it is observed that this tuning does not allow to perform an early discrimination of convective systems and compromises such as those obtained from GOES data after a tracking in two images (POD of 50% and a FAR of 5%) start to be observed after around 3 hours of tracking. However, it has to be underlined that there is not an exact correspondence between classes of age in these

  • two tunings (for instance, the class of first detection [;[ 150 of the tuning from GOES data does not corresponds to the class of first detection of the tuning from METEOSAT Rapid Scan images but spreads over classes of higher age values) due to the fact the detection and tracking algorithms are not the same and to the use of different values for the warm temperature threshold warmT (-10C for the tuning from GOES against 5C from METEOSAT Rapid Scan data). An attempt at tuning of the discrimination from the same METEOSAT Rapid Scan images using a warmT equal to -10C has been done. Results (see figure 4B) are comparable to those obtained when tuning the discrimination method with warmT = 5C.

    Figure 4A Figure 4B Figure 4: Quality of the discrimination for several age intervals. Tuning from METEOSAT Rapid Scan

    data. A): warmT = 5C. B): warmT = -10C

    Moreover, it is also observed that having a nearly perfect discrimination such as the one observed after 3 hours of tracking in the tuning from GOES data seems not feasible with the tuning based on METEOSAT Rapid Scan images. Indeed, after 9 hours of tracking the discrimination quality still not allow a POD of 90% or more together with a FAR of less than 5%. Figure 5A displays the detection and discrimination precocity as compared to the first occurrence of cloud-to-ground (CG) flash (for a given convective system, this time is the occurrence time of the first CG flash detected below a RDT object corresponding to this system) for the same sample of convective systems (1898 cases). More precisely:

    - the black line is the accumulated frequency of time difference between time of first detection and time of the first CG occurrence. For a given convective system, the time of first detection is the time of the infrared image in which a RDT object corresponding to this system has been detected and tracked for the first time. If this difference is negative then this means that first detection of this convective system by the RDT software occurred before it triggers its first CG flash.

    - the blue (respectively violet and green) line is the accumulated frequency of time difference between time of first discrimination with a relative cost FA/NDC of 0.1 (respectively 1 and 10) and time of the first CG occurrence. For a given convective system, this time of first discrimination corresponds to the first time it is discriminated as convective by the satellite-based discrimination method (using the above tuning) since its first detection. If this difference is negative then this means that the satellite-based discrimination method has allowed to discriminate this convective system before it triggers its first CG flash.

    It is then concluded that for around 77% of the convective systems, the current detection and tracking RDT algorithms have allowed a first detection before they trigger their first CG lightning flash, thus showing the good performances of the RDT software in term of precocity of detection (25% of the convective systems are detected more than 50 minutes before the occurrence of their first CG lightning flash). Concerning the precocity of discrimination of the convective nature of these systems, results are not as positive as for the first detection time, especially when using rather high values (leading to minimize the FAR)

  • of relative cost FA/NDC . For instance, less than 7% (respectively 13%) of the convective systems are discriminated as convective by the satellite-based discrimination method when using a FA/NDC of 10 (respectively 1) before they trigger their first CG flash. This percentage rises to 43% when using a FA/NDC of 0.1 (keeping in mind, see figure 4A, that the FAR is always more than 50% during the first three hours of tracking and that around 80% of the convective systems trigger their first CG flash before one hour of tracking).

    Figure 5A Figure 5B Figure 5: A): Detection and discrimination precocity as compared to the first occurrence of cloud-to-

    ground flash. B): total number of detected cloud-to-ground lightning flashes Nevertheless, even if this does not apply for the majority of convective systems, there are some cases for which a satellite-based discrimination is feasible before they lower their first CG lightning flash and sometimes with a significant anticipation. For instance, around 2% (respectively 4%) are discriminated as convective more than 30 minutes before they lower their first lightning flash (this percentage even reached 13% when using a FA/NDC value of 0.1). Figure 5B shows the accumulated frequencies of the total number of detected CG lightning flashes (for a given convective system, this total number of CG flashes is equal to the sum of all the CG flashes detected below the RDT objects of its trajectory) for different samples of convective systems: the whole sample (black line, 1898 convective systems) and samples of convective systems which are discriminated as convective by the satellite-based discrimination method before they lower their CG first lightning flash for three different values of FA/NDC (0.1, 1 and 10). These samples are respectively composed of 819, 243 and 126 trajectories. It is then observed that for the high values of FA/NDC (1 and 10), convective systems which are discriminated before they lower their first CG flash have a tendency to exhibit a higher electrical activity as compared to the whole sample of convective systems. For instance, one third of the convective systems discriminated before their first CG flash occurrence by the satellite-based discrimination method with a FA/NDC value of 10 lowered more than 67 CG lightning flashes against only 44 for the whole sample of convective systems. It was also observed (not shown) that the convective systems which are discriminated before they lower their first CG flash have a tendency to have longer durations compared to the whole sample of convective systems. For instance, around 50% of the convective systems discriminated with a FA/NDC value of 10 before they lower their first CG flash have a duration of more than 5 hours against only 25% for the whole sample of convective systems.

    4. CONCLUSION After having recalled the RDT satellite-based discrimination method of convective systems among all the detected and tracked cloud systems, the current state-of-the-art of the quality of this discrimination method over western Europe and Mediterranean Sea has been presented. In order to do so, this satellite-based method has been tuned with METEOSAT Rapid Scan infrared images available at 10-minute frequency and 6km pixel resolution.

  • Results of this tuning clearly show that the discrimination quality is weaker than from GOES infrared images. Reasons of this poor discrimination quality obtained from METEOSAT Rapid Scan images and when using up-to-date detection and tracking algorithms are not fully understood. Identified possible sources of this weaker quality are:

    - First, the better spatial resolution of GOES infrared images (around 4.5 km against 6 km for METEOSAT Rapid Scan images). Indeed, this could lead to observe larger values of cooling rate and peripheral gradient of brightness temperature and then could ease the discrimination of convective systems. The planned tuning of this discrimination method from MSG IR10.8 images (having a spatial resolution close to GOES images) will allow to check the sensitivity of the discrimination quality to spatial resolution of the infrared images.

    - Moreover, convective systems are not the same ! Indeed, tuning from METEOSAT Rapid Scan data has been done over western part of Europe and the Mediterranean Sea while the tuning from GOES images has been done on the central part of the USA, over the Great Plains region. The infrared characteristics (cooling rates and peripheral gradients) used to discriminate convective systems are perhaps not as high for European convective systems than those observed for the American ones.

    - Finally, the tuning of the discrimination method from METEOSAT Rapid Scan images has been done on convective systems which lower at least 5 CG flashes during their life-cycle while the tuning from GOES images has been done in a more restrictive way (considering only convective systems which trigger more than 30 CG during the first three hours of their tracking). As a consequence, having considered less electrically active convective systems (in order to keep a sound sample of convective systems) in the tuning of the discrimination method from METEOSAT Rapid Scan images could also contribute to the weaker obtained quality of discrimination as weaker electrically active convective systems seems more difficult to discriminate (the POD, using a FA/NDC = 1, of convective systems which lower more than 100 CG lightning flashes is 97% against only 73% for convective systems which lower more than 5 CG lightning flashes).

    Finally, a study of the relative times of first detection of cloud systems, first detection of CG lightning flash and first discrimination time has also been presented. It shows that the current detection and tracking RDT algorithms allows a very early tracking of convective systems (25% of the convective systems clouds are detected more than 50 minutes before the occurrence of their first CG lightning flash). Concerning the time of first discrimination, it appears that the satellite-based discrimination is often late compared to the detection of the first CG lightning flashes. However, even for large values of relative costs of false-alarms against misses which allow to keep a low FAR, it has been seen that the satellite-based discrimination can occur before first occurrence of CG lightning flash and that these convective systems have a tendency to be more electrically active and to have longer durations than the whole convective system sample.

    5. BIBLIOGRAPHIC REFERENCES CUMMINS, K., MURPHY, M., BARDO, E., HISCOX, W., PYLE, R. and A. PIFER, 1998: A combined TOA/MDF technology upgrade of the U.S. national lighting detection network. J. Geophys. Res., vol. 103(D8), pp 9035-9044. CRANE, R.K., 1979: Automatic cell detection and tracking. IEEE Transactions on Geoscience Electronics. vol. GE-17, No 4, pp 250-262. KOFFI, E. N., SENESI, S. and C. MOREL, 2001: Early convective development study using combined satellite and lightning data. Proc. The 2001 Meteorological Satellite Data Users Conference, pp. 669 676, Eumetsat and The Turkish State Meteorological service, Antalya, Turkey. MOREL, C. and S. SENESI, 2002: A climatology of mesoscale convective systems over Europe using satellite infrared imagery. I: Methodology. Q. J. R. Metorol. Soc.. vol. 128, pp 1953-1971. MOREL, C., SENESI, S. and F. AUTONES, 2002 : Building upon SAF-NWC products: Use of the Rapid Developing Thunderstorms (RDT) Product in Mto-France nowcasting tools. Proc. The 2002 Meteorological Satellite Data Users' Conference, pp 248-255, Eumetsat and Met Eireann, Dublin, Ireland. MOREL, C., SENESI, S., AUTONES, F. and L. LABATUT, 2000 : The Rapid Developing Thunderstorms (RDT) Product of the Nowcasting SAF. Prototyping activities and quality assessment using GOES images. Proc. The 2000 Meteorological Satellite Data Users' Conference, pp 698-705, Eumetsat and CNR, Bologna, Italy. TZANOS, D., MOREL, C. and S. SENESI, 2001: Report on WP262340: validation of a new detection algorithm. Technical note of the Nowcasting SAF. TZANOS, D., MOREL, C. and S. SENESI, 2002: Scientific Report on WP23/51.20: Introduction of a spatial tolerance in the tracking of small cloud systems. Technical note of the Nowcasting SAF.