An Agent Network for Microburst Detection - Semantic Scholar

1 downloads 0 Views 217KB Size Report
An Agent Network for Microburst Detection. Sandy Dance. Australian Bureau of Meteorology Research Centre. 150 Londale St. Melbourne, VIC 3000, Australia.
An Agent Network for Microburst Detection Sandy Dance

Australian Bureau of Meteorology Research Centre 150 Londale St Melbourne, VIC 3000, Australia [email protected]

Abstract Meteorological data provides a rich source of imagery which is well suited to the growing eld of agent networks based image processing. Here we discuss one such system which analyses Doppler and re ectivity radar volumes to detect microbursts (windshear events). The system employs a network of agents, each performing a subtask of the detection, with message passing between agents. The various agents detect re ectivity cores, high divergent shear zones, microbursts and track microbursts through time. It is implemented in the Java language, and is successfully detecting microbursts in data from the Kurnell radar near Sydney.

1. Introduction Meteorological data involves many datatypes ranging from weather station output of rainfall, etc, through radar images, lightning data to satellite images. These data are ingested by the various national weather services at the rate of gigabytes per day, and hence the pressing need to automate as much as possible their processing and interpretation. Here at the Australian Bureau of Meteorology Research Centre (BMRC) we are involved in a project to develop a system to do just that. One approach we are exploring is the use of agent networks. The work described in this paper is an application of this approach to microburst detection from radar data. In the rst section, we describe what a microburst is, and how they can be recognised. In the next section we talk about agent networks in general, and how their use is relevant to the current application. This is followed by a description of the architecture of agents we are employing here. We then describe in more detail the actual agents used for microburst detection. In the penultimate section we present the results of this work.

2. Microbursts A (wet) microburst [7] [17] [9] (Figure 1) is a mass of cold, wet air that drops out of cumulus clouds due to its density. When it hits the ground, it spreads out potentially forming wind shear. Both the descending air and the shear can be dangerous for aircraft near airports. For this reason, an automatic system to detect microbursts would be extremely useful. An earlier system, MAD, that performs this task used fuzzy logic on each pixel of the radar data to generate an 'interest' map [1]. This approach precludes the use of pattern matching (ie, detecting shapes) until the nal step of interpreting the interest map. In our approach, such processing is straightforward, although not performed in the system described here.

2.1. Microburst Signatures There are a number of signatures [14] attendant on microbursts that can be detected with radar. However, not all signatures are necessarily present simultaneously in the data, thus the need for an uncertainty calculus in the detector. The most reliable indicator is the divergent air ow formed on the ground by the microburst. Since Doppler radar only detects radial velocity, this shows up as a pair of \bulls eye" patterns of radial in ow and out ow symmetrically placed around the microburst axis. Of course, due to prevailing wind elds, these \bulls eyes" need not be symmetric. As the microburst drops from the cloud, mass continuity implies some convergent ow near the cloud base or in the cloud. This will also form a pair of velocity \bulls eyes", this time at cloud level, but weaker. However, this signature may not be detectable due to the complexity of wind elds in clouds. The microburst itself consists of cold air and rain, and thus is detectable by radar as a region of enhanced

re ectivity (a \re ectivity core"). Moreover, this core is descending, so if the radar is sampling at sucient frequency, say every 2{5 minutes, the core could be detectable in more than one frame. If the core in the two frames could be matched, and the resultant velocity fell within bounds, then this would provide a means of predicting microbursts. Microbursts are produced by convective cells (thunderstorms or showers) which in turn occur in areas of atmospheric instability. If such instability is known from the meteorological context, or other microbusts have already been detected in the area, these facts can be fed into the detector to enhance its likelihood of detection.

A

B

3. Agent Networks in Image Processing The use of agent networks in image processing is a relatively recent phenomena, spurred on partly by the increasing complexity required of the task. It does, however, have precedents, particularly in the ideas of Hewitt [8], Minsky [13] and Schank [15]. The 1990s was really the decade when agent networks came to the fore, with several workers involved in their application to images [11, 2, 4, 3]. Agent networks are especially appropriate in real time image processing, and when the processing is attempting to emulate the cognitive abilities of humans. In support of this, recently cognitive scientists have been proposing models for human vision (and cognition in general) that resemble agent networks [6, 10, 12]. Another argument for the use of agent networks is that it has been found by practitioners that designing and implementing systems using this approach is easier than the traditional single threaded programming paradigm. Here there is a sort of hierarchy of approaches, starting with assembler language programming, compiled language programming like Fortran, block structured language programming like Pascal and C, object oriented programming as in C++ and Java, and agent programming. In this hierarchy, there is increasing 'chunking' (or modularisation) of the problem as one ascends. It should be noted that the work described below is experimental, where the agent approach is being trialed with a fairly simple problem. If successful we intend to take this approach much further, applying it to our general \forecast database" in which all meteorological datatypes will be processed at di erent levels and different timescales and from di erent user perspectives to provide our 'clients' with a variety of forecast 'products'.

4. Agent Network Architecture C

D

Figure 1. Life of a typical microburst. In A, the microburst is emerging from the cloud, in B it is drawing in air from the upper levels. In C the burst has reached the ground, producing divergent shear, and in D has also produced a gust front.

According to Wooldridge and Jennings [18], an agent system has the following properties:  autonomy: agents operate without intervention,  social ability: agents interact with other agents,  reactivity: agents perceive their environment, and respond in a timely manner to it,  pro-activeness: agents exhibit goal-directed behaviour by taking the initiative.

In our system, agents are constructed from Java threaded objects (autonomy) that communicate with each other through a message system based on queues attached to each object instance (social ability). Their reactivity corresponds to a \bottom-up" or \data driven" mode of operation, in which agents react to changes in data derived from real-time weather observations (ie, radar), or messages from downstream agents, and inform other agents of these events. Proactiveness (\top-down" mode) is manifested by agents that are actively seeking to detect interesting weather phenomena (such as the micro-burst detector described below), by sending enquiry messages to other agents. Agents typically have various mixes of these modes. For instance, in our system several agents will react to a message, and respond by then seeking information elsewhere in the system.

5. Microburst Detector Figure 2 shows the set of agents for detecting microbursts. Each agent runs autonomously at its task. Upon detection they pass the coordinates and characteristics of the detected signature together with a con dence measure onto a higher agent. Temporal continuity is an important constraint, so while some agent's interest is limited to a single radar image (ie, CoreDetector), other agents are concerned with correlating signatures from a sequence of radar images (ie, BurstTracks). Because all signatures of a microburst are not necessarily present, and there is usually noise and ground clutter in the signal, the system uses DempsterShafer [16][5] evidential reasoning to put the signatures together. A microburst is detected if the support exceeds a threshold.

5.1. Read input radar data This agent, labelled ReadData in Figure 2, simply has the job of reading into the agent system the next available radar le, and passing components of it to other interested agents. Each dataset contains re ectivity and radial velocity (Doppler) data for all azimuth angles, and 19 elevation angles (scans). The full volume set of re ectivity data is sent on to the core detector, and the lowest level velocity data is sent to the shear detector.

5.2. Core detector agent This detector, labeled CoreDetector in Figure 2, nds volumes of connected high re ectivity in the radar

BurstTracks Display MBDetector meteorological context

ShearDetector

CoreDetector

radar

ReadData

Figure 2. The set of agents for detecting a microburst.

data. This is performed by thresholding the re ectivity data at 45dBZ producing a binary (2 valued) volume (ie, stacked set of binary images). This is in turn passed through a labelling routine which labels connected regions using 6-connectedness (boxels are labeled with their neighbours if joined through any of their 6 neighbours in 3 dimensions), and using a region joining algorithm based on trees of labels (two regions are joined by creating a new common root for them both). If a labeled region exceeds a size threshold (expressed in cubic kilometers) and a height threshold (in kilometers), then it is passed on to the microburst detector agent, together with details of its size, maximum re ectivity and maximum height. The core detector agent also reacts to enquiry messages from the microburst detector (which is responding to the shear detector). These messages are concerned with con rming hypotheses that a microburst exists based on a high shear event. If this event is associated with a region of high re ectivity over a threshold volume and height (less than the thresholds above since this an enquiry), then the message is con rmed.

5.3. Shear detector agent This agent, labelled ShearDetector in Figure 2, is sent the velocity data for the lowest level scan by ReadData, and is sent queries from the microburst detector for con rmation of shear. This agent searches for highly divergent shear regions in the data by convolving with a di erentiating kernel [1]. The gradient is returned as the shear for each point (expressed in

meter=second=meter, or seconds?1 ), regions above a

threshold shear are then labelled using a similar technique to Section 5.2 (in this case using 2-dimensional 4-connectedness), and labelled regions over a threshold area stored locally in a database and also sent on to the microburst detector giving the location and other details of the potential burst. The enquiry query from the microburst detector contains the location of a re ectivity core. The stored high shear regions are then scanned for a match between the core and the shear (based on location), and potential candidates are returned to the microburst detector.

5.4. Microburst Detector Agent This detector, labeled MBDetector in Figure 2, receives messages from the core detector and the shear detector, giving it the location and other parameters of possible microbursts. Upon receiving a core (shear) detector message, this agent sends an enquiry to the shear (core) detector agent with the potential burst's location. If it receives an armative reply, and the combined con dence measure (computed using DempsterShafer [16][5] evidential reasoning), which includes data from the meteorological context like atmospheric stability, and the presence of thunderstorms in the area, exceeds a threshold, then it sends messages to the display agent (which shows the burst as a symbol on images of the re ectivity and velocity elds), and to the microburst tracking agent. The center of a re ectivity core can be somewhat o set from the center of the corresponding high shear region, due to a couple of factors (Figure 3). The core's center can be several kilometers in altitude and it can descend at an angle, resulting in a shear zone some kilometers horizontally o set. The other reason is due to the fact that the radar takes approximately 5 minutes to complete a volume scan, from low elevation to high. In a high wind, the weather system can move several kilometers between the lowest scan and the highest. For this reason, to identify a core with a shear zone, we have allowed a maximum separation of 8 kilometers in our system. In this architecture, it is possible that the one microburst event could be found from two pathways, ie, shear detected rst and core detected rst. For this reason, microburst events are identi ed by their location (and time). Thus if two events are separated by less than the above distance threshold of 8 kilometers, then they are identi ed and reported as the one event. Note that the threshold in some cases is higher to start a detection than to con rm a detection, ie, if the

core

shearzone radar

A

B

C

Figure 3. Microburst-core offset is due to: AB cores can descend at an angle, A-C in the 5 minutes it takes a radar to scan from lowest to highest elevation, the storm system can move several kilometers. Situation at start of scan shown in solid lines, end of scan shown in dashes.

shear region is above 9 km2 then a detection is started, but to con rm an enquiry the threshold is 4 km2 . This is to allow for cases where there is such strong evidence for a burst from one source that the con rming evidence can be correspondingly weaker. The con dence measure for the detection as a whole is calculated from:  shear: threshold 0:002seconds?1,  shear area: threshold 9km2 for detection, 4km2 for con rmation,  core re ectivity: threshold 45dBZ,  core volume: threshold 10km3 for detection, 2:5km3 for con rmation,  core max height: threshold 4km,  distance core center to shear center: threshold 8km,  atmospheric instability (set detection level according to weather, currently input by hand). Note that the various thresholds above have been set through a combination of empiricism and consultation with meteorologists at BMRC.

5.5. microburst tracking agent This detector, labeled BurstTracks in Figure 2, receives messages from the microburst detector agent with details about each microburst found. This agent's job is to nd possible trajectories connecting bursts in

temporally contiguous radar volumes that satisfy continuity criteria reasonably well. This agent nds tracks by nding the potential velocity between the current burst and all bursts in the previous volume. If this velocity is less than a threshold (set at 40 meter/second), then the potential track is then extended back to the next previous volume, where a predicted burst location is calculated. If this lies within a threshold distance from a real burst, this track is retained. The track with the best con dence measure as computed from the thresholds is chosen. This track is then sent to the display agent. Unfortunately, at present, the radar is sending volume scans at 10 minute intervals, which is not frequent enough to tie microbursts together reliably. We hope soon to have the radar running every 5 minutes, which will facilitate tracking of bursts, not only horizontally, but vertically as well.

6. Experimental Results

Figure 5. Doppler velocity data (original in colour) corresponding to Figure 4, showing microbursts outlined with circles, old bursts in grey, tracks shown between centers. This image is in polar projection.

The output from the runs was compared with an analysis by a meteorologist of the same radar volumes, who detected 8 \events" altogether. To compare the human analysis with the agent system output, a hit was recorded if the human detected a microburst within 8 kilometers of an automatic detection. Our system scored 8 hits and 2 false alarms, with no missed detections. Following Albo [1], the commonly used statistical measures for algorithm performance, namely probability of detection (POD), false alarm rate (FAR) and critical success index (CSI) are de ned as follows: Figure 4. Reflectivity data (original in colour). Kurnell radar, showing microbursts outlined with circles, old bursts in grey, tracks shown between centers. This image is in polar projection.

The system described above was run on a series of radar volumes obtained from the Kurnell radar situated next to Botany Bay in Sydney. This radar is an Enterprise Electronics Corporation WSR81C-12 with a peak power of 500kW at a wavelength of 5.3cm, half power beam width 0.9 degrees, nominal range 150 km, and range resolution 250 meters. There were images obtained from two di erent days, the rst day contains two series of 3 and 4 volumes, the second day consisted of two volumes, 9 volumes in all. Contiguous volumes are 10 minutes apart.

X fd=(X fd + X h) X X POD = h= e X X X X CSI = h=( fd + h + md) FAR =

(1) (2) (3)

where h is hits, fd is false detections, e is events and md is missed detections.

Although our number of events is too low for meaningful statistical analysis, nevertheless, for comparison, the results for our system were: POD = 1.0, FAR = 0.2 and CSI = 0.8. This compares favourably with the system reported by Albo [1], which achieved POD of between 0.58 and 0.76 (depending on the radar con guration), FAR between 0.09 and 0.20, and CSI between 0.54 and 0.67.

7. Conclusion In this paper we have demonstrated the feasibility of using agent technology in the meteorological context, especially for detecting microbursts. We set out to emulate the processing that meteorologists perform in detecting microbursts, and succeeded to some degree in that objective. For emulating human abilities, the agent technology was found to be reasonably easy to modify and tune, although there is obviously much more that needs to be done to bring the system up to human standard. More pattern recognition at all levels is required. (Of course, since we de ned \ground truth" as the human interpretation, comparing the ef ciency of the two becomes moot. It may be that the automatic detector is closer to \truth" than the human.) Our agent network also proved comparable in eciency to a more mature system performing the same task, although less resources were expended on building our system. Of course, we need to run the system on more data to make this comparison more robust. Obviously, there are many improvements and re nements that can be made to the system described. For instance, we could add an agent to detect convergent shear at cloud level, and an agent to process the meteorological context to detect atmospheric instability. In the future we hope to expand the scope of the project to take in the recognition and tracking of other weather phenomena, for instance gust fronts and thunderstorms. Ultimately we wish to employ the agent technology for our \forecast database" which will be used to oversee the entire process from observation to forecast.

References [1] D. Albo. Microburst detection using fuzzy logic. Technical report, National Center for Atmospheric Research, Boulder, Colorado, 1994. [2] O. Baujard and C. Garbay. KISS - a multiagent segmentation system. Optical Engineering, 32(6):1235{ 1249, june 1993. [3] O. Boissier and Y. Demazeau. MAVI:a multi-agent system for visual integration. In IEEE Conference on Multisensor Fusion and integration for intelligent systems, Las Vegas, Nevada, October 1994. [4] S. Dance. SOO-PIN: Picture interpretation networks, a distributed agent-based approach. In T. Caelli and W. F. Bischof, editors, Machine Learning and Image Interpretation, chapter 6, pages 225{254. Plenum Press, New York, 1997. [5] A. P. Dempster. A generalization of Bayesian inference. Journal of the Royal Statistical Society, 30:205{ 247, 1968.

[6] D. Dennett. Consciousness explained. Little, Brown and Co., Boston, 1991. [7] T. T. Fujita. The Downburst. SMRP, Dept. of the Geophysical Sciences, Uni. of Chicago, Chicago, Illinois, 1985. [8] C. Hewitt. How to use what you know. In IJCAI-75, pages 189{198, Tbilisi, Georgia, September 1975. [9] M. R. Hjelmfelt. Structure and life cycle of microburst out ows observed in Colorado. Journal of Applied Meteorology, 27:900, 1988. [10] D. Hofstadter. Fluid Concepts and Creative Analogies: Computer Models of the Fundamental Mechanisms of Thought. Penguin Books, Ringwood, Victoria, Australia, 1998. [11] F. V. Jensen, H. I. Christensen, and J. Nielsen. Bayesian methods for interpretation and control in multi-agent vision systems. In Proceedings from Applications of Arti cial Intelligence X, Orlando, Florida, 1992. [12] G. Lako . Women, Fire and Dangerous Things. University of Chicago Press, 1987. [13] M. Minsky. A framework for representing knowledge. In P. H. Winston, editor, The psychology of computer vision. McGraw-Hill, New York, 1975. [14] R. Potts. Microburst precursors observed with Doppler radar. In 24th Conference on Radar Meteorology Preprints, pages 158{162, Tallahassee, Florida, March 1989. American Meteorological Society. [15] R. Schank and R. Abelson. Scripts, Plans, Goals and Understanding. Erlbaum, Hillsdale, N.J, 1977. [16] G. Shafer. A Mathematical Theory of evidence. Princeton University Press, 1976. [17] J. W. Wilson, R. D. Roberts, C. Kessinger, and J. McCarthy. Microburst wind structure and evaluation of Doppler radar for airport wind shear detection. Journal of Climate and Applied Meteorology, 23:898, 1984. [18] M. Wooldridge and N. R. Jennings. Intelligent agents: Theory and practice. Knowledge Engineering Review, 10(2):115{152, 1995.

Suggest Documents