The Australian Bureau of Meteorology has developed a new tool called the Thunderstorm Interactive Forecast System (TIFS; formerly known as ThunderBox) for interactively producing finished severe weather warnings and other forecasts from thunderstorm tracks, automatically diagnosed from radar data. TIFS is designed to apply recent advances in radar-based thunderstorm cell detection and tracking techniques to the efficient production of operational forecasts and warnings. The system ingests automated thunderstorm cell detections and tracks, allows graphical editing by forecasters, and produces graphical and text products from the edited data. The text generator uses a shallow, domain-specific approach. The graphical products include a map of areas that have been affected by storms, and are forecast to be affected by storms, as well as meteograms for selected locations.
1. Introduction and background
There is an unmet demand for weather warning products tailored to the particular needs of a range of recipients. The World Weather Research Programme (WWRP), Sydney 2000 Forecast Demonstration Project (FDP) impacts study (Anderson-Berry et al. 2004 in this issue) has revealed a need for the following: graphical warnings with more geographic detail than can be easily provided in text products, a variety of different graphical presentations, site-specific warnings, and displays simple enough for people without specialized meteorological training to interpret.
The impacts study also revealed that for some clients an accurate picture of the previous path, extent, and severity of storms is just as important or even more important than a forecast for the future location of the storm.
Advances in radar interpretation systems have increased the range and quality of guidance information available to severe weather forecasters. These systems help forecasters make better storm warning decisions, but there has been a weak link between these decision support systems and the preparation of the warning products.
Severe thunderstorm warnings currently generated by forecasters in the Australian Bureau of Meteorology (BoM) are produced manually with a text editor. Forecasters diagnose storm characteristics by subjectively integrating information from 10-min radar volumes, surface analyses, upper-air soundings, numerical model output, hourly satellite imagery, and storm spotter reports. This manual integration of guidance information into manually prepared warning products also occurs in many other weather services. It is a simple methodology for composing severe weather warnings, but it does have limitations. Compared to the achievable lead times for severe thunderstorm warnings, manually typing information into the text warning forms is a slow process. Each product the forecaster must prepare adds to the forecaster's workload when time is most critical. Under extreme time pressure, the manual translation of detailed information from decision support systems into the finished warnings can lead to mistakes, omissions, and inconsistencies in the finished warning products.
This paper describes the Thunderstorm Interactive Forecast System (TIFS; formerly known as ThunderBox) developed by the BoM. TIFS presents forecasters with a range of guidance information from some of the world's most advanced thunderstorm decision support systems in a common format. These systems include the Thunderstorm Identification, Tracking, Analysis, and Nowcasting system (TITAN; Dixon and Weiner 1993); the Warning Decision Support System (WDSS; Eilts 1997); the Canadian Radar Display Support system (CARDS; Lapczak et al. 1999); and the Auto-nowcaster (ANC; Wilson et al. 1998). TIFS allows forecasters to graphically select and edit the guidance. TIFS saves these forecast decisions and automatically generates a range of graphical and text warning products. It allows the generation of many tailored products and different representations of forecasts. It maximizes the reuse of forecast decisions as components of different products. Each product is just a different rendering of the same information in the database of forecaster decisions, so consistency between products is guaranteed. Similar forecast production philosophies have been outlined by authors in many other national weather services (summarized in Ruth 2000).
TIFS was used to produce warning products sent to clients during the Sydney 2000 FDP of the WWRP. The software was also used to graphically render processed radar data from participating FDP systems into a common graphics format for presentation on Web pages to forecasters in the Sydney Forecast Office.
2. TIFS data display and graphical editor
TIFS ingests guidance data in ASCII text format from multiple radar processing software packages and displays it in an interactive graphical environment (Fig. 1). This wide range of guidance information has been organized according to data type, rather than by originating system. The following are the data-type categories:
Tracks—These are thunderstorm cell tracks that include the storm history and a (linear) forecast track extending up to 60 min into the future.
Features—These are point detections of significant thunderstorm features. Those currently displayed by TIFS are mesocyclones, bounded weak echo regions, downburst and hail detections, tornado vortex signatures, as well as surface observations and lightning detections.
Winds—These are currently used only for gridded wind fields such as the radar-diagnosed wind from the Auto-nowcaster, but they are suitable for other gridded data types.
The forecast track is linear, assuming that there is currently insufficient forecast skill to justify the extra complexity of handling more sophisticated forecast tracks. Two thunderstorm tracking systems are currently used to feed TIFS with track guidance data. These are TITAN and the Storm Cell Identification and Tracking algorithm (SCIT; Johnson et al. 1998) from the National Severe Storms Laboratory's WDSS.
TIFS gives the forecaster direct access to cell and track statistics and feature detections (i.e., detections of mesocyclones, downbursts, and hail) to aid decision making. These are presented in tabular form in a separate window (Fig. 2). The forecaster can graphically edit the storm tracks, and select which tracks and cells should appear on the finished warning. The forecaster can change the speed and direction of the storm's forecast motion, as well as the storm size, shape, location, and intensity. Forecasters can also add new cells not captured by tracking systems. These graphical edits form the forecast and warning decisions and are stored as the TIFS “forecast database.” The original data are also stored to facilitate more comprehensive verification and so that forecaster corrections can be reset to their original values.
The forecast database is rendered as complementary graphical, text (Fig. 1), and site-specific meteogram (Fig. 3) products. The meteograms graphically represent the impact of storms cells at specified locations. Although they do not contain any new information, meteograms can make the storm information much easier for some clients to incorporate into their own decision making processes.
As the forecaster graphically edits the data, the text and meteogram products automatically and immediately update, reflecting the changes the forecaster has made. This immediate updating is an unusual feature of TIFS, made possible by the speed of the text generator. It is important because it lets forecasters fine-tune the output products by using the higher-level process of editing the information underlying them, rather than being tempted to edit the output products themselves. The importance of this seemingly subtle distinction is that by editing the underlying information, the consistency of the set of output products is preserved. Also, the underlying information is stored numerically in a database and is easy to verify. Text and graphical output products are much more difficult to verify and may need to be individually, and perhaps subjectively, verified by hand.
When editing is complete, the graphical and text products are disseminated to clients. The output products are all consistent with each other because they are just different views of the same forecast database, not independently generated products.
To simplify the graphical warning products and make them more easily understood by clients with little weather knowledge, storms are represented by tilted ellipses, rather than the more complicated polygon boundaries used by some other systems (e.g., TITAN, Auto-nowcaster). TITAN produces tilted ellipses as a standard output, in addition to polygon boundaries. SCIT does not produce tilted ellipses, but does output cell centroid, cell volume, and cell height, from which TIFS constructs an (always circular) ellipse.
These tilted ellipses can be defined by only six parameters: major axis, minor axis, and orientation; storm-top height, represented as a color or gray shade; and speed and direction of movement, represented as elliptical arcs drawn on the side of the storm cell toward which it is moving, with spacing between the arcs proportional to storm speed. This representation of movement is intuitive, clear, and minimizes clutter. It gives the impression of motion on a static display. These features are illustrated in Fig. 4.
Complicated algorithms operating on real data, such as those used for storm tracking, do not always perform perfectly. Detection thresholds may be set too high or too low for a given situation. Storm cells are sometimes misassociated between temporally adjacent radar images, resulting in anomalous cell motion vectors. TIFS copes with occasional cell misassociations or newly detected cells with incorrect motion vectors by allowing the forecaster to make the necessary corrections to the representation of the storm in the warning product. Rather than attempt perfect automatic tracking, the approach taken by TIFS is to use the forecaster's pattern matching and other meteorological skills to fine-tune automated guidance. TIFS allows forecasters access to several tracking algorithms from which they can choose the most appropriate in each situation to be the basis of the forecast. For example, if a forecaster using TIFS is focusing on the damage-producing core of a severe storm, then SCIT, which tracks cell cores, may be the best tracking algorithm to use. On the other hand, if the forecaster is producing a lightning alert or more generalized warning, then TITAN may be more appropriate. The forecaster's skills are complemented by machine-diagnosed cell speed calculations and geographic associations that forecasters do more slowly and sometimes less accurately.
Figure 1 shows the TIFS user interface running on a TITAN representation of a tornadic storm just west of Sydney, Australia, on 3 November 2000. The control panel and context-sensitive help are at the top left. The graphical representation of the storm is at the top right and the automatically generated warning text is below. The forecaster can graphically edit the representation of the storm using mouse clicks and drags. Alternatively, the storm representation can be edited via the keyboard on a spreadsheet-style data control form (Fig. 2), similar in appearance to the cell table in WDSS.
Figure 5 shows the same storms, using the SCIT tracker in WDSS as the data source. Note that SCIT has focused attention on the storm cores and, in this case, has better captured the left turn of the easternmost cell as well as more accurately capturing the direction of motion of the western cell.
3. Text forecasts and warnings
Computer-worded weather forecasts can be produced by systems ranging in complexity from simple template filling techniques, known as “shallow” text generators, through to complex artificial intelligence techniques (Driedger et al. 2000). Australian severe storm warnings are simple in their lexical structure, as are most severe weather warnings produced around the world. Consequently, it was considered that a shallow, domain-specific text generator would be sufficient to produce high-quality text forecasts and warnings from the TIFS forecast database. Indeed, real examples of TIFS text warnings are very similar in style and readability to their manually generated counterparts, but the TIFS products contain extra geographic detail that human forecasters do not have sufficient time to include in extreme time pressure warning situations. A TIFS-generated text warning is presented in Fig. 6, corresponding with the meteorological situation depicted graphically in Fig. 1.
The generation of forecast text begins with parsing the edited storm cell tracks from the TIFS database. Observed and forecast cell locations are compared with a geographic database and lists of local government areas, suburbs, and site names affected by storms are compiled. These names are included in text phrases along with time information also extracted from the storm cell database. Other phrases in the warning message are simply constructed from information in the storm database or extracted from a small library of standard warning phrases. A flowchart illustrating the process is presented as Fig. 7.
The text generator is simple and robust but produces text of appropriate sophistication for these simple warning messages. The text is always consistent in style and layout and is very clear. The simple and efficient production of text by TIFS is also very fast. Indeed it is fast enough to be running continuously and updated as forecasters graphically edit the storm warning.
Another application for the text generator is for it to be run within TIFS without human interaction. As each new radar volume becomes available, TIFS generates intelligent alerts for forecasters. So, rather than alerting on some simple radar thresholds, forecasters can receive an alert containing information about storm location, severity impact area, and timing. As TIFS is written in Java it can easily generate e-mail messages to the Bureau of Meteorology's Australian Integrated Forecast System (AIFS; Kelly and Gigliotti 1997), alerting subsystem or directly to the forecasting staff.
4. TIFS and the WWRP
TIFS played a pivotal role in the Sydney 2000 Forecast Demonstration Project (S2000 FDP) of the WWRP, by rendering, in a common graphics format, the output of many of the participating systems, and also by giving forecasters a mechanism for turning those outputs into severe weather warnings for dissemination to clients.
TIFS was designed to read a number of input data formats. Most useful of these has been the facility to read an eXtended Markup Language (XML) like self-describing text format called Aifs eXchange Format (AXF) used extensively at the BoM.
All participating systems in the S2000 FDP produced output in the AXF format. This enabled TIFS to read and display observed and forecast storm locations from WDSS, the ANC, and the BoM's TITAN implementation, and storm feature detections such as mesocyclones, microburst signatures, and hail signatures from CARDS and WDSS. Images generated by TIFS were displayed on interactive Web pages, used by forecasters in the Sydney Regional Forecasting Center of the BoM, the Sydney Airport Meteorological Unit, the Sydney Olympic Sailing Weather Office, and at the weather briefing office inside the headquarters of the Sydney Organizing Committee for the Olympic Games (SOCOG).
An example of TIFS rendering of the guidance information generated by the ANC on 3 November 2000 is shown in Fig. 8. This image corresponds in time with the TIFS rendering of TITAN and WDSS guidance information shown in Figs. 1 and 4. The polygons represent thunderstorm cell locations and forecast locations. The lines represent observed and forecast locations of convergence lines. Both are color coded for forecast lead time. The roughly north–south-oriented and stationary line is the sea-breeze front. The roughly east–west line, moving toward the north, is the outflow boundary of the main cell.
In this example, the ANC has forecast that the main cell (southwest of Olympic Park at 0455 UTC) will move to the northeast and decrease in size. This contrasts with both the TITAN and WDSS forecasts of more leftward movement to the north-northeast, which turned out to be more correct in this case. However, the ANC has also forecast cell development near the collision point of the sea-breeze front and the outflow boundary of the main cell. Rather than heading out to sea, the main cell intensified near the boundary collision zone where the ANC had forecast cell growth and tracked to the north-northeast, along the sea-breeze convergence zone. In this case, the cell tracking in TITAN and WDSS did a good job at predicting cell motion, and the ANC may have captured some of the mechanisms that lead to that cell motion.
TIFS rendering of the guidance information generated by CARDS at the same time as Figs. 1, 4, and 8 is shown in Fig. 9. The output of the CARDS hail, mesocyclone, and microburst detection subsystems are displayed using symbols color coded for intensity against a background of cell locations derived from TITAN. CARDS has analyzed 6.1-cm hail and a moderate strength mesocyclone in the main cell.
The main cell depicted in Figs. 1, 4, 8, and 9 produced large hail and a weak tornado over the western suburbs of Sydney. Forecasters used TIFS to view the guidance provided by a number of advanced nowcast systems to make decisions about cell development and movement, and quickly turn those decisions into a suite of consistent graphical and text products that were sent to SOCOG, emergency services, the aviation industry, and a tourism operator as detailed in the FDP impacts study (Anderson-Berry et al. 2004). The graphical outputs of TIFS use a deliberately simple “cartoon” style to make the warning products accessible to a wide range of clients who may have limited meteorological knowledge.
Twenty-three TIFS warnings were issued on 5 days during the FDP. They were generally based on TITAN cell tracks, with forecasters filtering out about 2/3 of the tracks that were considered unimportant or incorrect. This filtering resulted in a marked improvement of the mean errors of the forecast cell position on TIFS warnings (11.4 km) compared to the raw automatically generated TITAN tracks (19.6 km), illustrated in Fig. 10. The most dramatic improvements were for longer lead times and more intense cells. Forecasters rarely modified the speed and direction of cell tracks that they decided to leave in the warnings. On the few occasions when they did so, the modified tracks showed higher mean cell position error (8.50 km) than they would have if left unmodified (7.15 km). One could conclude that for strong, long-lived cells, automated linear extrapolation is hard to beat. On the other hand, the result could be due more to forecasters using the system operationally for the first time and to the small sample size. Detailed verification of the TIFS and other FDP products are contained in Ebert et al. (2004 in this issue).
The TIFS software described here is designed to maximize the utility of TITAN, WDSS, and other radar processing software by integrating information from them directly into the forecast production process. The design goal for TIFS is to streamline the forecast production process to allow forecasters to produce accurate, reliable, informative, and specific graphical and text forecasts, in several variants tailored for different clients, in the same or less time than forecasters currently take to produce a text warning. TIFS was originally developed to support the WWRP Sydney 2000 FDP. It has since been extended to support a wide range of thunderstorm warning products and has been installed in the Sydney Regional Office of the Bureau of Meteorology for operational use. Impact (Anderson-Berry et al. 2004) and verification (Ebert et al. 2004 in this issue) studies for the Sydney 2000 FDP showed that TIFS was well accepted by forecasters, resulted in warnings superior to those that could have been provided from unedited TITAN storm tracks, and showed that TIFS products were considered useful by clients.
The current version of TIFS is written in Java (version 1.1.8), which makes the code highly portable between computer systems and allows it to run as an applet on some Web browsers on some systems. The price paid for this unusual degree of portability is the relatively simple graphics application program interface (API) available under Java 1.x. As support for Java 2 becomes more widespread in the near future, the system will be able to use the more sophisticated graphics APIs available with Java 2D, 3D, and VisAD (Hibbard et al. 1997; additional information available online at http://www.ssecowisc.edu/~billh/visad.html).
TIFS is a realization of the streamlined forecast production philosophies being recommended by many developers (Ruth 2000), applied to the production of very short lead time warnings.
The author would like to acknowledge Rod Potts for his untiring efforts to keep TIFS supplied with real-time TITAN storm tracks, Andrew Treloar for his role in easing this new forecasting paradigm into operations at the Sydney office and for his helpful feedback on system design, and Tom Keenan and Jim Wilson for their encouragement.
Corresponding author address: John Bally, Bureau of Meteorology Research Center, GPO Box 1289K, Melbourne VIC 3000, Australia. Email: email@example.com