IBM Research | Weather Modelling Group

Lloyd A. Treinish
IBM Thomas J. Watson Research Center
Yorktown Heights, NY 10598
914-784-5038 (voice)
914-784-7667 (fax)
lloydt@us.ibm.com

Abstract
Efforts to create highly generic visualizations often fail when applied to non-research-oriented activities. Instead, specialized interfaces and tools matched to user goals and underlying visualization tasks are developed. To avoid the cost of addressing multiple requirements through independent development and training, the design of different visualization applications are matched to a set of tasks but built on top of a common framework with a similar approach to content. To promote high-level reuse of interface and content elements for each application, a general-purpose toolkit is employed. Such a package ordinarily would lack sufficient focus to be effective in operational efforts, but its direct facilities are hidden from the user. This approach is tested in detail by application to a demanding problem -- operational weather forecasting.

Visualization has matured sufficiently that the design of content can be driven more by user needs rather than the limitations of the enabling technology. Improvements in technology have often led to the implementation of very generalized systems as the preferred mechanism to address a diversity of visualization strategies from even a single data set. Such flexibility has been useful for research activities as well as application development. However, their inherent lack of focus makes them less suitable for direct application in environments with relatively fixed tasks or user goals. This is especially true in operational, mission-critical situations, where there is typically no identified need to master generalized interfaces, most of whose many facilities may be viewed as superfluous or merely distracting. There may also be a lack of resources to consider such generic tools compared to more specialized ones.
To overcome this apparent barrier, an understanding of user goals and how they map to visualization tasks is required. The effectiveness of a visualization must be defined by how well user goals are met. One commonly used decomposition implies three basic visualization tasks as justification for visualization:

In general, forecast meteorologists have a need to improve the accuracy of their local predictions. One solution is the usage of additional tools and data focused on the region of interest like mesoscale numerical modeling techniques and high-resolution radar observations. The introduction of such data challenges the typical operational visualization facilities and was an early motivation for this research. This work began with the project that enabled rapid generation and analysis of local weather simulations by the National Oceanic and Atmospheric Administration (NOAA) National Weather Service (NWS) at the 1996 Centennial Olympic Games [15]. It has been a collaborative effort with NOAA Forecast System Laboratory (FSL). A specific mesoscale model called the Regional Atmospheric Modelling System (RAMS), originally developed by Colorado State University was introduced [9]. To provide timely production, a parallelized version of RAMS was installed on a 30-node distributed memory supercomputer (IBM RS/6000 SP) at the NWS in Peachtree City, GA. To enable rapid presentation and analysis of weather simulations at different resolutions from this system, interactive three-dimensional visualization methods were introduced [17]. It provided a unique initial testbed for new visualization techniques, focused on meeting specific forecasting goals, yet affected by real, operational constraints. From that base, subsequent refinement of the visualization tasks and implementing of tools to address them proceeded.

Visualization in meteorology has a rich tradition and history, that predates computing when meteorologists drew contour maps of weather data by hand from observations to communicate a forecast. Consistent with the development of numerical models as among the earliest scientific uses of computer systems, the creation of computer-generated two-dimensional graphical representations of weather data was among the first applications of visualization. As a result, researchers in atmospheric sciences have been early adopters of modern three-dimensional visualization methods while operational applications like weather forecasting have focused on deployment of comprehensive, easy-to-use systems supporting two-dimensional visualization.

The majority of visualization systems today for meteorology are designed from the perspective of "one size fits all". (Systems in this context are turnkey software packages, not development environments that can be used to create such applications.) While there is variation among them, they individually provide one interface and one style of visualization independent of the task because they have been designed from the perspective of supporting a single class of users. Although such systems have been successful, there are user goals or operational efficiencies that are not addressed by the focused visualizations that they provide.

For mission-critical tasks like operational weather forecasting, improvements in speed and effectiveness have significant impacts. That is why weather agencies have invested in developing highly focused visualization tools. One major example is the Advanced Weather Interactive Processing System (AWIPS) deployed by the NWS, which provides two-dimensional visualizations via its D2D subsystem [8]. Conceptually similar tools are available from a plethora of other organizations worldwide.

This category of traditional weather visualization tools is termed Class I. It consists of conventional representations of selected meteorological fields for analysis tasks by forecasters with minimal direct (graphical) interaction at a specific "layer", either the ground or at a constant atmospheric pressure (isobaric level). Given a flat canvas for visualization design consistent with the display of a single layer, these tools can only show a few parameters simultaneously (e.g., overlaying wind as barbs or arrows, another scalar variable as line contours). Simplified versions of these representations have also been developed to support presentation to non-meteorologists, particularly for the media (e.g., [12]). Figure 1 is an example of a typical visualization produced by Class I tools. It shows a precipitation forecast for August 4, 1996 at 8 pm EDT generated by RAMS. In some cases, Class I may also provide simple film-loop-like animation.

Current sources of acquired and computed data imply the generation of large, three-dimensional volumes, for which improved facilities for visualization are required for timely assessment and utilization in forecasts. However, Class I techniques dominate in operational settings, even though their use can be burdensome with large data volumes. There are a few exceptions to this situation. Chief among them is Vis-5D developed at the University of Wisconsin. It has a fixed user interface with specific visualization tools typically tied to graphics hardware on Unix workstations, with support for a single class of data. The implementation focuses on manipulation of regularly gridded data, preferably compressed to byte precision to increase the speed of operation. This yields an highly interactive tool that maps well to many meteorological data sets [4], which is in use by several operational weather centers, primarily for analysis. However, for other forecasting tasks such as model assessment and dissemination, Vis-5D does not have the ideal interface or visualization content.

An independent effort by FSL provides three-dimensional visualization facilities for the Local Analysis and Prediction System (LAPS) as well as other data sources [14]. While it is used for operational analysis at FSL, it has the same limitation as Vis-5D for other tasks. To eliminate duplication of the capabilities of Vis-5D, this work has changed direction to build directly upon Vis-5D. The FSL efforts now concentrate on providing an interface consistent with other facilities (e.g., D2D) used primarily for analysis, based on evaluation of user preferences and tasks [7].

Fraunhofer Institut für Graphische Datenverarbeitung (FIGD) has taken a different approach. They have implemented independent systems that are highly focused for specific tasks yet designed to be operated by meteorologists. Their development has been in conjunction with the Deutscher Wetterdientst (DWD). The first was Triton, oriented toward generating two-dimensional visualizations for the non-meteorologist [12]. The second, TriVis, is based upon a related goal -- providing two- and three-dimensional visualizations for television broadcasts [13]. The third system, RASSIN, is designed to provide analysis facilities directly on the native grids of the meteorological data [6]. The latter two systems are in use by DWD.

To enable a set of design elements useful for a variety of tasks, visualization methods have been developed within a natural coordinate system to provide a context for three-dimensional display and interaction. They provide representations of the atmosphere consistent with the data source that are registered with ancillary or reference data (e.g., terrain and political boundary maps). This approach is one of correlative visualization, where each data set to be examined is processed independently and merging takes places at render time [16]. The design of the content and choice of coordinate system has been dictated by the particular task to support both conceptual and physical realizations.

Since color is a critical aspect of design, knowledge of human perception is applied via a rule-based advisory tool that is sensitive to the spatial frequency of data and the visualization task [10]. To eliminate complexity in the interface, this tool is not directly exposed to the user. Instead, it is employed in the design of specific visualization elements, which are integrated into the final composition provided to users. For example, relatively noisy data such as wind speed are primarily mapped into luminance, while relatively smoothly varying data such as temperature are primarily mapped into opposing saturation pairs to impart an isomorphic or continuous representation. For moisture-related data (e.g., humidity and precipitation), two colormaps are combined such that dry regions are mapped to brown ranging through yellow to green for modest values. At high levels, the data are mapped into blue, with decreasing luminance. When contouring is used to map the data onto a set of bands, a segmented colormap with perceived ordinality is applied. For discrete three-dimensional representations (e.g., cloud surfaces), uniform but complementary colors are chosen to minimize the effects of color mixing. For direct volume rendering of these data, the same hue is employed, but coupled with simultaneous mapping into luminance and opacity.

Several techniques are implemented for surface wind velocity, which are pseudo-colored by wind speed draped over a topographic surface. A continuous colormap is used ranging from a deep violet for very calm winds to white for the maximum speed specified. Vector arrows of fixed size are used to eliminate misleading motion cues during animation and to show gross atmospheric movement. In contrast, streamlines with directional arrows although visually more complex, are superior at capturing fronts, convergence zones, vortices, etc. On the other hand, waving flags that point in the direction of the wind have been effective to illustrate wind motion to the non-meteorologist. They can be either rigid or furled (i.e., straight at maximum speed and draped against a flag pole in the absence of wind).

The combination of these approaches provided a good base of techniques to present to forecasters allowing greater effort on development rather than progressive refinement. Subsequent iterations in the composition were relatively minor such as improvement of specific colormaps (e.g. base hue for wind, number of perceived segments for contouring) or the choice of visualization task for analysis (i.e., segmented vs. continuous realization). As a result, more effort was applied to strengthening the utility of the user interface.

The task decomposition leads to three other classes of visualization, each of which are described below. In each case, the user has a set of tools to design a visualization and interact with selected data. The available techniques are limited to focus on specific tasks, which are then supported by relatively simple interfaces:

Class II. Two-and 2-1/2-dimensional analysis

Class III. Three-dimensional browsing

Class IV. Three-dimensional analysis

Unlike FSL, the work herein considers a wider variety of user goals and underlying visualization tasks. FSL has focused primarily on interactive visualization for analysis. They have addressed the problem of user training and overall usability by developing an interface consistent with the two-dimensional visualization tools already in operation.

Although FIGD has developed task-specific visualization content and interfaces, their systems only share an underlying renderer as opposed to a common set of higher-level design and interface elements. As a result, a user of the FIGD systems that has multiple goals may need to utilize more than one of these tools. Since they present different user interfaces and design elements, additional effort in both development and training is required. The work at FIGD also differs in the identification of individual tasks, and thus, presents different composition and interaction.

Class II can be viewed as a superset of Class I by including enhancements into three dimensions and the ability to leverage modern workstation hardware. Its focus is for analysis by forecasters, particularly to support data comparison. Because the appearance of the visualizations may be complex, direct manipulation is provided.

As a result, up to five parameters may be visualized simultaneously. These two-dimensional variables may be any combination of surface or upper air layers from the same or different sources. They may be illustrated redundantly by applying multiple techniques (e.g., color and height). The variables and techniques can be independently selected interactively. An example is shown in Figure 2. Four scalar fields are visualized in the stereographic coordinates for December 5, 1995 at 11:00 pm EST from a forecast generated by RAMS. The primary features are precipitable water as the height of a shaded, deformed surface, pseudo-colored by temperature, and marked with the mean sea level pressure at discrete locations. Color-coded relative humidity contours at 10% intervals are overlaid along with streamlines of wind. The wind direction is indicated by arrows and the speed by color. The surface is also draped with local coastline (black), state boundaries (magenta) and river (blue) maps. For example, a correlation is visible between moderate levels of precipitable water and humidity with high temperature in south-central Georgia, and lower levels of precipitable water, wind and temperature along the coast.

In northern Georgia and southwestern North Carolina, there is a direct correlation between lower precipitable water and temperature with higher humidity. The effectiveness of this approach to examining multiple two-dimensional fields is further illustrated via time-based animation.

Class III enables forecasters to create qualitative three-dimensional representations for both interactive investigation and production of animation via browsing. The consumers may or may not be specialists but the interactive user is likely to be a meteorologist. The focus is placed on surface conditions and precipitation, which are of greatest interest for general forecasting. The visualizations are composed of a set of simplified qualitative techniques to utilize pattern recognition for general understanding as well as feature identification that serve two purposes: 1) gross assessment of the data and 2) source material potentially suitable for public dissemination to support the explanation of specific forecasts (e.g., media, World-Wide-Web, etc.). They are presented in a geographic coordinates consistent with the data source -- cartographically projected horizontally and terrain-following (i.e., true height) vertically. The techniques requires high-resolution data (temporally and spatially) to enable a coherent presentation. Figure 3 shows a result from such a browsing application, which enables interaction in geographic-altitude coordinates. The ability to create both time-based and key-frame (flyover) animations is available.

Figure 3 shows a three-dimensional representation of predicted cloud structure from RAMS as translucent, white isosurfaces of cloud water density at 10^-5 kg/kg for August 4, 1996 at 9:00 pm EDT. The cloud surfaces are registered with a terrain map overlaid with coastline (black) and state (white) boundary maps in a terrain-following, stereographic grid, where the cities of Atlanta and Savannah are marked. This representation can effectively show gross atmospheric motion and potential distribution of moisture. The terrain is pseudo-colored by total precipitation to indicate where and how much rainfall is predicted by the model, where heavy rainfall is shown as blue puddles. Translucent, cyan isosurfaces in the interior of the clouds are forecast radar reflectivities at a threshold of 25 dBz, approximating rain shafts. The correspondence between the rain shafts and the regions of relatively heavy precipitation is quite clear. The surface is also overlaid with vector arrows of surface wind velocity, color-coded by speed. The results were computed at 8 km horizontal resolution and are shown for the time during the closing ceremonies of the 1996 Centennial Olympic Games in Atlanta. The model, as illustrated through three-dimensional visualization, correctly predicted thunderstorm activity in the vicinity of Atlanta, but NOT over the city itself. This is quite evident through animation. Hence, forecasters were able to give Olympic officials an "all clear" for the Closing Ceremonies despite thunderstorms in the area. The visualization can be viewed in more detailed through interacting with a VRML representation of its geometry after simplification, or an image-based rendering via PanoramIX.

Class IV provides analysis, viewing, interrogation and interaction tools for standard data products designed for AWIPS. They are presented in geographic coordinates consistent with the data source -- cartographically projected horizontally but at standard pressure levels vertically. This class is similar to the visualization tasks addressed by the aforementioned Vis-5D and RASSIN packages, but with greater emphasis on direct manipulation and the introduction of new realization methods. Since these presentations can be visually complex even after application of complementary colormaps, facilities to interrogate and estimate data values are provided. The notion of a virtual met-station is introduced -- a realization of direct manipulation as a graphical analogue for a simulated atmosphere to the type of instrumentation that would be used to observe the real atmosphere. Figure 4 shows an example of a Class IV visualization, where the geographic-pressure coordinate system is annotated with an axes box and base maps.

Figure 4 illustrates data derived from the analysis of observations of the atmosphere by LAPS centered over San Jose on November 19, 1997 at 0900 GMT. A surface variable (total precipitation) has been selected for display as pseudo-color, which is overlaid on a topographic map. Rivers (blue) and coastlines (black) are draped on the surface. An upper air variable (relative humidity) has been selected for display via surface extraction. The surface at 90% is requested in translucent white as a representation of a cloud boundary. Another field (temperature) has been selected to show as a vertical slice, which is pseudo-color contoured. Any of the three-dimensional fields available can be visualized with either of these methods. The dynamics of this representation can be viewed through animation.

The upper air three-dimensional wind velocity is visualized via interactive marking of geographic locations for virtual soundings within the analyzed or modelled atmosphere. At each location (two in this case), a vertical profile is extruded through the atmosphere. Each profile is realized as a pseudo-colored tube, which is contoured by the variable selected for isosurface realization (i.e., humidity). The wind velocity along the profile is shown by a set of vector arrows that point in the direction of the wind. Horizontal speed at these points is indicated by the color and length of the arrows. Optionally, the locations on each virtual wind profiler can be used for seed particles for particle advection, which are realized as streamlines. These lines, which are also pseudo-colored by horizontal speed, indicate the instantaneous direction of the modelled wind from these locations.

A suite of tools invoked through c-shell scripts on a UNIX workstation or shortcut icons on an Intel/Windows (95, 98 or NT) workstation provides the facilities for all four visualization classes. They present user interfaces based upon XWindow/Motif for indirect interaction and OpenGL for direct three-dimensional interaction in cartographic coordinates. They provide forecasters interactive capabilities with three-dimensional representations of the state of the atmosphere (e.g., thermodynamics, moisture, clouds) derived from the data. They have been implemented with IBM Visualization Data Explorer (DX) [1]. DX is a portable, general-purpose software package for visualization and analysis. It employs a client-server architecture with an extended data-flow execution model

These tools are part of integrated system, which has been dubbed "Deep Thunder". It provides operational mesoscale numerical weather prediction in a wide variety of environments via the architecture shown schematically in Figure 5. Deep Thunder evolved from the aforementioned experiments to support forecasting at the 1996 Olympic Games [15]. The system leverages the available infrastructure of weather data to do localized forecasts -- observations for initialization via an assimilation step and synoptic-scale forecasts for boundary conditions. The simulation portion of the system is parallelized for rapid operational execution on an IBM RS/6000 SP distributed memory computer. All the aforementioned visualizations classes are integrated, with Class I being represented at the lower right with the others at the upper right.

Figure 5. Architecture of Deep Thunder.

Additional information about the system as well as numerous visualization examples are available on the World-Wide-Web ( http://www.research.ibm.com/weather).

6.1 Slicer -- Class I and Class II

After each model or analysis execution, all of the results are collected and reorganized into a form that can be used by standard meteorological analysis tools (e.g., AWIPS [8]). These same post-processed data are made available for interactive visualization and analysis (Classes I, II and IV) via two applications, Slicer and Viewer . This includes all of computed variables from the model at hourly resolution.

A single application, called Slicer, provides the Class I and Class II visualizations. A screen dump of the tool in action is shown in Figure 6, which shows the user interface of this application as applied to RAMS at 6.5 km resolution in a domain 650 x 650 km covering Hawaii. Figure 2 was also generated by this application. Slicer provides the ability to view and interact with the data in a latitude-longitude (from the model's stereographic grid) coordinate system. The coordinates are annotated with an axes box and base maps.

The display contents are driven by a Primary Controls panel, which enables the selection of a single variable by slice (surface or isobaric level) for each of the available pre-defined visualization techniques. The user may choose one of the available parameters to be independently realized as continuous pseudo-color, contours (either line or filled), deformed surface or numeric values. In additional, they may be overlaid with streamlines or vector arrows of winds. These may either be individual surface fields, including topographic height or an isobaric slice from any of the upper air variables. For RAMS, there are 34 surface, scalar, two-dimensional variables produced by the model. The model contains 15 upper air scalar, three-dimensional fields. For each of them, you may choose one of 21 isobaric levels. Additional information about these options is available via the Help button in the control panel. A secondary control panel, Input Controls , allows the user to select the data set of interest by the start time of a particular weather model run (e.g., from RAMS or LAPS).

There are two other windows of interest. The primary one is the Image window, within which you may view and directly interact with the model output. There are several options available, including changing viewing modes (Options pull-down, View-Control) and saving/printing images/animations (File pull-down). If the default view is "zoomed-in" or "zoomed-out" or the image is blank, then the last use may have been with a run over a different location. Hit the reset button (control-F) in the Image Window. Then ctrl-V or View Control from the Options menu and select, pan/zoom, rotate, etc. to get the desired view.

The other window is the Sequence Control, a graphical widget with the appearance of a VCR. It gives one the ability to specify a time step or frame within the model run currently being examined -- move forward or backward in sequence, single step, pause, loop continuously or loop back and forth. The default settings for the Sequencer are to have the program operate on each individual time step from the model as a frame in the animation. By clicking on the box in the upper right of the Sequencer, there is the ability to change the increment setting. By reducing the size of the increment, for example, an animation can be lengthened by skipping fewer time steps from the simulation. One can also manually control which time steps are available.

Figure 6. Screen dump of Class I and II Application -- Slicer.

A facility exists for probing the contents of the variable selected for color, which is illustrated in the image. If the probe button is pushed, the variable that was selected for color representation will be interrogated. The results will be displayed in a dialog box that will pop up on the screen. To change the location, go to Cursors Mode (Options pull-down in the Image window, View-Control) and select probe_area. A little probe point will be visible in the volume. With the left mouse button, one can drag the point around, which will show coordinates in the upper left. In execute-on-change mode, when the mouse button is released, the results will be shown.

6.2 Browser -- Class III

The visualization shown in Figure 3 was generated by the Browser application. It has the ability to view and interact with the data in a latitude-longitude (from the model's stereographic grid)-altitude (from the model's height calculations) coordinate system. The view is annotated with base maps. The vector/line maps are draped over a topographic surface and registered in this three-dimensional coordinate system with the model output. The focus of this application is qualitative assessment (browsing) of the model output. Unlike the Slicer (section 6.1) and Viewer (see section 6.3) applications, the browser works with data at three to twelve times the temporal resolution over a subset of the computed parameters. The ability to create both time-based and key-frame (flyover) animations is available. Figure 7 shows the user interface of this application as applied to the same RAMS run used in Figure 6. Additional information about these options is available via the Help button in each control panel.

The Primary Controls panel in the browse application allows selection of specific visualization techniques and data. One of several surface scalar parameters may be chosen to show on the map, which include temperature, dew point, relative humidity, precipitation, heat index, wind chill and pressure. The data are shown as a continuous pseudo-color field or color-filled contours using colormaps described earlier. In either case, the data are shown as a shaded colored surface that is deformed vertically by local topography, and overlaid with the vector maps. Optionally, one can select topography, then the surface will be colored like a topographic map, where color is a redundant way of indicating altitude with the height of the surface. There is also the ability to adjust the scaling range for the selected variable via a pair of stepper widgets. The Primary Controls panel also allows one to optionally visualize surface winds through one of several techniques, which will be pseudo-colored by speed as previously discussed. It also allows support of marking surface low and high pressure regions with an "L" and "H", respectively. In addition, one may indicate the locations of several major cities in the region, and control the display of basic annotation.

The model contains several parameters related to clouds. One may select an isosurface of total cloud water (sum of ice and liquid) to be generated, which can correspond to cloud boundaries. The isosurface is colored white with variation in darkness and opacity according to the data. Since these surfaces are three-dimensional analogues of contour threshold levels they do not necessarily represent a true "cloud surface". If the isosurface threshold value is low, then the surface can often be considered an outer cloud "boundary". One may indicate one or more threshold values in kg/kg (water/air) via a widget. Alternatively, the total cloud water density can be visualized via direct volume rendering, in which the larger predicted values are mapped to increased opacity and darker shades of white in a continuous representation. Forecasted reflectivities in dBz can also be selected using the same options as available for cloud water density, except cyan is used for coloring instead of white. Isosurface representations of reflectivity inside a cloud isosurface can correspond to internal rain shafts.

One has the ability to create a flyover key-frame animation along a path that is specified interactively with the mouse. To change the path, go to Cursors Mode (Options pull-down in the Image window, View-Control). A set of probe points in the volume is shown. With the left mouse button, individual points are displayed, which will show coordinates in the upper left. One can indicate a new location by double clicking with the left mouse button at the desired place. An extant location can be deleted by pointing at it with the cursor and double-clicking the left mouse button. The locations that are specified serve as control points for the animation. To view the animation, select the Flyover button and hit Execute (or Ctrl-O). Another image window will pop up and be the view of the model output of the current time step along the flight path. The main window will show the model output with the flight path and a little cursor corresponding to the location being viewed in the animation window, which is shown in Figure 7.

There are three other windows of interest. The primary one is the Image window, within which one may view and directly interact with the model output. There are several options available, including changing viewing modes (Options pull-down, View-Control) such as rotation or pan/zoom, and saving/printing images/animations (File pull-down). The coordinates can be optionally annotated with an axes box by turning on AutoAxes from the Options pull-down.

The Input and Outputs Controls panel enable the selection of the particular RAMS run of interest, including one that may still be executing. It also supports creation of products for distribution on the world-wide-web -- images, geometries, panoramas, and animations. The animations are time or geometrically sequenced, and stored by the application at workstation resolution but losslessly compressed. They can be converted to the MPEG format after their generation. The first three products are essentially snapshots of the content of the three-dimensional interactive Image window. The image is simply a static bitmap. The geometries are what is used to render the image but converted to the Virtual Reality Modeling Language (VRML). To improve the utilization of VRML in such applications, simplification methods are applied dynamically and individually to each geometric component of the visualization. Another class of web-based visualization can be provided through the use of image-based rendering methods (e.g., IBM PanoramIX). Based upon a set of images which describe the boundary of a geographic scene viewed from the interior, an interactive, non-immersive environment can be constructed. These images and a control file are generated on demand from the application. For both VRML and PanoramIX output, a simple plug-in to a web browser is available to enable navigation through the three-dimensional representation of the weather model output independently of the visualization software or the graphics workstation on which it operates. Although the VRML browser provides freedom for navigation it can require OpenGL hardware, a fast workstation and good bandwidth to the web server for effective utilization despite simplification of the geometry. PanoramIX enables a high-degree of interaction on low-end clients with limited bandwidth at the cost of constrained navigation.

The other window is the Sequence Control, a graphical widget with the appearance of a VCR, as described earlier.

Figure 7. Screen dump of Class III Application -- Browser.

Since the goals of this application include enabling tracking of the simulation during execution, not having to delay visualization until after the post-processing of model results, and permit higher temporal resolution animations, the output generation process by RAMS was modified. The workstations that provide interactive visualization share network-mounted filesystems with the RS/6000 SP server running the simulation code. Hence, there is a need to balance the amount of output generated with the available bandwidth, yet not slow down the model execution. Although direct interprocessor communication between the visualization and simulation software could have been used, a custom file format was designed instead associated with a set of specialized readers. To further minimize the volume of data that would have to be output, and then subsequently accessed over the network, several parameters (humidity, dew point, wind chill and heat index) were computed within the visualization application from more basic variables. Thus, limited variables generated at five to 30-minute intervals of simulated time optimized for network access resulted in little observable latency in interactive application.

6.3 Viewer -- Class IV

An application called Viewer provides the Class IV visualizations. A screen dump of the tool in action is shown in Figure 8. Figure 4 was generated by this application. This tool provides the ability to view and interact with the data in a latitude-longitude (from the model's stereographic grid)-pressure coordinate system. The coordinates are annotated with an axes box and base maps. Figure 8 shows the user interface of this application applied to RAMS over a domain 800x800 km in size at 8x8 km resolution centered over Dallas.

There are several windows of interest. The key one is the Image window, within which one may view and directly interact with the model output. There are several options available, including changing viewing modes (Options pull-down, View-Control) and saving/printing images/animations (File pull-down). Another window is the Sequence Control, as described earlier. There is an Input Control panel, which is not visible in the figure, that enables the selection of a RAMS run of interest. It is the same as the one shown in Figure 6 for the Slicer application.

The Primary Controls allows one to select specific visualization techniques and data. The surface may be pseudo-colored by any one of 34 surface, scalar, two-dimensional variables produced by the model. This may also include topographic height. RAMS produces several upper air, three-dimensional fields. For each of 15 upper air scalar fields, one may choose to realize the data as an isosurface, vertical slice and horizontal (isobaric) slice. For the isosurface, a specific threshold value must be specified via a stepper widget. The isosurface is colored according to a segmented colormap. When the variable of interest is changed, the default value for the isosurface is the mean. For the vertical slice, one may select a grid position, and whether the slice is meridional or zonal. The slice is color-filled, pseudo-colored with a segmented colormap and line contoured. For the horizontal slice, the pressure level must be specified. The slice is similarly displayed as the vertical slice but with a different segmented colormap.

Figure 8. Screen dump of Class IV Application -- Viewer.

One may probe the volume for specific values at selected locations within the data set. If the probe button is pushed, the variable that was selected for isosurface representation will be interrogated. The results will be displayed in a dialog box that will pop up on the screen. To change the location, go to Cursors Mode (Options pull-down in the Image window, View-Control) and select probe_volume. A probe point in the volume will be visible. With the left mouse button, one can drag the point around, which will show coordinates in the upper left. In execute-on-change mode, when the mouse button is released, the results will be shown. An example of this feature is shown in figure 8.

The upper air three-dimensional wind velocity may be visualized via interactive marking of geographic locations of interest, as discussed earlier. One may define one or more geographic locations for virtual soundings within the model atmosphere. This is also done in Cursors Mode by selecting profilers. The process is the same as the one for marking the control points for key-frame animation in the browse application. At the locations that have been specified, a vertical profile is extruded through the entire model atmosphere, which is realized as a tube. The sounding location is used to derive information about wind velocity. If a variable has been selected for realization as an isosurface, then the values along each profile of that variable are also shown as pseudo-colored, filled contour bands using the same segmented colormap as is employed for the isosurface. These profiles are also shown via a conventional pressure-profile plot as shown in a separate image window. They are numerically labelled in both windows.

6.4 Other Facilities

In addition to these three interactive applications there are a number of supporting facilities provided in this suite, with the following capabilities:

Either time-based or fly-over animations generated by the interactive applications are stored at workstation resolution, 24-bits deep and losslessly compressed. Tools are provided to play them back at full resolution as well as do some simple editing. The playback of full-fidelity animations from the Browser, typically at 10-minute time steps, provides both a dramatic display for the media and practical application by forecasters.

The stored animation often need to be converted to other formats for distribution. Utilities are available to convert them to the MPEG-1 format at both low- (352x240) and medium- (approximately VGA) resolution, typically for distribution on the World-Wide-Web. Other tools are available for conversion to animated GIF or digital broadcast formats (e.g., YUV).

The process that prepares the data that are used by the Slicer and Viewer application does not generate summary statistics, which are needed to properly scale and label the various visualizations that the applications support. A utility is provided to do these calculations on the model or analysis output, which optionally can be invoked by the visualization applications prior to interaction.

7. Utilization

The Browser application (Class III) enables model assessment. It is used to evaluate model results and create various products. Typically, an animation with 10-minute resolution over the full model run (24 hours long on a 3-hour duty/refresh cycle) is created after the forecaster selects the variables, techniques and geographic view. These would remain invariant throughout the animation. One or more animations are generated for local playback at workstation resolution to support media briefings. To aid in this selection process the forecaster would interactively move through the geographic scene, experiment with different displays and do limited animation. To illustrate this process, consider the montage of seven images in Figure 6, which are sequenced from left to right, top to bottom. Start with a three-dimensional representation of the local area -- a terrain map overlaid with state, county, coastline and river maps, and marked with major cities. The area is about 800x800 km in size centered over Dallas. Then, predicted temperatures for 9 AM CST on January 13, 1999 are overlaid. A value is calculated for each 8x8 km cell over the area. The temperatures are colored by the scale at the upper right to show their continuous variation. Surface winds are then displayed as a set of arrows pointing in the wind direction, while the color corresponds to speed as shown in the legend to the lower right. The lighter color implies faster wind (i.e., white is 30 mph). Next, three-dimensional representations of predicted clouds are added, which are illustrated as a white, translucent surface which as a "boundary" where the density of water (liquid + ice) in the atmosphere is above a certain threshold (i.e., an isosurface of 10^-4 kg/kg). From this information, the simulation can be examined in more detail. Rather than temperature, wind chill is shown since the data showed fairly windy conditions and low temperatures for Dallas in January. To illustrate regions where the wind chill might be particularly low, the data are shown as a set of bands, each with a distinct color that increase in value from dark to light, which can be seen with the color legend at the upper right. To examine the predicted winds in another way, they are shown as colored streamlines with arrows still indicating the direction. This illustrates a front moving through the area (i.e., lines bunched up at the lower right). Now, predicted wind chills for the specific time are shown for major cities. From the interaction employed to examine the prediction in a number of ways, a representation that might be useful to publish in a newspaper or on the web to illustrate a forecast is chosen. To simplify the visualization for public consumption, the wind data and some of the annotation are removed. In addition, the geographic viewpoint of the three-dimensional map is changed. Now only the terrain is shown, colored by bands of wind chill prediction and values at major cities along with the maps and cloud data.

Figure 9. A sequence of images illustrating a typical use of the Browser application.

For animations well-suited to support a particular day's forecast, they are MPEG-encoded and associated with a higher-resolution image for distribution on the World-Wide-Web. The contents of such a snapshot image can also be disseminated as a VRML geometry, key-frame flyover animation or as an image-based panoramic scene (i.e., via IBM PanoramIX). In addition, the ability to track the model during execution provides quality control and comparison with results from previously completed runs.

After each model or analysis execution, all of the results are collected and reorganized into a form that can be used by standard meteorological tools. These post-processed data are made available for interactive visualization. This includes all of computed variables, but at hourly resolution unlike the browser application that works with a subset of variables, typically at six times the temporal resolution. Analysis with the Slicer and Viewer applications (Class I, II and IV) must wait until this post-processing phase is completed. As a result, such interaction generally took place well after the next run was begun, which limited the ability to simultaneously compare output.

To illustrate the range of capabilities that these applications provide, consider the sequence of four images from left to right, top to bottom in Figure 10. The first three are produced by the Slicer application and the fourth by the Viewer for January 13, 1999 at 00:00 GMT for the same forecast generated by RAMS used in Figures 8 and 9. The capabilities of the Slicer and Viewer are complementary. The first image, which is a Class I visualization, shows two surface scalar fields, mean sea level pressure as a continuous color field, and line contours of relative humidity. These techniques and data can be viewed in animation. Another display shows the humidity contours as color-filled bands using the same segmented colormap, but now overlaid with 850 mb temperature values at specific locations and 750 mb winds as vector arrows, colored by speed.These fields are combined in the Class II visualization at the upper right, but only showing surface variables. The height of the deformed surface corresponds to lifted index, which reflects instability in the atmosphere. This representation is very effective, especially in animation, of showing the motion of a front. The image at the bottom, now shows contours of surface pressure, but combined with a three-dimensional representation of relative humidity, temperature and winds. The humidity data are shown as an isosurface at 75%, corresponding to a simple cloud boundary and sampled along two vertical profiles. The temperature data are visualized as a single vertical slice that is contoured. The wind data are shown as arrows as part of virtual wind profiler and as streamlines using the profile points as seeds. These techniques and data can be viewed in animation.

Figure 10. A sequence of images illustrating a typical use of the Slicer and Viewer applications for analysis.

8. Conclusions

Specialized interfaces and tools matched to user goals and underlying visualization tasks to support them is a promising approach to providing new visualization facilities for operational activities, especially in weather forecasting. Such successful facilities can be characterized as being easy to master via simple interfaces, even if the underlying capabilities may be quite sophisticated. Although a highly generalized system can be employed to provide similar functionality, the lack of focus in its interface increases learning time beyond what would be considerable acceptable in time-critical activities. This is in contrast to the flexibility that is often preferred in many research environments. Thus, a set of visualization tasks coupled with appropriate designs can be developed a priori, and then refined through modest iteration. Further, generalized approaches to these design elements can be employed to more efficiently develop specialized interfaces and tools matched to user goals.

An effective compromise between trying to use a general-purpose tool directly and implementing a set of highly customized packages has been developed. The generic tools are used for both prototyping new applications and efficient implementation of complete systems, particularly by promoting high-level reuse of underlying tools and design elements. Employing a generic toolkit (DX) also eliminated the need to implement a graphics and computational infrastructure. This is in contrast to the low-level reuse (renderer) in efforts by FIGD or code-level modifications to a turnkey tool (Vis-5D). In addition, unlike traditional meteorological graphics, DX enables the use of modern workstation hardware equipped with three-dimensional graphics accelerators and is parallelized for symmetric multi-processors. Since DX is built upon an unified data model that enables operations directly on the native grids without transformation or compression, customization for data types was not required as well as preserving fidelity during visualization. Further, such a toolkit is extensible to allow development to be focused on meteorological data and tasks, and reuse of tools between applications with similar user interface components. This simplifies training of users to employ the applications with different content matched to separate tasks. It also reduces the cost of development and maintenance, and enables more rapid iterative refinement with or adaptation to new users.

8.1 Application-specific

Class III visualizations proved to be more effective than initially expected by virtually eliminating the laborious evaluation of numerous Class I images via presentation of all the relevant information in an easy-to-interpret, four-dimensional display. Conceptual models that would normally require inference from a significant amount of two-dimensional data (e.g., the horizontal extent of cloud dissipation in the lee of the Appalachian mountains) are obvious in three-dimensional animations. Further, one could easily infer vertical motion based on a three-dimensional display of clouds forming. Although the data may not have indicated precipitation occurring in a specific location, the existence of clouds gives forewarning that precipitation may be possible in that vicinity. Originally intended for media displays, the Browser application quickly gained favor by forecasters as a valuable operational forecasting tool during the initial efforts at the 1996 Olympics.

The subsequent introduction of the Slicer (Class I and II) and Viewer (Class IV) applications into operations complements Class III, but uncovered problems inherent in utilizing the data for AWIPS or similar post-processed products. Although the user can easily select a data set of interest, the organization of the data is not ideal for the type of required access. Often the post-processed results (all variables and time steps) are collected into a single, large file. While convenient from the perspective of the data generator, access to specific arrays forces unnecessarily long seeks across a local-area network. This limits performance until requested data are loaded into memory. One could introduce an additional post-processing step to reorganize the data, but that would further delay access to generated data, and increase the need for disk storage. In addition, not all of the variables defined in this format are consistently populated, and the metadata describing the variables are incomplete. This can lead either to user error or increasing the complexity of the application and interface to compensate. An intermediate step of a simple post-processor to calculate statistics to enable consistent scaling for visualization has been implemented, which is invoked upon demand when a new model run is completed.

Although the deployment of AWIPS by the National Weather Service is on-track to support an impressive variety of complex analysis, interactive processing, display and rapid dissemination of forecasts from a number of diverse data sources, its software components do not include the type of visualizations discussed herein [8]. Both the Viewer and Browser applications represent capabilities complementary to AWIPS as well as the 2-1/2-dimensional features of the Slicer.

9. Future Work

User-driven visualization design was of immediate value at the 1996 Olympic Games, enabling the NWS to provide information for athletes, spectators and officials to plan around adverse weather conditions. This approach could be applied in other areas where precision forecasting shows promise like tourism, aviation, agriculture, broadcast, energy, insurance, pollution monitoring, and fire control and management. For effective utilization outside of general forecasting, a refinement of the task decomposition is necessary. Initially, that would imply customized interfaces, products and packaging, most likely for Class III. For aviation, that might include, for example, support for route planning, dispatch, etc. for both safety and efficiency, where time-varying predictions of prevailing winds, icing surfaces and clear-air turbulence are shown along a flight path. A different vertical coordinate system could be relevant for Class IV such as a normalized level above the ground. Additional derived quantities would be useful to visualize such as cloud ceiling and fraction, visibility on the ground and soaring index. In agricultural applications, introducing additional variables like transpiration and evaporation rate would complement the standard quantities. In many applications, the task refinement would be based upon the correlation of weather prediction with other sources of data relevant to decision making. This idea is illustrated in Figure 7, which shows the relationship between demographic data and the results of a numerical weather prediction. In this case, the visualization is a simple two-dimensional map, which shows a set of glyphs, colored by median house value. The glyphs are located at the centroid of the area associated with zip codes in the domain for a forecast model. These locations are only marked on the map when a set of conditions on house value, population and predicted wind speed are met, as indicated in the control panel widget. The user is free to set these thresholds and animate the results in time corresponding to the weather simulation in hourly time steps. Such an application may be useful for planning purposes by an insurance company or deployment of repair crews by a utility. An example animation is also available for viewing.

Figure 11. Correlation of a weather forecast with demographic data.

In other cases, the results from the simulation could be coupled to other models. In agriculture, that could include a hydrological model to evaluate the effect of runoff from predicted rainfall to help understand the impact on the application of pesticides and fertilizer or planting of new seed. In the energy industry, that could be input to a model used to predict load on a power-generation facility or transmission lines for efficient running of the facility or for trading.

Since there is preponderance of potential data sources that could be utilized with these tools, extensions to support them will be driven by the ability to leverage "standard" products (e.g., data formatted for AWIPS). Part of that effort will be to more tightly couple the interaction between Class III and the simulation to enable more efficient tracking/steering. Another aspect is to improve the organization of the model output, particularly for the post-processing (analysis) products.

Web-oriented visualizations can be generated by the current set of applications but an intermediate step of migrating the products to a web server is required. This has the advantage in an operational environment, where the forecaster has control over the content of the visualizations that may be disseminated to a variety of consumers. However, the task decomposition can be further refined by considering direct generation of visualizations within a web browser. Similar tasks could still apply, but the user interface and the content must be simplified to be effective. Thus, the current indirect interaction would be replaced with Java-based applets in the browser as a client that communicates with a DX server processing the data and generating the requested visualization. All of the aforementioned methods, MPEG animations, and image, VRML and PanoramIX snapshots are relevant. However, the generation time for the animation of a suitable length and the download time for VRML are current barriers.

The notion of task-driven customization of visualization content and interface has shown to be successful in weather forecasting. Although the results are discipline-specific, the concept is not. Many problem areas do imply different visualization tasks with various mappings to user tasks. The current literature on task definition, user modelling and content design generally employs examples from disciplines other than weather forecasting. Therefore, applying the approach discussed herein to other domains is desirable. Likely candidates include measurements collected from medical scanners, the output of data mining algorithms applied to relational data bases, and utilization of results from terascale physics simulations.

10. References

[1] G. Abram and L. Treinish. An Extended Data-Flow Architecture for Data Analysis and Visualization. Proceedings of the IEEE Visualization 1995 Conference, October 29-November 3, 1995, Atlanta, GA, pp. 263-270.

[2] S. K. Card and J. Mackinlay. The Structure of the Information Design Space. Proceedings of the IEEE Information Visualization 1997 Conference, October 20-21, 1997, Phoenix, AZ, pp. 92-99.

[3] G. O. Domik and B. Gutkauf. User Modeling for Adaptive Visualization Systems. Proceedings of the IEEE Visualization 1994 Conference, October 17-21, 1994, Washington, DC, pp. 217-223.

[4] W. L. Hibbard, B. E. Paul, D. A. Santek, C. R. Dyer, A. L. Battaiola, M. F. Voidrot-Martinez. Interactive Visualization of Earth and Space Science Computations . IEEE Computer , 27 , n. 7, July 1994, pp. 65-72.

[5] V. Jung. System for Guiding and Training Users in the Visualization of Geographic Data. Proceedings of the 1st Conference of GeoComputation. School of Geography, University of Leeds, September 17-19, 1996, Leeds, UK, pp. 470-482.

[6] M. Lux and T. Fruhauf. A Visualization System for Operational Meteorological Use. Proceedings of the Sixth International Conference in Central Europe on Computer Graphics and Visualization (WSCG'98), Plzen, Czech Republic, February 9-13, 1998, pp. 525-534.

[7] P. T. McCaslin, P. A. McDonald and E. J. Szoke. Developing and Testing a 3D VIsualization Workstation Application at FSL. Proceedings of the Fifteenth International Conference on Interactive Information and Processing Systems for Meteorology, Oceanography and Hydrology, American Meteorology Society, January 10-15, 1999, Dallas, TX, pp. 498-501.

[8] National Weather Service. AWIPS Program Information. http://www.nws.noaa.gov/msm/awips/awipsmsm.htm, March 1998.

[9] B. Rogowitz. and L. Treinish. How Not to Lie with Visualization. Computers in Physics, 10, n.3, May/June 1996, pp. 268-274..

[10] Pielke, R. A., W. R. Cotton, R .L. Walko, C. J. Tremback, W. A. Lyons, L. D. Grasso, M. E. Nicholls, M.-D. Moran, D. A. Wesley, T. J. Lee and J. H. Copeland. A Comprehensive Meteorological Modeling System - RAMS. Meteorology and Atmospheric Physics , 49 , 1992, pp. 69-91.

[11] H. Rushmeier, R. Lawrence and G. Almasi. Case Study: Visualization Customer Segmentations Produced by Self Organizing Maps. Proceedings of the IEEE Visualization 1997 Conference, October 19-24, 1997, Phoenix, AZ.

[12] F. Schroder. Visualizing Meteorological Data for a Lay Audience. IEEE Computer Graphics and Applications, 13, n. 5, September 1993, pp. 12-14.

[13] F. Schroder and M. Lux. TriVis -- Professional Television Weather Presentation. http://www.igd.fhg.de/www/igd-a4/projects/docs/trivis/trivis_e.html, October 1997.

[14] J. S. Snook, J. M. Cram, J. M. Schmidt. LAPS/RAMS, A Nonhydrostatic Mesoscale Numerical Modeling System Configured for Operational Use. Tellus A, Dyn. Meteorol. Oceanogr. (Sweden), 47A , n. 5, pt. 1, October 1995, pp. 864-875.

[15] J. S. Snook, P. A. Stamus, J. Edwards, Z. Christidis, J. A. McGinley. Local-Domain Mesoscale Analysis and Forecast Model Support for the 1996 Centennial Olympic Games. Weather and Forecasting, 13 , no. 1, 1998, pp. 138-150.

[16] L. Treinish. Visualization of Disparate Data in the Earth Sciences. Computers in Physics, 8, n 6, November/December 1994, pp. 664-671.

[17] L. Treinish and L. Rothfusz. Three-Dimensional Visualization for Support of Operational Forecasting at the 1996 Centennial Olympic Games. Proceedings of the Thirteenth International Conference on Interactive Information and Processing Systems for Meteorology, Oceanography and Hydrology, American Meteorology Society, February 2-8, 1997, Long Beach, CA, pp. 31-34.

[18] M. X. Zhou and S. K. Feiner. Data Characterization for Automatically Visualizing Heterogeneous Information . Proceedings of the IEEE Information Visualization 1996 Conference, October 28-29, 1996, San Francisco, CA, pp. 13-20.

lloydt@watson.ibm.com