By Edward Szoke, Herb Grote, Paula McCaslin, and Philip McDonald
IntroductionFor more than a decade, FSL has been developing and evaluating techniques for visualization of numerical model output in three dimensions. The development of D3D (Display 3D) has been driven by ideas posed by meteorologists who are searching for ways to better understand the three-dimensional motion of the atmosphere. The powerful three-space visualization tools can be used to examine any perspective of a storm, tilt around corners, zoom in, zoom out, loop, and change colors, parameter settings, and opaqueness of the field.
D3D is an extension of the interactive D2D, part of the Advanced Weather Information Processing System (AWIPS) now available to every National Weather Service (NWS) Weather Forecast Office (WFO). The core software for D3D is Vis5D, developed by the University of Wisconsin-Madison, with enhancements engineered by FSL in collaboration with the university. Vis5D was chosen from available software because it includes an API, a division between the system's user interface and its main core, enabling system developers to include Vis5D as a visualization subsystem of other systems. This free software also has superior animation capabilities, is designed specifically for atmospheric science, and has an established user base in the meteorological community.
This status report on D3D begins with a brief comparison of D2D and D3D, then testing and evaluation of D3D, uses of visualization tools (isosurfaces, plan views and cross sections, sounding tool and probe, trajectories, and volume visualization) in the forecast office, and last, D3D deployment issues and plans.
D2D and D3D SymbiosisAlthough not officially sanctioned as part of AWIPS, the D3D system was designed to be used with D2D software, which provides integrated access to meteorological data supplied by national and local weather services. D3D is used to investigate the complexities and 3-D structure of atmospheric parameters, and to explore its potential for adding value to the forecast. In developing D3D visualization for potential use in operations, our philosophy was to copy the D2D structure as much as possible to minimize the learning curve for using the application. This involved considerable changes to some of the Vis5D interfaces. Additionally, Vis5D was modified to provide access to the D2D basic and derived gridded datasets without the need to create intermediate Vis5D files.
One of the first changes to D3D was replacement of the main Vis5D interface (Figure 1), for selecting and modifying products, with an interface that mimicked the D2D product selector, the Volume Browser (Figure 2). Functionally the Vis5D matrix-like interface is effective, but this step made D3D appear physically much like D2D.
Figure 1. Vis5D product selection matrix.
Figure 2. D3D Volume Browser, with submenus open.
Requirements for displaying and editing the fields in D3D are different from those in D2D, necessitating additions to the Volume Browser. An example is shown in Figure 3, where editing options for a field displayed as an isosurface in D3D are accessed by clicking on the field in the Volume Browser list of products, which then opens up a graphical user interface (GUI) that allows changes to product attributes, some of which are not even options in D2D, such as transparency of an isosurface.
Figure 3. D3D Volume Browser in Edit Mode for editing an isosurface of vorticity.
Components of D3D that are actually more 2D in nature but are more powerful in their capabilities within D3D have generated much interest. These include the capabilities of horizontal and vertical cross sections that were expanded through the use of an easy-to-access slider bar, which enables one to quickly move the cross sections through the data, stopping at any arbitrary location. When used with an isosurface, this adds a level of quantitativeness to examining data, such as knowing more precisely at what level the isosurface existed and its value. (For more information, see the February 1999 issue of the FSL Forum. Also see Note at the end of this article for published papers on other aspects of the development of D3D.)
Meteorological Testing of D3DMeteorological testing and evaluation began after sufficient modifications were made to D3D that allowed it to access the real-time database, and most of the GUIs appeared more like those in D2D. Initially D3D was introduced to FSL meteorologists through the FSL Daily Weather Briefing, which has been used for 20 years as an informal testing ground for new technologies. Meteorologists from FSL and other laboratories leading the weather briefings were trained in the basics of D3D and encouraged to use it for at least part of the briefing. This informal evaluation resulted in suggestions for further improvements to D3D, but eventually it was decided that a more formal evaluation would be useful.
The more formal Forecast Exercise, using either real-time or case (displaced real-time) data, has been an integral part of D2D evaluation. Two such exercises for D3D involved inviting FSL meteorologists and operational forecasters from national and foreign forecast offices to participate. These real-time exercises in 1998 and 1999 led to many important suggestions for improvement to D3D. The exercises also stressed the need for training, because many participants encountered difficulties using the software or making sense of 3-D visualization, even though we had taken great care to try to reduce the learning curve by making many of the interfaces more D2D-like.
The second exercise, in 1999, was a major effort to stage a comprehensive evaluation from a diverse group of operational meteorologists. The NWS is organized into six regions, four in the continental United States. One WFO forecaster was invited from each region, and for a broader perspective, one from each of the National Centers within the NWS. The first week involved hands-on training and exercises using both displaced and real-time data. Various details of a weather situation were examined, and comparisons were analyzed as to how the participants went about finding the information using D3D to finding the same information using D2D. Ample time was available for the forecasters to record free-form comments through an electronic notebook setup, while other comments were gathered through more formal questionnaires (prepared by FSL's Evaluation Team), which were filled out at least daily. During the second week, the forecasters were tasked with attempting some specific weather forecasts using three displaced real-time cases. Forecasts were chosen for which they might not know the answer even if they were marginally familiar with the cases. As another twist, participants led a daily weather briefing using primarily D3D and real-time weather. Last, the forecasters completed a comprehensive written evaluation of the exercise, and were informally debriefed.
As expected, both exercises yielded an enormous amount of input, most of which became part of D3D, with some still in the queue for future development. The forecasters were overwhelmingly positive about D3D, and many wanted to take a version of the software back to their office. However, this was not possible, for instance, because the AWIPS workstations in use at that time were not powerful enough to run both D3D and D2D effectively. Some participants found isosurfaces somewhat confusing, but most thought they presented a great way to visualize the atmosphere. Together with other D3D tools, isosurfaces should provide a very quick way to peruse and better understand lots of model output. Many of the participants quickly learned how to better quantify the output they were scrutinizing by combining isosurfaces with other D3D tools such as cross sections or plan views, and/or by coloring the isosurface by another variable like height or pressure. D3D is very interactive (even more so than D2D), and forecasters enjoyed the ability to move cross sections around and get instant feedback, or to peruse the data with the very interactive and popular sounding tool.
Of course not every response was positive. For instance, the very interactive nature of D3D required that it be run on a fast enough machine, so naturally when participants spent time on slower machines, they were easily frustrated. Three-dimensional rendering of isosurfaces, the 3-D contour surface of a field at a particular value, was intriguing to use and most thought the end result was very revealing, yet there was still a feeling of uncertainty as to how one might best use them in operational forecasting. There was also an almost universal issue of georeferencing with isosurfaces. Other issues raised included the desire to see overlays of real data (observations, for example), the use of D3D to display radar data, and the ability to display multiple models at the same time.
A fairly common response was that training was necessary, not only in how to operate D3D (the "buttonology") but also in how to apply D3D meteorologically. We purposely did not make strong suggestions for the exercise in order to learn more about how the meteorologists actually used the various D3D tools. The training issue is still a priority consideration.
Using Visualization Tools in the Operational Forecast OfficeAfter learning so much from the two real-time exercises and making many improvements to the system, D3D has moved into a new phase. The intent has always been that D3D would be used in an AWIPS environment, or at least with an AWIPS data source. Until recently, the computational and graphics requirements of a 3-D display application exceeded the capabilities of typical AWIPS workstations. However, with increased performance and decreased cost of Linux PC workstations, coupled with a movement toward Linux AWIPS systems, the day of D3D in the forecast office is at hand. NWS has begun replacing two of the older workstations at each WFO with Linux workstations. Before this, individual WFOs often purchased machines on their own to run special applications. Generally it is on these extra machines that the D3D test version has been installed. The WFO managers who have accepted a test version recognize that access to the real-time database feeding the AWIPS workstations must not be compromised by D3D, and our testing at FSL shows that this should not be a problem.
During a session dedicated to D3D at the American Meteorological Society Annual Meeting in January 2002, we handed out CDs of the D3D software to those interested in trying it out in the operational setting. The WFO sites using this software are shown in Figure 4.
Figure 4. Map of NWS regions, with locations of WFOs. Ovals show the sites that have D3D availability, as indicated at the top.
The next paragraphs cover special operational uses of the D3D tools: isosurfaces, plan views and cross sections, sounding tool and probe, trajectories, and volume visualization.
Isosurfaces As noted earlier, isosurfaces, the main 3D visualization tool in D3D, have never been available to NWS operational meteorologists for use with real-time model output. Isosurfaces show the volume bounded by a particular value, allowing the user to visually depict the field's 3-D structure at any desired viewing angle. The examples shown in Figure 5 contrast a simple monocolored isosurface (5a) with how it appears when colored by another variable, in this case, height (5b), which adds quite a bit of information to the isosurface.
Figure 5a. A simple monocolor isosurface enclosing values of 90% and higher relative humidity from the Eta model.
Figure 5b. the same isosurface as in Figure 5a enhanced with another variable, height. Red values indicate 90% humdity near the surface; blue values indicate 90% relative humidity high in the atmosphere.
Optional features allow the display of multiple isosurfaces of the same variable at different values, and different values of transparency for displaying them together. It is also common to display isosurfaces from different variables together. We also modified the Vis5D isosurface display to allow a contour and a wind barb overlay of another field on an isosurface. This is especially useful for isentropic applications, where one could overlay winds on a theta isosurface that might be colored by pressure.
Plan Views and Cross Sections An example of plan views and cross sections is given in Figure 6, showing them combined with an isosurface. Plan views and cross sections are both highly interactive. They can be maneuvered through the volume of data by either grabbing at the edge of the plan view or the "handle" on the cross section (standard Vis5D features), or by using a slider bar that is accessed through the Volume Browser, much like the one illustrated in Figure 1. The advantage of the slider bar method is that while zoomed in on or tilted at any angle, the planes can still be maneuvered. Although plan views and cross sections are 2D tools, the highly interactive nature of these features in D3D makes them extremely powerful for rapid perusal of large volumes of data.
Figure 6. A combined D3D view from the FSL local model showing an isosurface of reflectivity, a plan view of 725-mb winds, and a cross section with an image of vertical velocity and contours of pressure.
Sounding Tool and Probe The sounding tool is another D3D tool that is a 2D depiction but available in a much more interactive mode than in D2D. Not only a standard thermodynamic sounding, this tool also has a hodograph display and a separate display on which one can plot up to three variables as a function of height (Figure 7). Although only the sounding and vertical plot options were available in Vis5D and during the 1999 Forecast Exercise, this was still an extremely popular tool. Forecasters especially enjoy the instant sounding depiction when roaming the cursor through the volume, or stepping through a time sequence.
The cursor used for this application (Figure 7) has a vertical extension to the surface, making it ideal for georeferencing. It has been suggested that such a cursor be available to use with isosurfaces for this purpose (without necessarily invoking the sounding tool). Based on forecaster input, the Vis5D sounding tool was modified to allow for a hodograph plot as well as a plot of convective parameters that can be read from the model's surface-based grid and shown as values (Figure 7). One has the option of choosing from a number of parameters, as well as the option of displaying all or any of the three displays. The other option known as the probe, as the name implies, allows the user to move a cross-hairs cursor in 3D space and get a readout of selected model values interpolated to that point, with the user determining the specific parameters to display. Through an added GUI, one can position the cursor at a location relative to the ground as well as at a vertical level and, if desired, can lock the probe to a specific level (either a pressure level or height AGL). This GUI makes the probe a potentially more useful tool, though its overall use has been somewhat limited, especially when compared with the sounding tool.
Figure 7. The D3D sounding tool showing all options displayed (see text). Plots are updated as the selection cursor (far right) is moved along with the mouse.
Trajectories Trajectories are usually discussed at some point in the education of a meteorologist, but operationally in the NWS, the AWIPS D2D streamline analysis has been the primary tool available for analyzing air flow trajectories. However, forecasters recognize that the 2-D streamline patterns neglect the vertical component of the flow and duplicate horizontal trajectories only if they are stationary. This can cause uncertainty regarding the actual direction taken by air parcels, leading to errors in temperature and moisture forecasts. So until now, there has been no adequate way to generate trajectories with model output in operations. Vis5D makes it easy to generate either backward or forward trajectories beginning at any point in the model output, and our modifications created a GUI that retains this ease of use and also allows quick generation of some potentially useful options. Four of the options along with the GUI are shown in Figure 8.
Figure 8. D3D trajectories showing four options (see text).
The simple trajectory calculation "point" plots a trajectory from a single point, which looks like the probe, and cursor can be located in 3-D space using the "put home cursor" GUI. The trajectory can be displayed as a thin line (e.g., white trajectory) or a thicker ribbon (e.g., yellow trajectory) and can be colored by another variable, such as pressure, for example. The calculation of the trajectory uses a method, developed by Bill Hibbard at the University of Wisconsin, with a simple yet effective way of interpolating between available model output times that may be 6 hours apart. (In research applications one would normally have many more model output times available to directly use for a trajectory calculation.) Our experience has shown that the air flow trajectories look very reasonable. Additional user interface options (Figure 8) include "Column," a script that launches a vertical column of trajectories at 50-mb intervals from a single latitude/longitude point (see the group of magenta trajectories), "Row EW," a script that launches trajectories in an east-west row at a user-selected spacing (see green trajectories), and "Row NS," which is similar but in the north-south direction, and "Series," a script that launches a trajectory from a single latitude/longitude point for each model time, resulting in a trajectory time series.
Since operational forecasters have not had much experience with trajectories, it will likely take sometime before their potential use is fully explored. Even with limited use, however, some forecasters are encouraged that this feature will provide better understanding of air flow in a given weather pattern. In an article presented at the AMS Meeting in January 2002, forecasters at the Boulder WFO expect that this visualization tool "will result in improved moisture, temperature, cloud, and precipitation forecasts, including significant weather events associated with well-developed cyclones, the onset of summer monsoon and convection in the desert southwest, and the development of upslope clouds and precipitation in and near mountainous terrain."
Volume Visualization The volume visualization tool has not been evaluated much at this point because it can be fairly computer-intensive and, under the current D3D implementation of software rendering of graphics, slow for many of the computers using D3D. However, volume visualization, like isosurfaces, is a true 3-D depiction (Figure 9), but instead of rendering a surface of a variable at a given value, as in the isosurface, the entire volume of data for that variable is rendered as a kind of "visual fog." The colors can be edited to enhance specific ranges of the data, as was done with the example here for the relative humidity display to make it appear somewhat like a satellite water vapor image. Our limited experience with volume visualization suggests that it is most effective when used in combination with an isosurface display.
Figure 9. Example of a model field of relative humidity rendered using Volume Visualization.
Deployment Issues and PlansWe are very encouraged by the operational use of D3D so far, but recognize problems associated with the fact that it is currently not in the official future plans of AWIPS. This creates its own set of potential operational issues since forecasters would like to use an integrated workstation. From feedback at a recent European conference, when one country attempted to introduce 3-D visualization to forecasters, the effort failed because the application was not integrated into the existing workstation. It is too early in our effort to determine if this factor will be a significant hindrance to operational D3D use in a testing phase. Our experience with other applications both support and contradict this idea. Some forecasters are willing to use separate displays (for examining other model data on the Web, for example), but our local model was not used extensively until it was integrated into AWIPS.
Given the unofficial status of D3D at this time, we have proceeded by presenting talks at conferences, placing papers on our homepage, and making the software readily available to interested forecasters. The software is fully contained, with some case study data, on a single CD; upgrades can be downloaded from our main Web page, d3d.fsl.noaa.gov/. As mentioned earlier, both the AWIPS and D3D software can be run on high-performance but inexpensive computers, which are available to replace current AWIPS workstations through the use of commercial off-the-shelf machines running Linux.
Our near-term plan is to acquire feedback from the various sites that now have D3D. We want to do this through our homepage using an electronic questionnaire and other feedback mechanisms. WFO managers at Portland, Maine, one of the original recipients of D3D, have developed their own questionnaire for surveying forecasters there. We will probably incorporate portions of this questionnaire into the one we create. Of course we plan to continue making improvements to D3D, based on current input and some from the 1999 Forecast Exercise, including multiple model context, displaying actual data, and continued development of a 3-D radar display.
Because D3D is in experimental status, funding is limited, but other agencies have expressed interest in the software. Recent work through a U.S. Space Agency funded project on a potential 3-D lightning display could result in additional data displays in D3D without much additional effort. We hope to also explore an interesting new Vis5D application of flow visualization developed by a Swiss laboratory.
Our encouragment by the number of sites interested in D3D is tempered somewhat by the level of actual use overall at many of the forecast offices. Discussion about the usefulness of 3-D in operational forecasting includes the observation that while meteorologists generally find the 3-D displays intriguing and useful, they often resort back to the more familiar 2-D depictions when developing their forecasts. We have observed similar behavior in general. While the atmosphere is clearly three-dimensional, most forecasters have been educated in the two-dimensional meteorological world. This suggests, as noted earlier, that training in how to use 3-D products should be a priority. An interesting suggestion is to feature a "Case of the Month" on our homepage, with contributions from various WFOs, to demonstrate how D3D is being used in the operational setting. The longer term solution might be to include 3D visualization in meteorological education at the college level, which would likely yield future forecasters who would naturally expect to see 3D visualization tools on the operational workstation.
We optimistically believe that with management support and comprehensive training in its useD3D will greatly benefit preparation of the weather forecast and result in more accurate prediction.
Note: A complete list of references and more information on this and related topics are available at the main FSL Website www.fsl.noaa.gov, by clicking on "Publications" and "Research Articles."
(The FSL D3D Team includes Edward Szoke (meteorologist),Herb Grote (Supervisory electronics engineer/Systems Development Division Chief), Paula McCaslin (mathematician), and Philip McDonald (research associate). Ed Szoke can be reached by e-mail at firstname.lastname@example.org.)