Vehicle data translator for road weather monitoring

Sheldon Drobot, Michael Chapman and Amanda Anderson, NCAR, and Paul Pisano, FHWA, detail latest results of testing of a vehicle data translator for road weather monitoring and information applications

The use of vehicle sensor data to improve weather and road condition products, envisioned as part of the 324 US Department of Transportation 321 Research and Innovative Technology Administration's (RITA's) IntelliDriveSM initiative, could revolutionise the provision of road weather information to transportation system decision-makers, including travellers. For example, vehicle-based probe data will significantly increase the density of weather observations near the surface and also provide unique datasets for deriving and inferring road condition information (Drobot et al. 2009).

With funding and support from RITA and direction from the 831 Federal Highway Administration's (FHWA's) Road Weather Management Program, the 2065 National Center for Atmospheric Research (NCAR) has spent the previous two years conducting research to develop a Vehicle Data Translator (VDT) that incorporates vehicle-based measurements of the road and surrounding atmosphere with other, more traditional weather data sources and creates road and atmospheric hazard products for a variety of users.

Test systems

The initial research was conducted at the Development Test Environment (DTE) facility located near Detroit, Michigan using six 1957 Jeep Grand Cherokees and three 278 Ford Edge vehicles. The researchers collected over 239,000 air temperature and pressure observations during 19 days of testing, spanning 28 January to 29 March 2010. These data were compiled directly from stock sensors. On both the Edges and Cherokees, the temperature sensors are located in the front bumper, where previous research indicated the least bias is expected (Mitretek 2006). For atmospheric pressure, the Jeeps rely on a computational formula based on the Manifold Absolute Pressure (MAP) sensor, and the Fords use a technique based on the Mass Air Flow (MAF) sensor.

In addition to these sensors, temperature, dewpoint and surface temperature data were collected by 144 Vaisala Surface Patrol HD sensors mounted to the left-front quarter-panel of each vehicle, providing observations that were collocated with individual vehicle observations as another source of ground truth data to compare with the vehicles. A fixed surface weather station, the Vaisala WXT-520, was also set up at the test facility to capture more representative surface weather data than the Detroit (KDTW) Automated Surface Observing System (ASOS) station, located about 30 miles away from testing.

Ensuring accuracy

One of the initial questions the researchers faced centred on the accuracy and bias of in situ vehicle temperature and pressure measurements. After quality checking the observations, and using well-calibrated meteorological instruments as a validation source, comparisons indicate strong correspondence between quality check-passed data from the nine test vehicles and the in situ sensor for air temperature measurements, but some disparity for barometric pressure.

Correlations are high in both cases (Table 1), but the bias and Mean Absolute Error (MAE) are far superior for temperature, indicating greater confidence in the temperature measurements. Without quality checking, the results degrade rapidly, particularly for pressure. In fact, for some vehicles more than half of the pressure observations failed the quality check procedure and if these data were used in the accuracy and bias computations, errors for some vehicles would exceed 30hPa with even larger bias numbers. Additional research is needed to ascertain precisely why the pressure data is subject to large errors but it clearly points to the need for rigorous quality checking of the raw data.

The researchers were also interested in determining whether variations in ambient conditions (for example, ambient air temperature variations, and precipitation versus no precipitation) or vehicle factors (such as vehicle speed or model) altered the overall results. If so, this would make the interpretation and use of mobile data a more complex operation.

Fortunately, there is little evidence that environmental factors alter the statistics shown in Table 1 (again, once the data are quality-checked). When stratified by ambient air temperature, time of day, day, wind direction and speed, or precipitation condition, there are more appreciable differences across the stratifications. Unfortunately, there is some evidence to suggest vehicles do introduce unique factors, though not for every condition. For example, error rates were consistent no matter what the vehicle speed, but when stratifying the statistics by vehicle (Figure 1 and Figure 2), some differences between vehicles are seen, particularly regarding bias.

The temperature bias shows variation not only between make of vehicle (Ford Edge versus Jeep Grand Cherokee), but also variation between vehicles of the same make. The MAE shows slightly less variation, with the Fords having a slightly higher MAE than the Jeeps. Correlation is high among all vehicles. For pressure, all vehicles exhibit a negative bias. Other non-environmental factors, such as vehicle colour and vehicle speed, had little demonstrable impact on the results.

Representativeness

One of the outstanding remaining questions revolves around the representativeness of these results for other makes and models. Additional research will be needed to determine that. Nonetheless, with confidence that these data are reliable, the researchers have spent time developing a suite of algorithms to provide drivers with road hazard warnings.

These include 'road precipitation', which uses a combination of ambient air temperature, radar reflectivity, windshield wiper status, headlight status and the ratio of vehicle speed to road segment speed limit to determine what type of precipitation is influencing the roadway. Other algorithms include 'pavement slickness potential', 'visibility', and an 'all hazards' algorithm combining output from the three individual algorithms.

Currently, the output is displayed on a website with a 1691 Google Maps underlay, and coverage extends over most of North America. In addition to the road hazards mentioned above, a variety of other information is available for the user. For example, raw vehicle output (speed, average temperature) is available on one-mile road segments with a five-minute update time. Gridded radar data and NWS storm reports are also selectable, as shown in Figure 3.

Now that the VDT has been successfully tested, NCAR and the USDOT are moving toward increasing the capabilities of the VDT. If you would like more information on VDT development, please contact Sheldon Drobot
(+1 303 497 2705; Email000oLinkEmaildrobot@ucar.eduSheldon Drobotfalsemailto:drobot@ucar.edutruefalse%>).The NWS storm reports consist of hail, wind and tornado reports gathered from External000oLinkExternalhttp://www.spc.noaa.gov/climo/onlineStorm Prediction Centerfalsehttp://www.spc.noaa.gov/climo/onlinefalsefalse%>. In addition, users can view Rapid Update Cycle (RUC) Surface Assimilation Systems (RSAS) air and dewpoint temperature, which are updated every 15 minutes on a 12km square grid, as well as the Geostationary Operational Environmental Satellite (GOES) cloud mask.

Another feature of the VDT system is the incorporation of 2066 National Weather Service (NWS) polygon warnings. Users can click on the link in the pop-up window and be redirected to additional details directly from the NWS. Moreover, the NWS watch/warning/advisory county-level warnings are also selectable. These warnings mirror the data from External000oLinkExternalhttp://www.nws.noaa.govNational Weather Servicefalsehttp://www.nws.noaa.gov/falsefalse%> but at a much higher resolution and with the other added features of the VDT. Finally, the NWS 'Story of the Day' is also selectable. This product might be used for daily planning by users.

Table 1. Comparison between vehicle and in situ measurements

	Temperature (°C)	Pressure (hPa)
Bias	-0.21	-4.33
MAE	0.84	4.95
Correlation	0.99	0.91

References
• Drobot, S.D., W.P. Mahoney III, P.A. Pisano, and B.B. McKeever, 2009: Tomorrow's Forecast: Informed Drivers.
• ITS International, July-August 2009, pp.NA1-NA2.
• Mitretek, 2006: Vehicles as Mobile Sensing Platforms for Meteorological Observations: Introductory Research during a Winter Season. 267 pp.