Quantifying Improvements in Road Weather Forecasting from Sensor Observations

Timely and accurate information on current and future road conditions is a critical element in maintaining safe autonomous vehicle operations and efficient planning of winter road maintenance actions. A wide variety of road sensors are now available to provide a wealth of information on conditions prevailing on the road surface and in its vicinity. Sensor observation data are not only used to maintain an up-to-date awareness of evolving road conditions, they can also be used as input into hyper-local forecasts of future roadway conditions. Available sensor types range from sophisticated optical devices providing a remote method to estimate pavement temperature and how much water, snow and ice is present on the road surface, to the more common in situ devices measuring air temperature and humidity in the vicinity of a roadway. The recent development of low-cost wireless sensors is enabling a more affordable deployment of observations across road networks. These sensors can be deployed as “gap-filling” devices to achieve greater observation coverage in between reference stations or can serve as the single source of observations in a more affordable sensor network. In either case, it is important to quantify the extent to which road forecasts are improved when observations from different sensors are used as input (i.e. the observation impact).

A quantitative assessment of how road weather forecasts are improved by observations from various sensors is presented. A series of twin forecast experiments performed with a road weather model are compared, in which selected input observations are withheld, providing the basis for estimating observation impact on forecast errors. We focus the validation on road surface temperature and road state (i.e. whether it is dry, wet, or icy) as the most important parameters in road weather applications. Data from sensors included in Road Weather Information System (RWIS) stations, as well as from the two recently developed wireless GroundCast and TempCast sensors are considered in this study. These sensors differ by the variables they observe and the underlying measurement principles they use to estimate them. GroundCast is embedded in the roadway and provides in-situ measurements of road surface temperature and sub-surface temperatures at different depths, along with a simple classification of road surface state (dry or not). TempCast is a roadside device measuring air temperature and humidity as well as providing remote optical measurements of road surface temperature.

Results from the twin experiments confirm that observations from all sensors considered contribute to making road forecasts more accurate. The impact of road surface temperature observations is found to be roughly equivalent between RWIS and the wireless embedded sensors, with 50 to 70% reductions in forecast errors up to a lead time of 6 hours. The largest differences in observation impact pertain to road state forecasts and the information content provided by road state sensors. The largest improvement in forecasts (skill is doubled) is obtained when the most sophisticated optical road sensor is used, providing detailed information on the thicknesses of water, ice or snow layers on the surface. The embedded GroundCast sensor with its simpler dry/not-dry classification also provides a positive albeit smaller impact on road state forecasts (improvement reduced by half compared to the more detailed observations). In the absence of direct information on road state, as in the case with the TempCast sensor, forecasts show modest skill improvements over a “no observation” baseline. In such cases, the road state forecast skill comes from the indirect constraints provided by ancillary observations, such as local measurements of other variables affecting the energy and mass budgets at the road surface (e.g. air temperature and humidity), or from road state observations from other nearby locations. Overall, the availability of road state observations lead to more accurate road state forecasts with improvements occurring in the first 4 to 6 hours. This contrasts with road surface temperature where the benefits extend over the entire forecast period of 72 hours.