PS 73-131
Geostatistical modeling of primary and secondary automobile traffic volume as an ecological disturbance proxy across the contiguous United States

Thursday, August 13, 2015
Exhibit Hall, Baltimore Convention Center
Sean L. McFall, The College of William & Mary, Williamsburg, VA
Matthias Leu, Biology, College of William & Mary, Williamsburg, VA
Benoit Parmentier, Sustainability Solutions Initiative, University of Maine, Orono, ME
Stuart E. Hamilton, Geography and Geosciences, Salisbury University
Marco Millones, Center for Geospatial Analysis, The College of William & Mary

   Roads have deleterious effects on ecosystem functions and landscape connectivity by providing habitat corridors for invasive plants and disrupting animal movement. To model negative effects of roads on ecological processes, researchers commonly use road density and distance to road as covariates but rarely include traffic volume as a proxy. Assuming homogenous effects of roads on ecological processes can lead to spurious results as traffic volume varies spatially. Our intent is to use traffic volume as proxy for ecological disturbance rather than an estimate of vehicle flow  for human transportation applications.

We created a nation-wide traffic-disturbance layer using Annual Average Daily Traffic (AADT) point data collected by the Departments of Transportation in the 48 contiguous states. Our product predicts traffic volume for primary roads, such as freeways or interstates, and less-traveled secondary roads, such as regional highways. We randomly selected 70% of each state’s AADT data for model fitting and used the remaining 30% for model validation. We interpolated the AADT data points with ordinary kriging in R (autoMap and gstat libraries). To standardize differences in traffic volume data collection among states, we included data from neighboring states within a buffer based on the adjacent states variogram range values.


   Preliminary results yield an expected and reasonable predicted surface values across the landscape. Also as expected,  the respective error surfaces communicate low error values around known points, and then become higher in error as distance from known points increases. However, states with few points tend to skew towards the average value of the entire layer. To account for this, we plan on combining AADT data points with a distance to road covariate or something suitably similar.

Validation for all training datasets showed that over 80 percent of states have R2 values greater than 0.8. RMSE and MAE metrics are all under 10 percent of each states  for training and testing datasets.

Similar datasets have been produced for primary roads, though none have included both primary and secondary roads, let alone for the lower 48 states. We are confident this novel dataset will prove a substantial improvement over available proxies for ecological disturbance caused by road activity.