After many years of client requests, we’re releasing the first cut of our StreetLight Volume: 2016 AADT Metric on the StreetLight InSight® platform. These beta Metrics provide a very robust estimate of 2016 Annual Average Daily Traffic (AADT) for almost any road in the US. To learn more, keep reading this blog post, or watch our recorded webinar on the new Metrics. Click here to watch the webinar.
We believe this Metric provides estimates that are comparable or better than most of the standard AADT estimation practices for three reasons:
In this blog post, we’ll walk you through our process for creating this Metric. It covers defining goals and developing our methodology as well as the validation and QA work that make us very confident in this Metric. In short, we predicted ~25,000 Virginia Department of Transportation AADTs with an R2 of 0.87 and average error of 22% (see Figure 1 below).
Figure 1: Scatter plot of StreetLight Volume: 2016 AADT values and VDOT AADTs for continuous loop VDOT counters (type A).
For this initial release, we’re focusing on simple AADT. Our goal for this beta release is to ensure the technique works on large and small roads, in rural and urban parts of the country. Collecting AADT data on far away, rural roads is particularly expensive for our clients, so we hope this is a high value place to begin. We’ll expand to other types of count estimates in the future, such as hourly, seasonal, and more.
This Metric is our most advanced yet. It combines both Location-Based Services (LBS) and navigation-GPS data, as well as the US census and a set of well-validated loop counters. We know others have previously tried to build volume analytics and estimated counts from probe data with varying levels of success. For us, these Metrics are only feasible because we have both GPS data (this data source excellent spatial precision and limited sample size) and LBS data (this data source has better normalization potential and larger sample size), and we are able to integrate these unique data sets.
First, we used the census data to normalize the LBS data (see our sample normalization blog for more on this). Next, we used the navigation GPS data and LBS data to balance each other and scale most of the way up to AADT.
Finally, we used a few high-quality AADT counters from some of our DOT partners and a subset of the TMAS database to fine-tune the algorithm. We determined that the algorithm works best for roads with over 400 AADT.
After we developed the algorithm, we looked for large sets of AADT to test it against. In this blog post, we will show you the test we performed on VDOT’s public AADT file.
To do this test, we used our algorithm to generate estimated AADTs for ~25,000 links in VDOT's public AADT file. These links cover the entire state: big roads and small, urban and rural. We ran the algorithm “blind,’ which means we did not use VDOT’s AADTs when creating the algorithm.
Note: To generate 25,000 links effeciently for use in StreetLight InSight, we had to make some assumptions when we uploaded the VDOT shapefile. (Check out our tutorial on "Making Road Segment Zones" to see why.) We hypothesize that this quick assumption-making to draw the location of the gates could be a source of error, and that these errors could be avoided by our clients who know their own links in detail.
We had to drop out a few segments for obvious flaws. For example, VDOT had a value of 0 or “NA” for some AADTs. Next, we calculated the error for each link (the difference between the VDOT AADT estimate and StreetLight Volume: 2016 AADT, as a percent of the VDOT AADT). We also took a deeper look at some of the outlier error locations.
For example, we had three Zones that were classified as “A”, or high-quality continuous loop counter data from VDOT, that were also outliers. These loop counters are located in Bristol, VA (see Figure 2 below). The pink lines are VDOT’s road segments. The blue bar is our AADT gate. VDOT has three different values on West Main Street (14000 at A, 8800 at B, and 23000 at C). StreetLight InSight estimated an AADT of 22,944 for the bidirectional segment.
For the A and B segments, we see two different values - even though there is no way for a car to get off the road between them. This immediately raises questions about the VDOT data. Perhaps one link is misplaced or mislabeled.
Next, all three VDOT values purport to be the bidirectional AADT. Thus, both sides of the street should show the same value, as is common in other split roads in the VDOT file. If we look at Segment C, our estimates are extremely close (less than 2% off). Thus, we considered the errors on A and B to be outliers, and we dropped them from the analysis. We eliminated points like this from the sample.
Figure 2: Bristol outliers -- as shown above, StreetLight Volume: 2016 Metrics match Segment C extremely closely.
Next, we compared our results to the cleaned VDOT data for validation purposes. We wanted to see three outcomes:
As shown in Table 1 below, the average absolute error and the correlation are very low and high, respectively. The worst average absolute error is for Class N- AADT of Similar Neighboring Link, which is the lowest quality form of (non) count.
As shown in Figure 3 below, the algorithm works similarly well for the smaller roads and the large highways. The slight staircase shape of the scatter plot is a result of the fact that we followed VDOT’s rounding conventions (to the nearest hundred or thousand).
Figure 3: Scatter plot of StreetLight Volume: 2016 AADT and VDOT AADT for ~2500 locations.
This is the first public iteration of our StreetLight Volume: AADT Metric. We’d like to hear from you:
Leave a comment below to share your feedback, or send us an email at email@example.com. For a more detailed overview of these Metrics, join our webinar on June 16th at 12pm PT. Click here to register for the webinar.
We know that our clients want other types of estimated counts – such as by time or day, weekday/weekend, season, and for origin-destination and other studies. These are coming soon! We want to collect feedback on our initial AADT Metric, and then use that feedback to guide upcoming product developments.