SafeGraph is excited to announce a collection of premium geometry rows depicting the shape and size of surface parking lots in the US. Parking Lots data shows the relationship between 6M US places and the surrounding parking lot(s) likely serving those POIs. Looking at parking lots as an extension of the places they serve, this data provides additional insight into the places our customers care about.
In this blog we will explore the methodology used in creating SafeGraph Parking Lots.
SafeGraph Parking Lots provides a 2D polygon that follows the boundaries of parking lots as precisely as possible. Creating these parking lot polygons was a two-step process: first, with the aid of a trained dataset, AI algorithms were able to recognize and delineate individual parking lots from satellite imagery covering the entire US. Next, we refined these results by cleaning up the data and managing edge cases. This was necessary as the results from the AI varied widely. Let’s have a look at some of these challenges and edge cases below.
While AI is able to “see” parking lots as 2D polygons and delineate them as such from satellite imagery, it is not good at matching the edges of drawn boundaries to where they are in reality. This would result in a single parking lot being split up into multiple polygons that in reality should have been merged together as a single geometry.
We solved this problem by filtering the data, using a minimum threshold value of 100 square meters for a single parking lot, which can host a maximum of six cars. This approach filters out many small, individual parking lots that got counted as separate parking lots but are in fact part of a larger geometry. Also, it turned out that some large parking lot features consisted of many smaller parking lots that were joined together incorrectly. Using a similar method, these were also filtered out of the dataset.
Similar to incorrectly delineating the edges of a parking lot, the AI would incorrectly create parking lot features with holes inside them, resulting in “donut” polygons. Examples are polygons drawn around parked cars on the parking lot itself or where a tree forms a shadow on the satellite imagery. This is obviously bad data that needs to be filtered out of the dataset. To do this, a statistical metric was calculated for each parking lot feature that measures the size of a parking lot relative to its parameter. Low values could indicate bad data, but not always.
Because the AI would often incorrectly mark smaller holes but correctly mark large ones, we decided to measure the size of each hole inside a parking lot polygon, and eliminate the ones that didn’t meet a specific threshold value, so that spatially correct holes would be kept - for example, holes indicating an apartment building structure with a parking lot around it.
Finally, drawing straight lines around the edges of a parking lot turned out to be a problem for the AI. While you only need two edge points to draw a straight line on a map to delineate one side of a 2D feature, the AI would instead draw jagged lines between extra, unnecessary coordinates. Applying simplification of these geometries resulted in nicer looking polygons and less data points, which makes it easier for customers to consume, process, transfer and visualize the Parking Lots data.
Here are a few examples of ways data on parking lots can provide greater context to specific use cases.
As with all of our datasets, we will continue to enhance our parking lot coverage and grow it to meet peoples’ use cases. Interested in checking out the data yourself? Download a free sample of SafeGraph Parking Lots.