Traffic flow is measured by counting cars on workdays over time. This is one of the more difficult indicators because cars are significantly smaller than buildings implying that this indicator requires relatively higher resolution imagery. Also, unlike buildings and roads which are mapped in OSM, cars are not static so there is no equivalent source of open training data -- training data must be manually labeled. We attempted to use GBDX’s built in car-counting algorithm, but found that it was not suitable for this project. Their current workflow is not designed to run on small regions -- instead it only accepts large catalog regions, which are vastly bigger than the zones we are interested in. It also gave very poor results with predicted cars ranging from a few to tens of meters long. This approach would not be accurate or cost effective when scaling up to hundreds of SEZs.

In the near future, we hope to create (and openly release) a training data set consisting of satellite imagery at 0.15 or 0.3 m/pixel resolution (zoom 19 and 18, respectively) with matching raster labels. Ideally, this data would be from developing regions, though we did discover an open dataset (Cars Overhead with Context; COWC) that would be useful for training a car detection model on Western countries. Regardless, we plan to develop a car counting model through deep learning object-detection methods in the near future by building off recent results. The CosmiQ group, for example, has shown that the Regional Convolutional Neural Network (R-CNN) and You Only Look Twice (YOLT) model can obtain good car detection performance with F1 scores of 0.9 or slightly better on zoom 19 satellite imagery. If we are able to develop these models, they will require cloud computing resources to scale up to multiple zones. Of all the indicators, this would likely be the most computationally intensive.