Development Seed has developed Skynet for extracting both road and building information using high resolution satellite imagery and deep-learning (https://developmentseed.org/projects/skynet/). For SEZ, we are using an existing Skynet model trained using OpenStreetMap data and Mapbox's satellite mosaic product. This trained model is used to predict roadway on multiple dates from discrete scenes sourced from Digital Globe's IDAHO image tiles. For SEZ, only the total km of roads for each scene is computed. The road network is generated by first calculating, for a set of zoom 17 tiles (~1.2 m per pixel) from Digital Globe, the likelihood that each pixel belongs to a road. Second, this image is binarized by thresholding the pixel values. Thirdly, the binary map is vectorized by skeletonization and converting the skeleton coordinates to GeoJSON linestrings. Finally, we compute the length (in km) of all roads (now represented in vector format) by simply adding the geographic length of every linestring. Because we can use a pretrained Skynet model, this indicator should scale to multiple zones assuming there are cloud computing resources to support it.

Conclusions

We found that we could obtain a rough estimate of roadway, with some areas performing better than others. However, there were a number of issues that we believe we could improve in future iterations. First, we noticed that purely quantifying road length can be misleading; particularly, when there are both asphalt and dirt roads in a zone. In the early development phase, some SEZs showed many winding dirt tracks before they were eventually replaced by a well-organized set of asphalt roads. In the future, we propose training a model that can differentiate between the two types of roads to provide a more informative measure of development. Second, we found that the model had some issues generalizing to SEZ zones after being trained on other regions. This was expected as the characteristics and features of roads common to one region of the world are unlikely to be exactly repeated in another. With time to prepare a better training dataset, we can likely better train a model that can recognize roadways in a wider range of environments. Finally, we found that Skynet often gave predictions with many disconnected (or severed) roads. Currently, we are testing new versions of the Skynet model with different model architectures and training configurations based on current deep learning research. As this research has demonstrated some impressive improvements recently, we are confident we can translate this into improvements in road detection in the near future.