Open Source Machine Learning

Quickly analyze massive amounts of satellite imagery using machine learning and open data.

Left column - Input images (infrared, true color), © DigitalGlobe; Middle - OpenStreetMap data; Right - Our model prediction


Detecting objects and change across the globe

There is a growing amount of satellite imagery available, but more open tools are needed to provide insights from this data. Skynet uses modern image segmentation techniques to identify objects from overhead images. The tool is versatile – besides roads, we’ve also experimented with Skynet to detect buildings, electricity infrastructure, and other features of interest.

“Skynet models perform better in developing countries, because they are trained there.”

Machine learning has tremendous potential to improve the work of development organizations. Skynet provides a set of tools broadly aimed at analyzing satellite imagery to determine land use, map various types of infrastructure, and quickly determine the extent of natural disasters – all in an automated fashion.

Live prediction of roads in Manilla, Philippines


Bringing modern machine learning algorithms to satellite imagery

Skynet is an application of Segnet, a convolutional neural network approach for semantic segmentation. We've built a suite of tools around Skynet to collect and manage training data, inventory trained models, produce useful vectorized data outputs, and optimize cleaning of that data. One key support tool, skynet-data, ingests data from OpenStreetMap and prepares it for use as training data.

Skynet is designed to support open algorithm development. The Skynet code is entirely open source and is available on Github. Our workflow links directly to OpenStreetMap, and we encourage using OpenStreetMap as part of any field effort to collect training data. When we create a model using Skynet, we provide that model under an open source license on our Docker Hub page.