Label Maker

Data Preparation for Satellite Machine Learning

Label Maker downloads OpenStreetMap QA Tile information and satellite imagery tiles and saves them as a file for use in machine learning training.

Building classification labels overlaid over Mapbox satellite imagery


Bringing Machine Learning to International Development

Most commercial machine learning algorithms are built to work in the US and Europe, where it is easy to find rich training data and deep archives of high-resolution imagery. These models often perform very poorly when generalizing to non-Western countries, where roads and buildings have different design and composition.

Label Maker is one part of an effort to make it easy to create AI models that work anywhere in the world. Label Maker quickly creates machine learning training data in any country – including developing nations. This ensures that we can train models specific to those areas’ unique features and get better results. It pulls label data from OpenStreetMap, an extremely rich map of the world and pairs that data with satellite imagery data.

Label Maker doesn’t completely solve the problem of AI in developing countries. There are still parts of the world that are undermapped in OpenStreetMap and labels, like crop that are sparse. But if you can get the data into OSM (and we can help with that) then you can be off to training a model in minutes.


How we Built This

Label Maker builds on the data preparation techniques started in Skynet. But it builds on those ideas to support more types of machine learning problems: it can generate classification labels, object detection bounding boxes, or segmentation masks. It also creates a framework-agnostic data package which allows you to drop your data into popular tools like TensorFlow, Keras, or MXNet.


Show this open source project some love.

Label Maker would love your help. Suggest new features or help update the documentation.

View Guidelines