Satellite image data used to train a computer vision model

Case study

Detailed scans of the Earth’s surface by satellites provide a rich source of image data used in industries ranging from agriculture to transportation. For this project, the simulation software company, Presagis, commissioned us to develop a computer vision model using satellite images of rooftops. Our goal was to produce an algorithm capable of assessing the structural integrity of roof surfaces and buildings in order to identify possible landing areas for a helicopter. Such capacities will streamline the planning of emergency responses where, for instance, responders will know if they can land a helicopter on a building to rescue people stranded by severe floods.

Here we provide examples of the dataset and how the trained convolutional neural network is capable of identifying basic structural features on the surface of buildings. Using the visual and intuitive environment of the Zetane Engine, we can assess in detail how each layer of our neural network processes each image file (e.g., view layer weights, biases and edge weights during inference passes). Members from Presagis later used the visual representation of the project in the Zetane Engine to present their technology at prominent trade shows.

This case study exemplifies the development an object recognition solution using the popular neural network VGG16.