DigitalGlobe has recently launched SpaceNet, an online repository of satellite imagery and associated training data, for users to experiment with machine learning and deep learning algorithms. Spacenet has been launched as a collaboration between DigitalGlobe, CosmiQ Works, and NVIDIA, and is available as a public dataset on Amazon Web Services (AWS). As a first step, SpaceNet will contain DigitalGlobe’s high resolution multispectral imagery from their premier WorldView-2 satellite at its industry-leading full 8-band spectral resolution and over 200,000 curated building footbrints across Rio de Janeiro, Brazil. This is unprecedented: Never before has satellite imagery at such high resolution of 50 cm been released publicly with building annotations. The released dataset contains over 7000 images over Rio de Janeiro. The satellite imagery is being delivered in GeoTIFF format while the building footprints are in GeoJSON format.
True color WV-2 high resolution imagery sample from the SpaceNet repository, along with corresponding building footprints. Source: NVIDIA
According to SpaceNet:
This dataset is being made public to advance the development of algorithms to automatically extract geometric features such as roads, building footprints, and points of interest using satellite imagery.
Scripts are already cropping up on GitHub for manipulating and using the satellite imagery data on SpaceNet: see code examples from Development Seed here and from CosmiQ Works here. NVIDIA has also released a detailed case study of analysis of SpaceNet data using their Deep Learning GPU Training System (DIGITS) platform, demonstrating the power and capability of GPU-based deep learning algorithms applied over high resolution satellite imagery. Application examples include detection of each building as a separate object and determining a bounding box around it, and semantic segmentation to partition the image into regions of pixels that can be given a common label, such as “building”, “forest”, “road”, or “water”.
SpaceNet plans a massive increase in both images and labeled features to be made available over the platform in the future. Incidentally, the name SpaceNet is inspired from ImageNet, a similar database of images created to help spur early advancements in computer vision.
To read more about the launch of SpaceNet, see coverage on GISCafe, TechCrunch, MIT Technology Review, and Popular Science.
SpaceNet datasets can be accessed on AWS here.