The SpaceNet 7 Multi-Temporal Urban Development Challenge: Dataset Release

Aug. 06, 2020

Adam Van Etten

Preface: SpaceNet LLC is a nonprofit organization dedicated to accelerating open source, artificial intelligence applied research for geospatial applications, specifically foundational mapping (i.e., building footprint & road network detection). SpaceNet is run in collaboration by co-founder and managing partner CosmiQ Works, co-founder and co-chair Maxar Technologies, and our partners including Amazon Web Services (AWS), Capella Space, Topcoder, IEEE GRSS, the National Geospatial-Intelligence Agency and Planet.

The SpaceNet partners are proud to announce the release of the SpaceNet 7 dataset. The SpaceNet 7 Multi-Temporal Urban Development Challenge dataset features a deep temporal stack of imagery and labels in over 100 locations across all six inhabited continents. Up-to-date high-fidelity maps are crucial for many applications (such as disaster response or humanitarian efforts, discussed in great detail in our previous blog), and this dataset will aid efforts to improve automated mapping and overhead change detection methods.

This is the first time that SpaceNet has used Planet imagery at ~4 meter resolution, as well as the first time that the temporal dimension will be explicitly incorporated into a challenge. The selected locations boast significant change over the two year data collection timeline. The dynamic nature of the data cubes permits tracking of urban development, specifically: building footprint evolution as well as address propagation. In the sections below we discuss the many facets of this unique open source dataset, distributed under a permissive CC BY-SA 4.0 license.

1. Geographic Coverage

The SpaceNet 7 dataset contains ~100 locations, spread out across the globe. Many of the locations were selected to highlight signifiant change, though some were selected to overlap with previous SpaceNet locales, and some were selected due to geopolitical interest. 60 of the data cubes are released as training examples (with both imagery and attendant labels), and 20 are released as test_public examples (imagery only). The remainder of the data cubes are reserved for final testing purposes. Figure 1 links to an interactive map that allows interested readers to explore the dataset locations.

Figure 1. Map of SpaceNet 7 locations (click map to visit interactive plot). Orange flags denote test_public locations, blue flags denote training locations, and the SpaceNet icon denotes collections over existing SpaceNet cities.

2. Data Cubes — Imagery

At each location, monthly mosaics are curated over a timespan of two years. This lengthy time span captures multiple seasons and atmospheric conditions, as well as the commencement and completion of multiple construction projects, see Figure 2.

Figure 2. Sample training data cube. Note the significant construction, seasonal changes, and varying atmospheric conditions.

Cloud cover or (rarely) sensor mis-calibrations render some regions of images unusable. Accordingly, each image has an attendant unusable data mask (UDM) in standard GeoJSON format, see Figure 3.

Figure 3. Left: Sample training image. Right: Unusable regions of the image masked out.

3. Data Cubes — Building Footprint Labels

For each monthly mosaic, the SpaceNet labeling team painstakingly outlined the footprint of each building. These GeoJSON vector labels permit tracking of individual building locations (i.e. addresses) over time, hence the moniker: SpaceNet 7 Urban Development Challenge. See Figure 4 for an example of the building footprint labels in one of the training cities.

Figure 4. Building footprint labels rendered as a mask. Note both the ongoing construction, as well as the frequent “flickering” of buildings due to cloud cover.

While building masks are useful for visualization (and for training deep learning segmentation algorithms) the precise vector labels of the SpaceNet 7 dataset permit the assignment of a unique identifier (i.e. address) to each building. Matching these building addresses between time steps is a central theme of the SpaceNet 7 challenge. In Figure 5 we display these building address changes.

Figure 5. Ground truth building footprints and addresses at different time steps.

4. Aggregate Stats

5. Dataset Access

As always, the SpaceNet data is freely available on AWS. As with previous data releases (e.g. SpaceNet 6) all you need is an AWS account and the AWS CLI installed and configured. Once you’ve done that, simply run the command below to download the training dataset to your working directory:

aws s3 cp s3://spacenet-dataset/spacenet/SN7_buildings/tarballs/SN7_buildings_train.tar.gz .

6. Conclusions

Detecting change in overhead imagery is a difficult task, greatly complicated by seasonal, atmospheric, and lighting effects. Yet the ability to localize and track the change in building footprints over time is an important facet in a number of applications, from disaster response to disease preparedness to environmental monitoring. Furthermore, this task poses interesting technical challenges for the computer vision community.

The SpaceNet 7 dataset provides a solid foundation for advances in these areas, via a large corpus of imagery and labels spanning over 100 distinct locations, greater than 40,000 square kilometers of observed area, and over 10 million labeled buildings.

Stay tuned for further updates on the upcoming SpaceNet 7 challenge and metric that will accompany this dataset (also see spacenet.ai for further resources). This challenge launches August 31, 2020, and will be featured as a competition at the 2020 NeurIPS conference in December.

* Special thanks to Jesus Martinez Manso for spearheading dataset creation.

2020-Aug-10 correction: The introduction was amended to specify the SpaceNet 7 dataset license (CC BY-SA 4.0)