The SpaceNet 8 Flood Detection Challenge: Dataset and Algorithmic Baseline Release
Editor’s note: SpaceNet is an initiative dedicated to accelerating open-source, artificial intelligence applied research for geospatial applications, specifically foundational mapping (i.e., building footprint and road network detection). SpaceNet is run by co-founder Maxar and our partners IEEE GRSS, Oak Ridge National Laboratory, Amazon Web Services (AWS) and Topcoder.
The SpaceNet partners are proud to report the release of the SpaceNet 8 dataset and algorithmic baseline.
The SpaceNet 8 Challenge launched July 12 and is off to a great start with more than 180 participants registered via Topcoder in the first two weeks. As explained in the previous blog post announcing SpaceNet 8, this challenge focuses on infrastructure and flood mapping related to hurricanes and heavy rains that cause route obstructions and significant damage.
SpaceNet 8 expands on past infrastructure mapping tasks for building footprints and roads to a multiclass feature extraction and characterization problem. This is the first SpaceNet dataset to also incorporate examples of such infrastructure affected by flooding. Like past SpaceNet challenges, a new dataset and baseline algorithm are provided. This blog post provides more detail on each.
Three areas of interest (AOIs) were selected for the dataset consisting of 12 Maxar satellite images of both pre- and post-flooding event imagery. Along with the imagery, hand labeled building footprints, road and flood attributes are provided for training and scoring. The AOIs include Germany with flooding from heavy rains in July 2021, Louisiana following Hurricane Ida in August 2021, and a “mystery” location that will be used to test the top 10 algorithms from the public leaderboard for final scoring after the challenge has concluded.
To download the data, you simply need a free Amazon Web Services (AWS) account and the AWS command line interface (CLI) installed and configured. The following commands can be used to download the training dataset:
- aws s3 ls s3://spacenet-dataset/spacenet/SN8_floods/
- aws s3 cp s3://spacenet-dataset/spacenet/SN8_floods/tarballs/Germany_Training_Public.tar.gz
- aws s3 cp s3://spacenet-dataset/spacenet/SN8_floods/tarballs/Louisiana-East_Training_Public.tar.gz
- aws s3 cp s3://spacenet-dataset/spacenet/SN8_floods/tarballs/Louisiana-West_Test_Public.tar.gz
In addition to the new training datasets provided for SpaceNet 8, prior SpaceNet datasets can be used as examples of building footprints and roads.
The Baseline Algorithm
The SpaceNet team also released a new baseline algorithm to provide a starting point for this challenge. SpaceNet 8 participants may choose to modify the baseline or to develop an entirely new algorithm.
Our baseline consists of two independently trained convolutional neural networks and post-processing steps to convert rasterized predictions into vector data suitable for submission. The first network focuses on foundational features, which works by segmenting buildings and roads from the pre-event imagery without any flood attribution. The second, a flood attribution network, predicts the flood attributes of roads and buildings by using both pre- and post-event imagery.
A final post-processing step merges the flood predictions into the vector representation of the foundation feature predictions, resulting in a suitable submission for the Topcoder challenge evaluation server.
The baseline algorithm is available via the SpaceNet GitHub repository here.
Our recent paper presented at EarthVision, a Computer Vision and Pattern Recognition Conference (CVPR) workshop, “SpaceNet 8 — The Detection of Flooded Roads and Buildings,” describes both the dataset and baseline algorithm in detail.
We hope you find this information helpful for SpaceNet 8 and wish you the best of luck throughout the competition. We can’t wait to review your submissions and share the challenge results!
Questions? Email us at firstname.lastname@example.org.