/stac-overflow

Winners of the STAC Overflow: Map Floodwater from Radar Imagery competition

Primary LanguageJupyter NotebookMIT LicenseMIT



STAC Overflow: Map Floodwater from Radar Imagery

Goal of the Competition

Flooding is the most frequent and costly natural disaster in the world. According to the World Health Organization, between 1998 and 2017, floods affected more than two billion people worldwide. As global warming continues to exacerbate rising sea levels, prolong rainstorms, and increase snowmelt, the frequency and severity of extreme weather is only expected to rise.

During a flood event, it is critical that humanitarian organizations be able to accurately measure flood extent in near real-time to strengthen early warning systems and target relief. Historically, hydrologists have relied on readings from rain and stream gauging stations to understand flood reach. Though helpful, these ground measures only measure water height, are spatially limited, and can be expensive to maintain. High resolution synthetic-aperture radar (SAR) imaging has strengthened monitoring systems by providing data in otherwise inaccessible areas at frequent time intervals. Specifically, C-band SAR from the Sentinel-1 mission provides an all-weather, day-and-night supply of images of the Earth’s surface.

The goal of this challenge is to build machine learning algorithms that are able to map floodwater using Sentinel-1 global SAR imagery. Microsoft AI for Earth has teamed up with DrivenData and Cloud to Street to investigate the applicability of machine learning models for detecting flood coverage in near real-time. Models that can effectively use remote sensing to improve flood mapping have the potential to strengthen flood risk assessment, relief targeting, and disaster preparedness.

What's in this Repository

This repository contains code from winning competitors in the STAC Overflow: Map Floodwater from Radar Imagery DrivenData challenge. Code for all winning solutions are open source under the MIT License.

Winning code for other DrivenData competitions is available in the competition-winners repository.

Winning Submissions

Place Team or User Public Score Private Score Summary of Model
1 Moscow Hares: sweetlhare, Belass 0.899 0.809 Create pixel-level vector representations of raster data by flattening VV, VH, and auxiliary bands. Train a set of CatBoostClassifiers over 1000 iterations using sequential, regional, k-fold, and stratified k-fold sampling, and average predictions. Finetune pretrained U-Net CNNs with EfficientNet-B0 and B4 backbones using samples of standardized, 3-channel images (VV, VH, & NASADEM) with flip, rotation, crop, and blurring augmentations. Train one model on weakly performing images. Ensemble CatBoost and U-Net predictions using the maximum output. Missing pixels are treated as zero.
2 Max_Lutz 0.908 0.807 Create 9-channel images by stacking VH, VV, and auxiliary input bands. Clip VV and VH bands, scale all values to 0-255, and concatenate additional flipped and rotated augmentations. Train 3 randomly initialized U-Net CNNs using different train-test-splits and a consistent loss function of dice loss with a squared denominator. Ensemble predictions using the mean.
3 loweew 0.898 0.804 Create standardized 9-channel images by stacking VH, VV, and auxiliary input bands. Apply augmentations including cropping, rotation, translation, scaling, blurring, affine transformation, grid and optical distortion, grid shuffle, and brightness contrast. Use a U-Net++ CNN with an EfficientNet-B8 backbone pretrained using AdvProp, which treats adversarial examples as additional input, to assign difficult floods to the training set. Finetune U-Net++ CNNs with an EfficientNet-L2 backbone pretrained with Noisy Student Training, and use Focal Tversky loss to address data imbalance. Take a “cross-validation as a jury” approach to average predictions.

Additional solution details, including tips for using subsets of the winning models, can be found in the reports folder inside the directory for each submission. To request access to pre-computed survey weights, please contact us.

Winners Announcement: Meet the Winners of STAC Overflow

Benchmark Blog Post: How to Map Floodwater from Radar Imagery using Semantic Segmentation

Competition Sponsor