Microsoft AI for Earth has teamed up with DrivenData and Cloud to Street to evaluate the applicability of machine learning models for near real-time flood detection. The goal of this challenge is to build machine learning algorithms that are able to map floodwater using Sentinel-1 global SAR imagery. Models that leverage SAR imaging will significantly improve flood risk assessment, relief targeting, and ultimately, disaster readiness and response.
This was our first experience with space imagery. Main requisition of the competition was to create a model based on SAR images.
SAR (synthetic-aperture radar) images are obtained using synthetic aperture radar C-SAR (developed by Astrium), which provides an all-weather, as well as round-the-clock delivery of satellite images.
Our data scientist team drew on the experience of colleagues who had already used Sentinel-1 radar imagery to simulate avalanches. To solve the problem, the Novel team used the Optuna hyperparameter optimization framework, the Catalyst prototyping framework and Neptune for prototyping.
“The choice of these tools allowed us to establish the interaction of team members, track model errors and quickly validate emerging ideas,” - says Nikolai Russkikh, head of ML department.
Novel team took the 14th place out of 664 participants. Excellent result!