Today, in India, Google maps will route you to the hospital, but it may steer you directly into a floodpath. We wanted to take daily satellite images and remote sensor network data to measure and predict water flow across roads in flood prone areas, and use that information to provide safe and effective routes for commuters and merchants.
Accomplishments that we're proud of
- We built a dynamic visualization of the flood conditions throughout the world that shows the the beating heart of the tropical monsoon.
- We trained a deep CNN to recognize flood patterns and ocean currents from satellite telemetry.
- We identified flood events in the past and present where our tech would be critical for safely moving around people and goods.
- We parsed about 2.5 million records of raw flood data stored as HTML tables like this, generating a huge dataset suitable for earth sciences and ML research.
What we learned
Processing satellite data requires lots of attention to detail. There are subtleties about many sources of earth systems data. For instance, clouds often interfere with satellite images. Different countries report water levels using different notation. On the other hand, there is a lot of earth data. It's possible to throw away lots of messy data and use what's well-understood to get started quickly.
What's next for EarthFlow
In its current domain, southeast Asia, Earthflow needs more sources of data and a pipeline for better models to predict and classify flood conditions at a level suitable for day-to-day use by individuals. We also need to enable dynamically loading data from the API as the user rescales their map. We plan to expand EarthFlow to other continents facing flooding issues in their transit systems.