Towards global flood mapping onboard low cost satellites with machine learning

Area of Science:

• Earth observation
• Satellite technology
• Disaster management

Background:

• Spaceborne Earth observation is crucial for flood response.
• CubeSats offer reduced revisit times for disaster monitoring.
• CubeSat data transmission is limited by power and bandwidth.

Purpose of the Study:

• Demonstrate onboard processing for CubeSat flood mapping.
• Reduce data transmission requirements using flood segmentation.
• Validate a flood segmentation algorithm on ESA's PhiSat-1 mission.

Main Methods:

• Developed a flood segmentation algorithm for onboard processing.
• Trained models on the WorldFloods dataset (119 global events).
• Tested algorithm efficiency on PhiSat-1's hardware accelerator.

Main Results:

• Algorithm efficiently generates accurate flood masks on the accelerator.
• Demonstrated successful data reduction from raw images to masks.
• Achieved fast processing times suitable for near-real-time applications.

Conclusions:

• Onboard processing is a viable solution for CubeSat data limitations.
• Flood segmentation on CubeSats enhances disaster response capabilities.
• The PhiSat-1 mission successfully proves the concept for efficient flood mapping.