In-orbit demonstration of a re-trainable machine learning payload for processing optical imagery.

Area of Science:

• Space Science
• Artificial Intelligence
• Earth Observation

Background:

• Cognitive cloud computing in space (3CS) is an emerging field driving innovation in planetary observation and deep space exploration.
• Machine learning (ML) payloads are crucial for extracting high-level information from onboard sensors in space.

Purpose of the Study:

• To demonstrate an operational ML payload, 'WorldFloods', on an orbiting satellite for flood detection and mapping.
• To evaluate ML model performance for onboard deployment and assess the feasibility of remote model retraining and updating.

Main Methods:

• Comparison of various segmentation models under critical onboard processing constraints.
• Generation of vectorized flood water polygons directly from satellite imagery.
• On-demand retraining of the ML model using downlinked sensor data.
• Uplinking updated models to the satellite for real-time processing of new acquisitions.

Main Results:

• Successful onboard generation of compressed flood maps using the WorldFloods ML payload.
• Demonstration of producing vectorized flood polygons from full Sentinel-2 tiles.
• Validation of the retrained model's performance on new satellite imagery after uplink.
• Proof of concept for updating ML models in orbit.

Conclusions:

• ML-based models can be effectively updated in orbit, enabling agile integration of onboard and ground processing.
• The study paves the way for continuous learning and adaptive ML applications in space-based Earth observation.