Training-free AI for earth observation change detection using physics aware neuromorphic networks

Area of Science:

• Earth and Space Science
• Artificial Intelligence
• Materials Science

Background:

• Earth observation satellites provide critical data for managing natural disasters.
• Low latency data downlink from satellites is essential for effective first responder decision-making.
• Bandwidth limitations in data downlinking pose a significant challenge for timely information dissemination.

Purpose of the Study:

• To develop a novel on-board processing method for prioritizing critical data from satellite imagery.
• To introduce a Physics Aware Neuromorphic Network (PANN) for detecting natural disaster-induced changes.
• To enable efficient data downlinking by processing and prioritizing relevant information.

Main Methods:

• A Physics Aware Neuromorphic Network (PANN) was proposed, inspired by memristor-based physical neural networks.
• The PANN utilizes dynamic network weights updated via memristor equations and circuit conservation laws.
• Change detection was performed using a distance-based metric on physics-constrained dynamical output features.

Main Results:

• The PANN successfully generated change maps from multi-spectral satellite image sequences.
• The model demonstrated comparable or superior performance against a state-of-the-art AI model in natural disaster detection.
• The PANN achieved comparable or better results in each natural disaster category tested.

Conclusions:

• The PANN offers a promising solution for resource-constrained on-board satellite data processing.
• The training-free nature of the PANN allows implementation with minimal computing resources.
• This approach addresses the challenge of low-latency data downlink for time-sensitive events like natural disasters.