Training-free AI for Earth Observation Change Detection using Physics Aware Neuromorphic Networks

TL;DR

This paper introduces a physics-aware neuromorphic network that detects natural disaster changes in satellite images without training, enabling low-resource, real-time on-board processing for Earth observation.

Contribution

The proposed PANN leverages memristor-based physics constraints to perform change detection without training, reducing computational requirements for satellite data processing.

Findings

Achieved comparable or better disaster detection accuracy than state-of-the-art models.

Operates with minimal computing resources due to training-free design.

Demonstrated effectiveness in resource-constrained on-board satellite processing.

Abstract

Earth observations from low Earth orbit satellites provide vital information for decision makers to better manage time-sensitive events such as natural disasters. For the data to be most effective for first responders, low latency is required between data capture and its arrival to decision makers. A major bottleneck is in the bandwidth-limited downlinking of the data from satellites to ground stations. One approach to overcome this challenge is to process at least some of the data on-board and prioritise pertinent data to be downlinked. In this work we propose a Physics Aware Neuromorphic Network (PANN) to detect changes caused by natural disasters from a sequence of multi-spectral satellite images and produce a change map, enabling relevant data to be prioritised for downlinking. The PANN used in this study is motivated by physical neural networks comprised of nano-electronic circuit elements known as "memristors" (nonlinear resistors with memory). The weights in the network are dynamic and update in response to varying input signals according to memristor equations of state and electrical circuit conservation laws. The PANN thus generates physics-constrained dynamical output features which are used to detect changes in a natural disaster detection task by applying a distance-based metric. Importantly, this makes the whole model training-free, allowing it to be implemented with minimal computing resources. The PANN was benchmarked against a state-of-the-art AI model and achieved comparable or better results in each natural disaster category. It thus presents a promising solution to the challenge of resource-constrained on-board processing.