Physics-Aware Neuromorphic Network Framework for Onboard Thermal Anomaly Detection Using Sentinel-2 Data
A novel framework known as the Physics-Aware Neuromorphic Network (PANN) has been introduced for the detection of thermal anomalies onboard, tackling issues such as sensor drift and domain shift in raw Earth Observation data. This lightweight design, which draws inspiration from principles of physical neural networks and neuromorphic computing, has been assessed using two datasets from Sentinel-2: the decompressed Level-0 (L0) raw data along with its metadata and the processed Level-1C (L1C) data. The goal is to enable swift and dependable early warnings for bushfires and volcanic eruptions, where delays in detection can lead to significant damage. Recent EO techniques indicate that thermal anomaly detection can occur directly on decompressed L0 data, circumventing costly preprocessing steps. However, challenges remain due to radiometric inconsistencies and limited labeled training samples. The PANN framework is elaborated in arXiv:2604.18606v1, where a performance comparison between raw and L1C data is presented to improve detection efficiency.
Key facts
- Physics-Aware Neuromorphic Network (PANN) framework proposed for onboard thermal anomaly detection
- Evaluated using two Sentinel-2 datasets: decompressed Level-0 (L0) raw data and Level-1C (L1C) data
- Addresses challenges like domain shift, sensor drift, radiometric inconsistencies, and scarcity of labelled training samples
- Aims to provide fast, reliable early warning for bushfires and volcanic eruptions to prevent escalating damage
- Lightweight architecture inspired by physical neural network principles and neuromorphic computing paradigms
- Recent Earth Observation approaches enable thermal anomaly detection directly on decompressed L0 sensor data
- Direct exploitation of raw data avoids computationally expensive preprocessing chains
- Detailed in arXiv:2604.18606v1, a cross-announcement abstract
Entities
—