E2E: Onboard satellite real-time classification of thermal hotspots events on optical raw data

What is it about?

In the last years, numerous researches explored the use of Artificial Intelligence (AI) on board satellites to process spaceborne data to detect wildfire, volcanic eruptions and other extreme events quickly and accelerate the transmission of early alerts. To this aim, it becomes very important to minimize the time requested to process satellite imagery on board. Typical approaches mimic the solutions designed for on ground processing, which rely on high-end data products, which undergo extensive corrective pre-processing. Such pre-processing chains are typically time-consuming and non-suited for onboard processing. Because of that, this paper investigates the use of AI to process directly "raw" data - i.e., low-level products with minimal corrective pre-processing - for classification of patches containing volcanic eruptions and thermal anomalies. To this aim, the paper proposes a processing chain who limits to register different image channels (like RGB in normal imagery), break the registered images into patches, and classify them as thermal anomaly / no thermal anomaly.

Featured Image

Why is it important?

Our findings demonstrate that processing directly raw data while limiting pre-processing to bands co-registration allows significantly reduce latency in anomaly patch classification. In our experimental setting, processing one image takes about 1.8s that is lower than the declared time required for the Sentinel-2 acquisition system to acquire and produce the corresponding image. At the same time, the model reports high classification performances.

Perspectives