ROMEO: Raw Onboard Multispectral Earth Observation — An Edge-AI Framework for Latency- and Bandwidth-Efficient Satellite Sensing

Abstract: Wildfire and volcanic-eruption alerts from orbit are delayed less by detection accuracy than by the ground processing chain. This paper reports a working onboard pipeline that analyses raw Sentinel-2 Level-0 granules without any ground-segment product: radiometric calibration, coarse spatial coregistration and coarse georeferencing are applied on board, a spectral gate discards cloudy scenes, thirteen bands are reduced to the three carrying thermal signal, and a 23,714-parameter network quantised to 29.9 KB scans the granule. On an 18-granule catalogue the system reaches 100 % granule accuracy with no false detections, and classifies 240 of 240 unseen patches correctly under leave-granules-out evaluation. Removing coregistration costs 16.67 percentage points, identifying it as the enabling step. Quantisation reduces inference latency eightfold, and transmitting alerts instead of imagery reduces 30.67 MB to 105.74 KB. Because the granules are synthesised, the results characterise pipeline correctness rather than field performance.

Keywords: Coarse spatial coregistration, convolutional neural network, edge computing, integer quantisation, onboard artificial intelligence, raw Level-0 data, Sentinel-2, thermal hotspot detection, wildfire detection.