A new AI pipeline lets Earth observation satellites spot oil spills, algal blooms, and sediment floods on their own, without waiting to beam raw imagery back to the ground.
Researchers built a two-stage system for satellites carrying multi- or hyperspectral sensors. A self-supervised neural network first compresses each image into a smaller representation, then an anomaly-detection model scans that compressed data for patterns that do not match normal sea conditions. The team benchmarked it against older methods - Isolation Forest, One-Class Support Vector Machines, and Local Outlier Factor - and designed it to run on modest onboard hardware, from embedded CPUs to AI accelerators. The pipeline has been integrated into two missions: the European Space Agency's Phisat-2 and the Microsoft/Thales Alenia Space IMAGIN-e.
Satellites generate far more imagery than they can realistically downlink, so most of it either waits in a queue or gets discarded. Flagging anomalies onboard means a satellite can prioritize sending down the handful of images that actually show a spill or bloom, cutting the lag between an event happening and someone on the ground finding out about it.
Worth noting: this is a research paper with two mission integrations, not a proven early-warning network - whether it catches real spills reliably, at scale, over years of operation is still an open question.