A system catches errors in satellite rainfall data using pre-trained vision models and rare human labels.

Private companies now launch smaller, cheaper satellites that track rainfall, but validating all this new data is a growing problem. Bad data or glitches, called artifactsartifactsErrors or anomalies in satellite data that don't represent real rainfall., can slip through before reaching the public. To solve this, researchers built an anomaly-detection system to catch flawed global satellite rainfall readings in near-real-time. Their system uses pre-trained computer vision modelspre-trained computer vision modelsAI systems that have already learned to recognize patterns from large image datasets before being used for a specific task., which are algorithms that recognize visual patterns in images. It also learns from a scarce amount of human-labeled examples to spot specific errors. Testing showed the tool successfully separated normal satellite orbits from orbits containing artifacts. It performs as well as current top methods, explains its reasoning, and can update and improve whenever it makes a mistake.

Why it matters

With more small satellites launching private rainfall missions, the public needs reliable data fast. This system catches errors early so people using weather forecasts or climate data aren't misled by bad satellite readings.

Questions answered