Satellite imagery is the backbone of modern earth observation — powering weather forecasting, climate monitoring, environmental assessment, and disaster response. But satellite data isn't perfect. Sensor artifacts, transmission errors, and calibration drift can introduce subtle errors that propagate through downstream analyses if left undetected. Our work with NOAA on the Earth Observation Digital Twin (EO-DT) project tackled this problem head-on with AI-driven binary error detection.
The Problem
NOAA's GOES (Geostationary Operational Environmental Satellite) and VIIRS (Visible Infrared Imaging Radiometer Suite) instruments generate massive volumes of imagery daily. Manual quality control at this scale is impossible — errors need to be caught automatically and flagged before the data enters analytical pipelines. The challenge is distinguishing genuine atmospheric phenomena from sensor artifacts and data corruption, which can look surprisingly similar in raw imagery.
Our Approach
We developed an AI-driven binary error detection system that classifies satellite image tiles as either valid or containing errors. The system processes imagery from both GOES and VIIRS instruments, each with different spectral bands, resolutions, and error characteristics. Rather than building hand-crafted rules for each error type, the model learns to identify error patterns directly from labeled training data.
The binary classification approach was a deliberate design choice — in operational settings, the first question is always 'is this data trustworthy?' before any finer-grained error characterization. By keeping the initial detection simple and fast, we ensured the system could operate at the throughput required for real-time satellite data processing.
Results and Publication
The system achieved 99.2% detection accuracy with sub-second inference times, enabling real-time quality control at the scale of NOAA's satellite data pipeline. This work was published in the AIES journal, contributing to the broader research on applying AI to earth observation data quality assurance.
Lessons Learned
- Binary classification as a first pass is operationally pragmatic — detect first, characterize second
- Satellite data quality is a prerequisite for every downstream application, making error detection high-leverage work
- Working with government scientific agencies requires deep understanding of their data pipelines and operational constraints
- AI for data quality assurance is an underappreciated but critical application area