A 99.2%-accurate anomaly model, running in NOAA production.

01The problem, stated plainly

NOAA's GOES (ABI) and JPSS (VIIRS) satellites stream terabytes of Level-1b imagery every day. A class of scene-level degradation — structural artifacts you can only recognize when a whole band is laid out spatially — was passing every pixel-level quality flag and flowing into the downstream weather and climate products built on that imagery.

02Why a CNN, not more rules

A single pixel looked fine; the defect was structural, so tightening signal-level thresholds could not catch it. We used a ResNet-18 — an 18-layer convolutional neural network, pre-trained on ImageNet — and transfer-learned it on 2–3 years of ABI and VIIRS Level-1b data across all channels. A set of 4,550 GOES ABI frames was labeled valid or invalid from the files' own quality flags, converted from netCDF to PNG, greyscaled, downscaled from 5000×5000 to 240×240 pixels, and augmented with random flips.

03Making a black box auditable

A CNN cannot say why it classified an image — and an operator will not trust a verdict it cannot see. We added Grad-CAM attention heatmaps that show exactly where the model looked before deciding, and a purpose-built Python validation GUI that let a data scientist inspect thousands of frames and swap models to compare. The opaque classifier became something a human could audit.

04Into production

The most accurate GOES-16 model was deployed on AWS, where it now checks every incoming ABI Level-1b file. Each verdict is written as timestamped JSON to S3, monitored through CloudWatch, and surfaced on a Grafana dashboard for NOAA decision-makers. Offline validation measured 99.2% accuracy on GOES ABI (April 2023 confusion matrix) and 89% on VIIRS, which was used for research only.

05Credit & source

This was Aeroxis's contribution — authored by David Daniel — to the NOAA/NESDIS Earth Observing Digital Twin (EO-DT) prototype: the machine-learning anomaly-detection subsystem, one of the program's three demonstrations. The broader digital-twin prototype was led by Science and Technology Corporation with a nine-author team, under contract 1332KP22CNEEP0013. The full final report is public domain (CC0): Rapid Prototype of Earth Observing Digital Twin of the NESDIS Ground System.