Anomaly Detection and Classification in NOAA Satellite Imagery with Deep Learning

There are visual artifacts, or anomalies, occur in satellite GOES ABI observation and NPP/J1/J2 VIIRS observation. Those image level anomalies are not sufficiently flagged in the pixel level QA and there is no enough information can be provided to downstream users. Some of these anomalies are predictable based on satellite maintenance schedules, observation time et al, but others may occur randomly. Many of these phenomena are so rare that most users, even ones who frequently analyze ABI imagery such as National Weather Service (NWS) field office forecasters, may not notice them. Unexpected and widespread data anomalies can disrupt NOAA operations and significantly impair hazard monitoring. The tradition method to detect and determinate of the anomalies is done by the manually observing visually and with the help of ancillary data, inevitably this process is prone to errors and very complex, laborious, and time-consuming. An Automatic anomalies identification and classification method for NOAA satellite imagery is gaining interest day by day.

A binary classification model basing on deep learning has been developed in NOAA/NEDSIS/OCS to detect the anomalies in GOES ABI and NPP/J1/J2 VIIRS observation. And this project is now being extended to develop a multilabel model for anomaly detection and classification. The extended model is developed to answer not only the question about whether anomalies exist but also the question about what types those anomalies are and reasons.

In this paper, the detail information about the training dataset used in the deep learning that contains anomalies, the evaluation result of the anomaly detection and classification model will be given. the problem existing in deep learning processing and possible solutions will also be discussed.