Operational Anomaly Detection in Satellite Imagery Using ResNet-18 Transfer Learning

Abstract The National Weather Service continuously uses satellite imagery to monitor natural hazards; however, data anomalies can disrupt operational forecast capabilities. To improve anomaly identification and alerting, we developed transfer learning models using residual neural network 18 (ResNet-18) to detect anomalies commonly encountered in the Advanced Baseline Imager (ABI) and the Visible Infrared Imaging Radiometer Suite (VIIRS) instruments. We implemented separate binary classification models for each instrument that classified images as anomalous or normal. For ABI, we used naturally occurring anomalies across all 16 channels. For VIIRS, we generated synthetic anomalies across all I- and M-band channels mimicking sensor degradation. All spectral channels were combined in the training datasets since anomalies exhibited similar characteristics across wavelengths. Transfer learning enabled effective anomaly detection despite limited training samples by leveraging pretrained ResNet-18 features and retraining only the final classification layer over 50 epochs. Training curves showed stable convergence for both models. The ABI model achieved 98% accuracy, and the VIIRS model achieved 95% accuracy. Model precision was lower (96% and 94%, respectively), reflecting a greater ratio of anomaly misclassifications than normal misclassifications. Class activation maps showed model focus concentrated over anomalous regions in an image for both instruments, whereas the attention scattered across normal images. This research provides operational satellite data quality teams with an automated, rapid anomaly classification capability across multiple channels with interpretable results. While transfer learning enabled training with smaller data sizes, synthetic training data offer additional advantages, such as preparation for anomalies that have not yet occurred. Significance Statement Satellite imagery is critical for hazard monitoring, but anomalies compromise prediction capabilities. Current anomaly detection is labor intensive, requiring numerous science teams and engineers to monitor the data. This study demonstrates that residual neural network 18 (ResNet-18) can automate anomaly detection with 98% accuracy in geostationary and 95% accuracy in low-Earth-orbiting satellite imagery. Our approach used naturally occurring and synthetic anomalies, demonstrating that both real and artificially generated training data can produce effective models. The approach successfully handles limited training data and utilized explainable artificial intelligence (AI) techniques to visualize influential image regions. This paper provides a successful approach for anomalous feature detection in satellite imagery and is adaptable to detecting more complex anomalies.