Speaker
Description
Radio Frequency Interference (RFI) poses critical challenges for radio astronomy,
particularly solar radio astronomy, corrupting observations of fundamental phenomena
like the quiet sun, solar radio bursts, solar flares, and coronal mass ejections. We
present a detection pipeline combining multi-domain feature engineering with
unsupervised machine learning to deal with unavailable labeled RFI data in solar
spectrographs. Our method extracts 19 interpretable features spanning temporal,
spectral, and statistical domains. We used these features to train four unsupervised
models (KMeans, DBSCAN, GMM, and Autoencoder) evaluated on physics-constrained
synthetic RFI with pixel-level ground truth. Results show clustering models
(KMeans/GMM) achieve 100% F1-scores in detecting simulated RFI—significantly
outperforming anomaly detection approaches (DBSCAN F1=0.39, Autoencoder
F1=0.10). Feature importance analysis reveals peak to average power ratio and
spectral kurtosis as optimal discriminators between RFI types. By eliminating
dependency on labeled data and adapting to telescope-specific RFI profiles, this
pipeline enables robust, RFI identification for solar radio observatories.
| Stream | Science or Engineering |
|---|