Enhanced RFI Detection in Imbalanced Astronomical Observations Using Weakly Supervised GANs

Radio astronomical observations are plagued by radio frequency interference (RFI) caused by a variety of man-made signals, e.g., cellular, automotive radar, satellite, GPS, etc. Both the wireless and radioastronomy communities have great interest in identifying RFI in the vicinity of the telescopes in real-time. This paper proposes the use of Generative Adversarial Networks (GANs) in the context of weakly supervised learning as a means to enhance the machine learning (ML) based detection of RFI in captured astronomical scans. Weakly supervised training is particularly appropriate when only a small subset of captured data is labeled, as is the case with many radioastronomical datasets. Such 'class imbalance' in the training dataset hampers the classifier's performance, particularly in terms of identifying the minority class samples with RFI, which is often the class of greater interest. Applying weakly or semi-supervised training to GANs addresses the class imbalance challenges. Our study is based on scans obtained from the 12-meter Alma-like Observatory at Kitt Peak, Arizona. We rely on the experience of radio astronomers to manually label the channels in a small fraction of the captured scans as 'clean' or 'dirty'. The remaining channels of 4 G Hz of the observed spectrum (95% of the scans) are unlabeled. We first use our human-labeled data as ground truth and train a baseline classifier in a supervised manner. Subsequently, we explore two approaches for weakly supervised learning. The first approach uses a combination of an autoencoder and conditional GAN, while the second approach uses a semi-supervised GAN (SGAN). Both techniques harness the features learned from the unlabeled dataset to train the generator and discriminator of a GAN. In the first approach, the trained generator is used to synthesize the dirty data, while in the second approach, the trained discriminator is modified to act as a clean/dirty RFI classifier. Simulations under extremely imbalanced training samples show that the SGAN approach can significantly improve the Fl-score and True Positive Rate (TPR) relative to the baseline classifier.