Along with the rapid development of telecommunication, radio frequency interference (RFI) generated from diverse human produced sources like electronic equipment, cell phones, GPS and so on can contaminate the weak radio band data. Therefore, RFI is an important challenge for radio astronomy. RFI detection can be regarded a special task of image segmentation. As for RFI signals, they appears in the form of point, vertical or horizontal lines. However, most existing convolution neural networks (CNNs) perform classification tasks, where the output is the single classification label of an image. The U-Net enables classification of each pixel within the image, which is suitable and competitive for image segmentation. Thus, in this paper, we implement the U-Net of 14 layers with framework of Keras to detect RFI signals. The U-Net can perform the classification task of clean signal and RFI. Also, the U-Net is a kind of extended CNN with symmetric architecture, which consists of a contracting path to capture context information and extract features and an expanding path to get precise localization. It extracts the features of RFI for learning RFI distribution pattern and then calculates the probability value of RFI for each pixel. Then we set a threshold to get the results flagged by RFI. We train the parameter of the U-Net with “Tianlai” data(A radio telescope-array, the observing time is from 20:15:45 to 24:18:45 on 27th of September 2016, the frequency is from 744MHz to 756MHz and the number of baseline is 18528). The experimental results show that, compared with the traditional RFI flagging method, this approach performs better with satisfying accuracy and takes into account the relationship between different baselines, which contributes to correctly and effectively flag RFI.
Link to PDF (may not be available yet): O1-4.pdf