Supervised neural networks for RFI flagging

Neural network (NN) based methods are applied to the detection of radio frequency interference (RFI) in post-correlation, post-calibration time/frequency data. While calibration does affect RFI for the sake of this work a reduced dataset in post-calibration is used. Two machine learning approaches for flagging real measurement data are demonstrated using the existing RFI flagging technique AOFlagger as a ground truth. It is shown that a single layer fully connect network can be trained using each time/frequency sample individually with the magnitude and phase of each polarization and Stokes visibilities as features. This method was able to predict a Boolean flag map for each baseline to a high degree of accuracy achieving a Recall of 0.69 and Precision of 0.83 and an F1-Score of 0.75.The second approach utilizes a convolutional neural network (CNN) implemented in the U-Net architecture, shown in literature to work effectively on simulated radio data. In this work the architecture trained on real data results in a Recall, Precision and F1-Score 0.84, 0.91, 0.87 respectfully.This work seeks to investigate the application of supervised learning when trained on a ground truth from existing flagging techniques, the results of which inherently contain false positives. In order for a fair comparison to be made the data is imaged using CASA's CLEAN algorithm and the UNet and NN's flagging results allow for 5 and 6 additional radio sources to be identified respectively.