This project developed a neural network to filter radio-frequency-interference (RFI) for astronomical radio interferometry. The code and networks published here are part of my Master's thesis, which will be linked here when available.
The network is trained to flag RFI, as well as dead and misfunctioning antennae. Therefore, it was trained on the GMRT telescope, with data taken from the GMRT archive. The flags, which are used as ground truth are generated by tfcrop, which is a flagging algorithm of the radio imaging software casa. It needs still to be checked if the GMRT data used for training can be published here.
The network is trained using supervised learning. The RFI flagging has such a quality that imaging is possible, with the resulting images having a 1.5 to 2 times SNR, compared to the images flagged with a combination of tfcrop and manual imaging.
The project can be seperated into three parts. The first part in bin contains
a mini-pipeline to flag more data of the GMRT. This can be employed for further
tests but also to use it as an RFI-flagger if the slight increase in SNR is
acceptable for the simplification of just running the application and additionally
gaining the ablity to flag dead antennae.
The second part is the training folder, which contains the code with which the
networks where trained and the third part which contains some code for graphicly
displaying the quality of the network comapred to the ground truth.
For further detail consult the Masters thesis, as soon as it is available.
The code is written with these modules and versions in mind and installed. It may work with newer versions of certain modules.
- Python 3.7.9
- TensorFlow 2.1.0
- Numpy 1.18.5
- Matplotlib 3.3.2
- Astropy 4.0.2
- Tqdm 4.50.2