Speaker: Hattie Stewart (University of Bristol)

Date: Thursday 18 May 2023

Time: 14:00

Location: NSQI Seminar Room

Radio Image Segmentation in the SKA Era with U-Net

Source detection and classification are challenges for large-scale radio surveys and will become more so as we move towards the era of the SKA. We have been experimenting with a machine-learning approach for detecting and classifying the full radio source population in the SKA Data Challenge 1 (SDC1) dataset, using U-Net to perform image segmentation. The trained network reconstructs the raw image data as a binary segmentation map that describes the angular size, ellipticity, position angle, and location of the sources. We recover the flux densities of sources from different observing periods and compare our results to those of current state of the art source finders PyBDSF and ProFound. We find that U-Net performs comparatively if not better for all observing periods when comparing the detection of real sources provided by the SDC1 truth catalogue. We determine the intersection over union of detections with true sources and find U-Net to produce the most accurate segmentation. PyBDSF and ProFound incorrectly associate independent sources as components of an extended source and worsen the normal confusion limit by a significant factor. This work serves as a proof of concept that U-Net can be used to detect and classify radio source populations from SKA-like data.