21cm Segmentation Network, SegU-Net
Deep learning approach for identification of HII regions during reionization in 21-cm observations.
In coming years, the square kilometre array (SKA) will come online, the largest telescope in the world.
The SKA will map the distribution neutral hydrogen on our sky at different redshift and produce tremendous amount of 3D image data. These tomographic images will
be prone to instrumental limitations, such as noise, limited resolution and foreground contamination. Hence a stable and reliable method for neutral/ionized
region identification is not trivial.
We therefore developed a convolutional neural network for 21-cm image segmentation, named SegU-Net.
Our network is able to identify the ionized region from noisy data with greater precision and is less sensitive to the limitation of previous methods. The animation below shows SegU-Net layers output for the test case of a sphere with a noise level that reflet the SKA1-Low setup (Ghara et al. 2017).
We therefore developed a convolutional neural network for 21-cm image segmentation, named SegU-Net.
Our network is able to identify the ionized region from noisy data with greater precision and is less sensitive to the limitation of previous methods. The animation below shows SegU-Net layers output for the test case of a sphere with a noise level that reflet the SKA1-Low setup (Ghara et al. 2017).