convolve_uv

pip install convolve-uv

import astropy.units as u
from convolve_uv import convolve_uv
from radio_beam import Beam
from spectral_cube import SpectralCube

cube = SpectralCube.read("my_cube.fits")
target_beam = Beam(major=0.86 * u.arcsec, minor=0.86 * u.arcsec, pa=0 * u.deg)

cube_conv = convolve_uv(image=cube, 
                        target_beam=target_beam,
                        )

convolve-uv is a small Python package that allows for image convolution without loss of resolution.

When using something like convolve_fft in astropy, you generate a kernel in the image domain, Fourier transform (FT) it and the input image, and then perform the multiplication in the Fourier domain to carry out the convolution. Whilst this works well for arbitrary kernels, when the kernel is small enough that the pixel scale does not properly sample it, artefacts can be introduced or the convolution can fail. The way around this is to pad out the kernel slightly so it is properly sampled, but this results in a loss of resolution.

convolve-uv takes a different approach, one that is useful for radio data when the beam of your data and the convolution kernel are both Gaussian (which is almost always the case). This instead builds a transfer function directly in $uv$-space, and multiplies the FT of the image by this transfer function. This means we can generate kernels that are not limited by the pixel scale of the image, and e.g. convolve an elliptical to a round beam without any unnecessary loss of resolution.

To illustrate this, the Figure below shows a demonstration of convolve-uv compared to the astropy convolve_fft. The original beam here is elliptical ($0.86^{\prime\prime} \times 0.65^{\prime\prime}$), and so the round beam is $0.86^{\prime\prime} \times 0.86^{\prime\prime}$. We see that convolve-uv successfully convolves the image to the round beam, while convolve_fft fails to do so due to the undersampling of the kernel. For larger kernels (convolving to a $1^{\prime\prime}$ and $1.5^{\prime\prime}$ beam), the two approaches give the same result.