Source total fluxes systematically overestimated at low S/N

[JonahDW]

This is something I've been running into for a while now, but haven't had the time to properly test. The results below are produced with PyBDSF v1.13, but I first saw it with older versions.

Essentially, there seem to be cases where the total flux (as in the Total_flux column) of a source is systematically overestimated by PyBDSF compared to its intrinsic flux. This is seen in simulations where you inject fake sources into (PyBDSF) residual images and rerun source finding. Simulating only point sources for simplicity, theTotal_flux (below, left) is overestimated by 20% at the detection threshold (5-sigma). As a comparison,Isl_Total_flux (below, right) is slightly underestimated, but only by 5% at the detection threshold.

Given that Total_flux is affected, it stands to reason that this might be coming from the Gaussian fitting. Checking those parameters (Peak_flux, Maj, Min) compared to expected values reveals that both Peak_flux (below, right) and Maj (below, left) are overestimated, both contributing more or less equally to the overall increase in Total_flux.

While it is difficult to diagnose to what extent this affects actual source finding on real data, it is very noticeable when comparing Total_flux and Peak_flux. The expectation here is that for point sources, Total_flux values are centred on the Peak_flux, with some scatter which increases with lower S/N. For the simulations (below, left) , this relation doesn't actually look to far from what we expect. But for an actual dataset (below, right), this distribution is completely skewed, so this effect might even be more pronounced on actual data! Though real data consists of not just point sources, the majority still should be, however nearly all sources have a Total_flux/Peak_flux > 1.

To ease concerns, as far as I am aware, this issue might affect only specific types of images. I have run these tests for MeerKAT L-band images as well as S-band images, and only see it in L-band. As such, it might somehow have to do with specific background noise patterns, so would not really classify as a 'bug'. Still, I think it is important to be aware. If a github issue is not the proper place to report this, I am happy to take it wherever it is appropriate.