Check literature for missing references

[sfarrens]

• [x] Latest deep transfer learning papers in astro
• [ ] Latest impacts of blends in WL

[sfarrens]

https://arxiv.org/pdf/2101.12628.pdf

[andrevitorelli]

https://arxiv.org/abs/2012.08567 - cite the impact of close blends on shear responsivity/biases

[andrevitorelli]

1st quick pass...

Deep Transfer Learning (most interesting):

• https://arxiv.org/abs/2011.12437 DeepShadows: Separating Low Surface Brightness Galaxies from Artifacts using Deep Learning
• https://arxiv.org/abs/2005.12039 (Bastien's paper)
• https://arxiv.org/abs/1909.02024 Deep Transfer Learning for Star Cluster Classification: I. Application to the PHANGS-HST Survey

Impacts of blends:

• HSC: 58% Bosch et al. https://arxiv.org/abs/1705.06766 (In the Wide layer of the SSP, 58% of all ob-jects are members of blends; this increases to 66% and 74%in the Deep and UltraDeep layers, respectively (and the latterhas not yet been observed to full depth). )
• LSST: 50% at d<=3' (Chang et al 2013)' But see: Bosch et al 18 (LSST’s 18,000deg2wide survey will exceed theSSP Wide depth before its fourth data release, and its deep-est fields may exceed SSP UltraDeep depths even sooner. As aresult, its blending problems will ultimately be even more severe.)

[sfarrens]

@andrevitorelli I have pretty much finished the introduction. We are still missing a couple of references:

• We need to update the LSST reference for the expected number of blends as the paper cited was not published at the time the first draft was written. I have asked François about this, he said to cite the HST paper.
• We should look to see if any of the Euclid papers mention how many blends are expected.
• We need a couple of reference to illustrate that deep learning is good for classification problems in general.

[andrevitorelli]

The relevant paper from euclid seems to be (in a first take) the prep IV (Martinet et al 2019 https://arxiv.org/abs/1902.00044 ) I'm checking for a number to quote.

The good for classification I'll try to find in refs from the paper about efficient nets. https://arxiv.org/abs/1905.11946

[andrevitorelli]

• Review Deep Learning (mentions successes in image classification - last paragraph) LeCun et al. 2015 10.1038/nature14539
• https://www.frontiersin.org/articles/10.3389/frobt.2015.00036/full
• (still some things to check...)

[andrevitorelli]

Possible paper of interest, recent. Effects of overlapping sources on cosmic shear estimation: Statistical sensitivity and pixel-noise bias Javier Sanchez, Ismael Mendoza, David P. Kirkby, Patricia R. Burchat (for the LSST Dark Energy Science Collaboration)

Full abs: The next generation of dark-energy imaging surveys – so called “Stage-IV” surveys, such as that of the Rubin Observatory Legacy Survey of Space and Time (LSST) – will cross a threshold in the number density of detected sources on the sky that requires qualitatively different image analysis and measurement techniques compared to the current generation of Stage-III surveys. In Stage-IV surveys, a significant amount of the cosmologically useful information is due to sources whose images overlap with those of other sources on the sky. We focus on the weak gravitational lensing probe, for which we expect the largest impact since the cosmic shear signal is primarily encoded in the estimated shapes of observed galaxies and thus directly impacted by overlaps. We introduce a framework based on the Fisher formalism to analyze the effect of the overlapping sources (“blending”) on the estimation of cosmic shear. This method gives concrete predictions for the minimum loss of information due to noise and blending for any choice of “deblending” scheme and shape-measurement algorithm. Our studies account for undetected sources but do not address their full effects and biases they may introduce.

We use simulated images and predict this impact of blending for three surveys: the Dark Energy Survey (DES), the Hyper-Suprime Cam Subaru Strategic Program (HSC-SSP), and the Rubin LSST. Our methodology successfully estimates the statistical sensitivity to weak lensing for DES and HSC early results.For LSST, we present the expected loss in statistical sensitivity for the ten-year survey due to blending. We find that for approximately 62% of galaxies that are likely to be detected in full-depth LSST images, at least1% of the flux in their pixels is from overlapping sources. We also find that the statistical correlations between measures of overlapping galaxies and, to a much lesser extent (0.2%) the higher shot noise level due to their presence, decrease the effective number density of galaxies,Neff, by∼18%. We calculate an upper limit on Neff of 39.4 galaxies per arcmin2 in r band. We study the impact of stars on Neff as a function of stellar density and illustrate the diminishing returns of extending the survey into lower Galactic latitudes.

We extend the simulation-based Fisher formalism to predict the expected increase in pixel-noise bias due to blending for maximum-likelihood (ML) shape estimators. We find that noise bias depends sensitively on the particular shape estimator and measure of ensemble-average shape that is used, and properties of the galaxy that include redshift-dependent quantities such as size and luminosity. Based on the magnitude of the estimated biases and these many dependencies, we conclude that it will not be possible to estimate noise biases of ML shear estimators using simulations, at the sensitivity required for LSST measurements of cosmic shear.The source code for these studies is available online.

https://arxiv.org/abs/2103.02078