Lahore-Sky SRCNN: Toward Flux-Conserving Super-Resolution for Urban Astrophotography

Amateur astronomers in large South Asian cities observe through skies of Bortle class 8–9, where artificial skyglow, aerosol extinction, ground-layer turbulence and the tracking errors of low-cost mounts degrade both angular resolution and photometry. We describe Lahore-Sky SRCNN, a super-resolution pipeline trained entirely in simulation for this regime, and report a full-scale test of it. A physical forward model converts calibrated Hubble Space Telescope (HST) images into the frames a 130 mm telescope would record from Lahore, Pakistan, with band-integrated Rayleigh, aerosol and ozone extinction drawn from AERONET climatology, a lamp-spectrum skyglow model with a light-dome gradient, long-exposure von Kármán seeing from Monte Carlo phase screens, worm-gear periodic error and polar-alignment drift, and a CMOS detector. The network, AstroSRCNN, is a 57,280-parameter convolutional network whose residual correction is projected onto the kernel of a 17×17 averaging operator; we prove that this confines any change of aperture flux to a ring within 12 px of the aperture edge. On six held-out fields the trained network did not improve on bicubic upsampling: the PSNR change was +0.1, +1.1 and -0.4 dB, too few point sources survived selection to measure resolution, and the 90th-percentile flux change in point-source apertures was 2.6% in M3 and 10.7% in M51, against a 2.5% limit. We trace the failure to the training data: background subtraction removes 15–96% of the light in the network's input that the target keeps, so the flux loss demanded flux that, by our own proposition, the network could add only at aperture edges. We give the correction, which remains to be tested.