Sensory systems are adapted to efficiently process natural stimuli. Such adaptation has been particularly well documented in the retina. Theoretical and experimental studies have demonstrated that multiple properties of retinal coding and architecture -- from receptive field (RF) shapes to RF mosaic alignment -- are optimized for efficient sensory coding. These adaptations have been primarily studied in small regions of the retina, however stimulus statistics vary systematically across the entire visual field (VF). It therefore remains unknown whether the collective structure of large neural populations, such as those covering the entire VF, could be shaped by global statistics of natural scenes. A particularly salient global property of the natural VF is the contrast inhomogeneity. Local contrast increases with elevation and rapidly drops-off at the horizon, inducing changes in signal-to-noise ratio (SNR) of the photoreceptor output. To understand how SNR variation could globally shape RFs of retinal ganglion cells (RGCs), we developed an encoding model related to the predictive coding (PC) theory. The theory postulates that RGCs recode photoreceptor output to minimize the metabolic cost of information transmission. Our model predicts two kinds of change in RF shapes induced by the contrast variation: increase of the surround strength along the VF elevation, and concentration of asymmetric RFs at the horizon. To test these predictions, we established a novel epi-fluorescent imaging approach that enables simultaneous recording of approximately 1000 neurons per field-of-view. Using this novel technique, we mapped RFs of 26558 RGCs, distributed across the entire VF. We found that large-scale retinal organization closely recapitulates theoretical predictions. The relative strength of the surround increases along the dorso-ventral axis, while the RGCs with vertically asymmetric RFs form a streak along the horizon line. Our results demonstrate that the statistics of natural scenes can shape the organization of neural populations at a previously unappreciated scale.