Deep learning-based semantic segmentation of night-sky clouds for operational telescope scheduling

Abstract

Abstract Ground-based optical telescopes necessitate prompt and spatially detailed information regarding dome-scale cloud coverage to facilitate target-specific shuttering and scheduling decisions. When only coarse or delayed atmospheric data are available, observatories risk inefficient use of scarce dark time and the irreversible loss of scientific exposures. To address this, we introduce the WOANC dataset, a pixel-annotated nighttime full-dome dataset acquired at an operational observatory, alongside NightCloudSegNet, a fisheye-aware segmentation framework specifically designed for low-light astronomical imaging. Evaluated on the WOANC test set, NightCloudSegNet achieves a mean intersection-over-union (mIoU) of 86.6% and a pixel-level F1 score of 92.8%. Furthermore, when tested on the external SWINSEG dataset, the model attains an mIoU of 86.2% and an F1 score of 92.6%, thereby demonstrating robust performance under conditions of fisheye distortion and low illumination. By translating pixel-level segmentation masks into per-target observability indicators, this approach has the potential to support informed shuttering and scheduling decisions, which is expected to enhance observational efficiency in automated telescope operations.