arXiv · 2611.14987
Sky Monitor AI: classifying night-sky conditions from an observatory all-sky camera with a CNN ensemble, label-free sky types and exact evidence maps
Abstract
Small and remotely operated observatories must decide every night whether the sky is clear enough to observe. We classify frames of the auxiliary all-sky camera of the Higashi-Hiroshima Observatory as clear or cloudy, using 8806 frames from 32 nights. A circular sky mask removes a printed date stamp that would act as a shortcut, and all evaluations use whole held-out nights. On six test nights, hand-crafted sky statistics reach 77.1% accuracy, a small CNN trained from scratch 84.3%, and a validation-selected ensemble of the small CNN and two ImageNet-pretrained ResNet-18 variants 86.7% (AUC 0.924) with the smallest night-to-night scatter. Moonlight is the dominant difficulty: across Moon altitude the ensemble's accuracy varies by 5 percentage points, the hand-crafted model's by 20. Fine-tuning deeper ResNet layers overfits with about 24 training nights, underlining the need to match model capacity to the number of independent nights. The labels were assigned in blocks of 1-2 h, and many remaining errors are mislabelled or partly cloudy frames. Without labels, clustering of pretrained CNN embeddings recovers the clear/cloudy split with 87% purity and isolates frames with water droplets on the camera dome. Gradient-based maps and a rim test indicate that the networks rely on sky regions. For networks ending in global average pooling, the decision decomposes exactly into local logits, giving gradient-free evidence maps and a cloud-cover estimate. We release Sky Monitor AI, an open-source application for single frames, whole nights and live cameras.
Explore related subjects
Keep this discovery
Explore connections, maps & timelines
Muhammad Waqas. 2026-11-11. Sky Monitor AI: classifying night-sky conditions from an observatory all-sky camera with a CNN ensemble, label-free sky types and exact evidence maps. https://searcharxiv.com/abs/2611.14987
Citations retain the original authors and DOI and include this site’s record URL. If this discovery service contributed to your research, you can also cite Searcharxiv by Indiciae: BibTeX · RIS.