arXiv Science⌕ Search

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

BibTeXRIS

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://arxiv.science/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 arXiv Science by Indiciae: BibTeX · RIS.

KEEP EXPLORING

Related papers

Optimal robust detection statistics for pulsar timing arrays

Pulsar timing arrays (PTAs) seek to detect a nano-Hz stochastic gravitational-wave background (GWB) by searching for the characteristic Hellings and Downs angular pattern of timing residual correlations. So far, the evidence remains below the conventional $5σ$ threshold, as assessed using the literature-standard "optimal cross-correlation detection statistic". While this quadratic combination of cross-correlated data maximizes the deflection (signal-to-noise ratio), it does not maximize the detection probability at fixed false-alarm probability (FAP), and therefore is not Neyman-Pearson (NP) optimal for the assumed noise and signal models. The NP-optimal detection statistic is a different quadratic form, but is not used because it also incorporates autocorrelations, making it more susceptible to uncertainties in the modeling of pulsar timing noise. Here, we derive the best compromise: a quadratic detection statistic which is as close as possible to the NP-optimal detection statistic (minimizing the variance of its difference with the NP statistic) subject to the constraint that it only uses cross-correlations, so that it is less affected by pulsar noise modeling errors. We study the performance of this new NPMV statistic for a simulated PTA whose noise and (putative) signal match those of the NANOGrav 15-year data release: GWB amplitude $A_{\rm gw} = 2.1 \times 10^{-15}$ and spectral index $γ= 13/3$. Compared to the literature-standard "optimal" cross-correlation detection statistic, the NPMV statistic increases the detection probability by 47% when operating at a $5σ$ false-alarm probability of $α= 2.9 \times 10^{-7}$.

astro-ph.IM↗

Lunar Reflective Interferometry

We present the method of lunar reflection interferometry (LRI) in which a virtual interferometer can be formed by using a spacecraft-borne antenna in lunar orbit that receives both direct rays and those reflected from the lunar surface. The technique exploits the method of images and is akin to the classic "sea cliff" interferometer, with the Moon's surface, primarily the lunar Maria, replacing the ocean surface. We describe the method in detail, demonstrate that significant portions of the Moon's surface are sufficiently smooth at low radio frequencies, $ν\lesssim 10\,\mathrm{MHz}$, for the technique to work, and outline a spacecraft instrument implementation. We describe potential systematic errors and how they could be mitigated. We present several astrophysics applications of the method.

astro-ph.IM↗

Wideband overlapped polyphase channelizer for kinetic inductance detector readout: split-path architecture and experimental characterization

Frequency-multiplexed readout of kinetic inductance detector arrays measures the complex transmission of thousands of superconducting resonators. The channelizer that separates them introduces systematic effects: a gain that depends on where each tone falls in the channel grid, leakage between channels and added noise. This article presents a 4096-channel polyphase channelizer with 2/1 overlap for a 4.096 GS/s input, implemented on a first-generation radio-frequency system-on-chip, and a procedure to characterize it on the board with two receive chains, one digitizing the excitation as a reference. A split-path architecture computes the two overlap states in parallel filter branches at 256 MHz and reorders them frame by frame, under three control rules; it keeps only the transform stage at 512 MHz and affords a prototype of 16 taps per polyphase path (65536 coefficients) at the multiplication rate of a published 8-tap bank. Each selected channel is exported as a complex series at 1 MS/s. The measured channel response follows the coefficient-table model within 0.15 dB down to -28 dB, with both overlapping channels at full gain across their common band. Leakage into every measured channel 1.875 MHz or more from the tone is -94 dBc or lower, apart from images attributed to converter interleaving. The signal-to-noise ratio per sample lies within 1.7 dB of the limit set by the data-sheet noise density of the converter, and an emulated detector pulse is recovered with a decay time of 14.65 $μ$s (standard uncertainty 0.04 $μ$s), within 0.7% of the expected value.

astro-ph.IM↗