Nighttime Cloud Detection, Tracking and Prediction with All-Sky Cameras

TL;DR

This paper introduces a real-time all-sky camera-based method for nighttime cloud detection, tracking, and short-term prediction to improve robotic telescope scheduling and reduce weather-related observation losses.

Contribution

The paper presents a novel algorithm that combines difference imaging, thresholding, and morphology for accurate cloud detection and prediction up to 15 minutes ahead, validated on extensive telescope data.

Findings

Achieved approximately 1% false positive rate

Enabled cloud movement prediction up to 15 minutes

Validated effectiveness on Liverpool Telescope data

Abstract

This paper presents a novel method for real-time nighttime cloud detection, tracking, and prediction using all-sky cameras, aimed at enhancing the efficiency of ground-based robotic telescopes. Ground-based telescopes are vulnerable to adverse weather conditions, particularly cloud cover, which can lead to the loss of valuable observation time and potential damage to the telescope. Existing methods for cloud detection have limitations in accuracy, particularly under varying illumination conditions such as gibbous moon phases. To address these challenges, we developed an algorithm that uses the temporal incoherence of image sequences from all-sky cameras. The method computes difference images to highlight moving cloud structures, applies Otsu thresholding to generate binary cloud maps, and uses mathematical morphology techniques to reduce noise from bright stars and other artifacts. Segmented cloud regions are then tracked across successive frames, allowing estimation of a velocity vector and enabling short-term predictions of cloud movement. Our approach achieves reliable cloud detection and tracking, providing predictions up to 15 minutes into the future - a capability critical for robotic telescopes that rely on look-ahead scheduling. The system was validated against extensive historical data from the Liverpool Telescope's Skycam A and T systems, achieving a false positive rate of approximately 1% and a similar false negative rate, depending on cloud thickness and speed. By improving cloud forecasting and observational scheduling, the system offers a valuable tool for enhancing the operational reliability of robotic telescopes.