This repository contains code built for Mommert (2020): Cloud Identification from All-sky Camera Data with Machine Learning published by The Astronomical Journal. For details on cloudynight, please refer to this publication.

The system consists of several parts:

• the cloudynight Python module, which contains tools for data handling and preparation, feature extraction, model training and prediction (see directory cloudynight/);
• a Python Django web server application for database management, data visualization, and manual labeling (see directory webapp/);
• example data used in this work (see directory example_data/);
• a number of scripts for testing the functionality on the example data (see directory scripts/).

Please that cloudynight only utilizes the more efficient lightgbm classifier. The ResNet code is also included in scripts/ for the sake completeness.

cloudynight requires the following Python modules to be available (version numbers in parentheses signify the version numbers that were used in building this code):

• numpy (1.16.3)
• scipy (1.2.1)
• matplotlib (3.0.3)
• pandas (0.25.3)
• sep (1.0.3)
• astropy (4.0)
• scikit-image (0.15.0)
• lightgbm (2.2.3)
• scikit-learn (0.22.1)
• astroplan (0.4)

The use of the ResNet implementation requires additional modules:

• pytorch (1.3.1)
• tqdm (4.36.1)

cloudynight contains all the parts necessary to build an automated cloud detector, but it is not intended as a plug-and-play software.

First, install the cloudynight module:

>>> python setup.py install

and run the provided example scripts to get familiar with the module.

Then, install the web application.

To use the software for real-time cloud detection, write a script that utilizes cloudynight.AllskyCamera.download_latest_data() to download data from the camera computer. Then, utilize cloudynight.AllskyCamera.process_and_upload_data to extract features and upload them to the database. Both tasks can be automated with cron jobs.

The label/ task of the web application can be used for manual labeling and training data generation. Once enough data are available, a modified version of scripts/model_lightgbm.py can be used to tune model parameters and fit a model. With a model being available, the check/ task can be utilized for faster manual labeling.

Cloud coverage can be predicted for the latest image obtained from the camera using the web API with task predictLatestUnlabeled.

If you decide to use code elements from this repository, please reference Mommert (2020): Cloud Identification from All-sky Camera Data with Machine Learning published by The Astronomical Journal.

The author would like to thank Ryan J. Kelly and the NAU/NASA Arizona Space Grant program for enabling a case study for this project.

This software is distributed under a 3-clause BSD license.