An AI Approach to Lunar Phase Detection: Enhancing the Identification of the New Crescent with Astronomical Data Integration

Murad Al-Rajab^1*Samia Loucif²Raed Abu Zitar³Mubarak Gwaza Abdu-Aguye⁴

• ¹Abu Dhabi University, Abu Dhabi, United Arab Emirates
• ²Zayed University, Dubai, United Arab Emirates
• ³Liwa College, Abu Dhabi, United Arab Emirates
• ⁴Mohamed bin Zayed University of Artificial Intelligence, Masdar City, United Arab Emirates

The observation of the lunar crescent has been important in astronomy, cultural traditions, and religious lunar observations. This paper introduces techniques for an image-based study of the birth of the new crescent detection using machine learning (ML) and deep learning (DL) techniques applied to space-borne imagery. Because earth-based images capturing the different phases of the moon within a single month and across multiple months, including the new crescent, are not available, this work employs orbital images based on data from NASA's Lunar Reconnaissance Orbiter (LRO) spanning over 13 years. The objective of this study is to investigate whether artificial intelligence (AI) models can learn and identify patterns related to the new crescent. In this paper, we investigated the performance of applying Convolutional Neural Networks (CNN) and traditional ML methods such as Random Forests (RF) and Support Vector Machines (SVM). The original new crescent identification was enhanced using a custom image preprocessing pipeline consisting of grayscale conversion, contrast-limited adaptive histogram equalization, and noise reduction techniques. This was followed by a CNN architecture that integrates lunar imagery with moon age data for more accurate predictions. Experiments were conducted on a temporally split dataset to mimic real-world conditions achieving high metrics of precision, recall, F-score, and even an accuracy of almost 98% across both ML and DL models. More interestingly, the RF and CNN models showed overall best performance results, outperforming the SVM model. The system was further evaluated using synthetically generated noise and occlusion applied to the orbital images in order to assess model robustness. The CNN model, in particular, maintained high accuracy in the presence of Gaussian noise and occlusions of up to 50%, demonstrating its practical applicability in lunar observation. This work contributes to the further advancement in the support of traditional methods for the new crescent identification, addressing long-standing issues of calendar discrepancies across different regions.