An enhanced MobileNetV2-191H architecture for accurate crescent moon detection using semantic segmentation

Abstract

This study proposes an enhanced lightweight semantic segmentation framework, called MobileNetV2-191H, for accurate crescent moon detection in astronomical observation images. Crescent moon detection is a challenging task because the crescent moon typically appears as a very thin and low-contrast object that is highly affected by atmospheric interference, cloud cover, image noise, and illumination variation near the horizon. The dataset used in this study was obtained from crescent moon observation videos provided by the Indonesian Agency for Meteorology, Climatology, and Geophysics (BMKG), resulting in 6,283 extracted image frames that were divided into training and testing datasets. The proposed MobileNetV2-191H framework was developed by adapting the original MobileNetV2 architecture through the integration of additional segmentation-oriented layers, including transposed convolution, feature refinement, and pixel classification modules, to improve pixel-level segmentation capability for thin crescent moon objects. Experimental evaluation was conducted using semantic segmentation metrics, including precision, recall, F1-score, and Intersection over Union (IoU). The proposed model achieved 84.81% precision, 85.32% recall, 84.07% F1-score, and 72.64% IoU, outperforming several baseline segmentation models such as U-Net, DeepLabV3, SegNet, and EfficientNet. These results demonstrate that the proposed lightweight framework effectively improves crescent moon segmentation performance while maintaining computational efficiency for real-time astronomical observation applications.