Juanda Airport Runway Visibility Modeling Using Gan Based on Imbalanced Dataset

Abstract

Abstract – Aviation safety and security are heavily influenced by airport visibility, as pilots require clear visual references for landing. However, poor weather conditions can reduce visibility and increase the risk of accidents. Therefore, an automated system is needed to classify visibility levels quickly and accurately, even when faced with the challenge of imbalanced datasets. This study employs a Generative Adversarial Network (GAN) approach, focusing on Vanilla GAN, DCGAN, and StyleGAN models. The data used is sourced from CCTV AWOS at Runway 10 of Juanda Airport, encompassing 14,458 images from the period of August 13 to 31, 2023. The models are evaluated using SSIM scores and feature extraction of color, texture, and HOG at various epochs. The results indicate that the Vanilla GAN model at 60 epochs is the most suitable for the minority class compared to the other models, based on feature evaluation, SSIM scores, synthetic image quality, and loss pattern outcomes. Its simple architecture aids in capturing low variation in the dataset, making it superior to more complex architectures like DCGAN and StyleGAN. Further optimization and architectural adjustments could enhance the results, especially for datasets with low variation like the one used in this study.
Keywords: Aviation, GAN, Dataset