CNN-based satellite localization: overcoming GPS absence-Egypt case study

Abstract This paper presents a robust approach to address the critical challenge of self-localization for flying vehicles, especially in environments with noisy data or an absence of GPS signals. Our proposed solution uses the power of deep learning, specifically Convolutional Neural Networks (CNNs), to accurately determine a vehicle’s geographic coordinates (latitude and longitude) by analysing satellite imagery. To train and validate our system, we have compiled a comprehensive dataset of Google Maps covering Egypt, ensuring a wide variety of visual terrains and features. The methodology leverages the inherent strengths of CNN architectures to efficiently process large-scale visual datasets while automatically identify and extract significant topographical features. This autonomous feature-extraction capability is crucial for real-world satellite operations, where conditions are often unpredictable and data can be imprecise. By employing a transfer learning strategy, we have adapted two powerful, pre-trained CNN models: SqueezeNet and GoogLeNet. These networks were fine-tuned using our custom Google Maps imagery dataset, enabling them to effectively learn and generalise from the distinct visual attributes present in the maps. The resulting system offers a reliable and precise localization alternative, proving essential for the consistent operation of autonomous systems in GPS-denied or challenging environments. The resulting system predicts position accurately, achieving an accuracy of 99.42% and 98.9% for GoogLeNet and SqueezeNet, respectively, for case 1(zoom level 9) images with a resolution of 5.6 km × 5.6 km. While the grayscale version for the same case gives 95.5% for GoogLeNet, and 96% for SqueezeNet. For case 2, zoom scale 11, where the network achieves a resolution of 395 m × 395 m, GoogLeNet reaches an accuracy of 97.7%, and SqueezeNet reaches an accuracy of 93%. Although the grayscale version results in 62% for GoogLeNet and 67% for SqueezeNet.