Inception-ResNet-V2 The U-Net Encoder for Road Segmentation using Sentinel 2A

Updating road network maps is essential for transportation services, as incomplete or inaccurate maps can lead to inefficiencies and diminish service quality. The online transportation industry generates vast amounts of GPS data as drivers navigate, which is valuable for mapping road networks and improving traffic management. However, since drivers do not cover all roads, satellite imagery plays a crucial role in identifying areas that are not mapped. By combining GPS data as labels with satellite imagery, the extraction of new road networks becomes more accurate. This research employs a deep learning Convolutional Neural Network (CNN) with the U-Net architecture for road segmentation, allowing for the identification of new paths. Two different encoders are tested in this research: Inception-ResNet-V2 and a pure U-Net encoder. The Inception-ResNet-V2 encoder achieves an accuracy of 91.3%, while the pure U-Net encoder achieves 90.7%. In terms of Dice Loss, the models record values of 0.051 and 0.08, respectively. The research highlights the effectiveness of different U-Net encoders in road network segmentation. With high accuracy and low Dice Loss, this approach provides a reliable method for automatically updating road maps. It has potential applications in navigation systems, urban planning, and AI-driven intelligent transportation systems.