Enhancing Road Segmentation in Satellite Images via Double U-Net with Advanced Pre-Processing

Roads play an important role in people’s lives by providing access to essential facilities such as schools and hospitals, thereby increasing people’s productivity. The use of satellite imagery and computer vision technology can speed up and improve the efficiency of road map creation. Deep learning methods such as U-Net have been used for road segmentation from satellite images, but the results have yet to be optimal. This study aims to examine the performance of DoubleU-Net and the effect of image pre-processing in satellite image-based road segmentation. This research implements the DoubleU-Net deep learning model for road segmentation based on satellite images. Pre-processing methods such as rotating augmentation, contrast stretching, and edge enhancement are applied to increase the contrast value and clarify the edges in the satellite images used. The results show that combining the two U-Net models into a DoubleU-Net model results in a significant improvement in the IoU and dice coefficient values, achieving an average increase of $\mathbf{2 \%}$ and integrated with image processing, the model performance experiences an average increase in IoU and dice coefficient values of $\mathbf{2 - 4} \%$. The highest performance was achieved by the DoubleU-Net model with rotating augmentation pre-processing, resulting in an IoU performance value of $\mathbf{4 5. 9 9 \%}$ and a dice coefficient performance of $\mathbf{6 1. 1 2 \%}$.