U-SPIN Pyramid: U-Net with Spatial and Interaction Space Graph Reasoning for Improving Road Extraction from Satellite Images

Remote sensing images are essential nowadays as they allow us to extract relevant information intended for various applications, notably road networks. Today, road extraction is crucial for updating maps, urban planning, developing autonomous driving and even quantifying the damage caused to roads in the event of natural disasters. However, this task presents many challenges, including the complex geometric paths of roads, occlusion by other objects like trees and buildings, and similarities to other objects such as car parks and house surfaces. All of this makes extracting roads using only simple Convolutional Neural Networks (CNNs) difficult. This study investigates the impact of the SPIN module, it makes use of interaction space and spatial space reasoning, on enhancing road extraction performance. For this purpose, the U-Net architecture was adapted to develop three SPIN-based variants, collectively termed U-SPIN. All three architectures surpass the performance of the U-Net model, achieving higher accuracy, recall, F1 score, and Jaccard Index (IoU) on the DeepGlobe dataset.