Multi-Task Road Extraction Leveraging Hybrid Global-Local Modeling: A Transformer-CNN Approach

Accurate road network extraction is essential for urban applications such as navigation and urban planning. Traditional methods relying on clustering or trace-merging often struggle with noisy GPS data and low sampling rates, while convolutional neural networks (CNNs) are limited by their inability to model global contexts. Transformers, on the other hand, excel at capturing long-range dependencies but suffer from high computational complexity. To address these challenges, we propose a hybrid framework that combines CNNs and Transformers to leverage their respective strengths. The framework features a Hybrid Attention Block (HAB) to efficiently integrate global and local contexts, enhancing the model's ability to handle complex road structures. Additionally, we introduce a multitask learning strategy, where road centerline extraction is the primary task, and road region segmentation acts as an auxiliary task to improve prediction accuracy.