High-resolution satellite imagery contains rich road network information, including urban main roads, narrow alleys, and rural branch roads. However, narrow roads are often missed in conventional segmentation approaches, and temporal changes in road networks—such as newly constructed or removed roads—are typically not captured. Recent methods, such as TrMamba, demonstrate strong performance in tubular structure modeling and road segmentation, yet they are mostly limited to static or main-road extraction and do not explicitly address narrow-road detection or temporal changes.

This work investigates high-precision narrow road segmentation and temporal road change detection, leveraging multi-scale feature fusion, tubular structure tracking, and Siamese-based temporal difference learning to achieve fine-grained narrow road extraction and identification of newly added or removed roads. The goal is to construct a unified framework capable of accurate narrow road extraction and dynamic road network updating, highlighting the novelty of combining fine road segmentation with temporal change detection.

Tasks to be performed by the student:

Review related works on road segmentation, tubular structure modeling, temporal change detection, and topological loss in remote sensing.
Implement a TrMamba-based backbone with multi-scale feature fusion and tube-tracking enhancement for narrow road segmentation.
Develop a Siamese temporal difference module for detecting newly added or removed roads across different timestamps.
Design loss functions combining pixel-level accuracy (Dice/BCE) and topology-aware constraints to improve continuity and connectivity of narrow roads.
Train and evaluate the model on publicly available high-resolution road datasets: Massachusetts Roads, DeepGlobe Roads, and SpaceNet Roads, using metrics such as Precision, Recall, F1-score, IoU, APLS, and connectivity metrics.
Conduct ablation studies on multi-scale fusion strategies, tube-tracking enhancements, and temporal difference detection schemes to quantify contributions of each module.

Note：Applicant will use publicly available datasets and receive technical support from SZTAKI.

Supervisor at the department: Dr. Chang Liu, Assistant Professor

External supervisor:   Prof. Tamás Szirányi, HUN-REN SZTAKI.