GLD-Road:A global-local decoding road network extraction model for remote sensing images

TL;DR

This work tackles fragmented road networks and slow connectivity in remote-sensing road extraction. It proposes GLD-Road, a global-local decoding model that first detects road nodes in parallel and then locally refines broken segments, aided by a 38-dimensional road node representation and a Connect Module. A denoising training strategy and two-stage loss further improve node accuracy and topology, delivering state-of-the-art TOPO and APLS scores on City-Scale and SpaceNet3 while maintaining fast inference. The approach balances global efficiency with local precision, enabling scalable, topologically coherent road-network extraction for mapping, autonomous driving, and disaster response.

Abstract

Road networks are crucial for mapping, autonomous driving, and disaster response. While manual annotation is costly, deep learning offers efficient extraction. Current methods include postprocessing (prone to errors), global parallel (fast but misses nodes), and local iterative (accurate but slow). We propose GLD-Road, a two-stage model combining global efficiency and local precision. First, it detects road nodes and connects them via a Connect Module. Then, it iteratively refines broken roads using local searches, drastically reducing computation. Experiments show GLD-Road outperforms state-of-the-art methods, improving APLS by 1.9% (City-Scale) and 0.67% (SpaceNet3). It also reduces retrieval time by 40% vs. Sat2Graph (global) and 92% vs. RNGDet++ (local). The experimental results are available at https://github.com/ucas-dlg/GLD-Road.

Figures (10)

• Figure 1: Comparison among the three types of road network extraction methods.
• Figure 2: Visualization results of road network extraction using global parallel and local iterative methods.
• Figure 3: Structure of the GLD-Road model. Each square or cube represents a road query. The arrows indicate the direction of data flow. In the predicted road network result map, the orange lines represent the predicted road network, the yellow dots represent the nodes, and the red line segments indicate the iterative retrieval results derivedfrom the Local Query Decoder. In the [id=R2]GlobalGlobe Query Decoder, the yellow dots indicate the positions of road nodes, while the red line segments represent the direction visualization results; the closer a line segment is to the circle's boundary, the higher the confidence in that direction.
• Figure 4: (a) Schematic diagram of road node modeling. (b) Node coordinates are shown on the left, the 36-dimensional directional descriptor vector (with zero values represented by ellipsis) is displayed in the middle, and the index numbers of the directional descriptor are indicated on the right.
• Figure 5: Structure of the Connect Module