Urban Computing Course Project: City-Scale Road Lane Type Extraction from Satellite Imagery

With the rapid pace of urban development and the pressing global environmental crisis, urban planners increasingly emphasize al ternative transit methods, such as dedicated bike and bus lanes, when designing road networks. Automating the extraction of road networks from satellite imagery presented significant challenges, particularly in European cities with diverse transportation. This study addresses the challenge of extracting road networks with detailed lane types by a dual-task learning model, combining pixel based multi-class semantic segmentation andgraph-basedinference using patch-wise road keypoint detection. Using satellite imagery and bike lane masks from six European cities, the model demon strated superior accuracy (90.45%) compared to a baseline U-Net. The results underline the efficacy of integrating pixel-based and graph-based techniques in road network extraction through the balance of global road morphology reconstruction and local connec tivity, overcoming limitations of previous pixel-only and graph-only approaches.

Code Directory

1. Get satellite image and road mask data

1. execute get_images.py to get satellite images from Google Maps api;
2. execute get_masks.py to gat corresponding road network masks from OSMnx library;
3. manually adjust the mask images in Adobe Photoshop (or other image editing applications) to make masks accurately overlay with roads in satellite images

2. Main model - dual-task learning

follow the readme instructions in Main_model folder

3. Baseline model - pixel-based U-Net

see readme file in Baseline_model folder