A Walkability Assessment Method for Urban Streets Based on Multi-source Urban Data and Deep Learning: A Case Study of the Central Urban Area of Beijing

Objective Walking is the most fundamental mode of urban transportation, and building pedestrian-friendly streets holds great significance in urban planning. Yet many metropolises remain auto-oriented and lack high-quality pedestrian environments. Advances in multi-source urban data and deep learning now make it feasible to evaluate street-level walkability with greater coverage, precision, and reproducibility than traditional audit or perception-based approaches. This study proposes a comprehensive framework that integrates multi-source urban data with deep learning to quantify urban street walkability, and demonstrates its utility through an application to Beijing’s central districts with external validation against on-site pedestrian ratings. Methods By reviewing relevant literature, a multi-criteria street walkability value framework was developed, covering five dimensions: imageability, convenience, vibrancy, comfort, and safety. A walkability evaluation indicator system comprising 23 indicators was established. To operationalize the measurement, we assembled and harmonized multi-source urban data: vector layers for roads, buildings, green spaces, and land-use types; high-resolution satellite imagery; 213,950 street-view images captured via the Baidu panorama application programming interface (API) at 100-meter intervals and four bearings per point; point of interest (POI) records (20 top-level categories); urban heat-map rasters sampled at six time slots across a workday−weekend cycle; field photos; and official statistics. The street network was segmented into a 100 m grid of 53,631 units after removing duplicates and very short fragments. A DeepLab v3+ semantic segmentation model, fine-tuned via transfer learning on urban streetscape data and optimized with cross-validation and learning-rate decay, produced pixel-level shares of sky, vegetation, roadway, buildings, and other salient elements for each viewing direction; a weighted fusion yielded panoramic proportions per street unit. A YOLO v5 detector, trained on 2,000 labeled images based on building-facade quality criteria and validated on a 10% hold-out set, identified facade attributes relevant to building quality. ArcGIS pipelines performed accessibility analysis, conducted geo-joins of POIs and heat-map intensities to the street grid, integrated building and population data through grid-based analysis, and visualized the final results; georeferencing of heat-map mosaics used a WGS-84 frame and second-order polynomial transformation. Indicator weights were derived using the analytic hierarchy process (AHP) based on judgments from 35 domain experts (consistency ratio CR < 0.10). Indicator values were min-max normalized and combined via weighted summation to obtain a composite walkability score for each street unit.The empirical application covers Beijing’s central urban districts (Dongcheng, Xicheng, Chaoyang, Haidian, Fengtai, Shijingshan), which occupy less than 10% of the city’s total land area but accommodate more than 50% of its population. This region is characterized by high density and intense daily travel demand. Composite walkability scores were discretized into five performance classes from I (0.80−1.00) to V (0.00−