Measuring nuanced walkability: Leveraging ChatGPT's vision reasoning with multisource spatial data

Recent advances in urban analytical tools, particularly street view image (SVI) data and computer vision (CV) algorithms, such as semantic segmentation, have enhanced walkability measurement by enabling the automated assessment of mesoscale features, such as greenery proportions. However, while SVI data contain rich environmental information, off-the-shelf CV models generally struggle to capture microscale features—design details attached to mesoscale elements, such as the quality of greenery or sidewalk maintenance. Moreover, because CV algorithms typically evaluate environmental features in isolation, they often fail to account for spatial arrangements and visual harmony among features, limiting their ability to support a holistic assessment of walkability. Recently, multimodal large language models (MLLMs), particularly ChatGPT, have introduced a transformative approach to image analysis by mimicking human perception. This study proposes a comprehensive walkability measurement framework that leverages ChatGPT's vision reasoning across multiple spatial data, including SVIs and GIS land use and road network maps. To validate this framework, we compare ChatGPT-generated walkability ratings with human assessments and examine their relationship with reported walking behavior data. Furthermore, by comparing ChatGPT-generated outcomes with evaluations from conventional walkability measurement tools, such as GIS and off-the-shelf CV models, we highlight the novel contribution of ChatGPT in walkability assessment. This research advances the literature by introducing a ChatGPT-based framework for a more comprehensive walkability assessment.