Exploring Built Environment Perception Pathways and Perception Profiles Using Vision–Language Models

Computer vision models applied to street-view imagery (SVI) can generate large-scale measures of perceived built-environment qualities relevant to urban planning and public health. Unlike convolutional neural networks (CNNs), vision–language models (VLMs) can provide both perception ratings and rationales. Using 121,764 participant-rated pairwise comparisons of SVI, we fine-tuned the Qwen2.5-VL model to predict five perceptions: beauty, nature quality, relaxation potential, safety from traffic, and safety from crime. We compared their accuracy with previously published Siamese CNN models and prompted the VLMs to generate a rationale for each prediction. The 32B VLM then predicted nature quality for 48,664 images across Corvallis, Oregon. We grouped model-generated rationales into 25 rationale topics using Top2Vec semantic topic modeling. On the held-out test set, VLM accuracy averaged 4.1 percentage points higher than the CNN model, with differences ranging from −1.0 percentage point (safety from traffic) to 8.0 percentage points. VLM topics associated with perceived nature quality included vegetation density, tree-canopy cover, landscaping, maintenance, and dappled sunlight. Mapping combinations of three example topics produced distinct perception profiles. VLMs can simultaneously predict urban perceptions and generate descriptions of visible characteristics associated with those predictions.