An Interpretable Multimodal AI Framework for Urban Park Assessment: Aligning Perception with Physical Environments

Urban parks have become critical urban green spaces that deliver ecological and health benefits. Visitors’ perceptions are central to evaluating park quality, which has traditionally been assessed using survey instruments. Survey instruments have limitations, however, such as high costs and self-selection bias, highlighting the need to employ multimodal data sources for more comprehensive assessment. This study develops a unified and interpretable multimodal artificial intelligence framework, aligning perception from review text with physical features from images. This framework enables more transparent interpretations of how visitors’ perceptions correspond to physical-environment conditions in parks. Using 1.56 million visitor reviews from 6,184 parks across China from 2006 to 2025, we derive five perceptual dimensions: accessibility, orderliness, comfort, aesthetics, and interactivity. We then align these five perceptual dimensions with observable physical features extracted from 1.23 million park images, remote sensing, and point-of-interest data using an XGBoost model. Empirically, intercity inequality in park perception declined from 2006 to 2025, whereas intraurban disparities showed a weaker, nonmonotonic, and dimension-specific pattern. Several dimensions exhibited a core advantage, a decline toward intermediate outward zones, and partial recovery in the outermost distance bin, suggesting that perceived park quality is shaped by both urban centrality and the presence of larger, scenic, or newly developed peripheral parks. The drivers of park perception have shifted from accessibility to aesthetics, particularly in eastern coastal cities. Although inequality of park perception decreased, older adults and children still exhibit poorer access to high-quality parks, posing a challenge for future planning. Overall, the study contributes a transferable but locally adaptable protocol for measuring perceived experiential quality, interpreting the physical-environment conditions associated with perception, and diagnosing quality-sensitive park equity across cities and social groups.