Heritage-sensitive urban vitality assessment

Historic districts present a “data–heritage paradox”: their cultural and morphological complexity demands nuanced, context-aware modeling, yet the scarcity of labeled data hinders robust, cross-regional generalization of data-driven methods. To address this challenge, we propose a heritage-sensitive transfer learning (HSTL) framework aligned with the United Nations Educational, Scientific, and Cultural Organization’s historic urban landscape approach. The framework integrates multi-source urban data—geographic information system/points of interest indicators, geotagged social media activity, and street view images—to construct a transparent neighborhood vitality index as the supervised learning target. Methodologically, HSTL enforces a structured separation between generic vitality drivers (Xg; e.g., density and accessibility) and heritage-sensitive attributes (Xh; e.g., façade chromatics and conservation zoning). A “heritage floor” constraint allocates a minimum aggregate weight of 30 percent to Xh to prevent high-variance generic features from overwhelming culturally salient signals. The feature blocks are standardized and processed separately prior to multimodal fusion, with knowledge transferred from a data-rich source domain (Beijing) to data-scarce target domains (Shanghai and Guangzhou) via fine-tuning. Using region-grouped cross-validation, HSTL achieves consistent target-domain gains, including approximately 25 percent reductions in mean squared error and ~0.15 increases in R2 relative to target-only baselines. Ablation tests show that removing Xh decreases target-domain R2 by 8–10 percent, confirming its functional contribution to vitality prediction. Overall, HSTL resolves the data–heritage paradox by providing an operationalizable, transferable approach for cross-regional vitality assessment that explicitly safeguards heritage-specific interpretability

Generalization · Pedestrian · Salient · Scarcity · Vitality · Advanced Neural Network Applications · Cultural Heritage Management and Preservation · Multimodal Machine Learning Applications