HP-BERT: Social media–based insights into differential impacts of urban environmental factors on objective and subjective heat exposure

Frequent extreme heat events have become a major environmental threat to urban individuals’ health and well-being, yet existing methods struggle to quantify individuals’ subjective heat perception at large scales, limiting urban climate adaptation and heat risk management. This study proposes a Transformer-based heat perception modeling framework, HP-BERT (Heat Perception BERT), which integrates manually annotated data with domain-adaptive pretraining (DAPT), unsupervised self-training, and low-rank adaptation (LoRA) fine-tuning. HP-BERT substantially improves semantic understanding and prediction of heat-related expressions in Chinese social media, outperforming mainstream large language models (R² = 0.835) and enabling large-scale urban text analysis. Using HP-BERT, we quantified individuals’ subjective heat exposure (SHE) during extreme heat events in Chengdu and, in combination with land surface temperature and urban environmental factors, employed interpretable machine learning to reveal nonlinear trends and local threshold effects of environmental factors on subjective and objective heat exposure (OHE). The results indicate that: (1) most factors exhibit similar overall effects on SHE and OHE, but local effects differ significantly; (2) some factors show mismatched directional effects, e.g., high NDVI areas reduce OHE but may increase SHE due to limited shading; (3) environmental factors such as building height and FAR contribute differently to SHE and OHE across value ranges. This study provides an efficient tool for large-scale urban heat perception modeling and offers scientific insights for understanding the divergence between subjective and objective heat exposure, supporting person-centred heat-health planning and optimizing urban climate adaptation policies.