Outdoor Thermal Comfort Research and AI-Based Prediction for Nightscape Planning and Design

Abstract

Focusing on AI-based prediction for nighttime outdoor thermal comfort, this research investigates how environmental, spatial, seasonal, and sensory factors influence people's thermal perception at night, and develops an integrated prediction framework to support evidence-based nightscape planning and design. To address the under-theorization of nighttime conditions in landscape architecture (LA) and the methodological gap between site-specific microclimate assessment and scalable perception-centered modeling, two complementary studies were conducted: Study A examined nighttime microclimate dynamics and subjective thermal perception through a four-season field experiment across twelve nightscape sites, and Study B developed an AI-based predictive modeling framework that integrates field observations, large-scale social media records, and standardized benchmark data. These studies aim to understand how spatial enclosure, vegetation configuration, waterside treatment, seasonal conditions, and soundscape shape perceived thermal comfort in nighttime urban environments. Based on the research findings, the benefits of AI-based predictive modeling for nightscape planning are discussed, and guidelines for spatial, vegetative, waterside, and soundscape design across seasonal contexts are developed. The findings from the two studies indicate that landscape spatial characteristics produce consistent and repeatable microclimate effects at the site scale, with the warmest semi-open area and the coolest waterside tree-around area maintaining a 2.0-2.6°C temperature gap across all four seasons, even as background air temperature shifted by more than 20°C. Subjective thermal perception data revealed a strong overall correlation between air temperature and thermal sensation vote (r = 0.885), with an estimated nighttime neutral temperature of 17.5°C, lower than the 20-25°C commonly reported in daytime studies. However, the relationship between thermal sensation and overall thermal comfort varied systematically across seasons, ranging from r = 0.419 in spring to effectively zero (r = 0.012) in summer, supporting a two-regime conceptualization of nighttime comfort: thermal sensation dominates comfort under cold-stress conditions, while non-thermal factors such as spatial enclosure, vegetation presence, and proximity to water dominate under thermal neutrality. The auditory experiment further demonstrated that sound type exerts a substantial cross-modal influence on nighttime thermal perception, with crowd sound and wind sound producing a perceptual spread of up to 1.15 TSV points under cold-stress conditions, approximately equivalent to a 7°C shift in air temperature. While water sound emerged as the most preferred condition in summer and fall. Emotional appraisal results confirmed that pleasant and calm responses were positively associated with comfort, while gloomy and anxious responses were negatively associated. For the AI-based predictive modeling, the Random Forest algorithm achieved the highest single-model performance (regression R² = 0.81; classification accuracy = 0.6354), and a hybrid model combining classification and regression Random Forests through a threshold-based switching mechanism achieved the highest overall accuracy (0.6695, R² = 0.85) among all fourteen modeling approaches evaluated, approximately doubling the PMV baseline. Feature importance analysis ranked air temperature (0.253) as the dominant predictor, followed by weather type, atmospheric pressure, cloud cover, humidity, and wind speed; when interpreted alongside site modifiability, wind speed, humidity, and wind direction emerged as the highest-leverage design variables. Based on these findings, evidence-based guidelines are proposed for nightscape planning, recommending spatial enclosure as the primary intervention in cold-stress seasons, vegetated enclosure over open turf in mild seasons, the conjunction of water-canopy-enclosure for waterside design, and soundscape programming as a legitimate non-thermal comfort strategy. A web-based prediction interface was further developed to translate the hybrid model into a practical planning tool for landscape architects, urban planners, and design-support professionals. By exploring nighttime outdoor thermal comfort and AI-based prediction within an ecologically valid research environment, this study contributes to the LA design process, practice, technologies, and theory. In terms of the design process, it identifies the two-regime structure of nighttime comfort as a framework for designers to prioritize interventions based on seasonal context and user scenarios, offering theoretical and practical suggestions to enhance the LA design process from a perception-centered nightscape planning perspective. In practice, it produces concrete guidelines for spatial enclosure, vegetation configuration, waterside design, and soundscape programming derived from real-world field measurements, and provides a web-based prediction interface that accelerates the design-evaluation cycle in real-world LA projects. Regarding design technologies, the research demonstrates how machine learning, natural language processing, and large language model methods can be integrated with traditional microclimate modeling and develops a hybrid prediction architecture and a data integration strategy that together provide a generalizable template for environmental perception research at scale. Lastly, this research advances LA theory by positioning nighttime as an independent microclimate-perception system, extending the cross-modal perception literature from indoor to outdoor nighttime settings, and integrating neuro-cognitive and adaptive comfort frameworks with AI-based modeling to enrich perception-centered environmental theory in landscape architecture.