Conference Agenda

Urban streetscape evaluation is evolving from objective, semantic-based analysis to perception-driven approaches that better reflect how people experience city environments. Traditional computer vision captures physical features like greenery and buildings but often ignores perceptual factors such as lighting and weather. This study proposes a new framework that uses generative AI and large-scale human feedback to assess perceptual responses to weather-altered street view images. Using the TSIT model, 88,000 images were transformed to simulate sunny, cloudy, and rainy conditions, with 38,525 participants providing subjective ratings across 22 dimensions. Results show that weather strongly impacts perception: sunny scenes scored highest in beauty, cleanliness, and openness, while rainy scenes rated highest in negative attributes like “boring” and “depressing.” Some qualities, like “greenery” and “interesting,” remained consistent. The framework offers a scalable tool for urban planning and inclusive design by capturing diverse perceptual responses. Future work may consider factors like time of day, air quality, and pedestrian activity to further enrich urban visual analysis.

Recent advancements in machine learning have enabled the modeling of people's perceptions of urban streets using large-scale image datasets such as Google Street View. However, such datasets are typically limited to images captured under clear daytime conditions, which constrain their ability to represent diverse environmental conditions in urban settings. This study aimed to quantitatively examine how weather and time of day influence the way people perceive streetscapes. Street images were collected by the author using a bicycle-mounted camera in four districts of Setagaya Ward, Tokyo, under various weather and lighting conditions. A web-based survey was conducted, in which participants evaluated these images along multiple dimensions, and a predictive model of street impressions was developed using the responses. It was found that: 1) the model identified streets where certain impression scores tended to be higher under nighttime or rainy conditions; 2) a regression-based factor analysis revealed that visual elements, weather, and time of day contributed significantly to impression evaluations; and 3) the characteristics of positively perceived streets varied depending on the environmental context. These findings provide a basis for considering the environmental conditions in streetscape evaluation studies and support context-aware evaluation and planning of streetscapes that reflect local environmental and social needs.

This study proposes a data-driven decision-support framework for identifying highway segments suitable for climbing and passing lane improvements using QGIS-based spatial analysis. Focusing on Thailand’s two-lane highway network, the methodology integrates multiple datasets, including digital elevation models (FABDEM), road geometry, and traffic volumes, with the Highway Capacity Manual (HCM) evaluation criteria. Road segments are standardized and segmented into fixed lengths, then aggregated into 500-meter and 3-kilometer groups for climbing and passing lane analysis. Key thresholds include road gradient, heavy vehicle percentage, and traffic volume filter for candidate segments. The results reveal distinct geographic patterns: passing lane opportunities are concentrated in flatter regions with high traffic flow while climbing lane needs are predominantly located in mountainous northern corridors. By combining open-source tools and national-scale datasets, the proposed framework enables scalable, objective, and transparent planning of auxiliary lanes, supporting safer and more efficient highway infrastructure development. The approach is adaptable and cost-effective, with potential for application beyond Thailand.