Assessing urban visual landscape perception with explainable machine learning: the mediating role of sentiment

Urban visual landscape is shaped by both environmental features and human perception. However, previous research has often overlooked the mediating role of sentiment in translating physical landscapes into subjective evaluations of visual quality. Moreover, most studies rely on linear models that are unable to identify effective thresholds or capture nonlinear and interaction relationships between landscape elements and visual quality. This study utilizes a dataset comprising images and text posts uploaded by users to the Weibo social media platform on Xiamen Island, China, in 2021. By employing the Light Gradient Boosting Machine and SHapley Additive exPlanations, the study explores nonlinear relationships, thresholds, and interaction effects between landscape elements and landscape preference. Mediation analysis further reveals the role of sentiment. The results demonstrate that: (1) the key thresholds for variables that significantly improve landscape preference are as follows: when sky and vegetation exceed 22.75% and 28.36%, respectively, and when road remains below 27.68%; (2) landscape elements interact, such that vegetation can offset limited sky openness, and low vehicle density can mitigate road impacts; and (3) sentiment serves as a significant mediator that cannot be ignored, with eight landscape elements having a significant mediating effect. These findings suggest that urban landscape perception and landscape preference are shaped not only by physical landscape elements but also by their capacity to evoke positive sentiment. This study offers quantitative, evidence-based guidance for designing urban environments that are both visually appealing and supportive of positive sentiment.