Let the landscape dance: Spatial dynamic changes of visual elements in urban park landscapes based on video panoramic segmentation

Abstract

Amid global efforts to promote healthy cities and high-quality public spaces, the dynamic variation in visual elements in urban park environments represents a novel and under explored research frontier. Traditional visual impact assessments (VIA) have largely relied on static images, overlooking the way visual compositions shift during movement. This study adopts Video Panoramic Segmentation (VPS) and uses walk-through video data alongside a TabPFN+SHAP deep learning regression model to examine how both static proportions and dynamic variation patterns of landscape elements relate to park popularity. The findings show that (1) dynamic features—such as amplitude, frequency, period, slope—exhibit stronger correlations with popularity than static composition ratios; (2) a novel methodological framework is proposed for dynamic VIA using VPS, Maximum Information Coefficient (MIC), and TabPFN; (3) Lower variation cycles for elements such as vegetation, roads, sky, grassland, and buildings are associated with higher park popularity. The frequency of these variations should not exceed 1.0 occurrences per minute, with an optimal value around 0.5 occurrences per minute; (4) Dispersed distribution of elements such as buildings and water bodies is preferable to concentrated arrangements. The accessibility of service facilities exerts a substantial influence on park popularity and (5) This study confirms that variations in the visual environment exert a more significant influence on park popularity than static visual features. This is the first study to quantify the dynamic variation indicators of visual elements and to assess their impact on VIA. These findings provide an evidence-based foundation for guiding spatial rhythm and variation strategies in environmental design.