Computer Vision-Based Identification of Riverfront Landscape Characteristics and Their Associations with Visual Perception: A Case Study of the Meishe River in Haikou, China

The 2026 public consultation draft of the Haikou Special Plan for Urban Renewal (2025–2035) calls for closer integration of blue–green spaces with urban life and underscores the need to improve the quality of urban waterfront public spaces. As a major urban river running through central Haikou, the Meishe River provides a representative setting for examining the relationship between waterfront landscape characteristics and public visual perceptions. Using field-collected images of the waterfront landscape, this study employed a fine-tuned DeepLabV3+ model to derive nine objective visual indicators and used TrueSkill-based pairwise image comparisons to obtain perceived attractiveness, comfort, and naturalness scores. Spatial autocorrelation analysis was used to characterize the spatial patterns of these indicators and perception scores, while spatially corrected correlation analysis and multivariable regression were conducted to examine their relationships. The results showed that: (1) The fine-tuned model achieved an mIoU of 0.712 on the independent test set, indicating satisfactory performance in identifying the principal landscape elements in the study area. All nine objective visual indicators and all three subjective perception dimensions exhibited significant positive spatial autocorrelation, indicating clear spatial clustering in both waterfront landscape composition and visual perceptions. (2) After accounting for spatial dependence and correcting for multiple comparisons, the green view index and color richness index generally showed significant positive associations with the three subjective perception dimensions, whereas the road pavement ratio, facility view index, and dynamic element view ratio predominantly showed negative associations. The building view index and visual entropy were primarily negatively associated with perceived naturalness, while sky openness and the water view index showed no stable associations. (3) Multivariable regression further indicated that the green view index was significantly and positively associated with perceived naturalness, whereas the facility view index, road pavement ratio, and dynamic element view ratio generally showed negative associations with subjective visual perceptions. Overall, the visual perceptions of the Meishe River waterfront were consistently related to vegetation, color characteristics, and built structures, with these relationships being most pronounced for perceived naturalness. The findings can inform waterfront greening, the improvement of built interfaces, and section-specific landscape optimization in the context of urban renewal in Haikou, while providing a reference for creating a vibrant waterfront along the Meishe River and promoting the sustainable renewal of waterfront landscapes.