Machine Learning Assessment of Visual Perception in Suburban Forest Trail Landscapes: a Case Study of Guangzhou Tianlu Lake Forest Park

Visual landscape perception is a pivotal pathway through which suburban forest trails deliver cultural ecosystem services (CES). To address the long-standing trade-off between subjectivity and scalability in traditional assessments, this study integrates Landsense Ecology with machine learning to propose a novel evaluation framework. Using Guangzhou Tianlu Lake Forest Park as a case study, we combine SegFormer-B5 semantic segmentation with Random Forest regression to spatially model CES supply derived from visual landscape perception. Based on 2100 public perception records, this model employs objective indicators such as the Green View Index (GVI) to generate predictions across five perception dimensions, including attractiveness and naturalness, and others. The results reveal a distinct “core—transition—periphery” perceptual gradient: Core forest zones exhibit high naturalness and minimal anthropogenic disturbance, whereas entry plazas act as perceptual depressions. Notably, tree cover and GVI emerge as dominant predictors of positive visual perception, while artificial structures inversely impact CES provision. By translating qualitative perception into spatially explicit indicators, this study advances the shift from descriptive CES assessment to fine-grained, data-driven “precision greening,” offering a transferable paradigm for identifying CES supply-demand mismatches and optimizing peri-urban forest management.