Modeling visual perception of Chinese classical private gardens with image parsing and interpretable machine learning

Understanding how spatial visual features shape perception is essential for interpreting the design logic of Chinese classical private gardens, yet perceptual thresholds and the role of spatial organization remain underexplored. This study develops a quantitative framework based on Kaplan’s preference matrix, integrating image parsing and interpretable machine learning to examine visual perception in three classical gardens. Visual features, including landscape elements, depth, color, and texture, were evaluated across four perceptual dimensions: coherence, legibility, complexity, and mystery. Results revealed uneven contributions of different features, with nonlinear and threshold effects, while spatial organization shaped perception by regulating the rhythm and intensity of visual features. Waterscapes enhanced coherence, legibility, and mystery. Mountain-view spaces balanced immersion with legibility, and different entrance types either reinforced coherence or encouraged exploration. The framework offers a reproducible approach for linking spatial visual features with perceptual responses in heritage gardens.