Interpretable assessment of human-perceived street-level urban greenery quality via panoptic segmented visual...

Street-level urban greenery shapes pedestrians' environmental experiences and psychological well-being, making accurate assessments essential. However, metrics such as the Green View Index primarily focus on the number of green pixels and often overlook the qualitative attributes that influence perception. Consequently, the automated assessment of perceived greenery quality remains limited and fails to explain the factors driving human preference. This study proposes an interpretable machine learning framework for predicting the perceived quality of urban greenery from street-view images. Greenery quality was defined as an aggregated perception score (0-10) based on expert evaluations. Using panoptic segmentation (OneFormer model trained on the ADE20K dataset), nine visual features describing the quantity, color, spatial structure, and morphological complexity of greenery were extracted from greenery segments across 1001 images. Regression and binary classification models were developed to predict perceived greenery quality using two perceptual thresholds that represent moderate and strict quality criteria. Model interpretability was examined using feature importance and SHapley Additive exPlanations (SHAP) analyses. The regression model achieved an R² of 0.60, whereas the classification models attained accuracies of 82.7% under the moderate-quality threshold and 77.4% under the stricter threshold. The quantity of greenery consistently emerged as the dominant predictor, followed by average vividness. Other features exhibited threshold-dependent effects: multiple structural and color-related attributes (e.g., shape complexity, spatial dispersion, and average brightness) contributed at lower thresholds, whereas several became negligible at stricter cutoffs. The SHAP results showed increasingly discontinuous feature contributions at higher thresholds, indicating that high-quality greenery is identified using a limited set of dominant visual cues rather than continuous modulation by multiple attributes. The results demonstrate that perceived greenery quality can be reliably predicted using interpretable visual features and clarify how perceptual criteria simplify under stricter quality judgments, offering a scalable tool for evaluating street-level greenery in urban design and planning.