Deep Learning-Based Assessment of Urban Visual Perception from Street View Images: Implications for Sustainable Urban Planning in the Macau Peninsula

Urban street-view imagery provides a scalable basis for examining visible environmental conditions relevant to human-centered sustainability. This study combines SegFormer-B5 semantic segmentation, six ResNet-50 classifiers trained using Place Pulse 2.0 annotations, spatial autocorrelation, XGBoost–SHAP, and Grad-CAM to analyze 20,256 images from 5064 sampling locations in the Macau Peninsula. We derived four objective visual indicators from the segmentation results. The mean classification accuracy across the six perception classifiers was 73.70% in the source-domain validation. The image-derived indicators and model-predicted scene categories exhibit significant spatial clustering, including global Moran’s I values of 0.673 for the Boring score and 0.637 for the Green View Index. XGBoost–SHAP further identifies nonlinear associations between 19 semantic scene components and the six model-predicted perception scores, while Grad-CAM visualizes the image regions that contribute to individual predictions. The integrated framework links streetscape measurement, spatial diagnosis, and model interpretation to support systematic identification of streets requiring detailed urban-design assessment and locally informed renewal planning.