Urban vegetation is a critical nature-based solution for mitigating the urban heat island effect, yet its cooling efficacy on road surfaces varies spatially due to differences in vegetation type, structure, and local urban morphology. This study investigates the spatially varying nonlinear cooling effects of differentiated street greenery, including trees, shrubs, and grass, on land surface temperature (LST) using Google Street View (GSV) imagery in Zurich, Switzerland. A fine-tuned Mask2Former deep learning model was employed for semantic segmentation to accurately extract five street view factors (trees, shrubs, grass, buildings, and sky) from GSV images collected via three methods: multi-view collection, panoramic image collection, and fisheye. These factors were then compared across methods, revealing methodological biases but limited influence on subsequent modeling outcomes. A geographically weighted random forest model integrated with SHAP (SHapley Additive exPlanations) analysis was developed to capture spatially non-stationary and nonlinear relationships between the five view factors and LST derived from Landsat 8. Results show that trees exert the strongest nonlinear cooling effect, followed by shrubs, while grass provides moderate and context-dependent benefits; buildings consistently drive warming, and sky view has minimal impact. The three GSV collection methods yield comparable predictive performance and feature interpretations, suggesting practical flexibility in method selection. This framework advances street-level urban climate research by combining pedestrian-perspective data with local machine learning interpretability, offering evidence-based insights for targeted greening strategies in heat-vulnerable urban areas.
Lambrecht et al. (Mon,) studied this question.