AI Summary & Key Takeaways
AI InsightCompared to YOLOv11, these results show notable improvements of 5.43%, 14.63%, 7.41%, and 9.20% across key metrics.
📌 3 Key Takeaways
Abstract
Accurate sky identification in street images is crucial for calculating the sky view factor (SVF) and understanding heat effects in urban environments. However, variable weather conditions—such as overcast, sunny, and rainy scenarios—pose significant challenges to this task. Additional difficulties arise due to limited visible sky regions, shadows, and color similarities between the sky and surrounding elements like buildings or walls. This study proposes an Improved YOLO model featuring a Swin Transformer backbone to address these issues, effectively capturing contextual information under complex and diverse weather conditions. The model incorporates SPPF as the neck to aggregate multi-scale features and GS-ELAN as the head to enhance information flow and feature sharing. A Segmentation Head is also integrated to provide precise pixel-level sky predictions. Data augmentation techniques simulating various weather and perspective conditions were employed to improve robustness. Experimental results reveal that the Improved YOLO achieves high performance with Precision, Recall, mAP, and F1 scores of 0.87, 0.94, 0.87, and 0.95, respectively. Compared to YOLOv11, these results show notable improvements of 5.43%, 14.63%, 7.41%, and 9.20% across key metrics. Despite the widespread use of semantic segmentation models for SVF estimation, YOLO-based architectures for sky view factor detection in perspective street images remain limited. This study addresses this gap with an improved YOLOv11 model. The model’s strong performance under diverse environmental conditions demonstrates its effectiveness for real-time sky detection, offering promising applications in urban planning and environmental research. Overall, this work contributes significantly to urban heat island research by enabling accurate and efficient assessment of sky view factors in urban areas.