关键词: 绿视率, 街景数据, 深度学习, 语义分割, 迁移学习

Abstract: The extraction of green vision index based on semantic segmentation model lacks applicability. Based on DeepLabv3+semantic segmentation pre-training model and manually-labeled samples, this study adopts transfer learning strategy to build a semantic segmentation model for street view image, and evaluates model performance. Then, based on the semantic segmentation model of street view image, GVI within the Third Ring Road in Beijing is extracted and calculated, and the spatial distribution characteristics of GVI at point and line scales are analyzed. The results show that:①Compared with DeepLabv3+green vision segmentation pre-trained model, F1 value and mIoU value of the transferred model are increased by 7% and 3%, respectively.②The GVI in the study area at the point scale has a clustering pattern which is high in the northwest and low in the southeast. There are 58.1% of street view sampling points with the GVI values ranging from 0 to 0.15 which indicates a relatively low green vision perception degree within the Third Ring Road in Beijing.③The GVI values at the linear scale are low on the ring road and high between the rings and witnesses a center diverges outward. There are 59.8% of roads with the GVI values ranging from 0 to 0.15. This study can provide an significant reference for improving the perception degree of urban street greening and urban spatial planning.

Key words: green visual index, street view data, deep learning, semantic segmentation, transfer learning