Segmentation of Green Roofs in High-Resolution Remote Sensing Images With GR-Net

Accurately recognizing the spatial distribution of green roofs is crucial for quantitatively assessing their ecological benefits in urban areas. Deep learning (DL) has been applied to this task using remote sensing images, reducing time and labor costs. However, challenges remain due to the irregular shapes, sparse distribution, homogeneity with ground vegetation, and high annotation costs of green roofs. To address these issues, we propose an end-to-end framework for urban-scale green roof segmentation, integrating the following: 1) a high-resolution attention-based convolutional neural network (GR-Net) to extract the contours of sparsely distributed green roof patches; 2) a building guided module (BGM) to reduce mis-segmentation of ground vegetation; 3) a remote sensing prior module (RSPM) to enhance vegetation feature discrimination; and 4) data augmentation and transfer learning to improve learning efficiency and model generalization. Taking Shenzhen, Beijing, and Shanghai as case studies, we construct a diverse green roof dataset with varying sources, spectra, and spatial resolutions. On the in-domain test dataset, GR-Net achieves an ${F}1$ score of 0.842 and an intersection over union (IoU) of 0.744. When applied to out-of-domain test dataset from three new cities, it maintains decent performance, with an ${F}1$ score of 0.756 and an IoU of 0.633. We also identify the optimal configurations for each module. Overall, this work presents a practical and reliable tool for quantitative green roof assessment. The code used in our study is publicly available at https://github.com/wangzhi123321/GR-Net