Segmentation of urban green space and water body based on high-resolution remote sensing images

Chen Yang, Ying Wei, Junwei Liu, Hao Yang, Zhoubang He, Jiaguang Li

Abstract: Urban green space and water body play important roles in urban planning and urban ecosystems. Accurate segmentation of urban green space and water body is one of the prerequisites for tasks such as land use/land cover (LULC) and ecological environment protection (EEP). Due to the high complexity, significant inter-class differences, and large temporal variations, challenges arise. With the rapid development of remote sensing technology in recent years, the high-resolution remote sensing image (HRRSI) from satellites provides tremendous assistance for the semantic segmentation of urban features. In this study, a dataset called NEU-RS1 is developed based on HRRSI. To our knowledge, this is the first dataset that includes various types of urban green space and water body from different periods, unified into two classes: urban green space and urban water body. The dataset aims to contribute to explore the interrelationships and combined impacts of green space and water body in urban environments. A data augmentation strategy is designed to generate more training samples and solve the problems in practical applications caused by map rasterization. To improve the accuracy of deep learning (DL) based semantic segmentation models, an improved adaptive threshold algorithm is proposed. The algorithm expands the number of thresholds and uses the Jaccard distance as the objective function. The thresholds with the best segmentation results for each class are solved based on the differential evolution algorithm. Ablation experiments are conducted on the NEU-RS1 dataset based on several widely used DL-based semantic segmentation models, namely FCN, UNet, PSPNet, UperNet, DeepLabv3+, HRNet, and SegFormer. The performance of the proposed data augmentation strategy and adaptive threshold algorithm is tested, and their effects on urban green space and water body segmentation are quantitatively and qualitatively analyzed. The results show that both proposed methods can effectively improve the accuracy of the DL-based models. The best mIoU and mAcc achieved by applying the proposed methods are 85.37% and 91.18%, respectively. The maximum improvement values of mIoU and mAcc are, respectively, 5.10% and 4.38%.