Swin-CFNet: An Attempt at Fine-Grained Urban Green Space Classification Using Swin Transformer and Convolutional Neural Network

Urban green space plays a critical role in contemporary urban planning and ecology as they provide recreational space for residents, promote ecological balance, and enhance the quality of the urban environment. However, the rapid development of urbanization poses increasingly complex challenges to the monitoring and management of these spaces. Previous studies have illustrated that semantic segmentation models based on convolutional neural network (CNN) perform well in classifying urban green space using high-resolution remote sensing images. However, there are still some deficiencies in CNNs model in capturing global information of green space and dealing with complex spatial relationships due to the special nature of urban environments, such as fragmentation of green space. Hence, swin transformer-CNN-fusion-network (Swin-CFNet) was proposed for urban green space classification, which overcomes the limitations of traditional methods in dealing with global green space information and complex spatial relationships by constructing a residual-swin-fusion (RSF) module for the fusion of multisource features. Experimental results demonstrated that the Swin-CFNet outperformed the UNet in urban green space classification, achieving an overall accuracy (OA) of 98.3% and improving the mean intersection over union (mIoU) compared to UNet and SwinUnet by 3.7% and 1%, respectively.