UGS-1m: fine-grained urban green space mapping of 31 major cities in China based on the deep learning framework

Urban green space (UGS) is an important component in the urban ecosystem and has great significance to the urban ecological environment. Although the development of remote sensing platforms and deep learning technologies have provided opportunities for UGS mapping from high-resolution images (HRIs), challenges still exist in its large-scale and fine-grained application due to insufficient annotated datasets and specially designed methods for UGS. Moreover, the domain shift between images from different regions is also a problem that must be solved. To address these issues, a general deep learning (DL) framework is proposed for UGS mapping in the large scale, and fine-grained UGS maps of 31 major cities in mainland China are generated (UGS-1m). The DL framework consists of a generator and a discriminator. The generator is a fully convolutional network designed for UGS extraction (UGSNet), which integrates attention mechanisms to improve the discrimination to UGS, and employs a point-rending strategy for edge recovery. The discriminator is a fully connected network aiming to deal with the domain shift between images. To support the model training, an urban green space dataset (UGSet) with a total number of 4544 samples of 512 x 512 in size is provided. The main steps to obtain UGS-1m can be summarized as follows: (a) first, the UGSNet will be pre-trained on the UGSet in order to obtain a good starting training point for the generator. (b) After pre-training on the UGSet, the discriminator is responsible for adapting the pre-trained UGSNet to different cities through adversarial training. (c) Finally, the UGS results of 31 major cities in China (UGS-1m) are obtained using 2179 Google Earth images with a data frame of 7 ' 30 '' in longitude and 5 ' 00 '' in latitude and a spatial resolution of nearly 1.1 m. An evaluation of the performance of the proposed framework by samples from five different cities shows the validity of the UGS-1m products, with an average overall accuracy (OA) of 87.56 % and an F1 score of 74.86 %. Comparative experiments on UGSet with the existing state-of-the-art (SOTA) DL models proves the effectiveness of UGSNet as the generator, with the highest F1 score of 77.30 %. Furthermore, an ablation study on the discriminator fully reveals the necessity and effectiveness of introducing the discriminator into adversarial learning for domain adaptation. Finally, a comparison with existing products further shows the feasibility of the UGS-1m and the great potential of the proposed DL framework.