The UGS-1m product provides fine-grained UGS maps of 31 major cities in China generated based on a deep learning (DL) framework. This generator is a fully convolutional network (UGSNet) designed specifically for UGS extraction, which integrates attention mechanism to improve the discriminative power of UGS and adopts point tearing strategy for edge recovery. The discriminator is a fully connected network designed to handle domain shifts between images. The detailed description of the dataset is as follows:
1. UGS-1m.zip: A fine-grained UGS map product for 31 major cities in China;
2. UGSet.rip: A large benchmark dataset that supports and promotes UGS research;
3. GUB_Datazip: Global city boundary data for each city;
4. GE_imagery_dataFrame.rip: Google Earth image grid data in ". shp" format, providing image composition for each city;
5. Other Zip files named after city names: Google Earth images of each city.
1. A total of 2179 Google Earth images covering the GUB area of 31 major cities in China were downloaded from Google Earth. Each data box has a longitude of 7 ′ 30 ′ 'and a latitude of 5 ′ 00 ′', with a spatial resolution of nearly 1.1 meters.
2. Collect 4544 images with a size of 512 × 512 and a spatial resolution of nearly 1 meter from 142 sample areas in Guangdong Province, China through GF2 satellite, providing a wide sample database for large-scale urban green space mapping and a benchmark for comparing deep learning algorithms.
3. Use the 2018 Global Urban Boundaries (GUBs; Li et al., 2020) data to occlude urban areas in each sample image to filter out green spaces in non urban areas.
The main steps to obtain UGS-1m can be summarized as follows:
Firstly, pre train UGSNet on UGSet to obtain a good starting training point for the generator;
After pre training on UGSet, the discriminator is responsible for adapting the pre trained UGSNet to different cities through adversarial training;
Finally, using 2179 Google Earth images, UGS results (UGS-1m) were obtained for 31 major cities in China, with a longitude of 7'30 "and a latitude of 5'00" in the data frame, and a spatial resolution close to 1 1 meter.
The evaluation indicators include OA, Pre, Rec, and F1. It can be seen that among the five validated cities, the average OA of all cities is 87.56%, and the OA of each city is higher than 85%. Among them, Changchun has the highest OA rate at 90.62%, while Beijing has the lowest OA rate at 85.86%, indicating that the UGS results in different cities are generally good. In terms of F1 scores, Guangzhou has the highest F1 score at 81.14%, followed by Beijing and Changchun at 79.23% and 77.10%, respectively. Although the F1 scores of Wuhan and Lhasa are relatively low, at 67.71% and 59.85% respectively, the average F1 score of the final UGS results also reached 74.86%. In addition, the average Rec of 76.61% also indicates that the missed detection rate of UGS extraction results is relatively low, which is very important in applications. In summary, the usability of UGS-1m has been preliminarily demonstrated through quantitative verification in multiple different cities.
This work is licensed under a
Creative
Commons Attribution 4.0 International License.
https://cstr.cn/31253.11.sciencedb.07049
10.57760/sciencedb.07049