In recent years, in order to meet the concept of ecological civilization and sustainable development, more and more attention has been paid to scientific urban green space planning and management in China. Therefore, improving the rationality of the UGS classification system and layout and building a green and livable city have been the focus of attention of the government and scholars in recent years. To this end, this paper selects 31 major cities in mainland China as the research area, aiming to construct a comprehensive UGS dataset for deep learning model training under the official classification system, and generate high-resolution green space maps for each city.
The study area includes four municipalities (Beijing, Shanghai, Tianjin, and Chongqing), the capitals of five autonomous regions (Hohhot, Nanning, Lhasa, Yinchuan, and Urumqi), and the capitals of 22 provinces in mainland China (Harbin, Changchun, Shenyang, Shijiazhuang, Lanzhou , Xining, Xi'an, Zhengzhou, Jinan, Changsha, Wuhan, Nanjing, Chengdu, Guiyang, Kunming, Hangzhou, Nanchang, Guangzhou, Fuzhou and Haikou)
The data was drawn by researchers including Shi Qian from Sun Yat-sen University through deep learning methods based on Google Earth images and city boundary data. The data format is raster format (.tif). The spatial resolution is 1m. Data collection and 2020.
The basic flow of this data production is as follows:
-
Using global urban boundaries (GUBs) data to crop remote sensing images, a high-resolution urban green space dataset (UGSet) is constructed, which contains a total of 4544 samples with a size of 512x512.
-
Build a deep learning model consisting of a generator (UGSNet) and a discriminator. UGSNet is a fully convolutional network for extracting green space spaces, which integrates an Enhanced Coordinate Attention (ECA) module to capture more efficient feature representations, and uses a nodding module to obtain refined green space results. The discriminator is a fully-connected network for adapting the spatial mapping of green spaces in different cities through adversarial training.
-
The main steps of the implementation process are as follows: a. First, pre-train UGSNet and use UGSet to obtain a good starting training point for the generator. b. After pre-training on UGSet, the discriminator is responsible for adapting the pre-trained UGSNet to different cities through adversarial training. c. Finally, use 2179 Google Earth images (a data frame with a resolution of about 1.1 meters, a longitude of 7′30′′, and a latitude of 5′00′′) to obtain the spatial distribution data of green space in 31 major cities in China (UGS -1m).
The basic production process of similar data is summarized as follows:
1. Data collection and preprocessing:
-
Download high-resolution remote sensing images (such as 1-meter resolution) corresponding to city boundaries from Google Earth Engine.
-
Obtain city boundary data from publicly available city boundary data sources such as OpenStreetMap.
-
Crop the downloaded remote sensing images and keep the data within the target city.
-
Perform data augmentation on imagery to improve model generalization, such as rotation, scaling, flipping, etc.
2. Label production:
-
The cropped remote sensing images are manually annotated to generate classification labels for green space and non-green space. This is the key data for training deep learning models.
-
You can also try to use semi-supervised or weakly supervised methods to generate labels to reduce the workload of manual labeling.
3. Model selection and training:
-
Choose an appropriate deep learning model, such as U-Net, SegNet, DeepLab, etc., which perform well in remote sensing image segmentation tasks.
-
Input the preprocessed remote sensing images and corresponding labels into the model for training, and adjust the hyperparameters to optimize the model performance.
-
Use cross-validation methods to evaluate the performance of the model, such as precision, recall, F1 score, etc.
4. Model reasoning:
-
Use the trained model to infer the remote sensing images of the whole city to obtain the green space distribution data.
-
It may be necessary to split the remote sensing image into smaller sub-images to fit the input size of the model. After inference is complete, the sub-images need to be stitched back to the size of the original image.
5. Post-processing and analysis of results:
-
Post-processing the green space distribution data output by the model, such as removing noise, filling holes, etc.
-
Quantitative analysis of green space distribution data, such as green space coverage, green space connectivity and other indicators.
-
Green space distribution data can be fused with other geographic information data for more in-depth urban planning and environmental protection research.
Throughout the process, attention needs to be paid to the balance of data quality, model performance, and computing resources. Continuously optimize the model and adjust parameters to achieve better data extraction results of green space distribution.
Data download address:
① Four municipalities directly under the Central Government ( Beijing , Shanghai , Tianjin and Chongqing );
②The capitals of the five autonomous regions ( Hohhot , Nanning , Lhasa , Yinchuan and Urumqi );
③Provincial capitals of 22 mainland provinces ( Harbin , Changchun , Shenyang , Shijiazhuang , Lanzhou , Xining , Xi'an , Zhengzhou , Jinan , Changsha , Wuhan , Nanjing , Chengdu , Guiyang , Kunming , Hangzhou , Nanchang , Guangzhou , Fuzhou , Taiyuan , Hefei and Haikou );