In recent years, in order to satisfy the concept of ecological civilization and sustainable development, scientific urban green space planning and management have been paid more and more attention in China . Therefore, improving the rationality of the UGS classification system and layout distribution to build a green and livable city has been the focus of government and scholars in recent years . To this end, this paper selects 31 major cities in mainland China as study areas, 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.

As Fig.  shows, the study area includes four municipalities (Beijing, Shanghai, Tianjin, and Chongqing), the capitals of five autonomous regions (Hohhot, Nanning, Lhasa, Yinchuan, and Ürümqi), 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 UGSNet contains two parts, namely a backbone to extract multi-scale features and generate coarse results and a point head module to obtain fine-grained results.

In order to transfer the prior knowledge from UGSet to images from other regions, a discriminator is adopted to obtain a well-adapted UGSNet for each city in an unsupervised way. As shown in Fig. , the discriminator consists of five convolutional layers with a kernel size of 4 and a stride of 2, each connected by a leaky ReLU layer. The output channels of each convolutional layers are 64, 128, 256, 512, and 1, respectively. Given an input of the softmax prediction map from the generator, P∈RH×W×C, the discriminator will output a discriminant result of the input, D(P)∈Rh×w×1. After that, the discriminator will optimize itself according to the discrimination accuracy through the cross-entropy loss in Eq. (5).

The training of the proposed deep learning framework can be divided into two steps, namely pre-training and adversarial training. At the beginning, the UGSNet will be fully trained on UGSet to obtain the initial parameters for the generator. After that, the discriminator will be adopted to help generalize the pre-trained UGSNet to different target cities through adversarial learning. Detailed information on the optimization process is described in the following.

During the pre-training process, the training set of the UGSet is used for parameter optimization, while the verification set was used to monitor the training direction and save the model in time. Five common semantic segmentation models are selected for comparison to prove the validity of UGSNet, including U-Net , SegNet , UPerNet , BiSeNet , and PSPNet . In addition, an ablation study is also conducted to further verify the effectiveness of the ECA modules and the point head. All models are fully trained for 200 epochs based on the Adam optimizer, with an initial learning rate of 0.0001, which begins to decline linearly in the last 100 epochs. A batch size of eight sample pairs is adopted due to the limitation on GPU memory. Data augmentation was applied during model training, including cropping, random rotation, and flipping. Specifically, cropping refers to randomly clipping an area of 256×256 size from the input sample of size 512×512. Rotation refers to randomly rotating the sample at a random angle within the range of [-30∘, 30∘] in the proportion of 20 %. Flipping refers to a mirror flip of samples from four different angles (up, down, left, and right) with a probability of 30 %. After training, all selected models were compared on the test set. The adversarial training process lasts for 10 000 epochs, in which the batch size is set to 2. Both the generator and the discriminator employ an initial learning rate of 0.0001. All experiments are implemented in PyTorch environments and are conducted in GeForce RTX 2080 Ti to accelerate model training.

Five indices are involved in the evaluation, including precision (Pre), recall (Rec), F1 score, intersection over union (IoU), and overall accuracy (OA). Suppose that TP, FP, TN, and FN refer to the true positive, false positive, true negative, and false negative, respectively. These indices can be defined as follows: 8Pre=TPTP+FP9Rec=TPTP+FN10F1=2precision⋅recallprecision+recall11IoU=TPFP+TP+FN12OA=TP+TNFP+TP+FN+TN. During the pre-training process, Pre, Rec, F1, and IoU are utilized to measure the model performance on UGSet, which is commonly used in semantic segmentation tasks. In the process of verifying the accuracy of the generated UGS maps (UGS-1m), Pre, Rec, F1 and OA are employed.

Figure  provides an overview of UGS-1m product, which can be downloaded from 10.57760/sciencedb.07049 . In order to show the detailed information of the UGS-1m, two groups of magnified snapshots are also provided in Figs.  and . It can be seen that the results of UGS-1m are consistent with the green space in the image, no matter which of the three developed municipalities in Fig.  or the three capital cities in development in Fig.  are evaluated. While these comparisons can initially display the good visualization effect of UGS-1m, the accuracy evaluation is also conducted to further test it from a quantitative perspective. Since there is no large-scale and fine-grained UGS ground truth for an accuracy evaluation, five different cities are selected to evaluate the reliability of the UGS results in the UGS-1m, including Changchun, Beijing, Wuhan, Guangzhou, and Lhasa, as shown in Fig. . In total, 17 sample tiles are collected from the five cities, among which Changchun, Beijing, and Guangzhou each contribute four tiles. Due to the relatively small built-up areas, three and two tiles are collected from Wuhan and Lhasa, respectively. The UGS annotations of all tiles are obtained by expert interpretation. The accuracy evaluation is conducted between the annotated reference map and the result in UGS-1m.

The evaluation results are summarized in Table , which are evaluated by OA, Pre, Rec, and F1. It can be seen that, in the five cities for verification, the average OA of all cities is 87.56 %, while the OA of each city is higher than 85 %. Among them, the highest OA is 90.62 % in Changchun, while the lowest OA also reaches 85.86 % in Beijing, indicating that the UGS results in different cities are basically good. In terms of the F1 score, Guangzhou has the highest F1 score of 81.14 %, followed by Beijing and Changchun with 79.23 % and 77.10 %, respectively. Though 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 reaches 74.86 %. Moreover, the average Rec of 76.61 % also denotes a relatively low missed-detection rate of the UGS extraction results, which is significantly important in applications. In general, after quantitative validation in several different cities, the availability of UGS-1m is preliminarily demonstrated.

The qualitative analysis is carried out in this section to further analyze the performance of UGS extraction and its relationship with external factors, such as geographical location, UGS types, phenological phase, etc. Therefore, visualization comparisons conducted in three cities, including Changchun, Wuhan, and Guangzhou, are displayed in Figs. –.

From the overview image of Changchun (Fig. ) and Guangzhou (Fig. ), it can be seen that the extracted UGS results are in good agreement with the reference maps, which is mainly reflected in the good restoration of UGS at various scales in each example image. The magnified area of each image further shows the details of UGS-1m for extracting different kinds of UGS, including parks, squares, green buffers, and attached green spaces. Specifically, the UGS-1m performs well in the extraction of green spaces attached to residential buildings, although they are complex and broken in morphology compared to other UGS types. Notably, though Changchun and Guangzhou are geographically far away, distributed in the northernmost and southernmost regions of China, respectively, the UGS results in these two cities are both good. This shows that the performance of the proposed UGS extraction framework is unlikely to be affected by the difference in geographical location, which may attributed to the adversarial training strategy to help model adaptation.

The visualization result of Wuhan is further provided in Fig.  for analysis. It can be seen that the UGS results in Wuhan are mainly influenced by the shadows of buildings. On the one hand, the UGS features are sometimes blocked by building shadows, resulting in relatively poor extraction effect, such as the magnified area in Fig. b. On the other hand, the building shadows can easily be extracted as attached green space, according to Fig. c. This shows that the result of green space extraction is related to the angle of the image taken. When the angle is larger, it is more likely to have building shadows in the image, thus affecting the subsequent UGS extraction, especially the green space that is attached to buildings. In addition, the results are also affected by the phenological phase, as shown in Fig. a. On the whole, the UGS with a higher and denser vegetation canopy is easier to identify accurately, and on the contrary, the lower and sparser UGS is more easily misclassified due to the similar appearance with other land types, such as bare land.

We compare the UGS-1m results with the existing global land cover products to verify the reliability of the results, including GlobeLand30 , GLC_FCS30 , and Esri 2020 LC (© 2021 Esri). Due to different classification systems, these products need to be reclassified into two categories. Specifically, forests, grasslands, and shrublands are reclassified as UGS, while the other categories are reclassified as non-UGS. Examples from six different cities in different geographical locations, including Changchun, Ürümqi, Beijing, Chengdu, Wuhan, and Guangzhou, are collected to give more comprehensive demonstrations of our UGS results. The visualization comparison among UGS-1m, GlobeLand30, GLC_FCS30, and Esri 2020 LC is shown in Fig. . Apparently, the three comparative products contain most large-scale UGS, among which the GlobeLand30 performs best with the most complete UGS prediction. However, many detailed UGS features are still missed due to the limitation in the spatial resolution of the source image. Remarkably, the UGS-1m provides UGS with a relatively large scale and detailed UGS information such as attached green spaces. The results and comparisons fully demonstrate the effectiveness and potential of the proposed deep learning framework for large-scale and fine-grained UGS mapping.

While the previous evaluations and comparisons have proved the advantages and validity of the UGS-1m, this section will explore the potential utility of the UGS-1m product. As mentioned in the introduction, green space equality is critical for achieving the Sustainable Development Goals, and this requires fine-grained distributions of UGS as basic data. At this point, compared with the traditional Statistical Yearbook data, our UGS-1m product can provide relatively objective and more detailed information about the distribution of green space. Therefore, statistics are conducted on the UGS-1m data to obtain particular information of each city, including the UGS area and the UGS rate in the GUB area, which are summarized in Table .

As shown in Table , the UGS area of the different cities varies greatly. For example, it is seen that the UGS area of Beijing is the largest among all 31 cities studied, reaching 1021.55 km2, while that of Lhasa is the smallest, with only 11.41 km2. The statistical information on the UGS area indicates that UGS-1m can provide a quick and intuitive comparison of the stock of UGS in different cities. However, a small UGS area does not always mean a lack of green space because the urban area needs to be considered at the same time. Therefore, the UGS rate in the GUB area is further calculated to measure the volume and distribution of green space in different cities. At this time, it can be seen that Yinchuan has the highest UGS rate, accounting for 25.35 %, while Beijing, which has the largest UGS area, has a slightly lower UGS rate of 20.74 %. Besides, the UGS rate of 9.27 % in Lhasa, which has the least UGS area, is also slightly better than that of Lanzhou, which has the lowest UGS rate of 8.64 %. The results and comparisons further show that the shortage and imbalance of green space cannot be reflected only from the perspective of the stock of green space, and more information is often needed for a more comprehensive analysis.

The above statistical analysis only shows the most simple and intuitive applicability of UGS-1m as a large-scale and refined green space product, but it is far more than that. Since the high-resolution UGS-1m can provide a detailed distribution of green space, it provides a possibility for much more research under fine-grained scenarios. When considering different datasets or materials, more comprehensive information of UGS can be obtained. For example, by combining high-resolution population (e.g., WorldPop data; ) with UGS-1m, the availability of residents to green space can be measured, that is, green space equity. Or, when combined with the distribution of urban villages and formal housing spaces, their differences in the green landscape pattern can be studied. We also hope to explore these research areas in the future.

As mentioned above, before the start of adversarial training stage, the generator will be initialized by a pre-trained UGSNet on UGSet. Therefore, in order to fully verify the advancement of UGSNet and its qualification to initialize the generator, this section introduces several state-of-the-art (SOTA) deep learning models as candidate generators for comparison. Note that the comparative experiment is completely conducted on UGSet, and no discriminator is introduced. After all the models have been fully trained on training set of the UGSet, the best-trained parameters of each model will be evaluated on the testing set of the UGSet. The comparative results are provided in Table .

As Table  shows, the proposed UGSNet outperforms all SOTA baselines with the highest F1 score and IoU of 77.30 % and 62.99 %, respectively. The second-ranked PSPNet obtains an IoU of 60.96 %, which is 2.03 % lower than that of UGSNet. The ablation study indicates that the integration of the ECA modules and point head can improve the base model by 0.17 % and 0.69 % on IoU, respectively, which proves their effectiveness on UGS extraction. The IoU of UGSNet is 1.22 % higher than that of the base model, indicating that the combination of ECA modules and point head has a greater gain effect. The quantitative results demonstrate the validity of UGSNet.

Figure  further demonstrates the performance of different methods on different kinds of UGS. It can be seen that, after fully training on the UGSet dataset, each model can identify the approximate region of various green spaces, including SegNet, which has shown poor performance in the quantitative comparisons. Therefore, the superiority of the green space identification results is mainly reflected in two aspects. One aspect is the ability to extract the UGSs of great inter-class similarity. As shown in the first row of Fig. , the UGSNet can accurately identify the yellow box area, in which the green space of a park has confused most comparative methods. Another aspect is the capability to grasp fine-grained edges, especially for small-scale UGS, such as the attached UGS in the last row of Fig. .

As we have mentioned above, the proposed framework is composed of a generator and a discriminator, which adopts the adversarial training to help the model transfer learning. In order to test the effectiveness of the proposed framework, this section further conducted ablation experiments with and without the discriminator, which, respectively, correspond to the following:

Our framework (G+D) contains a generator and a discriminator, in which the generator is initialized by the UGSNet pre-trained on UGSet, and the discriminator is employed at the adversarial training stage to overcome domain shifts and obtain a refined UGSNet for each target city before generating the UGS map for it.

At present, the availability of high-resolution images is still severely limited by factors such as temporal resolution, image distortion, and cloud occlusion. Therefore, it is hard to collect all the Google Earth images used to generate UGS-1m in the same period, making it difficult to keep the unity of phenology. In the proposed deep learning framework, we introduce a domain adaptation strategy to deal with this problem to some extent. A comparison of the results has shown the effectiveness of UGS-1m and the feasibility and potential of the proposed deep learning framework for large-scale, high-resolution mapping. However, the current DL techniques still have limitations. In view of the spatial and spectral diversity of high-resolution remote sensing images, we have not been able to fully evaluate the domain adaptation effect of the adversarial framework for all heterogeneous images with different dates and resolutions. With the emergence of more and more high-resolution satellite images, the adversarial learning of multi-source images remains to be explored. Future works will be dedicated to extracting UGS information based on data with higher temporal resolution, such as synthetic-aperture radar (SAR) images and unmanned aerial vehicle (UAV) images.

Besides, even though the UGSet has proven to be practicable for UGS mapping, we still have to point out that there may be a small number of missing labels in the dataset, especially for attached green spaces. As analyzed above, the extraction of attached green spaces can be more easily affected by external factors, like the angle of the image taken. This also affects the annotation process. Fortunately, despite the possible deficiencies, the DL model can still learn from a large number of accurate annotations and capture the characteristics of different types of green spaces due to the strong generalization ability. We still hope that, in the following work, more attempts can be made to deal with the problems of labeling and identifying hard UGS types.