Urban surface classification using semi-supervised domain adaptive deep learning models and its application in urban environment studies

Abstract

High-resolution urban surface information, e.g., the fraction of impervious/pervious surface, is pivotal in studies of local thermal/wind environments and air pollution. In this study, we introduced and validated a domain adaptive land cover classification model, to automatically classify Google Earth images into pixel-based land cover maps. By combining domain adaptation (DA) and semi-supervised learning (SSL) techniques, our model demonstrates its effectiveness even when trained with a limited dataset derived from Gaofen2 (GF2) satellite images. The model's overall accuracy on the translated GF2 dataset improved significantly from 19.5% to 75.2%, and on the Google Earth image dataset from 23.1% to 61.5%. The overall accuracy is 2.9% and 3.4% higher than when using only DA. Furthermore, with this model, we derived land cover maps and investigated the impact of land surface composition on the local meteorological parameters and air pollutant concentrations in the three most developed urban agglomerations in China, i.e., Beijing, Shanghai and the Great Bay Area (GBA). Our correlation analysis reveals that air temperature exhibits a strong positive correlation with neighboring artificial impervious surfaces, with Pearson correlation coefficients higher than 0.6 in all areas except during the spring in the GBA. However, the correlation between air pollutants and land surface composition is notably weaker and more variable. The primary contribution of this paper is to provide an efficient method for urban land cover extraction which will be of great value for assessing the urban surface composition, quantifying the impact of land use/land cover, and facilitating the development of informed policies.

Appendices

Appendix A

See Table 5

Table 5 Detailed information of the eight land cover categories in the GF2 dataset

Appendix B. The locations of meteorological and air quality monitoring stations

The locations of meteorological and air quality monitoring stations in Beijing, Shanghai, and GBA are demonstrated in Fig. 

Fig. 9

The location of meteorological monitoring sites (9 MT sites) and air quality monitoring sites (16 AQ sites) in Shanghai, and 12 MT sites and 24 AQ sites in Beijing

9, and

Fig. 10

The location of meteorological monitoring sites (12 MT sites) and air quality monitoring sites (66 AQ sites) in GBA

10. The meteorological monitoring stations are national surface meteorological monitoring stations and location information is provided by China National Meteorological Information Centre. The air quality monitoring network is built by The China National Environmental Monitoring Center (CNEMC).

Appendix C. the impact of the misalignment in Google dataset

The Google dataset is built with image from Google Earth and ground truth from GF2 dataset, as shown in Fig. 

Fig. 11

Difference between the Google dataset and the source dataset. The differences are highlighted using red frames. (a1–a2) Google image and GF2 image of the same area. (b1–b2) Google image and GF2 image with large land cover changes. (c1–c2) Google image and GF2 image has large difference caused by different off-nadir view angle. (a3–c3) The ground truth corresponds to the GF2 image

11(a1–a3). The misalignments exist between the Google Earth and GF2 images mainly because they were obtained at different acquisition times and off-nadir angles. Therefore, after being clipped into smaller tiles, tiles with severe misalignment (Fig. 11(b1–b3)) caused by the large land cover change were excluded. As a result, the target dataset now consists of 31 km² of Google images with ground truth. Misalignments caused by different shading or off-nadir view angle are hardly evitable, as shown in the red frame in Fig. 11(c1–c3). Therefore, the classification accuracy evaluated on the target dataset will be underestimated and can only be used for reference.

With minor land cover change and camera angle difference the classification overall accuracy can reach a high level (Fig. 12 (a1–a4)). However, severely underestimated score will occur when there is a land cover change or camera angle difference, even if the model has correctly classified the images. As shown in Fig. 

Fig. 12

The classification result on the Google image of the M2(3) model. Image differences are highlighted using red frames. (a1-a4) classification result on Google image with minor difference. (b1-b4) classification result on Google image with the difference caused by the off-nadir view angle. (c1-c4) classification result on Google image with large land cover change

12(b1–b4) and Fig. 12(c1–c4), the underlying land cover has been well recognized, but the overall accuracy score is significantly low because of the misalignment between Google image and ground truth caused by land cover change and difference caused by off-nadir view angles. Although we have excluded image tiles with significant land cover changes, misalignment caused by different camera angles still exists in most test images. This is because buildings in the study areas were mostly high-rise buildings with more than seven floors. Therefore, the evaluation result on Google dataset is underestimated.

Appendix D. Detailed settings for the training of deep learning models

The whole process was coded in the PyTorch (v1.10) (Mazza and Pagani 2020) framework. To reduce the dependence on the computer memory size, each training sample was furtherly divided into 64 patches [256m × 256m for one patch and 8 × 8 patches (2048/256 = 8)]. 80% of the training dataset was used for training and the remaining 20% was reserved for performance evaluation. Data augmentation is a common strategy used in deep learning training (Wong et al. 2016; Diakogiannis et al. 2020), especially when the training dataset is not sufficient. In this study, training samples were augmented with the transform modules implemented in torchvison(Mazza and Pagani 2020) including RandomHorizontalFlip, RandomVerticalFlip, ColorJitter, RandomAdjustSharpness, and RandomAutocontrast. Following Li’s work (Li et al. 2019b), the initial learning rate for training the classification model is 2.5 × 10⁻⁴ and decreased with a ‘poly’ learning rate policy with power as 0.9. The batch size is set to 4 and the CycleGAN model was trained for 15 epochs. The training process was conducted on a computer platform with an Intel(R) Core i7 9700K CPU, 64 GB of RAM, and two Nvidia RTX 2080Ti graphics processing units (GPUs).

Appendix E. Correlation analysis result in 2020

Correlation analysis was conducted using meteorological parameters from 2020. Seasonal and annual average values were used for analysis. Additional analysis for air quality parameters hasn't been conducted, because the available air quality stations in Shanghai are all located in compact urban areas which are inappropriate for analysis. As listed in Table

Table 6 Correlation coefficients between land cover and meteorological factors for each season in 2020. Bold values (with |r|≥ 0.3) indicate a linear correlation. Spring: Mar. 21 to Jun. 20; Summer: Jun. 21 to Sept. 20; Autumn: Sept. 21 to Dec. 20; Winter: Dec. 21 to Mar. 20

6, the results indicate a small seasonal variation for most of the correlation coefficients except for two cases: first, the relative humidity is correlated with artificial surfaces in the summer of Beijing which is not shown in other seasons; Second, in the autumn of Beijing and GBA, the correlation coefficients is different from other seasons.

Appendix F. Correlation analysis result for air pollution

The PM_2.5 and PM₁₀ concentrations are positively correlated with artificial land cover fraction in Shanghai, whereas this relationship is insignificant in Beijing and GBA (Table

Table 7 Correlation coefficients between land cover and air quality factors for two seasons in 2021. Bold values (with |r|≥ 0.3) indicate a correlation

7). The strong background PMs pollution and regional transport of pollutants can weaken the significance of intra-urban differences (Wang et al. 2014; Tao et al. 2017; Liu et al. 2017; Fan et al. 2020). Studies have shown that the air pollution in Beijing and GBA is significantly affected by its surrounding area (Liu et al. 2013; Xue et al. 2016; Tao et al. 2017). The local industrial activities and vehicle emission are the major contributors to PMs in Shanghai (Wang et al. 2014), which may explain its stronger correlation than the other two areas.

There is a ubiquitous correlation (0.26 to 0.38) between NO₂ concentrations and artificial land cover in all areas (Table 7). The traffic in urban areas with dense populations is usually heavier than that in rural areas as a major source for NO₂ in urban areas (Lee et al. 2014; Fan et al. 2020; Xue et al. 2020). The NO₂ concentrations of those sites vary significantly even with a similar artificial surface fraction (Fig. 

Fig. 13

Relationship between urban land cover indicators and air quality indicators (two-season average values, PM_2.5, PM₁₀, NO₂) in three regions

13(c, e)). This can be explained by the fact that the traffic intensity can be significantly different even when the fraction of artificial land cover is similar. Furthermore, other local factors like ship emissions can be an important factor affecting the correlation result (Xue et al. 2020).

Traffic is also a major factor in total CO emissions (Hrebtov and Hanjalić 2019). The results show that the CO concentration also has a distinct variation even in areas with a similar artificial surface fraction (Fig. 

Fig. 14

Relationship between urban land cover indicators and air quality indicators (two-season average values, CO, SO₂, O₃) in three regions

14 a, d). However, the correlation between CO concentration and artificial land cover in GBA is weaker (− 0.02 to 0.23) than that of NO₂. This is probably because the concentration of CO is also affected by agricultural and/or residential heating-related biomass burning in rural areas (Sokhi et al. 2021).

A negative correlation (− 0.42 to − 0.46) exists between the fraction of artificial surface and SO₂ concentrations in the summer of Shanghai and Beijing. The major sources of SO₂ are heavy industries and coal-fired power stations (Yoo et al. 2015; Xue et al. 2016, 2020; Li et al. 2017), and these heavy industries are normally located in suburban or rural areas. Therefore, the SO₂ emissions in rural industrial land can be larger than in urban residential areas. For instance, with a heavier industrial intensity, the SO₂ emission in the northern region of Shanghai is larger than in the denser urban central region (Xue et al. 2020). The local concentration of SO₂ also depends on the wind direction. In the summer of Shanghai, the prevailing southeast monsoon from the sea brings cleaner air, causing a lower SO₂ concentration in urban areas (Xue et al. 2020). However, in other seasons, the north wind from more polluted inland may lead to a similar SO₂ concentration throughout the city, making it a regional pollution problem and reducing the intra-urban variability. Therefore, the correlation in Spring can be weak. The correlation between SO₂ and artificial surfaces in GBA is weaker than in Shanghai and Beijing. Unlike Shanghai and Beijing, there still are heavy industries and electric power stations (Hu et al. 2021) in the center of GBA (Foshan city), which can weaken urban–rural differences.

There is no clear correlation shown for the O₃ concentration, which may be because O₃ has a more complex photochemical generation mechanism than other pollutants (Ren et al. 2021). The formation of O₃ can be affected by factors like air temperature and reduced solar radiation in rainy seasons. Therefore, the O₃ concentration shows a strong seasonal variation (Fan et al. 2020).