In the case of large-scale supervised classification, both the quantity and quality of samples are important (Foody and Arora, 1997). ISA is a highly variable object, and its attributes in the Sentinel-2 multispectral images are related to materials, viewing angles, and atmospheric conditions, while its response to the Sentinel-1 SAR instrument depends on dielectric properties, geometry, and surface roughness. Hence, a large number of training samples was required to address the aforementioned challenges that would be encountered at the global scale. Training samples are usually acquired by means of visual interpretation or automatic extraction from the existing datasets. However, the visual interpretation methods are labor- and time-intensive, even for small regions. Therefore, at a large scale, training samples are usually extracted from the existing datasets with similar temporal and spatial coverages. However, the sample quality is affected by the quality of the datasets. Theoretically, samples extracted from a single dataset will result in more errors and uncertainties, while multi-source data can improve the reliability of the training samples (Huang and Zhang, 2013). We therefore proposed to collect global training samples by incorporating the existing ISA datasets and the crowdsourced OSM database. To distinguish the two types of ISA samples, we named the ISA samples extracted from the existing satellite-derived ISA datasets as ISARS and those extracted from the OSM as ISAOSM.

The existing ISA datasets generally cover a broad terrestrial surface, but they differ in terms of their definition, spatial resolution, and temporal coverage. In this study, the GISA, FROM_GLC10, and GlobeLand30 products were chosen to extract the training samples for the following reasons: (1) the GISA is aimed at mapping the global ISA, which is consistent with GISA-10m; (2) the team behind the GlobeLand30 employed extensive visual interpretation to detect artificial surfaces, which can effectively reduce the false alarms from other datasets, i.e., GISA and FROM_GLC10 (Chen et al., 2015); and (3) the definition of FROM_GLC10 (impervious surfaces) is consistent with that of GISA-10m, and its spatial resolution is also 10 m. The GHSL 2018, WSF2015, and GAUD were not considered since they aim to outline human settlements or urban extent (Table 1). We then collected the eligible training samples according to the following rules.

Temporal rule. The GISA is a global ISA dataset covering 1972–2019, so we selected its results for 2016 to match the Sentinel data used in this research. GlobeLand30 documents global land cover for 2000, 2010, and 2020, so the 2010 map was chosen in this study. Although the 2020 map is more recent than the 2016 map, it contains ISA that was built after 2016, making it unsuitable for the GISA-10m mapping. Although there is a 6-year gap between GlobeLand30 and the other datasets (i.e., GISA and FROM_GLC10), we adopted the commonly used assumption that the transition from ISA to non-impervious surface area (NISA) rarely happens (P. Gong et al., 2020; Huang et al., 2021b, 2022), so that the GlobeLand30 for 2010 could be used for the GISA-10m mapping. The following spatial and spectral rules were used to remove the possible errors.

We extracted the ISAOSM samples from the OSM buildings and roads through the following rules.

Temporal rule. We chose the OSM data built on 2 January 2017 in terms of the time of GISA-10m. This version of the OSM data was employed to ensure that the buildings and roads were constructed in 2016 or before, and hence were suitable for the 2016 ISA mapping.

On the other hand, NISA (non-ISA) training samples are also important for accurate ISA mapping. We used the three existing datasets (i.e., GISA 30 m, FROM_GLC10, GlobeLand30) and the OSM to generate the NISA samples. Firstly, we took the intersection of the NISA regions in the three datasets as the initial NISA sample pool: 1NISA=NISAGISA∩NISAGlobeLand30∩NISAFROM_GLC10-ISAOSM. For GlobeLand30 and FROM_GLC10, NISA is defined as all the land-cover types other than ISA. We masked the initial NISA sample pool using the OSM buildings and roads to suppress the errors in the existing global datasets. To this end, we used the OSM version built in December 2020 (https://planet.openstreetmap.org/planet/2020/planet-201207.osm.bz2, last access: 13 March 2021), which documents more buildings and road networks than the 2017 version. In addition, we buffered the OSM roads with a 30 m buffer to mitigate the errors. Subsequently, 30 000 points were randomly selected in each hexagon as NISA samples. The distance between each NISA sample was kept larger than 200 m to ensure the diversity and irrelevance. Finally, we extracted 58 million training samples (51 674 533 NISA samples and 6 897 378 ISA samples) for the GISA-10m mapping.

The dedicated image pyramid of the GEE platform enabled us to perform pixel-wise feature extraction (Gorelick et al., 2017). Therefore, based on all the available Sentinel data for 2016, we constructed a set of spectral, phenological, textural, SAR, and topographical features with the temporal composite method (Table 2). This approach used all the available data and, at the same time, allowed us to reduce the feature dimension, preserve the temporal information, and minimize the effects of clouds and shadows (Yang and Huang, 2021). Firstly, we used the spectral signatures provided by the Sentinel-2 data to extract the ISA in the visible, red-edge, NIR, and infrared bands (Table 2). Moreover, considering that spectral indices can increase the differences between land covers, we also extracted a series of normalized spectral indices to enhance the discriminative ability between ISA and NISA (Yang and Huang, 2021) (Table 2). These indices were built according to the following criteria: (1) they were mainly constructed by the NIR and SWIR bands due to their better atmospheric transmission (Huang et al., 2021b; Yang and Huang, 2021); and (2) each index contained at least one 10 m band (i.e., visible and NIR bands) to ensure the spatial resolution of the features.

The complex spectral and spatial characteristics in urban environments increase the difficulty of ISA mapping. In this regard, textural features are usually employed to depict the spatial information of urban ISA (Huang and Zhang, 2013). To further exploit the textural information for the ISA mapping, we computed the gray-level co-occurrence matrix (GLCM) via the NIR band to depict the spatial information of urban ISA. Owing to the high redundancy among GLCM measurements (Clausi, 2002), we chose the contrast, dissimilarity, entropy, inverse difference moment (IDM), and angular second moment (ASM) for the texture extraction (Rodriguez-Galiano et al., 2012). The window size for the GLCM measurements was set to 7 × 7 as this is suitable for urban classification with an image resolution from 2.5 to 10 m (Puissant et al., 2005). In addition, we averaged the GLCM from different directions (0, 45, 90, and 135∘) to maintain the rotational invariance (Rodriguez-Galiano et al., 2012).

Given that the spectra and backscatter of some NISA (e.g., vegetation and water bodies) vary throughout time, the phenological information derived from the multi-temporal spectral and SAR data was utilized to depict the temporal fluctuations. We calculated the maximum NDVI and the standard deviation of the NDVI (Tucker, 1979), MNDWI (Xu, 2006), and normalized difference built-up index (NDBI) (Zha et al., 2003), to further enhance the temporal information. These temporal characteristics are useful in identifying NISA with temporal fluctuations. For example, the spectra of fallow cropland and ISA are similar, and even SAR data may not separate them well. However, the NDVI of cropland can describe the changes of crop growth, and hence its standard deviation can be used to distinguish between ISA and cropland. In addition, to increase the robustness of these temporal features, Sentinel-2 data from two adjacent years were also considered.

SAR data have the potential to reduce the false alarms caused by bare soil in optical images, and are more sensitive to buildings. In addition, SAR signals can penetrate clouds. Therefore, in this study, SAR data were combined with optical data for the ISA mapping. Specifically, the vertical–vertical (VV) polarization and the vertical–horizontal (VH) polarization backscatter coefficients from the Sentinel-1 images were selected. Based on all the available Sentinel-1 data, the annual mean and standard deviation of the VV and VH backscatter coefficients were calculated by the temporal composite method: 2σmean=1n∑i=1nσi3σstd=∑i=1n(σi-σmean)2n, where n denotes the total number of Sentinel-1 observations within a year, and σi represents the ith backscatter coefficient observation in the year. The temporal composite method can reduce the speckle noise in SAR imagery (Lin et al., 2020), while the annual standard deviation can reflect the temporal information. Topography-related features are also necessary for ISA mapping in order to reduce the confusion between complex terrain and buildings. For instance, topographical features can help to distinguish steep hills from buildings (Gamba and Lisini, 2013). Specifically, we used Shuttle Radar Topographic Mission (SRTM) digital elevation model (DEM) data in the areas below 58∘ latitude, and Global Multi-resolution Terrain Elevation Data 2010 (GMTED2010) in the areas above 58∘ (Huang et al., 2021b). Finally, a total of 41 features were constructed from the 2.7 million Sentinel images (2 613 180 Sentinel-2 and 122 156 Sentinel-1) and DEM data.

When dealing with global land-cover classification, the global terrestrial surface is usually divided into homogeneous sub-regions according to criteria such as climate, land cover, or administrative regions (Goldblatt et al., 2018). For global ISA mapping, regular square grids are commonly used (Table 1), such as 1 and 5∘ grids (e.g., WSF2015 and GLCFCS). In this study, we divided the terrestrial surface into 2∘ hexagonal grid cells (Fig. 1) due to the symmetry and invariance (You et al., 2021). Furthermore, there were no gaps or overlaps between hexagons, and the distance between adjacent hexagon centers was approximately equal (Richards et al., 2000).

The RF classifier has been widely used in global ISA mapping due to its robustness to erroneous samples, flexibility with high-dimensional data, and tolerance to noise (Bauer and Kohavi, 1999; Wulder et al., 2018) (Table 1). The RF classifier utilizes ensemble learning to obtain predictions by voting on categories through multiple decision trees (Breiman, 2001). Each tree uses a random subset of the input features to increase the generalization ability. In addition, trees are grown from different subsets of training data (i.e., bagging or bootstrapping) to increase the diversity (Rodriguez-Galiano et al., 2012). RF has been shown to outperform other classifiers when dealing with large-scale and high-dimensional data (Goldblatt et al., 2016). The ability of RF to handle multi-source data also makes it convenient when dealing with Sentinel radar and optical data. Therefore, together with the aforementioned multi-source features and global training samples, the RF classifier was used for the GISA-10m mapping. As suggested by Yang and Huang (2021), the number of trees was set to 200. We divided the global terrestrial surface into 1808 hexagons, and a local RF model was built for adaptive ISA classification in each hexagon. Therefore, a total of 1808 RF models were built. In terms of the features used to train each tree, the random forest uses a random subset of features to reduce the correlation between trees. In general, the diversity of the trees can be increased when fewer features are used for training each tree (Breiman, 2001). In the GISA-10m mapping, we set the number of features used for each tree to the square root of the total number of features, as suggested by H. Liu et al. (2020).

The accuracy assessment based on the visually interpreted samples is shown in Table 3. GISA-10m exhibits the highest OA of 86.06 %, representing an increase in OA of +2.73 %, +3.73 %, and +2.3 % over GHSL 2018, GLCFCS, and WSF2015, respectively (Table 3). The kappa of GISA-10m is 0.7165, which exceeds WSF2015, FROM_GLC10, and GAIA by 0.052, 0.1774, and 0.2039, respectively. GISA-10m also shows a higher accuracy than the 30 m resolution datasets (i.e., GISA, GAUD, GAIA), which suggests a better delineation of global ISA due to the higher resolution. Figure 6 summarizes the results of the accuracy assessment at the continent level, with the average and standard deviation of the OA for each continent shown in the box plots. Overall, GISA-10m exhibits a stable performance for each continent with an average OA of more than 85 %. Specifically, Oceania and South America show the best OAs of 87.25 % and 87.08 %, followed by Europe (86.45 %) and Asia (85.85 %). The results also show that the average OA of GISA-10m exceeds that of the existing datasets in Africa, North America, and Europe. In addition, it is apparent that the performance of GHSL 2018 and GLCFCS is relatively unstable in South America and North America, respectively.

GISA-10m obtains the best OA of 86.25 % on the ZY-3 samples, outperforming GHSL 2018, GLCFCS, and WSF2015 by 1.72 %, 1.69 %, and 0.81 %, respectively. The ZY-3 images employed by Liu et al. (2019) covered 45 major global cities, and therefore the ZY-3 samples were more inclined to reflect the accuracy in urban regions. Therefore, the accuracy difference between the various datasets is not significant (Table 3). Due to the relatively coarse resolution, the 30 m datasets usually tend to overestimate the ISA extent (P. Gong et al., 2020), resulting in a higher UA but lower PA (Table S1 in the Supplement). For example, the ISA UA of GISA is slightly higher than that of GISA-10m, but its PA is much smaller (Table S1). However, when the two metrics (PA and UA) are considered at the same time (i.e., the F1-score), GISA-10m outperforms GISA.

The population of rural regions is comparable to that of urban regions (https://data.worldbank.org/, last access: 9 August 2022). The existing studies, and their global ISA datasets, have usually focused on the performance in urban regions, and the accuracy of the rural ISA regions has not been sufficiently assessed. Hence, in this study, we paid special attention to the accuracy assessment in the global rural regions. Specifically, we divided the GISA-10m into urban and rural regions using the urban boundary defined by Li et al. (2020). In fact, due to the random sampling strategy, most of the visually interpreted test samples were located in rural regions.

In the case of the visually interpreted samples, GISA-10m exhibits a better OA of 86.19 % than GHSL 2018 (84.92 %), GLCFCS (83.25 %), FROM_GLC10 (78.83 %), and WSF2015 (83.81 %). As regards the three 30 m datasets (i.e., GISA, GAIA, GAUD), their ISA accuracy (F1-score) decreases significantly in the rural regions, while the NISA accuracy is relatively stable (Tables 2–3). Taking a closer look at the PA, it is apparent that the ISA PA decreases by more than 15 % for all three 30 m datasets (Table S2), which suggests that there are more omission errors in the rural regions (Fig. 11b). This demonstrates the deficiency of the 30 m datasets in depicting rural ISA, and also reflects the importance of 10 m global ISA mapping.

Furthermore, we also focused on the accuracy assessment in arid regions. In general, the OA of GISA-10m is higher than that of the existing datasets (Table 5). Although its ISA UA does not always outperform the other datasets, GISA-10m achieves the highest PA (Table S4). Specifically, GISA-10m exhibits a notably higher ISA PA than GLCFCS, FROM_GLC10, GISA, GAUD, and GAIA (Table S3), indicating its superior ability to detect ISA in arid regions (Fig. 7). Moreover, the accuracy of these global ISA products was assessed using the manually and randomly chosen rural building samples (see Sect. 3.2). It can be found that GISA-10m detects 15 % more buildings in arid regions than FROM-GLC10, GAUD, and GAIA (Table S5), which further verifies its superior performance in describing rural ISA.

In the case of urban regions, GISA-10m exhibits a satisfactory result, with an OA similar to that of the global assessment (Table S4). Note that urban ISA only accounts for one-third of global ISA, while nearly 70 % of ISA is located in suburban and rural regions. The existing datasets show relatively more ISA omissions in rural and arid regions, suggesting that global ISA mapping at a 10 m resolution (e.g., GISA-10m) is necessary. Moreover, we divided the visually interpreted samples located in cities into three levels (i.e., small, medium, and large cities) to assess the accuracy of GISA-10m for cities of different scales, i.e., Level 1 (population < 250 000), Level 2 (250 000 to 1 000 000), and Level 3 (> 1 000 000) (Yang et al., 2019). The OA of GISA-10m across the three levels of cities is 85.35 %, 87.43 %, and 85.42 %, respectively (Table S6). These results indicate that the performance of GISA-10m in different scales of cities is stable, and the results are also close to the global assessment result (OA of 86.06 %).

Based on the GISA-10m, we analyzed the global ISA distribution at a 10 m scale (Fig. S1). Global ISA is mainly distributed in Asia (41.43 %), North America (20.59 %), and Europe (18.93 %), followed by Africa (9.78 %) and South America (7.50 %). It is found that 67 % of global ISA is located in the Eastern Hemisphere, while 85 % of ISA is distributed to the north of the Equator. Rural ISA is more scattered than urban ISA (Fig. S1), and is mainly located in Asia (42.84 %), Europe (19.49 %), and North America (16.51 %). Asia has the largest urban ISA, which is more than twice that of Europe. Although North America only accounts for 20 % of global ISA, its urban ISA takes up more than 29 % of the global total. Taking a closer look at the ratio of rural and urban ISA (Table 6), it can be seen that rural ISA is 2.2 times larger than urban ISA. At the continental level, Africa possesses the highest “rural-to-urban ratio”, which is likely related to its large population but relatively poor economy.

At the country scale, China and the US account for 33 % of global ISA. Together with Russia, Brazil, India, Japan, Indonesia, France, Canada, and Germany, these 10 countries account for 58 % of the world ISA. The urban ISA of the top 10 countries (US, China, Russia, Brazil, Japan, India, Mexico, France, Germany, and the United Kingdom) makes up 69 % of the global total, while the top 10 countries in terms of rural ISA (China, US, Russia, Brazil, India, Indonesia, Japan, France, Canada, and Germany) account for only 54 % of the total (Fig. S3). In Africa, the Republic of South Africa has much more urban ISA than the other countries. However, Nigeria has a comparable rural ISA to South Africa (∼ 7738 km2). China ranks first in terms of rural ISA, most of which is located in the North China Plain (Fig. S1b). Indonesia also possesses a lot of rural ISA, since it ranks sixth for rural ISA but only 16th for urban ISA.

Roads are major anthropic footprints, so we attempted to analyze the global road area based on GISA-10m, courtesy of its high spatial resolution. Firstly, the road networks were extracted from the OSM database, and then the ISA regions in the GISA 10 m data within a 10 m buffer of the road networks were identified as road areas (Fig. 8). The results show that 82.84 % of the global road area is located in Asia (30.74 %), North America (27.17 %), and Europe (24.92 %), while the remaining 17.16 % is found in South America (8.26 %), Africa (7.47 %), and Oceania (1.44 %). Although Asia far exceeds the other continents with regard to ISA and rural road area, it possesses a smaller urban road area than North America. China and the US have the largest road area, together accounting for 29 % of the global total, followed by Brazil, Japan, Russia, Germany, India, France, Indonesia, and Mexico. The top 10 countries have more than half of the global road area. The global road area accounts for 14.18 % of the global ISA, and the rural road area is 1.5 times larger than the urban road area (Table 7). However, it should be noted that these estimates might be biased owing to the incompleteness of the OSM data. In addition, narrow roads might be partly detected or missed due to the limitation of the spatial resolution.