Impervious surfaces are characterized by sparse spatial distribution and complicated spectral and spatial heterogeneities; thus, impervious surface mapping should be treated independently. Its training samples are generated from five previous products: (1) the Global 30 m Impervious Surface Dynamic (GISD30) dataset, developed with the combination of spectral generalization and sample migration during 1985–2020 with the interval of 5 years, achieves a fulfilling accuracy of 90.1 % (Zhang et al., 2022). (2) The Global Impervious Surface Area (GISA 2.0) dataset, produced by considering the inconsistency among four existing products, is an annual time-series of impervious surface maps during 1985–2018 with an F1 score of 0.935 (Huang et al., 2022). (3) The 10 m impervious surface layer in ESA WorldCover dataset was generated by the supervision classification from Sentinel-1 and Sentinel-2 time-series imagery (Zanaga et al., 2021). It was demonstrated to achieve the great performance on impervious surfaces with a producer's accuracy of 82.99 % (Zhao et al., 2023). (4) The impervious surface layer in ESRI Land Cover is developed from Sentinel-2 time-series imagery and deep learning (Karra et al., 2021), and it achieves a high producer's accuracy of 88.42 % (Zhao et al., 2023). (5) The global urban boundary (GUB) dataset is generated by a combination of a cellular automata algorithm and a morphological approach and shows a good agreement with the results from human interpretation (Li et al., 2020).

Because almost all global land-cover products have large uncertainties in wetland identification (Zhang et al., 2023b), wetlands should also be treated as an independent land-cover type. The wetland training samples were also derived from three existing wetland thematic products: (1) the GWL_FCS30D is an annual global wetland product containing eight wetland subcategories (five inland and three coastal wetland subcategories) and achieves an overall accuracy of 86.95 ± 0.44 % (Zhang et al., 2024b). (2) The NWI (National Wetland Inventory) is a national wetland thematic product covering the whole United States and containing eight wetland subcategories (Wilen and Bates, 1995). (3) The GMW (Global Mangrove Watch) and GSM10 (global tidal marsh dataset) provide the spatial patterns of mangroves and salt marshes with an overall accuracy of 87.4 % and 85 % (Bunting et al., 2022; Worthington et al., 2024), and the GTF30 (Global tidal flat products), generated by the combination of a new low-tide spectral index and multisource classification method, achieved the overall accuracy of 90.34 % and covered the period 2000–2022 (Zhang et al., 2023a).

Except for wetlands and impervious surfaces, training samples for the remaining non-wetland natural land-cover types are generated from the GLC_FCS30D dataset. It was developed from the combination of a continuous change detection algorithm with an adaptive updating strategy and had 80.88 % (±0.27 %) accuracy covering the period of 1985–2022 with 35 fine land-cover subcategories. The dataset has high temporal stability in the European Union and the United States (Zhang et al., 2024c). In this work, we leverage this dynamic product to generate confident training samples for non-wetland natural land covers, as described in the Sect. 3.1.

It is noteworthy that the tree-cover cropland was only mapped in certain regions rather than globally in the GLC_FCS30D (Zhang et al., 2021; Zhang et al., 2024c); thus, global oil palm and plantation datasets are also used to identify tree-cover cropland. The global oil palm dataset is a time series covering 1990–2020 and exceeds 91 % accuracy for industrial plantations and 71 % accuracy for smallholders (Descals et al., 2024). A global plantation map was generated by combining some prior global plantation products and the LandTrendr method and has an F1 score of 86.83 % with a tolerance of ±5 years (Du et al., 2022).

Regarding the training samples of impervious surfaces, we combine four prior global 10 m or 30 m impervious surface products (GISA 2.0, GISD30, Imp-ESA_LC, and Imp-ESRI_LC) and one urban boundary dataset (GUB) to automatically generate the training samples. Specifically, because the previous studies have demonstrated high accuracies of three impervious surface products (Huang et al., 2022; Zanaga et al., 2021; Zhang et al., 2022) and a high producer's accuracy of Imp-ESRI_LC (Zhao et al., 2023), the areas marked as impervious surfaces by all four products (GISD30-2020, GISA2.0-2019, Imp-ESA_LC-2021, and Imp-ESRI_LC-2023) are selected as candidate areas for generating the training samples (TrainCanArea_imp) in Eq. (1). 1TrainCanAreaimp=GISD30∩GISA2.0∩Imp-ESALC∩Imp-ESRI_LC

Afterward, we further consider the uneven distribution of rural and urban impervious surfaces as well as their spectral variability. If random sampling is used to obtain training samples from the TrainCanArea_imp, rural impervious surfaces are underrepresented due to their sparse distribution; thus, the GUB urban boundary dataset for 2020 is further used to divide the TrainCanArea_imp into urban (TrainCanArea_urban) and rural areas (TrainCanArea_rural).

Beyond the confident impervious surface areas, it is equally important to identify high-quality natural land-cover types (Zhang et al., 2024a). To avoid confusion between natural land-cover types and impervious surfaces, the maximum impervious surface boundary (MaxBound_imp) is also generated. The training samples for natural land-cover types should be located outside of the MaxBound_imp; i.e., some inner-city areas, easily misclassified or confused with impervious surfaces, will be excluded because Imp-ESRI_LC exhibits extensive patches of the impervious surface and lacks spatial details (Wang et al., 2024; Xu et al., 2024b). To determine MaxBound_imp, the union of the four products is applied as 2MaxBoundimp=GISD30∪GISA2.0∪Imp-ESALC∪Imp-ESRI_LC.

In this work, wetlands are divided into four inland subcategories (swamp, marsh, river/lake flats, and saline) and three coastal wetland subcategories (mangrove, salt marsh, and tidal flats in Table 2). Because coastal wetlands have a more pronounced zonation, the global coastal wetland mapping has made great progress, while the work of global inland wetland mapping is still sparse (Wang et al., 2023). Thus, the generation of wetland training candidate areas further distinguishes between inland and coastal wetlands.

The time-series GWL_FCS30D wetland product covering the period of 2000–2022 is used to derive the inland wetland training candidate areas (Zhang et al., 2024b). Because temporally stable areas achieve higher accuracy (Yang and Huang, 2021), a temporally stable analysis is applied to the GWL_FCS30D, and only those stable areas where wetland subcategories do not change during 2000–2022 are retained (yielding the training area, TrainCanArea_Inwet). Then, because adjacent land-cover areas are more likely to suffer from the higher misclassifications (Radoux et al., 2014) and from the impact of the satellite geolocation error (Zhang and Roy, 2017), a spatial filter with a local window of 3 pixels × 3 pixels is applied to TrainCanArea_Inwet to retain spatially homogeneous areas as the training areas. Further, the integration of multiple wetland products can further improve sample quality, but there are few high-quality wetland products that have been publicly shared. In this study, only the National Wetland Inventory for 2019 is imported to optimize the swamp and marsh land-cover types in TrainCanArea_Inwet over the United States because the National Wetland Inventory does not identify river/lake flats or saline subcategories. Namely, the areas identified as swamp and marsh in TrainCanArea_Inwet and the National Wetland Inventory are retained.

As for the three coastal wetland subcategories, their training areas are generated from the combination of GWL_FCS30D and three coastal wetland products (GMW, GTF30, and GSM10 in Table 1). We identify temporally stable coastal wetland areas from GWL_FCS30D, GMW, and GTF30 through time-series analysis and label them GWL_stable, GMW_stable, and GTF_stable. Then, we intersect GWL_stable with GMW_stable to generate mangrove forest training areas and intersect GWL_stable with GTF_stable to generate tidal flat training areas. Next, as the GSM10 only provides salt marsh maps in 2020, the salt marsh training areas are selected as the intersection between GWL_stable and GSM10. The mangrove forest, tidal flat, and salt marsh training areas are grouped as TrainCanArea_Cowet.

Last, the maximum wetland boundary (MaxBound_wet) is also necessary for the subsequent identification of training areas for non-wetland natural land-cover types. MaxBound_wet is determined as the union of the five global wetland products: 3MaxBoundwet=GWL_FCS30D∪NWI∪GMW∪GTF30∪GSM10.

Many previous studies have emphasized that these spatiotemporally stable areas always performed at a higher mapping accuracy (Zhang and Roy, 2017; Zhang et al., 2024c). In this work, the time-series global 30 m land-cover dynamic product (GLC_FCS30D), covering the period 1985–2022, is used. Specifically, three measures are taken to identify spatiotemporally stable areas of non-wetland natural land-cover types from GLC_FCS30D: (1) a time-series consistency analysis is applied to the GLC_FCS30D, and only stable areas during 1985–2022 will be retained as TrainCanArea_NLCs. (2) The MaxBound_imp and MaxBound_wet are imported to mask the TrainCanArea_NLCs; i.e., the training areas for non-wetland natural land-cover types should be located outside of MaxBound_imp and MaxBound_wet. The aim of this step is to minimize confusion between non-wetland natural land-cover types and these two land-cover types. (3) A morphological erosion filter with a local window of 3 pixels × 3 pixels is used to find the spatially homogeneous areas for non-wetland natural land-cover types. In addition, it should be noted that the TrainCanArea_NLCs represents the stable areas during 1985–2022, and there is still a 1-year interval with the land-cover mapping year in 2023. Fortunately, the ongoing updating of GLC_FCS30D is still in progress. The land-cover change masks during 2022–2023 have been finished, which are used to guarantee the temporal consistency between prior land-cover products with the training areas.

As mentioned in Sect. 2.2.4, the training areas for tree-cover cropland (oil palm, orchards, etc.) from the GLC_FCS30D do not cover the globe. These cropland training areas are therefore divided into herbaceous rainfed cropland and tree- or shrub-cover cropland. Because the global oil palm and global plantation datasets provide the plantation years of oil palm and orchards at 30 m, we overlap the training areas of rainfed cropland, oil palm, and orchard plantation from the global plantation dataset to extract the training areas for tree-cover cropland. Then, to minimize error, the tree-cover cropland training areas are further filtered using a local window of 3 pixels × 3 pixels to ensure spatial homogeneity of tree-cover cropland training areas.

Although we take a series of measures to ensure training area quality (including TrainCanArea_urban, TrainCanArea_rural, TrainCanArea_Inwet, TrainCanArea_Cowet, and TrainCanArea_NLCs), how to generate training samples from the training areas needs to address the following two aspects.

First, the distribution and balance of training samples greatly affect the subsequent land-cover classification (Ghorbanian et al., 2020; Jin et al., 2014; Mellor et al., 2015; Pelletier et al., 2017; Zhu et al., 2016). There are two options to allocate the sample distribution: equal or area-fraction allocation (Zhang et al., 2021). Equal distribution means that all land-cover types have the same number of training samples; i.e., the sample sizes of sparse land covers will be augmented, while those of the abundant land covers will be suppressed. In contrast, area-fraction distribution allocates the sample size according to the land-cover area of each type; that is, abundant land covers have larger sample sizes, while sparse land-cover types have smaller sample sizes (Zhu et al., 2016). Because impervious surfaces and wetlands are sparser than natural land-cover types and are independently generated, equal-distribution allocation is suitable to enhance the training samples' ability to characterize these two land-cover types. As for the non-wetland natural land covers, the area-fraction allocation is more appropriate for the non-wetland natural land-cover types because we want to optimize results for all non-wetland natural land-cover types rather than a single land-cover type. Meanwhile, to avoid sample size imbalance in the area-fraction allocation, maximum and minimum sample sizes of 8000 and 600 pixels are chosen for the abundant and sparse land-cover types, respectively (suggested by the work of Zhu et al., 2016).

Second, most high-quality training samples (except for those for impervious surfaces) are derived from the 30 m training areas, so there is also a need to reduce the 30 m training samples to 10 m samples to achieve a global 10 m land-cover map. In this work, the “metric centroid” method is adopted, which had been used to downscale 500 m training samples from MCD12Q1 to 30 m in the work of Zhang and Roy (2017). Specifically, as each 30 m pixel corresponds to 3 × 3 10 m pixels, we first find the centroid from these nine pixels as Pcentroid through spectral averaging, and then the point with the smallest absolute distance with Pcentroid was chosen as the optimal downscaled 10 m sample point (Eq. 4). 4Pi=ρPi-ρPcentroidρPcentroid=19∑j=19ρPj9, where ρPi is the spectrum value of composited Sentinel-2 training features (see Sect. 3.3) at pixel Pi. If more than one point in the nine pixels has the same minimum absolute distance, then we pick randomly from among them.

To separate impervious surfaces and natural surfaces, we rely on the globally distributed training samples (Sect. 3.2.4) and the combination of multitemporal optical and SAR features. Specifically, this is because we divide impervious surfaces training samples into rural and urban samples and design the equal distribution to enhance the training samples' ability to characterize impervious surfaces. The ratio of urban samples, rural samples, and natural surfaces is 1:1:1 for each 5 × 5 geographical tile. Meanwhile, in terms of the sample size of each class, some previous studies have quantified the relationship between sample size and mapping accuracy (Foody, 2009; Li et al., 2014), and suggested a minimum size of 600 and maximum size of 8000 for these sparse and abundant land-cover types (Zhu et al., 2016). In this study, after considering the trade-off between sample representativeness with mapping efficiency, the sample size of each class was selected as 5000, which was also consistent with the work of Zhang et al. (2022) in monitoring the impervious surface dynamics.

Then, we split the globe into 984 5 × 5 geographical tiles (approximately 556 km × 556 km on the Equator, illustrated in Fig. S1 in the Supplement), because some studies emphasized that the local adaptive modeling usually achieves better mapping accuracy than single land-cover global modeling (Zhang et al., 2021), and previous studies of Zhang and Roy (2017) and Zhang et al. (2019) have explained that the training samples of sparse land-cover types in a small geographical grid were usually missed or greatly sparse. Thus, after balancing the training sample volume, mapping accuracy, and the limitation of GEE platform, the local modeling tile size of 5 × 5, similar to the studies of Zhang et al. (2021, 2024c), was used. When building the training model for each 5 × 5 geographical tile, we also import training samples within their spatial neighborhood of 3×3 tiles to ensure spatial consistency over the adjacent tiles. Since the MaxBound_imp (Eq. 2) provides the maximum potential areas of impervious surfaces because of the overestimation problem of Imp-ESRI_LC (Wang et al., 2024; Xu et al., 2024b), all identified impervious surfaces should be within the MaxBound_imp. Afterward, we can produce 984 instances of 5 × 5 impervious surface and natural land-cover maps using the local adaptive modeling strategy. In addition, although we divide the training samples into urban and rural samples, there is serious confusion between urban and rural areas in the classification maps because they share similar spectral and SAR characteristics. We therefore consider the two subcategories to be inseparable at the classification stage. Correspondingly, inspired by the work of Li et al. (2020), who used the cellular automata and morphological approaches to accurately capture urban boundaries, this method is also applied in this work to distinguish urban and rural impervious surfaces.

Notably, in terms of the selection of classifier at the local adaptive modeling, the random forest classification model (including the later Sect. 3.4.2 and 3.4.3) is used. The random forest has some advantages over other traditional classifiers, such as managing high-dimensional data more efficiently, having higher mapping robustness, being insensitive to parameter settings, and avoiding overfitting problems effectively (Belgiu and Drăguţ, 2016; Breiman, 2001). In terms of its parameter settings, the random forest only has two adjustable parameters, and the variations of these two parameters have little effect on the performance of the random forest model (Du et al., 2015; Gislason et al., 2006); thus, the default parameter settings are applied to train the random forest models on the GEE platform.

In terms of how to identify the fine wetland subcategories from natural land covers, we use the stratified wetland mapping algorithm to independently distinguish coastal wetlands and inland wetlands. Wetlands are divided into four inland and three coastal wetland subcategories (in Table 2), and equal-distribution sampling is used to enhance the training samples' ability to characterize wetlands. Additionally, some non-wetland land-cover types also reflected similar spectral characteristics with the wetlands; for example, the swamp and the forest/shrubland shared similar vegetation spectra during the peak growth period, while the marsh and cropland/grassland exhibited characteristics of herbaceous vegetation, and the river flats also performed the spectral characteristics of bare land during the dry seasons (Zhang et al., 2023b). Thus, the approximate ratio of inland wetlands, coastal wetlands, and non-wetlands (including water body, forest/shrubland, cropland/grassland, bare land, and others) is 4:3:5 in areas where they coexist. Then, because coastal wetlands have a more pronounced zonation, we can obtain their maximum coverage through the union of some previous coastal wetland products, as in Eq. (6). 6MaxBoundCos_wet=GWL_FCS30D_Coastal∪GMW∪GTF30∪GSM10

When building the wetland random forest classification models for each 5×5 geographical tile, we first train the coastal wetland classification model using the coastal wetland and non-wetland training samples within their spatial neighborhood of 3 × 3 tiles and combine multisourced training features to identify the spatial distribution of coastal wetlands within the MaxBound_Cwet; i.e., all coastal wetland pixels should be within the MaxBound_Cwet. Otherwise, they would be corrected.

Afterwards, the inland wetland and non-wetland training samples are used to train another random forest classification model, and the remaining areas are further classified as four inland wetland subcategories and non-wetland natural land-cover types using the inland wetland classification model.

After classifying the impervious surface and wetlands using the hierarchical land-cover mapping, we now need to classify the remaining non-wetland natural land-cover types. Like the previous mapping processes, the local adaptive random forest models are trained for each 5 × 5 geographical tile using the corresponding training samples within the spatial neighborhood of 3 × 3 tiles. The non-wetland natural land-cover types are classified through a combination of trained random forest models and multisource training features. Lastly, after overlapping the hierarchical maps for impervious surface, wetland, and non-wetland natural land-cover types, we can obtain 10 m land-cover maps with a fine classification system.

Table 3 presents the confusion matrix between GLC_FCS10 and the 56 121 globally distributed validation points for 10 major land-cover types (corresponding to the basic classification system in Table 2). Overall, GLC_FCS10 achieves an O.A. of 83.16 % and a κ coefficient of 0.789. For specific land-cover types, permanent snow and ice, water bodies, forest, impervious surfaces, and cropland perform the best, with the corresponding U.A. and P.A. values being approximately or exceeding 90 %. Their high accuracies stem from the distinct spectral properties inherent to water bodies and permanent ice and snow, the abundant coverage of cropland and forest, and hierarchical land-cover mapping for impervious surfaces. However, shrubland, grassland, tundra, and wetlands suffer obvious misclassifications, in which the shrubland has the lowest U.A. of 67.04 % and wetland has the lowest P.A. of 53.69 %. There is considerable confusion between shrubland, grassland, and bare areas because they share similar spectral characteristics and coexist in arid and semiarid areas. Wetlands have the lowest P.A. due to the confusion between wetlands, water bodies, forest, and grassland. Wetlands have complicated and heterogeneous spectral and temporal variations; thus, the swamp subcategory is easily confused with forest, and the marsh subcategory shares spectral characteristics with grasslands (Zhang et al., 2023b).

To intuitively understand the spatial distribution of the GLC_FCS10 accuracy metrics, Fig. 5 presents the spatial variations of O.A. and the κ coefficient among 30 climate zones from Köppen climate classification system. There is high consistency in the O.A. and κ coefficient within the spatial patterns; i.e., some climatic transition zones, land-cover heterogeneity zones, cloud-contaminated tropical zones, and small subdivisions tend to have lower accuracy (below 80 %). Conversely, some homogeneous zones (such as Greenland and the Sahara) and forest- or cropland-rich zones (such as the east-central United States, central Eurasia, and East Asia) achieve a high O.A. and a κ coefficient.

Table 4 further analyzes the confusion matrix between GLC_FCS10 and the global validation dataset with 16 land-cover types (refining the forest and cropland subcategories). Under this fine validation system, the GLC_FCS10 achieves an O.A. of 76.45 % and a κ coefficient of 0.736, which are reduced by 6.71 % and 0.053, respectively, from the metrics in Table 3. This reduction is due to confusion between the finer land-cover subcategories; e.g., forests have a U.A. value of 91.02 % in Table 3, but when broken down into five forest subcategories the average U.A. value drops to 68.52 %. Taking mixed forest as an example, it has a low accuracy of only 44.17 %, of which the proportions misclassified as evergreen broadleaved forest (EBF), deciduous broadleaved forest (DBF), evergreen needleleaved forest (ENF), and deciduous needleleaved forest (DNF) are 5.34 %, 7.77 %, 27.18 %, and 5.83 %, respectively. Higher likelihoods of confusion exist for closely related land-cover subcategories; e.g., the highest proportion of misclassification in EBF is in DBF, and rainfed cropland is easily misclassified as irrigated cropland, shrubland, or grassland. Previous research also demonstrated that considerable misclassifications occur between similar land-cover types (Homer et al., 2020; Wickham et al., 2021; Zhang et al., 2021; Zhang et al., 2024c).

Table 5 presents the confusion matrix for GLC_FCS10 based on the LCMAP_Val validation points over the United States. It should be noted that the LCMAP_Val dataset only contains eight land-cover types and merges shrubland and grassland into one mosaicked land-cover type (grass/shrub). The GLC_FCS10 achieves an O.A. of 85.09 % and a κ coefficient of 0.804 using these 16 082 national validation points. Regarding the U.A. and P.A., the cropland, forest, water, and grass/shrub land-cover types achieve balanced U.A. and P.A. values approximately or exceeding 80 %. In contrast, developed land has the lowest U.A. of 54.26 % with a high P.A. of 98.85 %, mainly because of the difference in definitions of developed land and impervious surfaces. The LCMAP_Val definition of developed land is broad enough to classify inner-city greenery as developed land as well (Xian et al., 2022), which is considered a vegetation land-cover type in the GLC_FCS10. Barren land has the lowest P.A. value of 31.93 %, indicating a high omission error of 68.07 %. Most of these misclassifications came from the confusion between barren land and grass/shrub land-cover types. It is noteworthy that the grass/shrub shares similar spectral characteristics with barren, and both of them co-exist in arid regions of the western United States. Thus, it is usually difficult to distinguish between the two with high accuracy.

Figure 6 illustrates the spatial distribution of O.A. and κ coefficient values among different Köppen climate zones using the LCMAP_Val validation points over the United States. There is notable consistency between O.A. and κ coefficient in terms of the spatial patterns; i.e., the GLC_FCS10 considerably outperforms in the Western United States and in the Eastern United States and has an optimal κ coefficient of more than 0.8 in the Northeastern United States. Combined with the climatic distribution, it performs relatively poorly in the arid and semi-arid zones of the Midwestern United States, mainly attributed to the difficulty in distinguishing between shrubs, grasses, and bare land within the region.

A principal difficulty in land-cover mapping is obtaining high-quality training samples (Li et al., 2023; Zhang et al., 2021). In this work, we integrate prior multisource global land-cover products to generate globally distributed training samples. To ensure the confidence of these derived training samples and minimize the classification errors of each prior product, we took the following actions: (1) spatiotemporal consistency checking was used to find homogeneous and stable areas. (2) The intersection of multiple land-cover products minimized the influence of classification errors in each product. (3) A morphological erosion filter was applied to reduce the impact of edge-mixing effects. The accuracy assessment partly demonstrates the reliability of these derived training samples; i.e., GLC_FCS10 achieves satisfactory accuracy metrics and outperforms several other land-cover products. Due to the large volume of these globally distributed training samples, we selected approximately 10 000 derived samples from the training sample pool in Sect. 3.2.4. Upon meticulous inspection, we determined that these chosen samples attained an overall accuracy (O.A.) of 92.18 %, with certain uncertainties existing for shrubland and grassland. This result was in accordance with the earlier analysis presented in Table 3.

Moreover, it is still uncertain whether this small number of erroneous training samples could impact the performance of land-cover mapping. Figure 10 illustrates the quantitative relationship between the erroneous training samples and the O.A. and κ coefficients for the basic classification system. Initially, O.A. and the κ coefficient remain stable as the number of erroneous training samples increases. However, a significant decline occurs when the proportion of erroneous samples exceeds 30 %. This indicates that the trained random forest model is robust to the erroneous training samples as long as their proportion remains below 30 %. In this work, if the fraction of erroneous samples was kept below 30 %, the difference in O.A. is approximately 2 %, and the decrease in the κ coefficient is approximately 3 %. Gong et al. (2024) also demonstrated that a small number of incorrect samples (approximately 20 %) did not affect the land-cover classification accuracy.

One of the novelties of this study is the adoption of the hierarchical land-cover mapping strategy. The accuracy assessment in Table 3 indicates that impervious surfaces have a high U.A. value of 92.70 % and P.A. value of 93.76 %. The wetland U.A. and P.A. values are 70.91 % and 53.69 %, which were superior to those of the other land-cover products in Table 6 and the cross-comparisons in Figs. 7–9. To intuitively understand the advantage of the hierarchical land-cover mapping strategy, a comparative experiment (ComExp) has been designed using the training samples from area–fraction allocation (explained in the Sect. 3.2.4) in Fig. 11; i.e., the impervious surfaces and wetlands are not classified separately in the middle reaches of the Yangtze River (the comparative site in Fig. 7). Overall, the ComExp is consistent with the GLC_FCS10 spatial patterns and shows some variations in the details. Specifically, in Fig. 10R1, the ComExp misclassifies some marsh wetlands as herbaceous rainfed cropland; i.e., some lake flats (red rectangles in Fig. 11R1) cannot be comprehensively captured when compared with the GLC_FCS10. Figure 10R2 gives the comparisons on the impervious surface areas. We can find that the ComExp has lower impervious surface areas because it misclassifies some bright impervious surfaces as bare areas and some residential areas as vegetated land. In summary, the hierarchical land-cover mapping strategy can increase the ability to characterize specific land-cover types. Similarly, the work of Sulla-Menashe et al. (2019) also used hierarchical mapping to generate annual global land-cover types for MCD12Q1 and demonstrated its better performance.