Finer-Resolution Mapping of Global Land Cover: Recent Developments, Consistency Analysis, and Prospects

Land-cover mapping is one of the foundations of Earth science. As a result of the combined efforts of many scientists, numerous global land-cover (GLC) products with a resolution of 30m have so far been generated. However, the increasing number of fineresolution GLC datasets is imposing additional workloads as it is necessary to confirm the quality of these datasets and check their suitability for user applications. To provide guidelines for users, in this study, the recent developments in currently available 30m GLC products (including three GLC products and thematic products for four different land-cover types, i.e., impervious surface, forest, cropland, and inland water) were first reviewed. Despite the great efforts toward improving mapping accuracy that there have been in recent decades, the current 30m GLC products still suffer from having relatively low accuracies of between 46.0% and 88.9% for GlobeLand30-2010, 57.71% and 80.36% for FROM_GLC-2015, and 65.59% and 84.33% for GLC_FCS30-2015. The reported accuracies for the global 30m thematic maps vary from 67.86% to 95.1% for the eight impervious surface products that were reviewed, 56.72% to 97.36% for the seven forest products, 32.73% to 98.3% for the six cropland products, and 15.67% to 99.7% for the six inland water products. The consistency between the current GLC products was then examined. The GLC maps showed a good overall agreement in terms of spatial patterns but a limited agreement for some vegetation classes (such as shrub, tree, and grassland) in specific areas such as transition zones. Finally, the prospects for fine-resolution GLC mapping were also considered. With the rapid development of cloud computing platforms and big data, the Google Earth Engine (GEE) greatly facilitates the production of global fine-resolution land-cover maps by integrating multisource remote sensing datasets with advanced image processing and classification algorithms and powerful computing capability. The synergy between the spectral, spatial, and temporal features derived from multisource satellite datasets and stored in cloud computing platforms will definitely improve the classification accuracy and spatiotemporal resolution of fineresolution GLC products. In general, up to now, most land-cover maps have not been able to achieve the maximum (per class or overall) error of 5%–15% required by many applications. Therefore, more efforts are needed toward improving the accuracy of these GLC products, especially for classes for which the accuracy has so far been low (such as shrub, wetland, tundra, and grassland) and in terms of the overall quality of the maps.


Introduction
Land cover is an essential climate variable and is widely used in numerous studies concerned with, for example, global climate change, Earth system modeling, natural resource management, food security, and conservation biology [1][2][3]. In recent decades, many global land-cover (GLC) products have been successfully generated using satellite remote sensing data, and GLC mapping is moving toward providing datasets with higher spatial resolution and finer classification [3][4][5]. The GLC products that have a coarse spatial resolution, such as the global land-cover classification product at 1 km (GLC_ 2000) [6] and IGBP (International Geosphere-Biosphere Programme) DISCover land-cover classification product (IGBP_DISCover) [7], the 500 m Moderate Resolution Imaging Spectroradiometer (MODIS) land-cover product (MOD12Q1) [8], and the 300 m global land-cover map (GlobCover) [9] and the European Space Agency (ESA) Climate Change Initiative land-cover product (CCI_LC) [10], have been used in many studies. However, these coarseresolution GLC products do not provide enough spatial detail for studies on, for example, urban expansion and resource management [11,12]. As pointed out by Herold et al. [13], more efforts based on the use of higherresolution satellite data should be made to improve the classification accuracy of heterogeneous landscapes, which is the most significant challenge for improving GLC mapping. Furthermore, fine-resolution (∼30 m) GLC products provide rich spatial information at the scale of most human activity, thus offering greater flexibility for application in a range of studies [11,12,14].
These fine-resolution GLC datasets have been generated as a result of various initiatives, so they differ in terms of classification systems, satellite data used, and classification methodologies [33,34]. For example, there are huge discrepancies between CGLS-LC100, ESA-S2, and FROM_GLC, which have an overall spatial agreement of 56.87% and a disagreement of 43.13% for Africa [35]. Consequently, it is difficult to assess which GLC dataset fits user application best [36,37]. Although many campaigns have been conducted with the aim of collecting reference datasets for validating fine-resolution GLC maps, their availability and role in applications outside their area of intended use have been very limited because of the inconsistencies between legends, sampling strategies, and response design protocols used [33,38]. According to reports relating to global validation campaigns, there are large differences between the overall accuracies of fine-resolution GLC products: 64.9% for FROM_GLC [12], 80.3% for GlobeLand30 [14], and 82.5% for GLC_FCS30 [16]. In spite of the validation assessments provided by the developers of these products, some third-party researchers have found considerably lower accuracies for some regions when verifying GLC products [39,40].
The need to assess the quality of the increasing number of fine-resolution GLC datasets and check their suitability for different user applications is leading to additional workloads. The objectives of this study were (1) to review recent developments related to current fineresolution GLC products, (2) to compare the available 30 m GLC maps and quantify their spatial consistency, and (3) to summarize the trends in and prospects for fine-resolution GLC mapping.

Consistency and Accuracy Assessment Metrics
The area-based and pixel-based consistencies are widely used to assess the agreement between different global products. The area-based analysis mainly focuses on the consistency of the areal proportions of each LC type between the different GLC products [39,41], whereas the pixel-based analysis is concerned with the spatial (dis)agreement for each LC type between the different GLC products [34,41,42]. In this study, both an analysis of the consistency of the current 30 m GLC products and an analysis for different classes (impervious surface, forest, cropland, and water body) of current global 30 m thematic products were performed to give a more comprehensive understanding of the process of global land-cover mapping.

Analysis of the Consistency of Current 30 m GLC
Products. First, the area percentage of each land-cover type in each GLC product was calculated to give an overview of the consistency between the current GLC products. Then, in order to intuitively illustrate the spatial consistency between the different products, the spatial superposition method [41] was employed to obtain the pixel-by-pixel spatial correspondence for the different products. A maximum value (equal to 3) was assigned for highly consistent pixels when the target land-cover types for the three LC products were exactly the same; a value of 2 was assigned for moderately consistent pixels when only two land-cover types agreed; a minimum value (equal to 1) was assigned for lowconsistency pixels when the three LC products all had different target land-cover types [39]. To obtain more detailed information about the spatial consistency between the different products, the spatial superposition method was also applied to each pair of products. Pixels for which the target land-cover type was the same for both products were considered "consistent," whereas pixels with different target landcover types were labeled "inconsistent." The spatial consistency distribution for all three products and for each pair of products is given based on a spatial resolution of 0.05°. In the spatially aggregating step, the consistency metric was firstly calculated for each 30 m pixel; then, the proportion of highly consistent (PHC) pixels, proportion of moderately consistent (PMC) pixels, proportion of low-consistency (PLC) pixels within each 0.05°grid were calculated and stored in three channels. Therefore, the spatial consistency was assessed at a resolution of 30 m, while illustrated at a resolution of 0.05°.

Analysis of the Consistency of Current Global 30 m
Thematic Products. The analysis of the consistency of the current global 30 m thematic products focused on the impervious surface, forest, cropland, and water body classes. First, the fractional areas for each thematic class were calculated at a spatial resolution of 0.05°. The total meridional and zonal fractional areas were then calculated for each class for each 0.05°longitude and latitude bin. To quantitatively evaluate the consistency between each pair of thematic products, the linear regression approach was applied, and scatter plots for each pair of products were made. Based on the linear regression, the root-mean-squared difference (RMSD) and coefficient of determination (R 2 ) were calculated as follows: where frac xðiÞ is the fractional thematic class area for thematic product x in the i th 0:05°× 0:05°grid cell; frac yðiÞ is the fractional thematic class area for thematic product y in the i th 0:05°× 0:05°grid cell; and n is the total number of grid cells in the thematic product. The RMSD value was used to reflect the discrepancy between the two thematic products in each pair: where frac x is the average fractional thematic class area for thematic product x and frac y is the average fractional thematic class area for thematic product y. The higher the value of R 2 , the better the fit between a pair of thematic products. There are three 30 m global land-cover products that are currently available free of charge: GlobeLand30 [14], FROM_ GLC [12], and GLC_FCS30 [16]. Some basic information about these three products is listed in Table 1. The GlobeLand30 product, which is one of the most widely used land-cover products, was developed by using the pixel-object-knowledge-based (POK) method and multi-temporal Landsat and HJ-1 A/B imagery [14]. This product has been validated as having an overall accuracy of 80.3% based on 154,586 validation samples from across the globe [14]. FROM_GLC was the first global 30 m land-cover product. Currently, the 2015 version (which has a 30 m resolution) and the 2017 version (which has a 10 m resolution) are available. FROM_GLC30-2015, which was produced using multitemporal Landsat imagery combined with over 90,000 visually interpreted training samples and random forest classifiers [12], has been validated as having an overall accuracy of 80.6% [43]. The GLC_FCS30 product, which contains a large variety of land-cover types, was developed on the GEE platform by combining time series of Landsat imagery, a global prior training dataset obtained from GSPE-CLib [44], and local adaptive random forest models [16]. GLC_FCS30-2015 has been validated as having an overall accuracy of 82.5% and a kappa coefficient of 0.784 based on 44,043 global validation samples [16].

Review of the Current Global 30 m Products
There are also some differences between the classification systems employed by the three different GLC products. Glo-beLand30 consists of 10 basic land-cover types, whereas FROM_GLC30 uses a two-level hierarchical classification system with 10 land-cover types at level 1, which is the same classification system used by GlobeLand30. GLC_FCS30 employs the Climate Change Initiative land-cover dataset (CCI_LC) classification system, which has 16 level 1 landcover types. As a uniform classification system is the basis of the comparative analysis of different GLC products, these different classification systems were transformed into a target classification system with 10 basic land-cover types based on the study of Gao et al. [45] before the consistency analysis was carried out. Details of the classification systems used in the three GLC products and the correspondence between them are shown in Table 2. 3.1.2. Review of the Accuracy Assessments of the Three GLC30 Products. There have been numerous validations of the 30 m GLC products at both the regional and global scales. The classification of 10 land-cover types is carried out using a split-and-merge strategy. Robust integration of pixel-and object-based classification is applied to classify each land-cover type; finally, a knowledge-based interactive verification procedure is applied to improve the mapping accuracy. Zhang et al. [16] 3 Journal of Remote Sensing summarizes the reported accuracies. Generally, the accuracies vary according to the validation dataset used. All producers provide accuracy reports for their own products based on different global validation datasets. The reported accuracy of GLC_FCS30 is the highest-an overall accuracy of 82.5% [16]-as against 80.3% for GlobeLand30 [14] and 64.9% for FROM_GLC [12]. Since GlobeLand30 and FROM_GLC were released earlier, numerous validations of these products have been conducted, and the reported accuracies have ranged from 46.0% for Central Asia [46] to 88.9% for the European Union [45] and from 57.71% for Indonesia [39] to 80.36% for Beijing, China [47], for GlobeLand30 and FROM_GLC, respectively. Recently, two sets of reported results for GLC_FCS30-2015 gave an accuracy of 65.59% for Indonesia [39] and 84.33% for the European Union [45].
Sun et al. [46] Central Asia GlobeLand30-2010 data have an overall accuracy of 46% and a kappa coefficient of 0.283. Wang et al. [84] China The overall accuracy of GlobeLand30-2010 for China is 84.2%. Yang et al. [34] China The overall accuracy of GlobeLand30-2010 is 82.39%. See et al. [88] Kenya The GlobeLand30 gave an overall accuracy ranging from 53% to 61%.
Fonte et al. [89] Nepal The overall accuracy is 61% and 54% in Tanzania and Kathmandu, Nepal, respectively.
Brovelli et al. [90] Italy The overall accuracy is higher than 80% according to a validation covering eight regions across Italy.

Balogun et al. [91] Malaysia
Overall accuracies of 63.45% and 65.70%, respectively, were obtained from the error matrix using sample counts and the new unbiased area estimator with GlobeLand30.

3.2.
Review of the Global Impervious Surface Maps 3.2.1. Overview of the Global Impervious Surface Maps. Impervious surfaces are a special land-cover type that usually relates to surfaces that are covered by anthropogenic materials such as roads, driveways, sidewalks, parking lots, and rooftops which prevent or retard the entry of water into the soil [48]. Mapping the global distribution and dynamics of global impervious surfaces using satellite remote sensing data is extremely important to improve our understanding of the intensity of human activity and of global change.
In this study, eight global 30 m impervious surface products were acquired and investigated. Some details of these products are presented in Table 4. Most global impervious surface products cover the long time period stretching from 1985 to 2020 and include multiepoch impervious surface maps; the exceptions to this are MSMT_IS30, HBASE, and GUF. In this study, an analysis of the spatial consistency between these eight impervious surface products in 2015 was conducted. Three products were not available-Globe-Land30 and HBASE/GMIS for 2010 and GUF for 2012. The temporal variations between 2010 and 2015 were, therefore, neglected, which may have led to a certain amount of bias, especially in the case of developing countries. Most of the impervious surface products (MSMT_IS30, NUACI, GAIA, FROM_GLC, GlobeLand30, and GHSL) use similar definitions of "covered by anthropogenic materials which prevent water penetrating into the soil" for impervious surfaces. HBASE defines impervious surfaces as the sum of all surfaces constructed by humans and the area surrounding these surfaces, including urban green spaces; GUF defines impervious surfaces from the perspective of spatial structures-"impervious surfaces should show obvious vertical components"-and so does not define roads, parking lots, airport runways, etc. as impervious surfaces. Lastly, almost all impervious surface products are produced using classification-based methods, including independent random forest classification (MSMT_IS30 and FROM_GLC), a combination of segmentation, classification, and prior knowledge (GlobeLand30 and HBASE), or a combination of unsupervised classification, supervised classification, and prior knowledge (GHSL and GUF). The NUACI is different as it uses a novel built-up index to generate a global multiepoch impervious surface product; GAIA uses normalized scores (the exclusion-inclusion method) and knowledge-driven temporal checking methods to produce annual global impervious surface products.

3.2.2.
Review of the Accuracy Assessments of the Global Impervious Surface Maps. Table 5 lists the accuracy metrics produced by various researchers at different scales for eight global 30 m impervious surface products. Firstly, from the perspective of the global accuracy of each product, the overall accuracy of the impervious surface class in GAIA is 89%, which exceeds the values of 80% in GlobeLand30, 81%-84% in NUACI, and 84.18% in GHSL using the corresponding validation points. At the regional scale, MSMT_IS30 has been validated as having an overall accuracy of 95.1% and a kappa coefficient of 0.898 for 15 test sites of size 1°× 1° [22] and as having the user's and producer's accuracies of 73.61% and 99.07%, respectively, in Indonesia [39] and 73:96% ± 0:08% and 85:46% ± 0:08%, respectively, in the European Union [45]. As for the GlobeLand30 impervious layer, an overall accuracy of 88.4% was found for 15 1°× 1°t est sites [22], along with accuracies of 91.37% in Beijing [49], 80.54% in Malaysia [50], and 67.86% for 450 1°× 1°test sites [51]. GHSL has also been widely used and assessed and has been validated as having an overall accuracy of 90.3% for 15 1°× 1°test sites [22], as well as accuracies of 89.21% in Malaysia [50] and 70.64% for 450 1°× 1°test sites [51]. This product has been shown to have an accuracy of 96.28% and a kappa coefficient of 0.3233 in Europe [52] and a kappa coefficient of 0.562 for 26 counties of the US [53]. Lastly, using 15 1°× 1°test sites, Zhang et al. [22] found that the overall accuracies of NUACI, FROM_GLC, and HBASE were 85.6%, 89.6%, and 88.0%, respectively. Kang et al. [39] and Gao et al. [45] found that the producer's and user's accuracies of FROM_GLC were 81.31% and 52.41%, respectively, in Indonesia and 51:68% ± 0:05% and 69:21% ± 0:09%, respectively, in the European Union. Wang et al. [48] analyzed the performance of HBASE and found that it had an overall accuracy of 97.9% and a kappa coefficient of 0.90.
In summary, although the global 30 m impervious surface products have improved markedly in recent decades, there is still a lack of a comprehensive accuracy assessment for these global products that uses a common validation dataset. Furthermore, the accuracy metrics of these global impervious surface products vary according to the geographical region and the validation datasets used.

Review of the Global Forest Maps
3.3.1. Overview of the Global Forest Maps. Forest is an important natural land-cover type that usually describes surfaces covered by trees or by a canopy that exceeds a certain treecover percentage and height [30]. Monitoring global forestcover distribution and disturbances using remote sensing techniques is critical for understanding the intensification of human activities, climate change, and the global carbon cycle. In recent decades, as the intensity of human activities has increased and free access to medium-and highresolution remote sensing imagery has become available, global forest mapping and monitoring has experienced a transition from coarse-resolution to medium-and highresolution mapping. In this study, seven global 30 m forest products were acquired and investigated. These products are listed in Table 6.
All of these forest products define the forest land-cover type using the tree-cover percentage (TCP) and tree height (TH); however, they do not all use the same thresholds for the TCP and TH. For example, a TCP of 15% is used as the An unsupervised classification scheme is used to identify impervious surfaces based on the backscattering amplitude and texture; the results are used together with postediting and mosaicking to produce the final urban footprint products.

Xing et al. [49] Beijing
The GlobeLand30-2010 impervious layer was validated as having an overall accuracy of 91.37%, producer's accuracy of 92.76%, and user's accuracy of 98.19% using 66,058 validation points in Beijing.

Liu et al. [19] Global
The global overall accuracy, producer's accuracy, and user's accuracy for The overall accuracies of GHSL and GlobeLand30 were 89.21% and 80.54%, respectively, the corresponding producer's accuracies were 87% and 73%, and the user's accuracies were 86% and 77%.
Liu et al. [107] Global (45 cities  TreeCover 2010 Forest cover is defined using certain tree canopy cover thresholds without attribution to specific land-cover categories and land use. The forest cover class is identified using a threshold ≥ 49% tree canopy cover. Dataset link: https://glad.umd.edu/dataset/ global-2010-tree-cover-30-m A regression tree model is applied to estimate the maximum (peak of the growing season) treecanopy cover for each pixel from cloud-free annual growing season composite Landsat-7 ETM+ data from c.2010.
Hansen et al. [30] Potapov et al. [110] GLADForest 2000-2019 (annual) Trees are defined as vegetation taller than 5 m in height and are expressed as a percentage per output grid cell as "2000 Percent Tree Cover." "Forest Cover Loss" is defined as a standreplacement disturbance or a change from a forest to a nonforest state, during the period 2000-2019. "Forest Cover Gain" is defined as the inverse of loss or a nonforest to forest change entirely within the period 2000-2012. Dataset link: http://earthenginepartners.appspot .com/science-2013-global-forest/download_v1. 7 .html Tree-cover percentage, forest loss, and forest gain training data are related to the time series metrics using a decision tree. For the tree-cover and change products, a bagged decision tree methodology is employed. Forest loss is disaggregated to annual time scales using a set of heuristics derived from the maximum annual decline in the tree-cover percentage and the maximum annual decline in minimum growing season NDVI.
Hansen et al. [30] GFC30 2018 Land spanning more than 0.5 ha with trees higher than 5 m and a canopy cover of more than 10%, or trees able to reach these thresholds in situ Dataset link: http://www.chinageoss.cn/geoarc/ data/2019-GEOARC-01.zip The Earth is divided into 45 forest ecological zones. Each zone is classified using the random forest model and time series of Landsat imagery.
Zhang et al. [32] 9 Journal of Remote Sensing global 30 m forest map, Table 7 lists the accuracy metrics of each forest product based on the results of different studies. It should be noted that the forest land-cover type exists as an independent layer in GLC_FCS30, FROM_GLC, and Glo-beLand30 and also as a single thematic element in the GFCC30TC, TreeCover-2010, GLADForest, and GFC products.
For the products where forest is an independent layer (GLC_FCS30, GlobeLand30, and FROM_GLC), the producer's accuracies are 94.0%, 92.4%, and 76.5%, respectively, and the user's accuracies are 90.4%, 84.1%, and 80.5%, respectively, at the global scale. Kang et al. [39] and Gao et al. [45] comprehensively assessed the forest layers in these three different products at a regional scale, and the results indicated that GlobeLand30 and GLC_FCS30 have the same producer's and user's accuracies and that these values were higher than those of FROM_GLC. Furthermore, Moreno-Sanchez et al. [54] found that the overall accuracy of "forest" in GlobeLand30 was higher in tropical forest areas than in temperate forests, and Tsendbazar et al. [55] confirmed that the accuracy of the forest layer in GlobeLand30 is poorer in Africa than in other regions. Overall, the accuracy of the forest land-cover type in GlobeLand30 and GLC_FCS30 was found to be better than that in FROM_GLC.
Secondly, the global thematic forest products GFCC30TC and GFC30 have been validated as having accuracies of 90% and 90.94% at the global scale. GFCC30TC performs differently in different geographical regions: for example, it underestimates the high-canopy forest (canopy cover > 70%) but overestimates the low-canopy forest (canopy cover < 30%) in the boreal forest in Finland [56] and has a higher accuracy over North America (82% in Maryland) [57] than in South America (where the producer's and user's accuracies are 73:4% ± 0:09% and 94:7% ± 0:11%, respectively) [58]. TCC-2010 has been shown to have a higher accuracy than other forest products: for example, the producer's and user's accuracies were shown to be 88:4% ± 0:07% and 91:3% ± 0:07%, as against 73:4% ± 0:09% and 94:7% ± 0:11% for GFCC30TC, in South America [58]. In Kyrgyzstan, TCC-2010 was found to have an overall accuracy of 94.50% against 88.27% for GlobeLand30 [59]. Finally, GLADForest, which is the most widely used forest change product, has been validated to have an overall accuracy of 99:6% ± 0:7% for forest loss and 99:7% ± 0:6% for forest gain based on only 1500 validation points [30]. However, regional validations have indicated that this product has an overall accuracy of 81% in Maryland [57], 66%-56% in Indonesia [60], and 80:8% ± 2:3% in Costa Rica [61]. Therefore, as shown by the accuracy metrics in Table 7, the global thematic forest products (GFCC30TC, Tree-Cover-2010, GLADForest, and GFC30) generally have a higher overall accuracy than the global "forest layer" products (where forest is an independent layer), and the regional accuracies are usually lower than the global values. security. Cropland has a close relationship with human activities and usually consists of land cultivated with plants that are harvested for food, animal feed, and fibers. In recent decades, as the global population has rapidly grown, the growing demand for food has put added pressure on the food supply system, particularly because of the demand for grains used as livestock feed in regions where food security remains elusive [62]. Therefore, accurate global mapping of cropland is of great importance to assessing and monitoring global food security and production potential. In this study, six global 30 m cropland products were acquired. These products are summarized in Table 8.
Most global cropland products (GLC_FCS30, Globe-Land30, GCAD30, GFSAD30, and FROM_GC) contain single-epoch cropland information; the exceptions are the FROM_GLC and GlobeLand30 cropland products. In this study, the nominal year of 2015 was selected for analyzing the consistency between the global products. However, the GCAD30 stopped sharing after GFSAD30 was released and two of these products (FROM_GC and GlobeLand30) were available for 2010. The temporal variations between 2010 and 2015 were, therefore, neglected, and this may have led to a certain amount of bias in the comparison. Furthermore, from the point of view of the definition of cropland, there are obvious differences between these global cropland products. For example, GLC_FCS30 defines cropland using only human management as the criterion and FROM_GC excludes "permanent pastures." GlobeLand30, FROM_GLC, and GFSAD30 define cropland from a functional perspective as being land cultivated with plants that are harvested for food, animal feed, and fiber; however, FROM_GLC excludes fruit trees and classifies them as belonging to "forest." Finally, in terms of the methodology used, these products can be divided into three groups: GLC_FCS30 and FROM_GLC are produced using pixel-based supervised classification; GlobeLand30, GCAD30, and GFSAD30 use a combination of pixel-based classification and object-based segmentation; and FROM_GC is composited from four previously produced cropland maps.  Table 9. Similar to the forest land-cover type that was discussed in Section 4.2, cropland exists as a single layer in the GLC_FCS30, FROM_GLC, and GlobeLand30 land-cover products and also as an independent thematic product in FROM_GC and GFSAD30. (Note that GCAD30 has not been available since the release of GFSAD30.) Looking first at the products in which cropland is an independent layer, in GLC_FCS30, GlobeLand30, and FROM_GLC, cropland has been shown to have the producer's accuracies of 88.0%, 88.2%, and 47.7% and the user's accuracies of 83.9%, 88.7%, and 74.7%, respectively [16]. The accuracy of the cropland class in GlobeLand30 and FROM_ GLC varies according to the geographical zone and validation dataset used; for example, Zhong et al. [63] found an overall accuracy of 79.61% and kappa coefficient of 0.58 for FROM_GLC as against 76.23% and 0.52 found by Lu et al.  [111] The Loess Plateau, China The forest class in GlobeLand30, FROM_GLC, GLADForest, and GFCC30TC had an overall accuracy of 97% ± 0:2%, 95% ± 0:2%, 94% ± 0:3%, and 93% ± 0:3%, respectively. The producer's and user's accuracies of these four forest products were 93% ± 0:9% and 84% ± 0:1%, 92% ± 0:1% and 72% ± 0:1%, 63% ± 0:1% and 94% ± 0:8%, and 79% ± 0:1% and 75% ± 0:1%, respectively.

Moreno-Sanchez et al. [54] Mexico
The accuracy of the forest class of GlobeLand30 is higher in tropical forests than in temperate forests (around 90% compared with around 77%).

Xu et al. [35] Africa
The producer's and user's accuracies of the forest class of the second version of the African land-cover mapping in FROM_GLC are 82.4% and 72.6%, respectively.

Townshend et al. [109] Global
The overall accuracy of GFCC30TC is above or near 90%; the average user's accuracy and producer's accuracy for persisting forests are 92.5% and 95.4%, respectively. For persisting nonforest, forest loss, and forest gain, the corresponding accuracies are 94.8% and 91.1%, 87.1% and 91.1%, and 84.7% and 81.5%, respectively.

Sexton et al. [31] Global
The overall accuracy of the static forest-cover layers in GFCC30TC was 91%, and the overall accuracy of forest-cover change was >88%-these are among the highest accuracies reported for recent global forest-and land-cover products. 11 Journal of Remote Sensing [64]. GlobeLand30, which is the most widely used land-cover product [65], was demonstrated to have an overall accuracy of 80% in Africa, America, and Asia [66], an accuracy of 79.61% and a kappa coefficient of 0.58 in China [64], an accuracy of 80.61% in Shaanxi Province, China [67], an accuracy of 86.98% and a kappa coefficient of 0.692 in East Africa [68], an accuracy of 47% ± 1% in Tanzania [69], and an accuracy of 79.5% over five West African countries [70].
FROM_GC and GFSAD30 are two widely used thematic cropland products-this applies to GFSAD30 in particular. GFSAD30, which was derived and optimized from GCAD30 [71], was assessed for different continents and countries. Globally, an overall accuracy of 91.7%, a producer's accuracy of 83.4%, and a user's accuracy of 78.3% were found using 19,171 validation points [27,72]. Regionally, this product was found to have an overall accuracy of 94% in Australia and China [73], an overall accuracy of 90.8% and a weighted accuracy of 93.8% in Europe, the Middle East, Russia, and Central Asia [74], a weighted overall accuracy of 94.5% in Africa [75,76], an overall accuracy of 88.1% in Southeast and Northeast Asia [77], and an average user's accuracy of 73.1% over five West African countries [70]. As for the FROM_GC cropland product, it was assessed to have a significant linear relationship with FAOSTAT data (with an R 2 of 0.97) globally [29], and overall accuracies of 77.67% and 36% ± 1% were demonstrated for Shaanxi Province, China [67] and Tanzania [69], respectively. FROM_GC and GFSAD30 matched for 68.8% of cropland samples from Africa [75,76].
Comparisons between GlobeLand30 (in which cropland is a single land-cover layer) and GFSAD30 (in which cropland is an independent thematic product) have also been conducted. The results have indicated that the two products have an overall similarity of 88.8% and a kappa coefficient of 0.7 in Europe, the Middle East, Russia, and Central Asia [74]. GlobeLand30 was shown to have a higher overall accuracy than GFSAD30 over five West African countries [70]. GFCC30TC underestimated the high-canopy cover forest (cover > 70%) and overestimated the low-canopy cover forest (cover < 30%) in boreal forests, giving an R 2 of 0.53 and a bias of -2.1% for the comparison with the field-measured data.
Arjasakusuma et al. [60] Central Kalimantan Province, Indonesia The overall accuracy of GLADForest is 66%-56% Cunningham et al. [61] Costa Rica The overall accuracy of GLADForest is 80:8% ± 2:3%, and the producer's and user's accuracies are 79% and 81%, respectively, based on 1154 validation points. Inland water refers to fresh and brackish water bodies such as lakes, reservoirs, and rivers [78]. These water bodies play a vital role in the global and regional hydrological and biogeochemical water cycles [23,79]. However, due to their sensitivity to changes in the environment and climate, the distribution of water bodies varies over time. The resulting variations in water bodies have significant implications for the water balance and regional biodiversity [80]. In this study, three 30 m global inland water thematic products (G1WBM, Land used for planting crops including paddy fields, irrigated dry land, and rainfed dry land; land used for growing vegetables, herbage, greenhouses, or fruit trees and other economically valuable trees; also land used for planting shrub cash crops such as tea and coffee Dataset link: http://www.globallandcover.com/ GLC30Download/index.aspx Developed based on both crop phenology and the regular distribution patterns of cultivated land; object-based segmentation was then overlaid onto potential images of cultivated land. Using virtual interpretation, only objects displaying regular man-made patterns such as circles or rectangles were identified as cultivated land.
Chen et al. [14] FROM_GC 2010 The definition of cropland in this study is consistent with FAO's definition of "arable lands and permanent crops"; "permanent pastures" is not included. Dataset link: http://data.ess.tsinghua.edu.cn/data Composited from four existing cropland maps, including FROM_GLC and FROM_GLC_agg, and two 250 m masked cropland layers Yu et al. [29] GCAD30 2010 There are several rules for defining cropland including the following: (1) the minimum mapping unit of a particular crop is an area of 3 × 3 Landsat pixels (0.81 hectares); (2) all cultivated plants harvested for food, feed, and fiber, including plantations (e.g., orchards, vineyards, coffee and tea, and rubber plantations), are included; (3) >50% of the pixel is cropped; and (4) irrigation is defined as the artificial application of any amount of water to overcome crop water stress. Dataset link: not available An ensemble of methods was employed, including spectral matching techniques, the Automated Cropland Classification Algorithm (ACCA), and the hierarchical segmentation (HSeg) algorithm based on the Landsat 30 m Global Land Survey 2010 (GLS2010) dataset and a suite of secondary data (e.g., long-term precipitation and temperature data and a DEM).
Teluguntla et al. [71] Teluguntla et al. [114] GFSAD30 2015 Lands cultivated with plants harvested for food, feed, and fiber, including both seasonal crops (e.g., wheat, rice, corn, soybeans, and cotton) and continuous plantations (e.g., coffee, tea, rubber, cocoa, and oil palm). Cropland fallows are lands uncultivated during a season or a year but are farmlands and are equipped for cultivation, including plantations. Dataset link: https://lpdaac.usgs.gov/news/releaseof-gfsad-30 meter-cropland-extent-products/ Integration of pixel-based classifiers (random forest and support vector machines) and an object-based classifier (Recursive Hierarchical Image Segmentation) to obtain the cropland map for each ecological zone Oliphant et al. [27] Phalke et al. [74] 13 Journal of Remote Sensing Xu et al. [35] Africa The FROM_GLC cropland layer had a user's accuracy of 44.55% and a producer's accuracy of 69.71% using 6819 validation points.
Perez Hoyos et al. [66] Africa, America, and Asia GlobeLand30 generally provided adequate results to monitor cropland areas, and the overall accuracy was around 80% for the three continents studied.
Lu et al. [64] China The results of this study indicated that GlobeLand30 performed better than the other four datasets including FROM_GLC30, regardless of the cropland area and the spatial location. Specifically, the overall accuracy and kappa coefficient of GlobeLand30 were 79.61% and 0.58, respectively, whereas the overall accuracy and kappa coefficient of FROM_GLC30 were 76.23% and 0.52, respectively.
Yu et al. [115] China GlobeLand30 shows that there is little loss of the cropland area but increasing fragmentation of cropland in China. Specifically, the results show that 703 out of 2420 countries experienced both cropland loss and increased fragmentation.
Jacobson et al. [68] East Africa GlobeLand30 was found to have an overall accuracy of 86.98% and a kappa coefficient of 0.692 using the Google Earth grid dataset.
Oliphant et al. [27] Global GFSAD30 had an overall accuracy of 91.7%, a producer's accuracy of 83  Samasse et al. [70] Five West African countries The average user's accuracies for GFSAD30 and GlobeLand30 in the five countries were 73.1% and 79.5%, respectively. The omission and commission errors were about 25% and 7%, respectively, for GFSAD30 and about 29% and 5%, respectively, for GlobeLand30.
Yu et al. [29] Global An analysis of the agreement between FROM_GC and FAOSTAT data was conducted, and the results indicated that FROM_GC had a significant linear relationship with the FAOSTAT data, giving an R 2 of 0.97.
15 Journal of Remote Sensing JRC_GSW, and GLADWater) and three 30 m GLC products (GLC_FCS30, FROM_GLC, and GlobeLand30) were acquired for carrying out a spatial consistency analysis of inland water bodies for 2015. Details of these products are shown in Table 10.
Apart from JRC_GSW and GLADWater, for which annual water maps are available for periods of a number of years (1984-2015 and 1999-2018, respectively), these products were only produced in certain individual years. Specifically, G1WBM was only produced in 2010, GLC_FCS30 is only available for 2015, FROM_GLC is only available for 2010, 2015, and 2017, and GlobeLand30 was only produced in 2000 and 2010. Because 2015 was chosen as the nominal year for comparison and because there was no G1WBM and GlobeLand30 available for 2015, any discrepancies between 2015 and the actual years used may also have caused some discrepancies in the results. Regarding the definitions of "water body" used by the different products, the three thematic water products were distinguished between the permanent and seasonal water types based on the frequency of the presence of water occurrence, whereas the three GLC products were not distinguished between the permanent and seasonal water bodies. In addition, there were also differences between the definitions of permanent water and seasonal water used by the three thematic products. G1WBM defines permanent water as corresponding to pixels where water is present more than 70% of the year; however, JRC_GSW defines pixels as permanent water only if the water is present throughout the year. GLADWater also labels as permanent water the pixels where the difference between the maximum frequency and the minimum frequency of the presence of water over three consecutive years is less than or equal to 33% and the average frequency over these years is greater than 90%. In order to reduce the discrepancies caused by these different definitions of water bodies, we combined the permanent and seasonal water types in the thematic products into one water class. The three GLC products are all derived using supervised classification methods. However, two of the thematic water products (G1WBM and JRC_GSW) are generated based on classification rules and globally consistent thresholds; only GLADWater is produced using a supervised classification method.

Review of the Accuracy Assessments of the Global Water
Maps. The results of the assessment of the accuracy of the six products that include global inland water maps are summarized in Table 11. Among the three GLC products, the results for GlobeLand30 and FROM_GLC include regional and global results, whereas the assessment of GLC_FCS30 was performed for Indonesia only. Among the global thematic inland water products, global results were available for JRC_GSW and GLADWater, whereas, for G1WBM, only the report for the United States could be found.
For GlobeLand30 and FROM_GLC, the water class has a satisfactory global accuracy. The overall accuracies of this class in GlobeLand30-2000 and GlobeLand30-2010 are 96.51% and 96.48%, respectively [81]; the user's and producer's accuracies are 81.97% and 85.66%, respectively, in FROM_GLC-2010 [82] and 81.97% and 86.39%, respectively, in FROM_GLC-2017 [18]. At the regional scale, the accuracy was found to vary with the location. For example, Guo et al. [83] found that the overall accuracy for water in FROM_ GLC-2017 was 95.87% in West Asia-Northeast Africa, 89.75% in Central and Eastern Europe, 87.01% in Central Asia, and 15.67% in South Asia. Also, the user's and producer's accuracies for GlobeLand30-2010 were 93.44% and 54.42%, respectively, in Indonesia [39] and 94.00% and 79.00%, respectively, in China [84]. Additionally, it was found that the producer's accuracy (which is related to the omission error) was generally much lower than the user's accuracy at the regional level. For instance, the user's and producer's accuracies of the water class in GLC_FCS30, Glo-beLand30, and FROM_GLC in Indonesia were 97.74% and 60.19%, 93.44% and 54.42%, and 89.97% and 68.52%, respectively [39]. Also, the user's and producer's accuracies of the water class in FROM_GLC-2015 in Africa were 93.50% and 66.90%, respectively [85]. These results indicate that these GLC products are not sufficiently accurate to fully support the analysis of water bodies at the regional scale and that they are more likely to omit water bodies in specific regions.
In the three thematic inland water products, the accuracies of the permanent water body class are quite high. The global commission and omission accuracies of this class in JRC_GSW are greater than 99% and 97%, respectively [25]. Also, for the United States, more than 85% of the permanent water bodies in G1WBM were proved to be accurately mapped [86]. For GLADWater, the global user's and producer's accuracies are 97:8 ± 1:8% and 85:8 ± 2:4%, respectively [26]. However, the accuracies for the seasonal water body class are much lower than those for permanent water bodies. Pekel et al. [25] reported that the omission accuracy of the seasonal water body class in JRC_GSW was only about 75%. The accuracy of seasonal water in JRC_GSW was found to be even lower in the research of Pickens et al. [26], which gave the user's and producer's accuracies of 17:4% ± 12:1% and 36:3% ± 8:3%, respectively. Similarly, the user's and producer's accuracies for GLADWater were only 44:0 ± 7:1% and 73:0 ± 5:6%, respectively, based on the time series of stable seasonal water bodies [26]. These results show that it was still difficult to accurately detect seasonal water bodies.

Assessment of the Spatial Consistency of the Current Global 30 m Products
4.1. Spatial Consistency of the Three GLC30 Products. All these 30 m GLC datasets were aggregated to a spatial resolution of 0.05°for visual comparison of the spatial pattern, in which the land-cover type in each 0.05°grid was determined by the land-cover type with the largest proportion. Figure 1 shows the spatial distributions of the different land-cover types in the three fine-resolution GLC maps both at the global scale with a resolution of 0.05° (Figure 1(a)) and at four typical inconsistent regions with a size of 0:5°× 1°and a resolution of 30 m (Figure 1(b)). Overall, the three products are consistent and roughly capture the actual distribution of land-cover types at the global scale. Specifically, bareland is mainly found between latitudes 15°N-50°N and 15°S-35°S, including parts of Central 16 Journal of Remote Sensing Gong et al. [12] GlobeLand30 2000, 2010, 2017 Water bodies include clear water, green water, and turbid water and display spectral diversity. Clear water has a lower reflectance in all bands, while turbid water has a higher reflectance due to the blend of mud and sand; green water is caused by eutrophication and to some extent displays spectral features similar to green vegetation. Pickens et al. [26] 17 Journal of Remote Sensing Asia and the western United States. In the northern hemisphere, cropland is mainly found in India and southern China in Asia, mid-to high-latitude parts of Europe, and the southeastern United States; in the southern hemisphere, cropland is mainly found in southern Brazil and northern Argentina. The forest class is concentrated in rainforest areas (including the Amazon rainforest, African rainforest, and Indian-Malay rainforests) and high-latitude areas. However, the three GLC30 products show large inconsistency in the transition zones or heterogeneous regions (as illustrated in Figure 1(b) A-D). For example, the disagreement is noticeable in the transition zones from bareland to grassland or cropland (Figure 1(b) A, B) and in the heterogeneous mountainous areas with complicated land-cover types (Figure 1(b) C, D). Figure 2 illustrates the percentage areas covered by each class according to the three GLC products. The percentages show roughly the same pattern for the different products: forest has the highest percentage area, accounting for about 30% of the global total, followed by grassland, bareland, and shrubland.
The area percentages of grassland, shrubland, and tundra in the GLC_FCS30-2015 product are very different from those in the other two products. The shrubland percentage in GLC_FCS30-2015 (11.38%) is higher than that in FROM_GLC30-2015 (7.47%) and GlobeLand30-2010 (5.18%); in contrast, the grassland and tundra percentages are significantly smaller than those in the other two products. In particular, the tundra percentage (4.79%) in GLC_FCS30-2015 is about half of the area in the other two products. Furthermore, the area percentages of cropland and wetland in FROM_GLC30-2015 are much lower than those in GLC_ FCS30-2015 and GlobeLand30-2010. In FROM_GLC30-2015, wetland accounts for only 0.02% of the global total, whereas in GLC_FCS30-2015 and GlobeLand30-2010, the corresponding percentages are 2.24% and 2.52%, respectively. In FROM_GLC30-2015, cropland covers only 7.44% of the total area-much lower than the figures of 13.96% for GlobeLand30-2010 and 15.64% for GLC_FCS30-2015. Figure 3 illustrates the spatial consistency between the three GLC products. It is obvious from Figure 3 that the  18 Journal of Remote Sensing spatial distribution of the three products is relatively consistent in homogeneous regions such as the northern part of the African continent (mainly bareland) and the northern region of South America (forest). However, the spatial consistency is low in heterogeneous or transition zones such as Australia and southern Africa and the transition zone from the vegetation to nonvegetation classes in Africa and also central North America. It can be seen from Figure 1(b) A-C that the disagreement in these areas mainly occurs in areas labeled as grassland, shrubland, or bareland. This is mainly due to the maximum fraction of vegetation coverage that the different classification systems permit in areas defined 19 Journal of Remote Sensing as not being covered by vegetation. For example, the maximum fraction of vegetation coverage in bareland, as defined by the GlobeLand30 product, is 10%, whereas the maximum fraction defined by GLC_FCS30 is 15%. Therefore, using the former definition, sparse shrubland or grassland will be classified as shrubland, whereas using the latter definition, sparse shrubland and grassland will be wrongly classified as bareland.
To further analyze the spatial consistency between each pair of GLC products, the superposition method was applied to each pair of products. Figure 4 shows the spatial distribution of the proportion of consistent pixels within each 0:05°× 0:05°grid cell for each pair of the compared GLC products (Figure 4(a)) and the spatial consistency at four typical inconsistent regions with a size of 0:5°× 1°and a resolution of 30 m (Figure 4(b)). Among all of the pairs, GlobeLand30-2010 and FROM_GLC30-2015 have the highest consistency with the consistent area accounting for about 63.35% of the total area; in terms of consistency, this is followed by GLC_FCS30-2015 and GlobeLand30-2010 (59.73%) and then GLC_FCS30-2015 and FROM_GLC30-2015 (57.11%). These results may have been affected by the conversion between the different classification systems. Also, similar to the results for the spatial consistency between the three different products, there is a low degree of consistency in the areas with complex land-cover types and in the Area proportion (%) Figure 2: Comparisons of the area percentages for each class for the different GLC products.  Figure 3: The spatial consistency between the three GLC products. PHC: proportion of highly consistent pixels (whose consistency metric equals 3) within each 0:05°× 0:05°grid cell; PMC: proportion of moderately consistent pixels (whose consistency metric equals 2) within each 0:05°× 0:05°grid cell; PLC: proportion of low-consistency pixels (whose consistency metric equals 1) within each 0:05°× 0:05°grid cell.

20
Journal of Remote Sensing transition zones from the vegetation to nonvegetation classes. For example, there is a low degree of consistency in the heterogeneous regions with complex land-cover types and in the transition zones from the vegetation to nonvegetation classes. For example, there is significant disagreement between each pair of products in the transition zones from bareland to grassland or cropland in Africa (Figures 1(b) A and 4(b) A). The same phenomenon can be seen in highlatitude heterogeneous mountainous areas with complicated land-cover types (Figures 1(b) D and 4(b) D). Further

Spatial Consistency of Global 30 m Impervious Surface
Maps. Figure 5 illustrates the fraction of impervious surfaces across the globe according to the seven global 30 m impervious surface products using a grid size of 0.05°. Overall, all these products accurately capture the global patterns of impervious surfaces. For example, impervious surfaces can be seen to be mainly found between 20°N and 60°N (including central and eastern parts of North America, the whole of Europe, and eastern parts of China). This agrees with the general patterns of global economic development and population distribution. However, considering the patterns shown in more detail, NUACI-2015, FROM_GLC-2015, and GAIA-2015 clearly produce lower estimates of the amount of impervious surface cover over America and India compared with other products.
To quantitatively assess the spatial consistency between the seven global impervious surface products, the total meridional and zonal areas classified as impervious surfaces were calculated using 0.05°steps ( Figure 6). It can be seen from the meridional histogram that all of the products share similar curve shapes that exhibit significant volatility. Four peak intervals in the longitude direction are also captured; these include the region between 60°W and 100°W covering the central and eastern parts of North America, the region between 10°W and 40°E covering the whole of the European Union, the region between 60°E and 90°E, which covers India, and the region between 100°E and 120°E, which covers the east of China and Southeast Asia. Looking at the different products in more detail, using a 0.05°step, HBASE-2010 and GAIA-2015 have the biggest and smallest impervious surface areas among these products, respectively. The values for GAIA-2015 in the 10°W-40°E region (Europe) are clearly low, which is consistent with the results shown in Figure 5.
Similarly, the zonal histogram illustrates that all the global products accurately capture the global distribution patterns of impervious surfaces and shows that this cover type is mainly concentrated between 20°N and 60°N, which is where most of the world's population and economic activity are also concentrated. HBASE-2010 has the biggest area classified as impervious surfaces in almost all latitude regions except for 45°N-55°N, where GHSL-2015 has a bigger area. In terms of the total area classified as impervious surfaces, HBASE-2010 has the largest area, followed in order by MSMT_IS30-2015, GHSL-2015, GlobeLand30-2010, FROM_GLC-2015, NUACI-2015, and GAIA-2015 (see Figure 6).
Together with the area-based analyses, scatter plots were also used to analyze the spatial consistency between the different global impervious surface products (Figure 7). In this study, MSMT_IS30-2015 was selected as the reference

22
Journal of Remote Sensing because it has been demonstrated [22] to have a better performance than most other products. First, it was found that the two products that are most consistent with each other are MSMT_IS30-2015 and GHSL-2015: the relationship between these two products gave an R 2 of 0.783, an RMSD of 0.038, and a regression slope of 0.921, with the points in the scatter plot being distributed around the 1 : 1 line. The second best agreement was found between FROM_GLC-2015 and MSMT_IS30-2015, for which R 2 was 0.740 and the RMSD was 0.042; in this case, there was a better agreement between the points in the "high-density" regions than in the "low-density" regions, which resulted in a regression slope of less than 1. In GlobeLand30-2010, there is a temporal interval of 5 years between 2010 and 2015 and a minimum mapping unit of 4 × 4 pixels, and so it was expected that GlobeLand30-2010 would have a smaller impervious area than MSMT_IS30-2015; in this case, the points in the scatter plot are mainly located below the 1 : 1 line. HBASE-2010 is the only dataset that has a bigger impervious surface area than MSMT_IS30-2015-the slope of the regression is 1.07. This is especially the case in "high-density" regions because, in HBASE, impervious surfaces are defined as comprising all types of man-made built-up surfaces and the areas surrounding them that are functionally linked to those surfaces (e.g., urban green spaces) [102].

23
Journal of Remote Sensing forest maps accurately capture the real patterns in the forest land-cover type. For example, forests can be seen to be mainly concentrated in rainforest areas, including the Amazon rainforest, African rainforests, and Indian-Malay rainforests, and in high-latitude areas where needleleaf forests are found. In high-latitude areas, the forest cover according to GFCC30TC-2015 is obviously sparser than the other forest products; TreeCover30-2010 and GLADForest-2018 give higher densities than the other forest products in the African rainforest areas. Figure 9 illustrates the distribution of forest in the meridional and zonal directions in steps of 0.05°according to the seven global forest products. Specifically, in terms of the meridional direction, all the forest products capture the global forest spatial patterns, including three peak intervals between 90°and 50°W (the Amazon rainforest), 0°and 40°E (which includes the African rainforests), and 60°and 150°E. There is greater consistency between these products in the western hemisphere than in the eastern hemisphere. It can also be seen that GFCC30TC-2015 generally exhibits the smallest forest areas while TreeCover-2010 exhibits the largest areas, particularly between 0°and 40°E (which includes the African rainforests).
The consistency between the seven products is also illustrated by the forest area curves for the zonal direction. Over-all, there is a greater agreement between the products in the southern hemisphere than in the northern hemisphere. The greatest difference occurs between GFCC30TC and the other products around 60°N (where the total forest area in TreeCover-2010 is nearly twice that in GFCC30TC-2015). Looking at the patterns in the curves, the seven forest products can be divided into three groups: GFCC30TC-2015 on its own, GFC30-2015 and GlobeLand30-2010, and the remaining four products as the final group.

24
Journal of Remote Sensing concentrated around the 1 : 1 line. Finally, from the regression slope coefficients, it can be seen that the spatial consistency between GLC_FCS30, TreeCover-2010, and GLADForest is more significant than that between other pairs of products (the slopes of TreeCover-2010 and GLAD-Forest are closer to 1.0).   Figure 14 shows the spatial distribution of the water class in the six global 30 m inland water maps after aggregation of the data to a resolution of 0.05°. Overall, the spatial patterns in the inland water maps are reasonably consistent. For instance, there is a relative abundance of water bodies in the eastern whereas water bodies are relatively sparsely distributed in Africa. However, some discrepancies still exist between these products in specific regions. For example, there are obviously more water bodies in the western part of South America in FROM_GLC than in the other five products.

Spatial Consistency of Global 30 m Cropland Maps.
The spatial consistency between the six global inland water maps can also be seen from the meridional and zonal distributions of this class based on 0.05°intervals that are shown in Figure 15. These six products have similar trends in the latitudinal and longitudinal directions. However, there are still some obvious disagreements in specific regions.
The distribution of inland water between 15°W and 30°W, 60°W and 80°W, and 50°E and 120°E is quite different in the different maps. Between 15°W and 30°W and 60°W and 80°W, the size of the water area in FROM_GLC-2015 is much greater than in the other products. Also, in both FROM_ GLC-2015 and GLADWater-2015, between 50°E and 120°E, there is a much greater water area than in the other products. These patterns are also noticeable in Figure 15, where it can be seen that, in FROM_GLC-2015, there is a greater number of water bodies in northwestern Africa (15°W to 30°W), Iceland (15°W to 30°W), western South America (60°W to 80°W), and eastern North America (60°W to 80°W). Also, in GLADWater-2015 and FROM_GLC-2015, there are more water bodies in mainland Asia (50°E to 120°E).
There are also differences in the latitudinal distribution of inland water between the products. In FROM_GLC-2015 and GLADWater-2015, there is a greater area of water than in the other products from 25°N to 35°N and above 70°N. This can also be seen from Figure 15. In mid-and high-latitude regions, FROM_GLC and GLAD-Water exhibit higher proportions of water than the other four products (Figure 15). In order to quantitatively analyze the differences between the six global inland water maps, scatter plots of the GLC_ FCS30-2015 water map against the other products were made using a 0:05°× 0:05°grid and the results are shown in Figure 16. The results show that, for all five pairs of products, the value of R 2 is greater than 0.82 and all the products have a high degree of consistency with the water map in GLC_ FCS30-the values of R 2 range from 0.823 (G1WBM-2010) to 0.913 (JRC_GSW), and the RMSD values range from 0.106 (G1WBM-2010) to 0.068 (JRC_GSW). The inconsistency between GLC_FCS30-2015 and GlobeLand30-2010 and G1WBM-2010 may be partly because these three products were produced in different years. Moreover, since the algorithm used to develop G1WBM-2010 is mainly intended to accurately detect permanent water bodies [24], the uncer-tainty in the detection of seasonal water bodies may also be a cause of some of the differences between G1WBM-2010 and GLC_FCS30-2015.

Perspectives
Firstly, the synergy between the spectral, spatial, and temporal features derived from multisource satellite datasets can improve the classification accuracy of fine-resolution GLC products. Over recent decades, the volume and number of types of satellite data have grown due to various breakthroughs, providing great opportunities for accurate landcover mapping. Taking the mapping of impervious surfaces as an example, due to the complicated spectral and spatial structures associated with these surfaces, it is very difficult  Journal of Remote Sensing to produce accurate large-area impervious surface products using only optical imagery. For example, Liu et al. [19] used Landsat imagery to generate a global 30 m impervious surface product, but this map suffered from serious omission errors, especially in rural areas and small cities [22,106]. Furthermore, as free access to Sentinel-1 RADARSAT and ENVI-SAT synthetic-aperture radar (SAR) imagery has become available, the combination of optical imagery and SAR imagery has been demonstrated to significantly improve the accuracy of the mapping of impervious surfaces. This is because optical imagery can capture surface reflectance characteristics, whereas SAR data can be used to provide information on the structure and dielectric properties of surface materials [22,106,118]. Recently, the rapid development of nighttime light remote sensing has also helped to promote accurate impervious surface mapping. For example, Goldblatt et al. [119] combined nighttime light data and Landsat imagery to characterize the built-up land cover in three geographically diverse countries: India, Mexico, and the US. Zhang et al. [22] combined multisource (optical, SAR, nighttime light, and DEM) datasets and spectral and textural features from time series stacks to generate a global 30 m impervious surface product that had an overall accuracy of 95.1% and a kappa coefficient of 0.898. Secondly, the local adaptive stratified classification scheme is essential for accurate large-scale land-cover mapping. Due to the spectral heterogeneity of land-cover types, it is very difficult to generate an accurately classified product using a single classification model [120]. The stratified classification scheme is more appropriate for regional or global mapping that has a complicated classification system [19,44]. For example, to improve the mapping accuracy, Chen et al. [14] presented a split-and-merge strategy for handling the classification of 10 different land-cover types: i.e., first, each class in a prioritized sequence was identified, and then the ten types were merged together. Zhang et al. [16] directly used an existing global 30 m impervious surface map as an independent land-cover type in the global land-cover product, which was superimposed over the global land-cover classifications. In addition, as the distribution of the land-cover types varies with the region, the use of a local adaptive classification scheme is more appropriate [19,44]. For instance, Liu et al. [15] divided the global land surface into 5°× 5°geographical tiles and developed a local random forest model for each tile using the corresponding training samples, then generated a global impervious surface map (MSMT_IS30) by mosaicking 948 geographical tiles. This map, developed using local adaptive classifiers, has a much higher accuracy than the single classifier-based products: MSMT_IS30 has an overall accuracy of 0.95, which compares with 0.89 for FROM_GLC-2015, 0.903 for GHSL-2015, 0.884 for Globe-Land30-2010, and 0.880 for HBASE-2010 [22]. Furthermore, the classification accuracy of fine-resolution GLC maps can also be improved using deep learning algorithms. For 31 Journal of Remote Sensing example, using a deep convolutional neural network-(CNN-) based classifier, an accurate 30 m large-scale land-cover map was generated, achieving a classification accuracy of 84.4%, compared to 79.9% for RF and 80.2% for SVM 80.2% [121].
Thirdly, over the last four decades (since the launch of the first Landsat satellite in 1972), land-cover updates based on a time series change detection approach have made it possible to obtain information about land-cover types and changes at a high resolution. Land-cover change (LCC) is, in fact, the key information needed for global change research. Although the existing global land-cover products partially meet users' requirements in terms of performance, none of these global land-cover products meets the requirements in terms of interannual stability due to the limited classification accuracy of results if the GLC map is generated independently for each year. Thus, these temporally independent land-cover products cannot be used to support LCC analysis [4,8,[122][123][124]. To deal with this issue, change-based land-cover mapping has been adopted by the remote sensing community to generate temporally consistent GLC products. Using the traditional approach, first, an annual land-cover dataset is generated, and then some time series postclassification processing steps are implemented to ensure temporal consistency and logical transition, which requires a large number of subjective rules [125][126][127]. Recently, the continuous change detection approach has been widely applied to the detection of land-cover changes using time series of remote sensing images. Using this approach, changed pixels are first identified using dense time series of images; the changed regions are then updated using temporal segments instead of specific epochs [128][129][130]. In this way, an opportunity for generating accurate and consistent annual fine-resolution GLC products that can be used for the analysis of global land-cover changes is provided.
Finally, cloud computing platforms greatly facilitate global fine-resolution land-cover mapping. The volume of data required for generating a fine-resolution GLC product is huge, and this task was previously very labor-and computationally intensive [131]. Figure 17 outlines the development of some representative fine-resolution land-cover products over the years, and it can be seen that the early land-cover products (e.g., the National Land-Cover Database (NLCD) [132] and China's land-use and land-cover (LUCC) product [133]) were mainly based on manual, visual interpretation, which is very laborious and time-consuming. Later, the Landsat dataset became freely available to the public and more fine-resolution remote sensing datasets (e.g., Landsat-8 and Sentinel-2) appeared. In order to promote the automation of land-cover mapping, sample-based classification strategies were employed to reduce the amount of labor required compared to the manual editing approach. Some fine-resolution GLC products, such as GlobeLand30 [14] and FROM_GLC [12], were then developed. However, the amount of labor required for the collection of training samples was still a limitation [44], and the amount of computation required remained a challenge for scientists until cloud-based computing platforms, such as GEE, became available. GEE has proved its potential in many different applications, especially in land-cover mapping, and the Landsat and Sentinel datasets have been extensively utilized by GEE users [44]. Due to GEE's powerful computing ability [134], the production of fine-resolution GLC products can now be completed in a few hours or days [134]. Also, because of the massive data reserves held within GEE, the training datasets can be automatically derived from existing landcover products and multisource remote sensing datasets without any labor input [135]. Gong et al. [18] automatically developed the 10 m FROM_GLC10 by transferring an existing 30 m resolution sample set to the 10 m Sentinel-2 imagery using the GEE platform. Zhang et al. [16] used the freely available multisource data and the programming environment within GEE to construct a spatiotemporal spectral library to automatically generate GLC_FCS30. Therefore, we can expect more and more GLC products with higher resolutions, finer classification systems, more accurate classification, more real-time releases, and more frequent updates. However, perhaps too many GLC products are being generated by the remote sensing community too quickly. Does having so many datasets really benefit or mislead data users? It is a particularly urgent task to assess the quality of the increasing GLC datasets and their suitability for different user applications. As also reported in numerous research studies, the mapping accuracy varied noticeably for different validation datasets. Therefore, rigorous and transparent validation is much more important than ever.

Conclusions
Most human impacts on the land cover can be captured at a resolution of 30 m, and the 30 m GLC products have been widely used in many applications. In this study, the accuracy   98.50% (JRC_ GSW) for the inland water products. Given that the target maximum error is 5-15% either per class or for the overall accuracy [136], most of the current land-cover maps still do not meet the accuracy demands of many applications. In addition, the consistency and comparability of different GLC maps show a good overall agreement in terms of overall spatial patterns but a limited agreement for some vegetation classes (such as shrub, tree, or grassland) in specific areas, mostly transition zones. Unfortunately, land-cover change primarily occurs in these transition zones, which means that it is extremely essential to improve mapping accuracy in these zones. So far, all these temporally independent GLC products cannot be used to support LCC analysis due to the limited classification accuracy. Therefore, more efforts are needed toward improving the classification accuracy and generating temporally consistent GLC products, especially for classes for which the accuracy is currently low (such as shrub, wetland, tundra, and grassland) and for regions with low accuracy (heterogeneous or transition zones, tropical rainforest regions).

Conflicts of Interest
The authors declare that there is no conflict of interest regarding the publication of this article.