Annual Dynamics of Global Land Cover and its Long-term Changes from 1982 to 2015

Land cover is the physical evidence on the surface of the Earth. As the cause and result of global environmental change, land cover change (LCC) influences the global energy balance and biogeochemical cycles. Continuous and dynamic monitoring of global LC is urgently needed. Effective monitoring and comprehensive analysis of LCC at the global scale are 15 rare. With the latest version of GLASS (The Global Land Surface Satellite) CDRs (Climate Data Records) from 1982 to 2015, we built the first record of 34-year long annual dynamics of global land cover (GLASS-GLC) at 5 km resolution using the Google Earth Engine (GEE) platform. Compared to earlier global LC products, GLASS-GLC is characterized by high consistency, more detailed, and longer temporal coverage. The average overall accuracy for the 34 years each with 7 classes, including cropland, forest, grassland, shrubland, tundra, barren land and snow/ice, is 82.81 % based on 2431 test sample units. 20 We implemented a systematic uncertainty analysis and carried out a comprehensive spatiotemporal pattern analysis. Significant changes at various scales were found, including barren land loss and cropland gain in the tropics, forest gain in northern hemisphere and grassland loss in Asia, etc. A global quantitative analysis of human factors showed that the average human impact level in areas with significant LCC was about 25.49 %. The anthropogenic influence has a strong correlation with the noticeable vegetation gain, especially for forest. Based on GLASS-GLC, we can conduct long-term LCC analysis, improve 25 our understanding of global environmental change, and mitigate its negative impact. GLASS-GLC will be further applied in Earth system modeling to facilitate research on global carbon and water cycling, vegetation dynamics, and climate change. The GLASS-GLC data set presented in this article is available at https://doi.pangaea.de/10.1594/PANGAEA.913496 (Liu et al., 2020).


Introduction
In addition, in order to enhance the discriminating capacity, we also used terrain data from Global Multi-resolution Terrain Elevation Data 2010 (GMTED2010). Based on the elevation data, the slope information can be further calculated to reflect the terrain and help distinguish different vegetation types growing on steep slopes to those on level ground. The dataset comes 125 from the GEE platform and contains Earth Elevation data collected from various sources (USGS, 2020). The primary source is the Shuttle Radar Topography Mission (SRTM) Digital Terrain Elevation Data (DTED) (void-filled) 1-arc-second data.
Other sources are used for filling the gaps in areas outside the SRTM coverage. As the terrain is relatively stable over years, using the data as input for multiple years is plausible. The spatial resolution of the GMTED2010 data used is 7.5 arc seconds and it has been upsampled to 5 km in subsequent analyses. 130

Classification system
The classification system in FROM-GLC Version 2 (FROM-GLC_v2) defines eleven main classes that can be easily mapped to the Food and Agricultural Organization of the United Nations (FAO) LC Classification System and the International 6 training sample in each class is shown in the inner pie chart.

Input data collection
We constructed an input data set with a strong discrimination ability to detect LC from multiple aspects such as terrain, 155 phenology, spectrum, and spectral index, etc. The annual percentiles (including 0, 10, 25, 50, 75, 90, 100) of all bands of the GLASS CDRs and the mean and standard deviation of the NDVI between two adjacent percentiles are calculated, as an annual input data collection from GLASS CDRs. Among them, the percentile that represents specific phenological information can provide simplified time series information, reduce the noise of annual time series, and help improve the classification accuracy (Hansen et al., 2013). By extracting the statistical information between adjacent percentiles, the time series information can be 160 further supplemented. Due to the systematic deviation of AVHRR products (Song et al., 2018b), in order to ensure the interannual consistency of the GLASS data, we used the processing method developed for generating the VCF products, with the corresponding MODIS products for end-member correction, where desert and intact forest are regarded as the end-member of each pixel (Song et al., 2018a). After the correction, the inter-annual inconsistency of input data collection from the GLASS CDRs is improved. Figure S1 shows the time series of the global median value of the GLASS ABD_WSA_VIS band, where 165 the orange one represents the curve before the correction and the grey one is the result after the correction. It can be seen that after the correction, the fluctuations of the input data become smaller, and the individual abnormal values are also adjusted.
Taking into account the time span of the GLASS CDR-based input data collection, the VCF products from 1982 to 2015 are used, with the missing 1994 and 2000 data supplemented by calculating the average of the adjacent years. The VCF consists of percentages of tree cover (TC), short vegetation (SV), and bare ground (BG) for each year. Based on the GMTED2010 170 dataset, the slope information is calculated and finally included to obtain an average slope value for each 0.05 ° pixel. In addition, the central latitude and longitude information of each 0.05 ° pixel is also recorded as part of input data. Finally, an annual collection of 81 kinds of input for the period of 1982 to 2015 was constructed, including the annual GLASS CDR percentiles (7×9), the mean and standard deviation of the NDVI annual adjacent percentiles (6×2) and VCF percentages (3), slope information (1), latitude (1), and longitude (1) information (Table 2). 175

Classification method and temporal consistency check
We used a random forest classifier for global LC mapping following the good performance of the random forest classifier in the machine learning field (Rodriguez-Galiano et al., 2012;Pal, 2005). The number of trees was 200 with out-of-bag mode turned on. The number of variables per split was set to 0, as the square root of the number of variables. The minimum size of a terminal node, the fraction of input to bag per tree, and random seed were set to be 1, 0.5, and 0, respectively. The classifier 180 was trained using the training sample with an annual data collection constructed as the input. The global LC maps from 1982 to 2015 were obtained using the trained classifier. The out of bag accuracy reached 87.12%.
In order to further ensure the temporal consistency of the mapping results, we used the "LandTrendr" method (Kennedy et al., 2010;Cohen et al., 2018) and implemented a linear regression-based algorithm for the constructed annual input data collection to find the breakpoints in the time series (Li et al., 2018). The class labels in the time series between adjacent breakpoints will 185 be updated to the mode values of the class label time series for the time period. Through this strategy, we can smooth the time series of the mapping results, avoid noise interference as much as possible, and finally get the adjusted GLASS-GLC.

Accuracy assessment
To verify the reliability of GLASS-GLC CDR products from multiple perspectives, we performed accuracy assessments and uncertainty analyses. The test sample was extracted from the 30 m resolution FROM-GLC_v2 (Li et al., 2017) to evaluate the 190 2015 LC mapping results. First, we dropped those sample units whose classes were not included in our classification system.
The remaining test sample units were then overlapped with the abovementioned aggregated 0.05 ° FROM-GLC_v2 mapping result, and only those whose class labels were consistent were kept. These were regarded as huge homogeneous sample units  Fig. 2 (b), where the percentage of test sample for each class is shown in the inner pie chart.
In addition to obtaining the classification confusion matrix in 2015 based on the above test sample, in order to identify regions where classification is difficult, an uncertainty analysis was carried out. The incorrect test sample locations are marked as 1, while the correct test sample locations are marked as 0. The spatial distribution map of the uncertainty of the LC mapping 200 result in 2015 is depicted based on a Kriging interpolation method (Oliver and Webster, 1990). The search radius parameter of Kriging interpolation is set to 12 nearby sample units, the other parameters as default. The value of the uncertainty ranges from 0 to 1. A value near 0 indicates a lower uncertainty, while a value near to 1 indicates a higher uncertainty and a higher possibility of misclassification.
Other than these, we collected an independent test sample and performed accuracy assessment. Specifically, we 205 collected 2431 randomly distributed 5 km sample points in different years around the world. According to the majority principle, we manually interpreted the land cover class of each sample as an independent test sample. To prove the impact of change detection, we further compared the accuracies with and without change detection. The geographical distribution of the independent test sample is shown in Fig. 2 (c), and the temporal distribution is shown in the inner chart.

Data inter-comparison
To better reflect the product quality, We inter-compared GLASS-GLC with other available global land cover products with a relatively long time series. Land cover products from MODIS and the ESA-CCI were used. The MLCT global land cover products come from Collection 6 (C6) MLCT products (Sulla-Menashe et al., 2019), and are supervised classification results from 2001 to 2016. Considering the comparability to our classification system, the FAO-Land Cover 215 Classification System land use (LCCS2) layer was used. The corresponding relationships of classes are listed as follows, and the class names we used are the latter: barren -barren land, permanent snow and ice -snow/ice, all kinds of forest -forest, forest/cropland mosaics and natural herbaceous/cropland mosaic -cropland, natural herbaceous and herbaceous cropland -grassland, shrubland -shrubland. The ESA-CCI land cover products (Bontemps et al., 2013) are 300m resolution yearly products ranging from 1992 to 2015. The products were developed using the GlobCover unsupervised 220 classification chain and merging multiple available Earth observation products based on the GlobCover products of the ESA (Liu et al., 2018). Referring to the class relationships in (Liu et al., 2018), we cross-walked classes including cropland, forest, grassland, shrubland, barren land, and snow/ice. Apart from land cover products, we also compared GLASS-GLC with the Food and Agricultural Organization of the United Nations statistical data (FAOSTAT) on cropland and forest (forest land) classes, which are the main sources of 225 country-level land cover data for many applications. The annual FAOSTAT data set on cropland we used ranged from 1982 to 2015, and that on forest we used ranged from 1990 to 2015.
We made an inter-comparison between classes, including cropland, forest, grassland, shrubland, barren land and snow/ice.
The main inter-comparison is the area corresponding to the top 50 countries in each class. Besides, to compare the accuracy of different products, test samples from FLUXNET site data in 2015 are given for independent accuracy 230 assessment.

Statistical analysis of LCC
To extract the area of LCC, we estimated the trend of change through statistical analysis and avoided the influence of abnormal fluctuations from the obtained time series LC products. The annual area for each class on the scales of latitudinal zones, continents are summarized. A time series of the annual area for each class was generated. The boundary data of countries and 235 continents were obtained from the Bureau of Surveying and Mapping of China. Eco-region data were obtained from the FAO global eco-region dataset (Simons et al., 2001;FAO, 2018).
Although the inter-annual consistency has been ensured as much as possible in the above mapping framework, the effects of inter-annual changes due to climate conditions and phenological changes may still exist. To estimate the long-term trend of change, we fitted a linear trend (Theil-Sen estimator (Sen, 1968)) in area for each class, where the slope of annual change and 240 the 95 % confidence interval of the slope is given. In addition, a Mann-Kendall test (Mann, 1945) was used to test the trend of time series with a p-value given. If p < 0.05, it is considered that the trend of change is significant.
Further, we obtained the change mask where all pixels showed a significant change trend (Wang et al., 2016). First, we downscaled the grid from 0.05 ° to 0.25 °, and the time series of the area ratio of all classes in each 0.25 ° grid was summed.
Using the Mann-Kendall test, those grids showing a significant change (p < 0.05) were obtained. Then the slope of annual 245 change based on area ratio for each grid with an increasing or decreasing trend was found using a Theil-Sen estimator. The change ratios were then summarized at regional scales to estimate the corresponding significant areas of change from 1982 to 2015.
In order to quantify the magnitude of global LCC between 1982 and 2015 and reveal the global temporal LCC pattern, we calculated the ratio of annual global LCC to the global total terrestrial LC area by different time periods. To ensure the 250 quantified LCC to be non-accidental, we limited the computation area within the change mask in which all grids show a statistically significant loss or gain trend. We then summarized the annual LCC by 5-year and 10-year time intervals, respectively.
To further identify the direct causes of LCC, we assessed the LC conversion from 1982 to 2015. Based on the 0.05 ° LC mapping results for 1982 and 2015, a map of LC conversion can be obtained. The computation was also limited to the change 255 mask to ensure statistical significance. The conversion sources and destinations of LC classes were separately computed, to assess the direct causes of change in various LC classes.

Human impact process
To further explore the role of human impact in regions with significant LCC, the results are evaluated based on data from the

Accuracy assessment
First of all, to evaluate the magnitude of the errors introduced by our training samples, we randomly selected 500 samples from the training sample for manual interpretation and evaluation, and the assessment accuracy was 92.26 %. It shows that the 275 training sample we generate from 30m FROM-GLC_v2 is sufficient for our data production.
The global LC mapping result in 2015 is shown in Fig. 6. Its accuracy was tested with the H-homo sample in 2015 to obtain a confusion matrix (Table 3). The overall accuracy for the year 2015 reached 86.51 %. As for each class, the accuracies for forest, barren land, and tundra are relatively high, where the user's accuracies and producer's accuracies are over 90 %. The accuracy of cropland is also high, with the user's accuracy and producer's accuracy reaching 73.54 % and 78.62 %, respectively. The 280 user's accuracy of shrubland reached 83.62 %, while that of grassland is 67.58 %. Grassland is mainly mixed with cropland and shrubland. Table 4 shows the testing results of the FLUXNET test samples in which the number of sample units for shrubland, tundra, barren land, and snow/ice are relatively small. The overall accuracy of all classes is 82.10 % with the FLUXNET sample. Among them, the user's accuracy and the producer's accuracy for forest reach 91.01 % and 88.04 %, respectively. The producer's accuracy for cropland is 69.23 %, while its user's accuracy is 73.26 %. 285 Putting the test results from FROM-GLC_v2 and FLUXNET together, a spatial distribution map of the uncertainty of the 2015 LC mapping result was generated. As can be seen from Fig. 7, most of the world is shown in green color, which means that the mapping result for most regions is most likely to be correct, and the result for 2015 is highly credible. There are still some regions showing a yellow or orange color, and a smaller number of regions showing red, representing those regions that may have been misclassified. Since there are no test samples in Greenland., the interpolation results are ignored. In general, the 290 places with high uncertainty are Africa, East and South America, South Alaska, North and East Australia and Southwest Indonesia.
The assessment result with independent test samples is shown in Table 5 and Table 6. It shows that the overall accuracy of the GLASS-GLC without change detection is 81.28 %, and that with change detection is 82.81 %. This reflects the reliability of GLASS-GLC since the test samples are randomly distributed along the spatial and temporal dimensions, and also confirm the 295 significance and effectiveness of the change detection method.

Data inter-comparison
The assessment results of MLCT products and ESA-CCI land cover products based on test samples from FLUXNET site data are shown in Table 7 and Table 8, respectively. The overall accuracies of ESA-CCI products and MLCT products are 73.90 % and 80.38 % in 2015, respectively. Compared to these, The overall accuracy of GLASS-GLC (82.10 %, Table 4) is better. 300 Although the cross-walk of the different classification systems may be slightly different, It can still reflect the high accuracy of our GLASS-GLC products. Figure 5 shows an inter-comparison with MLCT products, Figure 6 with ESA-CCI products, and Figure 7 with FAOSTAT. The scatter plots and the linear fit lines reflected the results in 2015, and the box plots represent the distribution of R 2 of the annual linear fit lines for each class. It can be seen that various classes in several different products are relatively equivalent, although 305 they are under different classification systems. In comparison with MLCT products, the results of 2001-2015 for cropland, forest, and snow/ice have high R 2 . In comparison with ESA-CCI products, the mean R 2 of the linear fit lines of forest, grassland, and snow/ice from 1992 to 2015 reach 0.99, 0.82, and 0.98, respectively, while the R 2 for shrubland is low. The intercomparison of some other classes is poor, which may be caused by differences in the class definition in various classification systems. For instance, our classification system incorporates tundra, while the other two did not. Compared with FAOSTAT, 310 the mean R 2 of the linear fit lines of cropland and forest is 0.82, and 0.87, respectively. In general, our GLASS-GLC products have a reasonable consistency with other products and statistics, and the difference is not significant. What is more, the duration of GLASS-GLC, 34 years, is much longer than MLCT and ESA-CCI land cover products (as shown in Fig. 8). The comparison with other data illustrates the reliability and accurateness of GLASS-GLC.

Patterns along latitudinal gradients
The global distribution of 0.25 ° grids with significant LCC from 1982 to 2015 is shown in Fig. 11 and Fig. S3 for the whole world, where the color depth represents the estimated change in area ratio per year. The distribution of significant LCC along latitudes is shown in the right, where the red curve represents a significant increase, green a significant decrease, and blue a 330 net change.
The distribution pattern of LCC along latitudes is different, especially for cropland and forest, where it can be seen that cropland has increased significantly in the northern tropics and the southern hemisphere. It is confirmed that the significant increase in cropland has occurred mainly in the tropics and southern hemisphere (Gibbs et al., 2010). Forest has decreased significantly in the southern hemisphere and has increased significantly in the northern hemisphere, showing regional differences. In 335 particular, in the high latitudes of the north, forest has increased significantly with a decrease of tundra. However, the increase in forest area in the northern hemisphere is significantly larger than that in the southern hemisphere, reflecting an overall increase in total forest area.
The grassland area has reduced at almost all latitudes. There might exist an increasing trend in global vegetation coverage, where shrubland and forest expansion led to a reduction in the grassland area. It can be seen that shrubland has increased 340 significantly in the southern hemisphere, corresponding to the reduction in the grassland area there. The area of barren land is decreasing, especially in the middle and high latitudes of the north, which further reflects the increase in vegetation coverage.
The area of snow/ice in the northern high latitudes has reduced.

Continental patterns
The statistical results for each class at the continental scale are shown in Table 9, Table 10, Table 11, Table 12, Table S1, Table  345 S2, and Table S3, where the slope and p-values are estimated according to the class area time series. At the same time, gain and loss are the computed values from 0.25 ° grids with significant LCC.
Cropland significantly increased in South America, with a growth rate of 9.1×10 3 km 2 /year (p = 0.0108). The area of significantly increased cropland in Asia and Africa reached 67×10 3 km 2 and 23×10 3 km 2 , respectively. Corresponding to the increase in cropland, forest decreased significantly in South America, at a rate of 10.8×10 3 km 2 /year (p = 0.0242). Meanwhile, 350 the area of forest in Africa has significantly decreased by 29×10 3 km 2 . The area of forest in Asia has increased at the fastest speed. The area of forest in Europe and North America has also increased significantly. Meanwhile, the tundra area in Asia, Europe, and North America decreased significantly by 132×10 3 km 2 , 12×10 3 km 2 and 22×10 3 km 2 , respectively. Shrubland has increased significantly in Africa at a rate of 47.4×10 3 km 2 /year (p = 0.0030). However, as shown in Fig. 3, the LC mapping result in 2015 in Africa is of high uncertainty, the trend here should be treated carefully. Shrubland also increased significantly 355 in Oceania, by an area of 38×10 3 km 2 . The decrease of grassland in Asia is serious, and the area of grassland in Asia decreased significantly by 315×10 3 km 2 . Barren land in Asia also significantly decreased by 82×10 3 km 2 . The global snow/ice area has decreased significantly, at a speed of 19.2×10 3 km 2 /year (p = 0.0003).

Characteristics of LC class conversion
We attempted to find out some high-frequency LC class conversions for the period 1982 to 2015 (Table 13). Besides, the 360 conversion sources and destinations of each LC class are computed separately, as shown in Fig. 12.
Among land converted to cropland in 2015, grassland was the biggest source ( Fig. 12 (b)), accounting for 67.58 %. 6.61 % of cropland was converted from forest ( Fig. 12 (b)), showing the process of forest destruction. Among land converted to forest, the proportion of cropland reached 21.74 % (Fig. 12 (b)). Barren land and grassland were respectively the large sources of grassland and barren land (Fig. 12 (b)), reflecting the dynamic transformation between the two classes. Grassland accounted 365 for 35.00 % of the increasing source of barren land (Fig. 12 (b)), indicating the process of grassland loss (Bai et al., 2008).
The most frequent direction of conversion from cropland in 1982 was forest ( Fig. 12 (a)), which reached 78.22 %. At the same time, forest was also the main cause of loss of grassland and shrubland ( Fig. 12 (a)). The conversion of forest to grassland accounted for 59.04 % of all conversions from forest ( Fig. 12 (a)). The main conversion direction of tundra was forest, reaching 64.60 % ( Fig. 12 (a)). 370 Overall, the increase in forest accounted for the highest proportion of all conversion processes, reaching 44.17 % (Table 13).
The increase of grassland and cropland were second and third highest, reaching 19.79 % and 13.64 %, respectively (Table 13).
In addition, the proportions of grassland to shrubland and barren land to grassland were 7.73 % and 5.75 %, respectively (Table   13). Cropland gain and vegetation gain were the main phenomena reflected by the changes in global LC from 1982 to 2015.

Human impact evaluation 375
Figure 13 (a) shows different human impact (HI) levels among different LCC areas. Overall, the average HI level in regions with significant changes in all LC classes is 25.49 %, indicating that human activity has a great impact on LCC. The highest HI level was found in those regions with significant increases in cropland, reaching an average value of 51.38 %. Meanwhile, the HI level of cropland loss reached 48.02 %, while the HI level for forest loss was 26.91 %. In addition, in any change of vegetation, such as forest, grassland and shrubland, the HI level in regions of vegetation loss is higher than that of gain. 380 The HI levels along continents can be found in Fig. 13 (b). The highest level of HI is found in Europe and the lowest in Oceania.
The HI in Europe reached 46.86 %, indicating that human activity played a relatively important role in regions with significant LCC. Asia came second, with a HI level of 32.07 %. In South America and Oceania in the southern hemisphere, the overall HI level in the LCC regions is small.
As shown in Fig. 13 (c), the polar regions and the boreal conifer forest regions at high northern latitudes with significant LCC 385 have lower HI levels. The level of HI in subtropical regions is high, among which HI levels in subtropical steppe and subtropical humid forest regions reached 38.23 % and 43.90 %, indicating that the role of LC conversion caused by human activity in subtropical climate areas is significant. In addition, in the temperate steppe regions, the HI level in the regions of significant LCC is also high, reaching 39.87 %. In the tropics, the average HI level in dry forest regions is highest among

Local hotspots of LCC
Regarding LC, more attention tends to be paid to global and regional LCC. At the local scale, we can further explore the hot spots of LCC. The main regions of LCC hotspots are shown in Fig. 14, where the depth of color represents a significant change.
In the north of Eurasia, forest has increased significantly ( Fig. 14 (a)), and that in Siberia has moved northward to the tundra regions. In northern North America, such as Alaska and the north of Canada, forest has also increased but the extent of the 395 increase is weaker than that in North Eurasia. In the Great Plains of Central North America, grassland has decreased and cropland has increased (Fig. 14 (b)). In most countries of South America, croplands have expanded substantially (Fig. 14 (d)) and forests have decreased significantly (Fig. 14 (c)), especially in the southeastern part of the Amazon rainforest. In Southeast Asia, such as Cambodia, Vietnam, Indonesia, and Malaysia, forest has also decreased significantly, and cropland has increased.
While our LCC analysis shows these trends in the Asian tropics, higher resolution data and more specific land cover mapping 400 are needed to explicitly determine the reasons for LCC in this region (Cheng et al., 2018). In Africa, forest in the northern part of the Congo Basin has expanded while forest in the southern Miombo forest belt has decreased (Fig. 14 (e)). In China, forest has increased (Fig. 14 (f)). Some grassland in Mongolia and Inner Mongolia of China showed a trend of decrease ( Fig. 14 (g)).
There is an obvious increase in grassland areas in the eastern part of the Qinghai-Tibet Plateau (Fig. 14 (h)) and a decrease of grassland in central Asia and parts of Western Asia (Fig. 14 (i)).In some parts of the former Soviet Union in Eastern Europe, a 405 decrease of cropland (Fig. 14 (j)) and an increase of forest can be observed.

Discussions
Based on the accuracy assessment and data inter-comparison results, it can be seen that the global LC mapping products of For example, the reduction of much cropland is due to urbanization, and the expansion of cities is usually sporadic. Although 415 those changes are large at the global scale, they can hardly be reflected with 0.05 ° pixels. Moreover, due to the synthesis principle, the classification result of each pixel can only represent the class with the largest proportion in area, and the information of remaining classes is ignored even though they can sometimes be more than 50 % in total. Such a neglect, due to the famous "Scale Effect" (Turner et al., 1989) can also cause great deviations in the final statistical summary of the LC Second, our sampling strategy for training has certain limitations. On the one hand, since the training sample is generated from 30m resolution maps of more than 73 % accuracy, this will inevitably propagate and accumulate error to 5 km resolution. Of course, due to the higher signal-to-noise ratio of the high-resolution data, the sampling is still satisfactory compared to the direct visual interpretation of the coarse resolution images. On the other hand, the training sample used is only from a single year of circa 2015. Although we have implemented a time series correction for the original input data and performed a time-425 consistent post processing on the classification results, the effects of inter-annual fluctuations in data cannot be completely avoided (Song et al., 2018a). On the other hand, according to the stable classification with limited sample theory (Gong et al., 2019), a representative sample collected in one year with less than 20 % in error should suffice in multiannual use at the global scale. Therefore, a multi-year sample set may not be as critical for multiannual classification provided the sample is better than 80 % accurate. In our case, although the source training data has an accuracy of 73.17 %, we are not certain if the aggregated 430 sample set exceeds an accuracy of 80 %. While this needs further assessment, the expected loss of accuracy should be within a couple of percents (Cheng et al., 2018). According to the 92.26 % test accuracy reported in 3.1.1, the aggregated sample set can be satisfactory.
For the generation of test sample units in 2015, we have adopted the scale-up approach. That is to say, we first upscaled the 30m test sample set to 5 km by maximum area synthesis, which contains unavoidable errors because of scale transformation. 435 Due to the difficulty of visual interpretation in coarse scale and field investigation (Gong et al., 2013), establishing a sample library at 5 km resolution is not easy. Thus, instead, we adopted the data aggregation method based on the 30m FROM-GLC_v2 results. Since mixed pixel problems for remotely sensed data are unavoidable at any scale, choosing one category for mixed pixels is inevitable and the cost of simplification in a traditional classification process. The development of LC ratio mapping products (similar to VCF products), rather than hard classification, especially for the case of coarse resolution, should be 440 considered and further assessed. However, the independent interpreted 5 km test sample set alleviates the problem.
We have eliminated wetland and impervious surface in our classification system. This is a tradeoff when working at the 5 km scale. Patches of wetland and impervious surface are usually small, and it is difficult to achieve a pixel size of 0.05 ° for many situations, so the classification of the two types is extremely difficult. However, both are important LC types. Wetland is a transitional zone between terrestrial ecosystems and aquatic ecosystems (Davidson, 2014). The impervious surface can 445 represent the urban area. In recent years, urban expansion has been a relatively significant phenomenon in global environmental change (Seto et al., 2011). Urban expansion reflects an important type of human activity, so the impervious surface is also one of the essential components to reflect anthropogenic influence. However, the total area of its change is usually small.
It should be pointed out that at a coarse resolution of 0.05 °, our definition of forest is more inclined to the tree canopy cover.
Thus the changes in internal density of trees can also be reflected in the area change of forest instead of just the stand-unavoidable to include internal density change of various classes, which in turn will further affect the classification and change area calculation of forest class.
In addition, because we are mainly depicting the natural biophysical properties of vegetated areas with limitation in resolution, some artificial characteristics cannot be distinguished, such as plantations (rubber, oil palm, and various fruit trees) and natural 455 forest, which are uniformly included as forest in our classification system.
In the statistical analysis, although we have already conducted post-classification time-consistency processing for the original LC mapping results as much as possible, it is inevitable that there are still large fluctuations and interferences from various unknown factors unfavorable to the extraction of long-term trend of LCC. In order to ensure that the trend of the resulting time series is significant, we have to scale up the classification result from 0.05 ° to 0.25 °, converting the original class label of 460 each 0.05 ° pixel to the class area ratio of 0.25 ° grid. The long-term time series of the area ratios are tested for statistical significance. However, in some cases, this procedure will also be influenced by the "Scale Effect".
In the analysis of anthropogenic influences, indirect effects of many human activities were ignored because the main objective was to include the effects of directly visible human activities. For example, human activities increase the concentration of carbon dioxide in the atmosphere, which in turn affects the global climate, leading to higher temperature, and thus increasing 465 vegetation coverage (Piao et al., 2006;Bonan, 2008). This pathway of action is indirect, but it is difficult to reflect in the human impact data we use, which results in an underestimation of the assessment of anthropogenic influences.
GLASS-GLCs contain more detailed LC classes, longer temporal coverage (34 years), high consistency, which meets the requirement for CDR. GLASS-GLC CDRs are the first collection of global LC dynamics of 5 km, and fill the existing gap for high-reliability and consistency of long-term general purpose global LC products. In addition, our strategy of generating 470 samples from high-resolution classification products can greatly reduce the cost and investment of sample collection. It can flexibly and effectively be extended to other coarse-resolution LC mapping tasks in the future.
In the future, with the advancement of technology and the accumulation of remote sensing datasets, the use of remote sensing products for LC mapping with higher resolution and longer time series will undoubtedly better reflect the global LC and its changes. However, under limited conditions, we can consider using coarse-resolution satellite data to determine the locations 475 of potential rapid change, and then use high-resolution data in these hotspots to accurately estimate the rate and mode of change.
Moreover, it is necessary to establish a multi-year sample library to assess the impact of inter-annual fluctuations in input data on the accuracy of change characterization and analysis. Wetland and impervious surface are LC classes that have extremely high value. It would be useful to supplement the mapping and change analysis of these two classes when suitable data become available. For the analysis of global LCC, systematic and in-depth attribution analysis and research can be further carried out.
GLASS-GLC products at 5 km resolution from 1982 to 2015 are available to the public in the GeoTIFF format at https://doi.pangaea. de/10.1594/PANGAEA.913496 (Liu et al., 2020

Conclusions
In order to better reflect the global land changes, continuous and dynamic monitoring of global LC is necessary. We built GLASS-GLC, the first CDRs for global LC on the GEE platform. It can capture the global LCC information from 1982 to 2015. Compared to previous global LC products, GLASS-GLC products cover a longer time period and have higher consistency and more detailed classes. Our entire mapping framework is based on FROM-GLC_v2, including the classification 495 system and high-quality H-homo sample generation.

Class
Subclass with reference to (Li et al., 2017)    evapotranspiration, gross primary production, broadband emissivity, respectively. ABD_WSA_VIS, ABD_BSA_NIR, and ABD_WSA_shortwave represent white-sky albedo in visible band, near infrared band, and shortwave band, respectively. TC, SV, and BG stand for tree canopy, short vegetation, and bare ground cover, respectively.