Abstract
Land change is a cause and consequence of global environmental change1,2. Land-cover and land-use change significantly alter the Earth’s energy balance and biogeochemical cycles, contributing to climate change, which in turn affects land surface properties and the provision of ecosystem services1–4. Yet quantification of global land change is lacking. Here, we analyze 35-years of satellite data and provide the first comprehensive record of global land change dynamics during 1982–2016. We show that contrary to the prevailing view that forest area has declined globally5, tree cover has increased by 2.24 million km2 (+7.1% relative to 1982 level). This overall net gain is a result of net loss in the tropics outweighed by net gain in the subtropical, temperate and boreal climate zones. Global bare ground cover has decreased by 1.16 million km2 (−3.1%), most notably in agricultural regions in Asia. Of all land changes, 60% are associated with direct human land-use activities and 40% with indirect drivers such as climate change. Land-use change exhibits strong regional dominance, including tropical deforestation and agricultural expansion, temperate reforestation/afforestation, cropland intensification, and urbanization. Consistent across all climate domains, global montane systems have gained tree cover over the past 35 years, whereas many arid and semi-arid ecosystems have lost vegetation cover. The global land change quantified in our study and the driver attribution results reflect a human-dominated Earth system. The freely available dataset may be used to improve the modeling of land-use change, biogeochemical cycles and vegetation-climate interactions to further advance our understanding of global change1–4,6.
Humanity depends on land for food, energy, living space and development. Land-use change, a traditionally local-scale human practice, is increasingly affecting Earth system processes, including the surface energy balance, the carbon cycle, the water cycle, and species diversity1–4. Land-use change is estimated to contribute a quarter of cumulative carbon emissions to the atmosphere since industrialization3. As population and per capita consumption continue to grow, so does demand for food, natural resources and consequent stress to ecosystems. Recent research suggests that human-induced perturbations to the Earth system, especially the climate system, have exceeded natural variability and that we have entered a new geologic epoch referred to as the Anthropocene7.
Because of their synoptic view and recurrent monitoring of the Earth’s surface, satellite observations contribute substantially to our current understanding of the global extent and change of land cover and land use. Previous global-scale studies were mainly focused on annual forest cover change (stand-replacement disturbance) for the time period after 20008 or at sparse temporal intervals9,10. Long-term gradual changes in undisturbed forests as well as areal changes in cropland, grassland and other non-forested land are less well quantified.
For the time period 1982 to 2016, we create an annual global vegetation continuous fields product11, consisting of tall vegetation (≥ 5m in height) hereafter referred to as tree canopy (TC) cover, short vegetation (SV) cover and bare ground (BG) cover, at 0.05° × 0.05° spatial resolution (see details of definitions in Methods). For each year, every land pixel is characterized as percent of TC, SV and BG cover, representing the vegetation composition at the time of the local peak growing season. The dataset is produced by combining optical observations from multiple satellite sensors, including the Advanced Very High Resolution Radiometer, the Moderate Resolution Imaging Spectroradiometer, the Landsat Enhanced Thematic Mapper Plus and various very high spatial resolution sensors. We employ non-parametric trend analysis to detect and quantify changes in tree canopy, short vegetation and bare ground over the full time period at pixel (0.05° × 0.05°), regional and global scales. Observed changes are attributed to direct human activities or indirect drivers based on a global probability sample and interpretation of high resolution images from Google Earth.
Total area of tree cover increased by 2.24 million km2 from 1982 to 2016 (90% confidence interval (CI) [0.93, 3.42] million km2) representing a +7.1% change relative to 1982 tree cover (Extended Data Table 1). Bare ground area decreased by 1.16 million km2 (90% CI [−1.78, −0.34] million km2) representing a decrease of 3.1% relative to 1982 bare ground cover. Area of short vegetation cover decreased by 0.88 million km2 (90% CI [−2.20, 0.52] million km2) indicating a decrease of 1.4% relative to 1982 short vegetation cover. A global net gain in tree canopy contradicts current understanding of long-term forest area change; the Food and Agriculture Organization of the United Nations (FAO) reported a net forest loss between 1990 and 2015. However, our gross tree canopy loss estimate (−1.33 million km2, −4.2%, Extended Data Table 1) agrees, in magnitude, with FAO’s net forest area change estimate (−1.29 million km2, −3%), despite differences in the time period covered and definition of forest (FAO defines forest as tree cover ≥ 10%; see details in Methods).
The mapped land change (Fig. 1) consists of all land-cover and land-use changes induced by natural or anthropogenic drivers. Land-cover and land-use change themes are also inherently linked in the tree cover – short vegetation – bare ground nexus. For example, deforestation for agricultural expansion is often manifested as tree canopy loss and short vegetation gain, whereas land degradation may simultaneously result in short vegetation loss and bare ground gain. Pairs of ΔTC, ΔSV and ΔBG show strong coupling and symmetry in change direction but vary substantially over space (Fig. 1b and Extended Data Fig. 1). That is, the globally dominant, coupled land changes are ΔTC co-located with ΔSV and ΔSV co-located with ΔBG.
The overall net gain in tree canopy is a result of net loss in the tropics outweighed by net gain in the subtropical, temperate and boreal climate zones (Extended Data Table 2). A latitudinal north (gain)-south (loss) contrast in tree cover change is evident (Fig. 2a). Conversely, for short vegetation, tropical net gain is exceeded by extratropical net loss. The latitudinal profile of ΔSV largely mirrors that of ΔTC, most obvious in the northern mid-to-high latitudes (45°N-75°N) and low latitudes (30°S-10°N) (Fig. 2b). For bare ground, subtropical net gain partially offsets losses in all other climate domains. In the northern low-to-mid latitudes (10°N-45°N), the profile of bare ground loss (Fig. 2c) closely corresponds to that of short vegetation gain (Fig. 2b).
Changes were unevenly distributed across biomes (Fig. 3, Extended Data Fig. 2 and Extended Data Table 2). The largest area of net tree canopy loss occurred in the tropical dry forest biome (−95,000 km2, −8%) (Extended Data Fig. 2a), closely followed by tropical moist deciduous forest (−84,000 km2, −2%) (Fig. 3c) [all percent net changes are expressed relative to the benchmark of the area of the cover class in 1982]. Tree canopy in major forest biomes outside the tropics increased over the past 35 years, with temperate continental forest experiencing the largest gain (+726,000 km2, +33%) (Fig. 3d), comparable to the next two biomes combined: boreal coniferous forest (+463,000 km2, +12%) (Extended Data Fig. 2m) and subtropical humid forest (+280,000 km2, +18%) (Extended Data Fig. 2e).
Short vegetation loss mirrored tree cover gain dynamics, but with smaller magnitudes: temperate continental forest (−610,000 km2, −14%), boreal coniferous forest (−430,000 km2, −10%) and subtropical humid forest (−249,000 km2, −9%). In contrast, tropical forest biomes all gained short vegetation, with tropical shrubland experiencing the largest areal increase (+417,000 km2, +10%) (Fig. 3e), twice the amount of short vegetation gain in tropical dry forest (+246,000 km2, +5%). Tropical shrubland also experienced the largest bare ground loss (−408,000 km2, −10%).
Subtropical desert, the second largest dryland biome on Earth, had the largest gain in bare ground (+154,000 km2, +4%) (Fig. 3f), followed by subtropical steppe (+107,000 km2, +5%) (Extended Data Fig. 2h).
Consistent across all climate domains, mountain systems experienced net bare ground loss, net short vegetation loss, and net tree canopy gain (Extended Data Fig. 2c, f, i, n and Extended Data Table 2). In the high-latitude boreal tundra woodland and the polar ecozone (Extended Data Fig. 2o, p), bare ground decreased and tree canopy increased in both biomes, while short vegetation decreased in tundra woodland, but increased in the polar ecozone.
Based on the data from the global probability sample, an estimated 60% of all changes were associated with direct human land-use activities and 40% with indirect drivers such as climate change (Extended Data Figs. 3 and 4; see Methods). Direct human impact varied from 36% for bare ground gain to 70% for tree canopy loss. At the continental scale, land-use activities account for the majority of observed land changes in Europe (86%), South America (66%), Asia (62%) and Africa (50%), but play a smaller role in North America (47%) and Oceania (35%). The specific land change drivers are diverse, multi-scale and interactive1,13 (see below for detailed discussion). However, changes collectively induced by the various drivers at the global scale appear gradual over time (Extended Data Fig. 5).
Expansion of the agricultural frontier is the primary driver of deforestation in the tropics14. The three countries with the largest area of net tree cover loss during 1982–2016 are all located in South America: Brazil (−385,000 km2, −8%), Argentina (−113,000 km2, −25%) and Paraguay (−79,000 km2, −34%) (Supplementary Information Table 1). The “arc of deforestation” along the southeastern edge of the Amazon has been well documented8,9,14,15. Clearing of natural vegetation for export-oriented industrial agriculture also prevailed in the Cerrado (Fig. 4a) and the Gran Chaco (Fig. 4b). Spatially clustered hotspots of deforestation are also found in Queensland, Australia and Southeast Asia, including Myanmar, Vietnam, Cambodia and Indonesia, diminishing the already scarce primary forests of the region16. In sub-Saharan Africa, tree cover loss was pervasive across the Congolian rainforests and the Miombo woodlands (Fig. 4c), historically related to small-holder agriculture, and increasingly for commodity crop cultivation17. Forests in boreal Canada, eastern Alaska and central Siberia exhibited large patches of tree canopy loss and short vegetation gain, similar to the tropics (Fig. 1b). However, these are the result of persistent disturbances from wildfires and subsequent recovery of natural vegetation18.
Discernible impacts of climate change on vegetation change are also revealed at regional scales. In the Western United States (Fig. 4d), forests are suffering from increasing stress from insects, wildfires, heat and droughts due to regional warming19. But in the temperature-limited Arctic, warming is facilitating woody vegetation growth in northeastern Siberia, western Alaska and northern Quebec20 (Fig. 4e). Land-use activities are rare in these boreal tundra and polar ecosystems, contributing less than 1% to observed land changes (Extended Data Fig. 3e). In water-limited savannas in Central and West Africa (Fig. 4f), forest expansion and woody encroachment, observed both from space and in the field21, are likely driven by increases in precipitation and atmospheric CO2 22. Extreme high rainfall anomalies also contributed to the greening of the Sahel22 (Fig. 4f). Altitudinal biome shift is also expected in a changing climate. Global treeline positions have been advancing since 1900 AD as a result of climate warming23. The aforementioned bare ground loss, short vegetation loss and tree canopy gain in global mountain systems further suggest that an enduring transformation is occurring with regard to the distribution, structure and composition of montane vegetation.
Political, social and economic factors can influence vegetation in conjunction with climate drivers. Tree canopy in Europe, including European Russia, has increased by 35%—the greatest gain among all continents (Extended Data Table 1). Spatially contiguous hotspots of tree canopy gain were found in European Russia and Carpathian montane forests (Fig. 4g). Natural afforestation on abandoned agricultural land is a common process in Eastern Europe after the collapse of the Soviet Union24. Our satellite record confirms the effectiveness of China’s large-scale reforestation and afforestation programs, particularly in the Loess Plateau and the Qin Ling–Daba Mountains25 (Fig. 4h). Increasing area of plantations in southeastern China has also led to China’s tree canopy gain (+34%). Tree canopy also increased in the United States (+15%), mostly in the eastern U.S (Fig. 1b). Unlike declining forest cover in the western U.S. (Fig. 4d), southeastern forests are recovering from historical disturbances or are under intensive forestry management26.
The world’s arid and semi-arid drylands exhibited large areas of decrease in short vegetation and increase in bare ground, indicating long-term land degradation. Hotspots of vegetation loss include southwest U.S., southern Argentina, Kazakhstan, Mongolia (Fig. 4i), Inner Mongolia, China, Afghanistan (Fig. 4j) and large areas of Australia. The decrease of short vegetation cover in eastern Australia is likely the consequence of long-term precipitation decline in the local growing season27. Rising surface temperatures, reduction in rainfall, and overgrazing caused extensive grassland deterioration in the Mongolian steppe28. A U.S. nationwide ground survey revealed degradation of soils and vegetation along with increased dominance of invasive species in the southwest29.
Human activities undoubtedly play the dominant role in agricultural and urban landscapes, where lands have been continually modified through human history. India and China had the largest bare ground loss among all countries (India: −270,000 km2, −34%; China: −250,000 km2, −7%). India also ranked second in short vegetation gain (+195,000 km2, +9%), after Brazil (+396,000 km2, +12%). While the short vegetation gain in Brazil is mainly due to the expansion of agricultural frontiers into natural ecosystems, short vegetation gain in India is primarily due to intensification of existing agricultural lands—a continuation of the “Green Revolution”30. Some of the observed bare ground gain can be attributed to resource extraction and urban sprawl, most notably in eastern China (Fig. 4h). However, at the global scale, the growth of urban areas accounts for a small fraction of all land change31.
Previous studies have found a greening Earth based on trends in satellite-based vegetation properties (e.g. leaf area index or LAI) and have linked the greening trend to a number of climatic and ecological factors25,32–35. Recently, using ecosystem models, Zhu et al. (2016) attributed 70% of observed global LAI increase to the CO2 fertilization effect and 4% to land-use change. Our finding that global bare ground cover has decreased over the past 35 years suggests a net increase in vegetation cover and thus agrees with the greening trend. However, our results differ from previous studies by quantifying the prominent role of land use in global vegetation change. Using a global probability-based sample, we attribute 60% of observed land changes to land-use activities (Extended Data Fig. 3). Our empirical approach is based on observations of high-resolution satellite data (Extended Data Fig. 4), avoiding the challenges of modeling the underlying drivers of land change1,13. Additionally, our TC-SV-BG land-cover product is thematically more advanced than vegetation indices in characterizing land surface change. For example, differentiating long-term changes in tree cover from other vegetation can facilitate improved understanding of global fluxes of water, carbon and energy11. Our study provides observational evidence of increasing tree cover in northern continents, which may constitute the missing carbon sink3,36. In contrast, tropical tree cover loss is associated with higher biomass forests and responsible for carbon emissions from deforestation3,15.
Results of this study reflect a human-dominated Earth system. Direct human action on landscapes is found over large areas on every continent, from intensification and extensification of agriculture to increases in forestry and urban land uses, with implications for the maintenance of ecosystem services2. However, human-induced climate change has been documented as an indirect cause of many of the quantified large-scale regional change dynamics, including woody encroachment in Arctic and montane systems and vegetation loss in semi-arid ecoregions. Continuing land-use change and the increasing role of climate change in modifying land cover warrants continued monitoring of the Earth’s land surface from space.
Methods
Definitions
Vegetation continuous fields (VCF) represent land surface as a fractional combination of vegetation functional types that can be remotely sensed from satellites11. Consistent with previous research37–41, the VCF product developed in this study consists of percentages of tree canopy (TC) cover, short vegetation (SV) cover and bare ground (BG) cover. Trees are defined as all vegetation taller than 5 meters in height. TC refers to the proportion of the ground covered by the vertical projection of tree crowns42,43. SV characterizes the proportion of the ground covered by vegetation other than trees, including shrubs, herbaceous vegetation, and mosses, while BG represents the proportion of the land surface not covered by vegetation. TC, SV and BG are quantified from nadir view at top of canopy and are mapped during the local annual peak of a growing season31,41,44. TC is not equivalent to forest cover, although forest cover may be defined based on TC. For example, the FAO defines forest as a parcel or unit of land of at least 0.5 hectares in size which is covered by 10% or more trees that are 5 meters or taller5. Gain or loss in TC, SV, or BG refers to net increase or decrease in each respective cover over the study period due to any anthropogenic or natural factors, excluding temporary changes attributable to within-year vegetation phenology or year-to-year rotations.
Generation of AVHRR VCF
The Advanced Very High Resolution Radiometer (AVHRR) instruments on-board NOAA satellites remain an important data source for studying long-term changes in land surface properties as they provide the longest time-series of global satellite measurements45–47. We used the version 4 Long Term Data Record (LTDR) to generate the annual VCF products47,48. The LTDR was compiled from AVHRR observations through a series of processing steps including radiometric calibration, geolocation correction, atmospheric correction and bi-directional reflectance effect correction47. The daily LTDR surface reflectance data contain 5 multi-spectral layers of AVHRR channels 1–5 and the normalized difference vegetation index (NDVI) layer computed from channels 1 and 249. Each pixel is 0.05° × 0.05° in size. We implemented an improved version of the operational Moderate Resolution Imaging Spectroradiometer Vegetation Continuous Field (MODIS VCF) approach to convert daily LTDR to yearly VCF38 (Supplementary Information Fig. 1a).
Daily AVHRR was first aggregated into monthly composites based on the maximum NDVI value in the month. Maximum NDVI composition can minimize cloud contamination, reduce bi-directional and off-nadir viewing effects, minimize band-correlated atmospheric effects and enhance vegetation discrimination50. The technique has been widely adopted for generating NDVI and land-cover products from daily satellite data for sensors such as AVHRR, MODIS and VEGETATION46,51–54.
Monthly composites were subsequently converted to annual phenological metrics8,38,55–57 (Supplementary Information Fig. 1b). Metrics are statistical transformations of pixel time-series that can capture the salient features of vegetation phenology while maintaining high spatial and temporal data consistency. Metrics thus provide a unique advantage to large-area land cover mapping and monitoring. We created a total of 735 annual metrics from a combination of 5 multi-spectral bands and one NDVI layer, each available as time-series of 12 months.
An empirical normalization procedure was applied to enhance the year-to-year consistency of the AVHRR metrics (Extended Data Fig. 6). Time-series data from AVHRR are known to have systematic discrepancies due to different satellite platforms, orbital drift, changes in sensor design and sensor degradation45,46,58. The systematic differences are particularly pronounced before and after year 2000; beginning with NOAA-16 in 2000, satellite orbits were stabilized and a major improvement was introduced in the sensor design to increase sensitivity at the low end of radiance45. Research has also shown that the varying observational solar zenith angle as a result of orbital drift affects reflectance more than NDVI and is negatively related to leaf area or positively related to soil exposure59. That is, dense vegetation is less affected than sparse vegetation. Additionally, remaining atmospheric effects in the AVHRR surface reflectance can also cause inconsistency between years. The normalization was designed to remove these artifacts unrelated to actual surface change.
A rich literature exists on calibration of AVHRR time series. One commonly used method is to apply calibration coefficients estimated from “stable targets” such as deserts, oceans, clouds or rainforests60–65. For example, earlier works by Myneni et al.32,63 used the Sahara desert as reference to adjust global NDVI. Gutman (1999)64 used global deserts and rainforests to correct reflectances as well as NDVI. Recently, data from well-calibrated sensors such as MODIS and SPOT were used as reference for anchoring AVHRR-based NDVI time series45,46.
To normalize annual metrics, we designed a two-step approach, using MODIS data as reference. The first step was to apply a dark object subtraction (DOS) to remove systematic biases for vegetated surfaces, especially forest. DOS is also a simple and effective method of removing atmospheric contamination in remotely sensed data66–70. We used the intact forest landscapes (IFL)71 of the tropical rainforest biome (i.e. the minimally disturbed tropical rainforests, average tree cover 97%; Extended Data Fig. 6c) as the dark stable target, which was also considered a spectral end-member. The second step was to apply a slope-based adjustment for pixels that contained visible bare ground. This step involved the use of tropical, subtropical and temperate deserts with 100% Landsat-based bare ground cover28 (Extended Data Fig. 6c) as the bright stable target, or the other spectral end-member. Biases over other land surfaces are assumed to be within these two extreme end members64. To create the MODIS reference data, an identical procedure (Supplementary Information Fig. 1a) was applied to daily MODIS LTDR44 to derive annual metrics for years 2000 through 2016. The 17-year median values for each metric were subsequently derived and used as reference.
DOS was conducted by applying the following equations:
(1) |
(2) |
where, xm,t,i is the original AVHRR value of metric m in year t and pixel i, ym,t,i is the DOS-adjusted AVHRR value, is the mean bias of metric m over a total of nIFL IFL pixels indexed by j, rm,j is the MODIS reference value of metric m in IFL pixel j.
The soil-induced bias was then corrected relative to the desert end-member, which has maximum residual bias after DOS correction, as well as the IFL end-member, which has minimum residual bias. Dense vegetation is largely immune to this correction. The correction is summarized by the following equations:
(3) |
(4) |
(5) |
(6) |
where, zm,t,i is the slope-adjusted AVHRR value of metric m in year t and pixel i, ym,t,i, is the DOS-adjusted value from equation (1), is the mean bias of metric m over a total of nDES desert (DES) pixels indexed by k, 𝑣t,i is the peak growing season NDVI value of pixel i in year t, is the mean peak growing season NDVI value of all IFL pixels, is the mean peak growing season NDVI value of all desert pixels, and rm,k is the MODIS reference value of metric m in desert pixel k. Here we use peak growing season NDVI, which is one of the metrics and computed as the mean of all NDVI values between 75 and 100 percentiles, in the slope term instead of the annual mean NDVI as used in Gutman (1999)64, because our annual VCF represents the vegetation state of the local peak growing season. Using this annual metric (before any correction) dynamically optimizes AVHRR data for the growing season of each year.
Adjusted annual metrics were used as input to supervised regression tree models to generate the annual TC and BG product. This non-parametric machine learning method was chosen as it can accommodate nonlinear relationships between the dependent variable (percent TC or percent BG) and independent variables (AVHRR metrics); in addition, the decision rules are easily interpretable72–74. Training data for TC were obtained by spatially aggregating the circa-2000 Landsat-based percent TC product from 0.00025° × 0.00025° to 0.05° × 0.05°, which was in turn trained using very-high spatial resolution images8. For each 0.05° × 0.05° grid cell, we computed the average value of all Landsat TC pixels that fall in the grid cell and derived the percentage of TC per grid cell. Likewise, training data for BG were obtained by spatially aggregating the circa-2000 Landsat-based percent BG product31. Model training and prediction were performed separately for TC and BG. We pooled two years of AVHRR metrics before and after 2000 (i.e. 1999 and 2001) as input features to train 21 bagged regression tree models to account for the remaining inter-annual bias of AVHRR metrics, if any, as well as to avoid over-fitting of the regression tree algorithm. The 21 trained models were applied to annual AVHRR metrics to generate percent TC and BG for each year. Due to missing data in years 1994 and 2000, TC and BG maps in these two years were not produced from AVHRR, but were linearly interpolated using antecedent and subsequent annual TC or BG estimates on a per pixel basis. Following the MODIS VCF approach38, annual SV was derived as the residual term by subtracting TC and BG percentages from 100. Permanent water surfaces were excluded based on the Landsat-derived permanent surface water product8.
Accuracy assessment
Validating a global land-cover product spanning multiple decades is a challenge. The primary obstacle is the lack of sufficient ground observations that match the spatial extent, the temporal frequency and the thematic content of a satellite-derived product. Satellite observations with higher spatial and temporal resolutions can characterize land cover and change with higher accuracy75. Thus, higher-resolution satellite or aerial imagery is often employed to replace ground observations when determining the reference condition for validation76. Here we leverage the established validation protocols77,78 and the best available reference datasets to evaluate the accuracy of our VCF product. Specifically, we used a sub-meter resolution, global land-cover validation sample developed by the United States Geological Survey (USGS)79 as the primary reference for TC. We also used the 30-m resolution Landsat-based TC, SV and BG estimates as reference to evaluate the AVHRR-derived TC, SV and BG layers.
The USGS reference dataset is a stratified random sample of TC estimates produced from n = 475 sample blocks distributed across the globe77–79 (Extended Data Fig. 7a). Each sample block was 5-km × 5-km (~0.05° × 0.05°) in size. Sub-meter resolution commercial images including QuickBird, WorldView, IKONOS and GeoEye between years 2002 and 2014, depending on each block, were classified to categorical land cover classes including tree cover79. The percent TC for each block was computed from these data to provide the reference values for comparison to the AVHRR percent TC. The USGS reference data were developed in Universal Transverse Mercator (UTM) projection and the footprints of the 5-km × 5-km reference sample blocks did not exactly overlap with AVHRR pixels, which were in Geographical Latitude / Longitude projection (Extended Data Fig. 7b-c). This geolocation mismatch inevitably introduced some error in the validation results. Thus, we also evaluated AVHRR TC using the Landsat-based TC estimates. Because the spatial units of the Landsat estimates were spatially aligned with the AVHRR pixels, this comparison is free from geolocation error. For BG and SV, due to the lack of reliable high-resolution reference data, we used Landsat-based BG and SV (computed as 100% – Landsat-based BG% – Landsat-based TC%) estimates at the USGS sample locations as reference data for estimating accuracy. These BG and SV reference data were obtained for the same stratified sample of blocks used to evaluate the AVHRR TC product77,78.
The paired AVHRR and reference VCF values were used to calculate four accuracy metrics including root-mean-square-error (RMSE), mean absolute error (MAE), mean error (ME) and r2 78,80:
(7) |
(8) |
(9) |
(10) |
where pi, ri and wi are estimated VCF, reference VCF and sample weight (inverse of inclusion probability of the sample block for the stratified design) at a location i in a sample of size n; is the estimated mean of the reference values.
We also computed the conventional confusion matrices including overall accuracy (OA), user’s accuracy (UA) and producer’s accuracy (PA) using the paired AVHRR and reference VCF values and a general ratio estimator78,81:
(11) |
where, H is the total number of strata; Nh is the total number of 5-km × 5-km blocks within stratum and are the sample means of variables y and x in stratum h and the specific identity of y and x depends on the accuracy metric being estimated. To estimate OA, y = area of agreement between AVHRR and reference for a VCF class c in each sample block (i.e., overlapped area) and x = area of the sample block. To estimate UA, y = area of agreement between AVHRR and reference for a VCF class c and x = area of class c mapped by AVHRR. To estimate PA, y = area of agreement between AVHRR and reference for a VCF class c and x = area of class c given by reference.
The estimated variance of is:
(12) |
where is the number of sample blocks selected from stratum h, and are the sample variances of y and x for stratum h and sxyh is the sample covariance of x and y for stratum h. The standard error of is the square root of the estimated variance. As noted above, the identity of x and y depends on the accuracy metric being estimated. A summary of accuracy results for TC, SV and BG is provided in Extended Data Fig. 7.
Trend analysis
Per-pixel TC, SV and BG percentages were aggregated to a series of spatial scales including global, continental, climate zone, biome and country scales to obtain annual total area estimates at these aggregated scales. For example, for the trend analysis of Africa, the per-pixel values of each cover type were aggregated to produce a single value for each year in the time series. We used the FAO ecological zones boundary shapefile to report VCF area estimates per biome and per climate zone12. We also used the Global Administrative Areas (GADM) country boundary shapefile (http://www.gadm.org) to report VCF area estimates per country.
The approach to change analysis was predicated on using a linear trend (Theil-Sen estimator) to smooth the annual time series of data when determining net change82. Although the classification methodology (monthly compositing, annual metrics calculation, inter-annual bias adjustment and multi-year model training) was constructed to ensure year-to-year consistency to the degree possible, the smoothing approach was still necessary because of the annual variation in the percent TC, SV, and BG values attributable to a variety of sources including different weather conditions, varying vegetation phenology, and image misregistration. For TC, SV and BG time series in each aggregated spatial unit (e.g., a biome or a country), we applied the Theil-Sen estimator to derive the slope (annual change) of trend and provide the estimate of net change between 1982 and 2016 (i.e., slope times 34 years). The upper and lower change estimates based on the 90% confidence interval for the slope were also derived (Extended Data Tables 1 and 2, Supplementary Information Table 1). It is important to point out that the derived Theil-Sen trend represents long-term land-cover changes as the effect of changes in sensor capabilities has been effectively removed.
We further imposed the objective constraint of statistical significance of the trend to define net change at the pixel level. A Mann-Kendall test was applied to the TC, SV, and BG time series in each pixel83. If the Mann-Kendall test was not statistically significant (p ≥ 0.05), we defined net change as 0. If the trend test was significant (p < 0.05), we applied the Theil-Sen estimator to estimate the per-pixel net change between 1982 and 2016. These non-parametric statistical methods were chosen due to their robustness for trend detection and insensitivity to outliers. They have been applied to detect the greenness trend of land surface using AVHRR-based NDVI and leaf area index datasets34,84,85 as well as the microwave-based vegetation optical depth data86. Six global VCF gain (positive slope) and loss (negative slope) layers were derived: (i) tree canopy gain; (ii) tree canopy loss; (iii) short vegetation gain; (iv) short vegetation loss; (v) bare ground gain; and (vi) bare ground loss (Fig. 1b and Extended Data Fig. 1). Subsequently, per-pixel loss (gain) were aggregated to global, continental, climate zone, biome and country scales to derive gross loss (gain) estimates for each aggregated spatial unit (Extended Data Tables 1 and 2, Supplementary Information Table 1).
Driver attribution
Drivers of land-cover and land-use change are diverse, multi-scale and interactive1,13,87–89. Different drivers can be most broadly classified into two groups: anthropogenic and natural. Anthropogenic drivers are mainly related to land-use activities (e.g., deforestation, agricultural expansion, agricultural intensification, infrastructure construction and resource extraction), which are in turn driven by a number of underlying demographic, economic, technological, institutional, and cultural factors. Natural land-change drivers also include a variety of agents such as wildfire, drought, flood, windthrow, landslide, disease, insect attack, natural vegetation growth and glacial retreat, many of which are related to long-term climatic variation. Different drivers interact with each other in complex ways and the interactions are even evident at the broadest level in the Anthropocene7. With substantial human perturbations to the climate system, human-induced climate change and natural climatic variation and their effect on terrestrial ecosystems are intertwined. Disentangling human-induced climate change from natural climatic variation is a challenge, which can be studied using Earth system models35. Our objective for the global driver attribution was to provide a statistical, observation-based estimate of the relative contribution of direct human activities versus indirect drivers (including the combined effects of natural and human-induced climate change) to the observed global land change. Regionally dominant, specific land-change drivers were not explicitly quantified, but were identified and summarized through a comprehensive literature review.
We used a global probability sample and interpretation of high resolution images from Google Earth to estimate the proportion of changes attributable to drivers90,91, separately for each VCF change type: (i) tree canopy gain; (ii) tree canopy loss; (iii) short vegetation gain; (iv) short vegetation loss; (v) bare ground gain; and (vi) bare ground loss. For each VCF change type, 250 sample pixels (a pixel is a 0.05° × 0.05° grid cell) were selected with probability proportional to each pixel’s absolute change area (−1 * change area in the case of loss) of the target VCF change type, where the area of change was obtained from the global change layers described above. A total of 1500 sample pixels were selected (Extended Data Fig. 3a). For each sample pixel, we created a polygon feature representing its boundary and imported it in Google Earth (Extended Data Fig. 4). Each polygon was also divided into 25 0.01° × 0.01° grid cells to aid photo interpretation (Extended Data Fig. 3b-h). We used high-resolution images and the time slider tool in Google Earth to estimate the proportion of a pixel under human land use, including forestry and agricultural landscapes, cities, villages, houses, roads and other artificial objects. This proportion value was defined as the direct human impact associated with land-cover and land-use changes within the pixel. The impact of indirect drivers was defined as the residual of direct human impact. Areas of long-term land degradation resulting from the combined effects of land use and climate change were labeled as indirect if no signs of land use, for example fence lines or grazing paddocks, were observed. We estimated the direct human impact for each VCF change type as well as all land changes, using the following equations:
(13) |
(14) |
where Hc is the direct human impact of each of the 6 cover change types indexed by c, ℎj is the proportion of pixel j that is under human land use, nc is sample size (nc= 250), OH is the overall direct human impact of all land changes, and wc is the weight of each cover change type, given by the proportion of its global area over total absolute change area of all types (Extended Data Table 1). Similarly, we also estimated the overall direct human impact for all land changes within a continent and a biome. Attribution results are summarized in Extended Data Fig. 3.
Uncertainty analysis
The uncertainties of the area estimates of net land-cover change were characterized as statistical bounds (Extended Data Tables 1 and 2). Here we conducted an additional uncertainty analysis on gross change estimates to investigate whether the overall VCF trends hold true.
We first varied the statistical significance level in the Mann-Kendall trend test for defining change. Compared with change area estimates resulted from the p < 0.05 threshold, using p < 0.1 to define change, the estimated TC, BG and SV change area would differ by 6%, 2% and 14% respectively, whereas using p < 0.01 to define change, the estimated TC, BG and SV change area would differ by 16%, 1% and 31% respectively. Moreover, the signs of TC, BG and SV change were consistent at all significance levels — net gain in TC, net loss in BG and net loss in SV.
We further investigated the effect of VCF mapping uncertainty on change characterization. We employed the deviance value (i.e., the sum of squared difference between predicted value and training reference value) of each leaf node of the bagged regression tree models and computed a root-mean-square-deviation (RMSD) layer as VCF prediction uncertainty72,92. This per-pixel uncertainty layer was produced for each year between 1982 and 2016. Since RMSD is a quantitative indicator of land-cover uncertainty, we compared it with the magnitude of land-cover change by constructing a “signal-to-noise” ratio. The uncertainty of change for a given pixel i is then represented by the ratio of land-cover change to RMSD, summarized using the following equations:
(15) |
(16) |
where, for each pixel i in year k, the annual mean (unit: percent land cover) is the average value of N years; the ratio metric ri for each pixel i is computed as 1982–2016 VCF change within the pixel (ΔVCFi in units of percent land cover) to 1982–2016 average model prediction uncertainty .
A greater absolute value of the ratio metric ri indicates lower uncertainty of land-cover change and vice versa (Extended Data Figure 8). The density distributions (Extended Data Fig. 8c and 8d) suggest that for any threshold (dashed lines), the proportion of area under the frequency curve for tree cover gain always exceeds tree cover loss and similarly the proportion of area under the frequency curve for bare ground loss always exceeds bare ground cover gain. Hence, the overall trends in ri corroborate the main findings of our study, which are that there is a net gain in tree cover and a net loss in bare ground cover over the study period of 1982 to 2016.
Extended Data
Extended Data Table 1.
Tree canopy cover | Short vegetation cover | Bare ground cover | |||||||||||||||||||
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
Annual net change |
Gross change |
Annual net change |
Gross change |
Annual net change |
Gross change |
||||||||||||||||
Continent | Area 1982 (103 km2) | Slope (103 km2 yr−1) | Lower (103 km2 yr−1) | Upper (103 km2 yr−1) | p | loss (103 km2) | gain (103 km2) | Area 1982 (103 km2) | Slope (103 km2 yr−1) | Lower (103 km2 yr−1) | Upper (103 km2 yr−1) | p | loss (103 km2) | gain (103 km2) | Area 1982 (103 km2) | Slope (103 km2 yr1) | Lower (103 km2 yr−1) | Upper (103 km2 yr−1) | p | loss (103 km2) | gain (103 km2) |
Africa | 4672 | −1.9 | −7.6 | 3.6 | 0.609 | −267 | 262 | 11653 | 14.8 | 6.5 | 23.2 | 0.016 | −268 | 571 | 13413 | −12.4 | −19.9 | −4.7 | 0.020 | −371 | 105 |
Asia | 8457 | 37.5 | 28.0 | 45.3 | 0.000 | −178 | 1170 | 21774 | −22.9 | −34.5 | −9.6 | 0.008 • | −1261 | 760 | 13926 | −15.1 | −23.1 | −7.4 | 0.002 | −798 | 358 |
Europe | 2719 | 28.3 | 20.4 | 32.8 | 0.000 | −17 | 758 | 6320 | −22.0 | −27.3 | −14.7 | 0.000 | −673 | 50 | 668 | −4.3 | −5.8 | −2.6 | 0.000 | −92 | 9 |
North America | 5815 | 15.6 | 3.5 | 24.2 | 0.020 | −205 | 583 | 12921 | −12.7 | −4.1 | −2.2 | 0.031 | −594 | 286 | 4847 | −2.5 | −7.3 | 2.2 | 0.363 | −186 | 140 |
South America | 8767 | −14.1 | −20.5 | −7.4 | 0.001 | −621 | 190 | 7165 | 14.8 | 8.1 | 21.0 | 0.002 | −224 | 655 | 1717 | 1.9 | −1.2 | 3.8 | 0.307 | −92 | 102 |
Oceania | 680 | 0.1 | −1.4 | 1.7 | 0.887 | −40 | 56 | 4600 | −4.4 | −12.1 | 3.4 | 0.349 | −132 | 50 | 2772 | 5.2 | −3.9 | 12.5 | 0.280 | −35 | 113 |
Global | 31628 | 66.0 | 27.3 | 100.5 | 0.008 | −1331 | 3039 | 64539 | −26.0 | −64.8 | 15.2 | 0.244 • | −3170 | 2380 | 37412 | −34.0 | −52.3 | −10.0 | 0.023 | −1582 | 830 |
Extended Data Table 2.
Tree canopy cover | Short vegetation cover | Bare ground cover | |||||||||||||||||||
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
Annual net change |
Gross change |
Annual net change |
Gross change |
Annual net change |
Gross change |
||||||||||||||||
Biome / climate zone | Area 1982 (103 km2) | Slope (103 km2 yr−1) | Lower (103 km2 yr−1) | Upper (103 km2 yr−1) | p | loss (103 km2) | gain (103 km2) | Area 1982 (103 km2) | Slope (103 km2 yr−1) | Lower (103 km2 yr−1) | Upper (103 km2 yr−1) | p | loss (103 km2) | gain (103 km2) | Area 1982 (103 km2) | Slope (103 km2 yr−1) | Lower (103 km2 yr−1) | Upper (103 km2 yr−1) | p | loss (103 km2) | gain (103 km2) |
Tropical rainforest | 10519 | −1.9 | −4.4 | 1.6 | 0.443 | −332 | 315 | 3721 | 2.1 | −1.0 | 5.1 | 0.307 | −292 | 326 | 236 | −0.4 | −0.7 | −0.1 | 0.025 | −19 | 15 |
Tropical moist deciduous forest | 3569 | −2.5 | −8.5 | 2.2 | 0.460 | −373 | 285 | 6912 | 5.8 | 0.3 | 10.9 | 0.078 | −236 | 386 | 492 | −2.2 | −3.0 | −1.3 | 0.001 | −71 | 29 |
Tropical dry forest | 1236 | −2.8 | −5.1 | −1.1 | 0.018 | −184 | 99 | 5386 | 7.2 | 4.0 | 10.2 | 0.001 | −70 | 246 | 821 | −3.8 | −5.9 | −1.9 | 0.010 | −121 | 32 |
Tropical mountain system | 1333 | 3.5 | 2.4 | 4.5 | 0.000 | −23 | 118 | 2092 | −1.4 | −3.1 | 0.2 | 0.118 | −106 | 65 | 1092 | −1.7 | −2.7 | −0.9 | 0.002 | −61 | 17 |
Tropical shrubland| | 149 | 0.3 | −0.2 | 0.7 | 0.349 | −15 | 20 | 4010 | 12.3 | 6.8 | 18.5 | 0.001 | −41 | 371 | 4137 | −12.0 | −19.0 | −6.1 | 0.003 | −379 | 43 |
Tropical desert | 19 | 0.0 | 0.0 | 0.0 | 0.532 | 0 | 1 | 692 | 1.6 | 0.2 | 3.4 | 0.061 | −31 | 87 | 10846 | −1.5 | −3.4 | −0.1 | 0.057 | −88 | 31 |
Tropical climate zone | 16837 | −4.1 | −14.4 | 3.6 | 0.320 | −927 | 837 | 22691 | 30.0 | 14.7 | 43.0 | 0.002 | −775 | 1480 | 17617 | −25.5 | −34.7 | −12.0 | 0.002 | −740 | 167 |
Subtropical humid forest | 1566 | 8.2 | 4.4 | 12.0 | 0.002 | −48 | 268 | 2866 | −7.3 | −10.5 | −3.7 | 0.003 | −236 | 46 | 196 | −0.7 | −1.4 | −0.3 | 0.012 | −38 | 22 |
Subtropical mountain system | 516 | 3.1 | 2.3 | 3.8 | 0.000 | −20 | 116 | 2571 | −2.8 | −4.3 | −1.2 | 0.008 | −153 | 68 | 1756 | 0.0 | −1.7 | 1.8 | 0.932 | −79 | 79 |
Subtropical dry forest | 198 | 1.6 | 0.9 | 2.3 | 0.001 | −8 | 49 | 1107 | 0.2 | −0.5 | 0.8 | 0.755 | −37 | 33 | 266 | −1.2 | −1.9 | −0.6 | 0.002 | −45 | 10 |
Subtropical steppe | 179 | −0.8 | −2.0 | 0.2 | 0.191 | −27 | 12 | 2594 | −1.4 | −4.2 | 1.8 | 0.460 | −84 | 64 | 2106 | 3.2 | −0.9 | 6.7 | 0.201 | −61 | 106 |
Subtropical desert | 29 | −0.2 | −0.4 | 0.0 | 0.118 | −3 | 2 | 2606 | −4.4 | −10.3 | 1.3 | 0.233 | −128 | 45 | 4001 | 4.5 | −1.3 | 10.6 | 0.233 | −46 | 133 |
Subtropical climate zone | 2453 | 12.1 | 6.5 | 16.8 | 0.004 | −105 | 448 | 11741 | −14.0 | −23.1 | −5.6 | 0.013 | −639 | 257 | 8323 | 5.6 | −6.8 | 17.0 | 0.443 | −269 | 350 |
Temperate continental forest | 2172 | 21.4 | 15.1 | 26.0 | 0.000 | −11 | 591 | 4451 | −17.9 | −22.5 | −11.3 | 0.000 | −528 | 28 | 277 | −2.7 | −3.4 | −2.2 | 0.000 | −61 | 7 |
Temperate mountain system | 1552 | 5.9 | 3.6 | 7.4 | 0.001 | −53 | 198 | 3459 | −2.0 | −4.0 | 0.5 | 0.211 | −213 | 172 | 2175 | −2.9 | −5.1 | −1.1 | 0.023 | −161 | 62 |
Temperate oceanic forest | 551 | 3.8 | 2.1 | 5.3 | 0.001 | −6 | 101 | 1162 | −3.3 | −4.9 | −1.8 | 0.003 | −92 | 8 | 61 | −0.4 | −0.5 | −0.2 | 0.000 | −8 | 2 |
Temperate steppe | 320 | 2.2 | 0.1 | 3.4 | 0.069 | −18 | 56 | 4191 | −2.9 | −6.7 | 0.0 | 0.105 | −130 | 72 | 1338 | 2.3 | −1.5 | 5.8 | 0.363 | −86 | 108 |
Temperate desert | 61 | −0.1 | −0.2 | 0.1 | 0.514 | −5 | 4 | 1661 | −0.3 | −2.9 | 3.5 | 0.955 | −101 | 135 | 3642 | 0.3 | −3.6 | 3.1 | 0.887 | −135 | 103 |
Temperate climate zone | 4681 | 33.5 | 21.0 | 41.9 | 0.000 | −92 | 951 | 14814 | −24.3 | −37.3 | −12.2 | 0.006 | −1064 | 414 | 7491 | −4.3 | −14.0 | 1.8 | 0.268 | −451 | 282 |
Boreal coniferous forest | 3938 | 13.6 | 6.7 | 18.7 | 0.003 | −75 | 415 | 4239 | −12.6 | −17.1 | −6.3 | 0.002 | −369 | 71 | 205 | −1.4 | −1.8 | −0.9 | 0.001 | −23 | 2 |
Boreal mountain system | 2035 | 6.6 | 3.9 | 9.9 | 0.005 | −61 | 225 | 3909 | −6.2 | −8.8 | −3.3 | 0.003 | −193 | 64 | 341 | −1.2 | −1.7 | −0.6 | 0.002 | −33 | 11 |
Boreal tundra woodland | 971 | 1.3 | −1.5 | 3.7 | 0.363 | −58 | 82 | 2723 | −0.9 | −2.9 | 1.2 | 0.478 | −63 | 52 | 228 | −0.7 | −1.2 | −0.2 | 0.044 | −16 | 5 |
Boreal climate zone | 6796 | 21.0 | 7.9 | 31.0 | 0.009 | −194 | 723 | 10857 | −20.3 | −27.5 | −11.7 | 0.002 | −625 | 187 | 772 | −3.4 | −4.5 | −2.2 | 0.001 | −71 | 19 |
Polar | 236 | 2.2 | 0.9 | 3.1 | 0.009 | −7 | 55 | 4109 | 0.4 | −1.3 | 2.1 | 0.712 | −43 | 30 | 3080 | −2.6 | −3.6 | −1.1 | 0.010 | −41 | 8 |
Supplementary Material
Acknowledgments:
This study was funded by the NASA Making Earth System Data Records for Use in Research Environments (MEaSUREs) program (NNX13AJ35A). We thank Tom Loveland, Bruce Pengra and Pontus Olofsson for making their tree cover validation data available. We thank Charlene Dimiceli for assistance on VCF development and Zhen Song for assistance on AVHRR calibration.
Footnotes
Data availability: The generated AVHRR VCF products will be distributed through Land Processes Distributed Active Archive Center (LP DAAC, https://lpdaac.usgs.gov/). VCF change layers are provided at http://glad.geog.umd.edu/dataset/long-term-global-land-change for view and download.
Code availability: The code used for deriving land-cover change from time-series VCF layers is available upon request from the corresponding author.
Competing financial interests: The authors declare no competing financial interests.
Supplementary Information is linked to the online version of the paper at www.nature.com/nature.
References
- 1.Turner BL, Lambin EF & Reenberg A The emergence of land change science for global environmental change and sustainability. Proc. Natl. Acad. Sci. USA 104, 20666–20671 (2007). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2.Foley JA et al. Global consequences of land use. Science 309, 570–574 (2005). [DOI] [PubMed] [Google Scholar]
- 3.Le Quéré C et al. Global carbon budget 2016. Earth Syst. Sci. Data 8, 605–649 (2016). [Google Scholar]
- 4.Alkama R & Cescatti A Biophysical climate impacts of recent changes in global forest cover. Science 351, 600–604 (2016). [DOI] [PubMed] [Google Scholar]
- 5.FAO. Global Forest Resources Assessment 2015. (UN Food and Agriculture Organization, Rome, Italy, 2015). [Google Scholar]
- 6.Bonan GB & Doney SC Climate, ecosystems, and planetary futures: The challenge to predict life in Earth system models. Science 359, eaam8328 (2018). [DOI] [PubMed] [Google Scholar]
- 7.Steffen W, Grinevald J, Crutzen P & McNeill J The Anthropocene: Conceptual and historical perspectives. Philos. Trans. A Math. Phys. Eng. Sci 369, 842–867 (2011). [DOI] [PubMed] [Google Scholar]
- 8.Hansen MC et al. High-resolution global maps of 21st-centry forest cover change. Science 342, 850–853 (2013). [DOI] [PubMed] [Google Scholar]
- 9.Hansen MC, Stehman SV & Potapov PV Quantification of global gross forest cover loss. Proc. Natl. Acad. Sci. USA 107, 8650–8655 (2010). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10.Feng M et al. Earth science data records of global forest cover and change: Assessment of accuracy in 1990, 2000, and 2005 epochs. Remote Sens. Environ 184, 73–85 (2016). [Google Scholar]
- 11.DeFries RS et al. Mapping the land surface for global atmosphere-biosphere models: Toward continuous distributions of vegetation’s functional properties. J. Geophys. Res. 100, 20867–20882 (1995). [Google Scholar]
- 12.FAO. Global ecological zoning for the global forest resources assessment 2000. (UN Food and Agriculture Organization, Rome, Italy, 2001). [Google Scholar]
- 13.Lambin EF et al. The causes of land-use and land-cover change: Moving beyond the myths. Global Environ. Change 11, 261–269 (2001). [Google Scholar]
- 14.Gibbs HK et al. Tropical forests were the primary sources of new agricultural land in the 1980s and 1990s. Proc. Natl. Acad. Sci. USA 107, 16732–16737 (2010). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15.Song X-P, Huang C, Saatchi SS, Hansen MC & Townshend JR Annual carbon emissions from deforestation in the Amazon basin between 2000 and 2010. PLoS One 10, e0126754 (2015). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16.Margono BA, Potapov PV, Turubanova S, Stolle F & Hansen MC Primary forest cover loss in Indonesia over 2000–2012. Nat. Clim. Change 4, 730–735 (2014). [Google Scholar]
- 17.Ordway EM, Asner GP & Lambin EF Deforestation risk due to commodity crop expansion in sub-Saharan Africa. Environ. Res. Lett. 12, 044015 (2017). [Google Scholar]
- 18.Hicke JA et al. Postfire response of North American boreal forest net primary productivity analyzed with satellite observations. Glob. Change Biol. 9, 1145–1157 (2003). [Google Scholar]
- 19.van Mantgem PJ et al. Widespread increase of tree mortality rates in the western United States. Science 323, 521–524 (2009). [DOI] [PubMed] [Google Scholar]
- 20.McManus KM et al. Satellite-based evidence for shrub and graminoid tundra expansion in northern Quebec from 1986 to 2010. Glob. Change Biol 18, 2313–2323 (2012). [Google Scholar]
- 21.Mitchard ET & Flintrop CM Woody encroachment and forest degradation in sub-Saharan Africa’s woodlands and savannas 1982–2006. Philos. Trans. R. Soc. Lond., Ser. B: Biol. Sci 368, 20120406 (2013). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.Brandt M et al. Human population growth offsets climate-driven increase in woody vegetation in sub-Saharan Africa. Nat. Ecol. Evol 1, 0081 (2017). [DOI] [PubMed] [Google Scholar]
- 23.Harsch MA, Hulme PE, McGlone MS & Duncan RP Are treelines advancing? A global meta-analysis of treeline response to climate warming. Ecol. Lett 12, 1040–1049 (2009). [DOI] [PubMed] [Google Scholar]
- 24.Potapov PV et al. Eastern Europe’s forest cover dynamics from 1985 to 2012 quantified from the full Landsat archive. Remote Sens. Environ 159, 28–43 (2015). [Google Scholar]
- 25.Piao S et al. Detection and attribution of vegetation greening trend in China over the last 30 years. Glob Chang Biol 21, 1601–1609 (2015). [DOI] [PubMed] [Google Scholar]
- 26.Birdsey R, Pregitzer K & Lucier A Forest carbon management in the United States. J. Environ. Qual 35, 1461–1469 (2006). [DOI] [PubMed] [Google Scholar]
- 27.Donohue RJ, McVicar TR & Roderick ML Climate-related trends in Australian vegetation cover as inferred from satellite observations, 1981–2006. Glob. Change Biol 15, 1025–1039 (2009). [Google Scholar]
- 28.Liu YY et al. Changing climate and overgrazing are decimating Mongolian steppes. PLoS One 8, e57599 (2013). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 29.Herrick JE et al. National ecosystem assessments supported by scientific and local knowledge. Front. Ecol. Environ 8, 403–408 (2010). [Google Scholar]
- 30.Evenson RE & Gollin D Assessing the Impact of the Green Revolution, 1960 to 2000. Science 300, 758–762 (2003). [DOI] [PubMed] [Google Scholar]
- 31.Ying Q et al. Global bare ground gain from 2000 to 2012 using Landsat imagery. Remote Sens. Environ 194, 161–176 (2017). [Google Scholar]
- 32.Myneni RB, Keeling CD, Tucker CJ, Asrar G & Nemani RR Increased plant growth in the northern high latitudes from 1981 to 1991. Nature 386, 698–702 (1997). [Google Scholar]
- 33.Forkel M et al. Codominant water control on global interannual variability and trends in land surface phenology and greenness. Glob. Chang Biol 21, 3414–3435 (2015). [DOI] [PubMed] [Google Scholar]
- 34.Zhu Z et al. Greening of the Earth and its drivers. Nat. Clim. Change 6, 791–795 (2016). [Google Scholar]
- 35.Mao J et al. Human-induced greening of the northern extratropical land surface. Nat. Clim. Change 6, 959–963 (2016). [Google Scholar]
- 36.Pan Y et al. A large and persistent carbon sink in the world’s forests. Science 333, 988–993 (2011). [DOI] [PubMed] [Google Scholar]
- 37.DeFries RS, Townshend JRG & Hansen MC Continuous fields of vegetation characteristics at the global scale at 1-km resolution. J. Geophys. Res 104, 16911–16923 (1999). [Google Scholar]
- 38.Hansen MC et al. Global percent tree cover at a spatial resolution of 500 meters: First results of the MODIS vegetation continuous fields algorithm. Earth Interact. 7, 10 (2003). [Google Scholar]
- 39.Hansen MC et al. Towards an operational MODIS continuous field of percent tree cover algorithm: examples using AVHRR and MODIS data. Remote Sens. Environ. 83, 303–319 (2002). [Google Scholar]
- 40.Hansen MC, Townshend JRG, DeFries RS & Carroll M Estimation of tree cover using MODIS data at global, continental and regional/local scales. Int. J. Remote Sens 26, 4359–4380 (2005). [Google Scholar]
- 41.Hansen MC et al. Continuous fields of land cover for the conterminous United States using Landsat data: First results from the Web-Enabled Landsat Data (WELD) project. Remote Sens. Lett. 2, 279–288 (2011). [Google Scholar]
- 42.Jennings SB, Brown ND & Sheil D Assessing forest canopies and understorey illumination: Canopy closure, canopy cover and other measures. Forestry 72, 59–74 (1999). [Google Scholar]
- 43.Korhonen L, Korhonen KT, Rautiainen M & Stenberg P Estimation of forest canopy cover : A comparison of field measurement techniques. Silva Fenn. 40, 577–588 (2006). [Google Scholar]
- 44.Hansen MC et al. Monitoring conterminous United States (CONUS) land cover change with web-enabled Landsat data (WELD). Remote Sens. Environ. 140, 466–484 (2014). [Google Scholar]
- 45.Pinzon J & Tucker C A non-stationary 1981–2012 AVHRR NDVI3g time series. Remote Sens. 6, 6929–6960 (2014). [Google Scholar]
- 46.Tucker C et al. An extended AVHRR 8-km NDVI dataset compatible with MODIS and SPOT vegetation NDVI data. Int. J. Remote Sens 26, 4485–4498 (2005). [Google Scholar]
- 47.Franch B et al. A 30+ year AVHRR land surface reflectance climate data record and its application to wheat yield monitoring. Remote Sens. 9, 296 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 48.Pedelty J et al. Generating a long-term land data record from the AVHRR and MODIS Instruments. IEEE Int. Geosci. Remote Sens. Symp. Proc, 1021–1025 (2007). [Google Scholar]
- 49.Tucker CJ Red and photographic infrared linear combinations for monitoring vegetation. Remote Sens. Environ. 8, 127–150 (1979). [Google Scholar]
- 50.Holben BN Characteristics of maximum-value composite images from temporal AVHRR data. Int. J. Remote Sens 7, 1417–1434 (1986). [Google Scholar]
- 51.Loveland TR et al. Development of a global land cover characteristics database and IGBP DISCover from 1 km AVHRR data. Int. J. Remote Sens 21, 1303–1330 (2000). [Google Scholar]
- 52.Hansen MC, DeFries RS, Townshend JRG & Sohlberg RA Global land cover classification at 1 km spatial resolution using a classification tree approach. Int. J. Remote Sens 21, 1331–1364 (2000). [Google Scholar]
- 53.DeFries RS, Hansen M, Townshend JRG & Sohlberg R Global land cover classifications at 8 km spatial resolution: The use of training data derived from Landsat imagery in decision tree classifiers. Int. J. Remote Sens 19, 3141–3168 (1998). [Google Scholar]
- 54.Bartholomé E & Belward AS GLC2000: A new approach to global land cover mapping from Earth observation data. Int. J. Remote Sens 26, 1959–1977 (2005). [Google Scholar]
- 55.Lloyd D A phenological classification of terrestrial vegetation cover using shortwave vegetation index imagery. Int. J. Remote Sens 11, 2269–2279 (1990). [Google Scholar]
- 56.Reed BC et al. Measuring phenological variability from satellite imagery. J. Veg. Sci 5, 703–714 (1994). [Google Scholar]
- 57.DeFries R, Hansen M & Townshend J Global discrimination of land cover types from metrics derived from AVHRR pathfinder data. Remote Sens. Environ 54, 209–222 (1995). [Google Scholar]
- 58.Latifovic R, Pouliot D & Dillabaugh C Identification and correction of systematic error in NOAA AVHRR long-term satellite data record. Remote Sens. Environ 127, 8497 (2012). [Google Scholar]
- 59.Kaufmann RK et al. Effect of orbital drift and sensor changes on the time series of AVHRR vegetation index data. IEEE Trans. Geosci. Remote Sens. 38, 2584–2597 (2000). [Google Scholar]
- 60.Kaufman YJ & Holben BN Calibration of the AVHRR visible and near-IR bands by atmospheric scattering, ocean glint and desert reflection. Int. J. Remote Sens 14, 21–52 (1993). [Google Scholar]
- 61.Vermote E & Kaufman YJ Absolute calibration of AVHRR visible and near-infrared channels using ocean and cloud views. Int. J. Remote Sens 16, 2317–2340 (1995). [Google Scholar]
- 62.Los SO Estimation of the ratio of sensor degradation between NOAA AVHRR channels 1 and 2 from monthly NDVI composites. IEEE Trans. Geosci. Remote Sens 36, 206–213 (1998). [Google Scholar]
- 63.Myneni RB, Tucker CJ, Asrar G & Keeling CD Interannual variations in satellite-sensed vegetation index data from 1981 to 1991. J. Geophys. Res. Atmos 103, 6145–6160 (1998). [Google Scholar]
- 64.Gutman GG On the use of long-term global data of land reflectances and vegetation indices derived from the advanced very high resolution radiometer. J. Geophys. Res. Atmos 104, 6241–6255 (1999). [Google Scholar]
- 65.Donohue RJ, Roderick ML & McVicar TR Deriving consistent long-term vegetation information from AVHRR reflectance data using a cover-triangle-based framework. Remote Sens. Environ 112, 2938–2949 (2008). [Google Scholar]
- 66.Hansen MC et al. A method for integrating MODIS and Landsat data for systematic monitoring of forest cover and change in the Congo basin. Remote Sens. Environ 112, 2495–2513 (2008). [Google Scholar]
- 67.Potapov PV et al. Quantifying forest cover loss in Democratic Republic of the Congo, 2000–2010, with Landsat ETM+ data. Remote Sens. Environ 122, 106–116 (2012). [Google Scholar]
- 68.Chavez PS Jr Radiometric calibration of Landsat Thematic Mapper multispectral images. Photogramm. Eng. Remote Sensing 55, 1285–1294 (1989). [Google Scholar]
- 69.Song. C, Woodcock CE, Seto KC, Lenney MP & Macomber SA Classification and change detection using Landsat TM data: When and how to correct atmospheric effects? Remote Sens. Environ 75, 230–244 (2001). [Google Scholar]
- 70.Woodcock CE, Macomber SA, Pax-Lenney M & Cohen WB Monitoring large areas for forest change using Landsat: Generalization across space, time and Landsat sensors. Remote Sens. Environ. 78, 194–203 (2001). [Google Scholar]
- 71.Potapov P et al. The last frontiers of wilderness: Tracking loss of intact forest landscapes from 2000 to 2013. Sci. Adv 3, e1600821 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 72.Breiman L, Friedman JH, Olshen RA & Stone CJ Classification and regression trees. (Chapman & Hall/CRC, 1984). [Google Scholar]
- 73.Hansen M, Dubayah R & Defries R Classification trees: An alternative to traditional land cover classifiers. Remote Sens. Lett 17, 1075–1081 (1996). [Google Scholar]
- 74.Friedl MA & Brodley CE Decision tree classification of land cover from remotely sensed data. Remote Sens. Environ 61, 399–409 (1997). [Google Scholar]
- 75.Mascorro VS, Coops NC, Kurz WA & Olguin M Choice of satellite imagery and attribution of changes to disturbance type strongly affects forest carbon balance estimates. Carbon Balance Manag. 10, 30 (2015). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 76.Strahler AH et al. Global land cover validation: recommendations for evaluation and accuracy assessment of global land cover maps. (Office for Official Publications of the European Communities, Luxembourg, 2006). [Google Scholar]
- 77.Olofsson P et al. A global land-cover validation data set, part I: Fundamental design principles. Int. J. Remote Sens 33, 5768–5788 (2012). [Google Scholar]
- 78.Stehman SV, Olofsson P, Woodcock CE, Herold M & Friedl MA A global land-cover validation data set, II: Augmenting a stratified sampling design to estimate accuracy by region and land-cover class. Int. J. Remote Sens 33, 6975–6993 (2012). [Google Scholar]
- 79.Pengra B, Long J, Dahal D, Stehman SV & Loveland TR A global reference database from very high resolution commercial satellite data and methodology for application to Landsat derived 30m continuous field tree cover data. Remote Sens. Environ. 165, 234–248 (2015). [Google Scholar]
- 80.Willmott CJ Some comments on the evaluation of model performance. Bull. Am. Meteorol. Soc 63, 1309–1313 (1982). [Google Scholar]
- 81.Cochran WG Sampling Techniques (3rd edition). (John Wiley & Sons, 1977). [Google Scholar]
- 82.Sen PK Estimates of the regression coefficient based on Kendall’s tau. J. Am. Stat. Assoc 63, 1379–1389 (1968). [Google Scholar]
- 83.Mann HB Nonparametric tests against trend. Econometrica 13, 245–259 (1945). [Google Scholar]
- 84.Guay KC et al. Vegetation productivity patterns at high northern latitudes: A multi-sensor satellite data assessment. Glob. Change Biol. 20, 3147–3158 (2014). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 85.Fensholt R et al. Greenness in semi-arid areas across the globe 1981–2007 — an Earth observing satellite based analysis of trends and drivers. Remote Sens. Environ 121, 144–158 (2012). [Google Scholar]
- 86.Andela N, Liu YY, van Dijk AIJM, de Jeu RAM & McVicar TR Global changes in dryland vegetation dynamics (1988–2008) assessed by satellite remote sensing: comparing a new passive microwave vegetation density record with reflective greenness data. Biogeosciences 10, 6657–6676 (2013). [Google Scholar]
- 87.Geist HJ & Lambin EF Proximate Causes and Underlying Driving Forces of Tropical Deforestation. Bioscience 52, 143–150 (2002). [Google Scholar]
- 88.DeFries RS, Rudel T, Uriarte M & Hansen M Deforestation driven by urban population growth and agricultural trade in the twenty-first century. Nat. Geosci. 3, 178–181 (2010). [Google Scholar]
- 89.Liu S et al. Grand Challenges in Understanding the Interplay of Climate and Land Changes. Earth Interact 21, 1–43 (2017). [Google Scholar]
- 90.De Sy V et al. Land use patterns and related carbon losses following deforestation in South America. Environ. Res. Lett 10, 124004 (2015). [Google Scholar]
- 91.Tyukavina A et al. Types and rates of forest disturbance in Brazilian Legal Amazon, 2000–2013. Sci. Adv 3, e1601047 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]
- 92.Breiman L Bagging Predictors. Mach. Learn. 24, 123–140 (1996). [Google Scholar]
- 93.Gessner U, Machwitz M, Conrad C & Dech S Estimating the fractional cover of growth forms and bare surface in savannas. A multi-resolution approach based on regression tree ensembles. Remote Sens. Environ. 129, 90–102 (2013). [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.