Main

Stand-replacing disturbances are natural or anthropogenic events that cause the sudden death of a group of trees1,2,3. These events occur at different frequencies and intensities and are generally temporary, forming an integral part of the natural cycle of forest dynamics2. The patches left by disturbances also vary greatly in terms of size, shape and spatial arrangement. These patterns of damage are the result of a complex combination of factors, involving the disturbance agents, the state of the forest and the broader environmental context4. Because tree growth and succession happen over timescales of decades to centuries, disturbance patterns have a long-lasting impact on forest structure and composition4,5, with consequent impacts on both carbon cycling6 and habitat diversity7. The latter, in turn, is then crucial for the diversity and abundance of fauna and flora8,9. Understanding disturbance dynamics is therefore a fundamental building block for understanding the form and function of the world’s forests.

While a growing body of works focuses on characterizing disturbance rates at a large scale3,10,11,12,13, work on characterizing the structure of disturbances has generally been limited to the landscape or regional scales14,15,16. The few studies that have assessed structural patterns over the scale of one or more biomes have reported substantial consistency in the shapes and sizes of disturbance patches between unmanaged forests in temperate and boreal biomes, but with large variations within biomes17,18. Whether such consistency holds more broadly across biomes, particularly given management pervasiveness19,20, remains unknown.

Several mechanisms affect the structure of disturbance patches. The nature of a disturbance agent and its intensity can fundamentally influence the form of the marks left on forest landscapes. For example, rapidly and linearly moving processes, such as tornadoes or flooding, may produce patches that are elongated or linearly arranged14,21,22,23,24. By contrast, an enduring contagion process, such as a biotic outbreak or a smouldering fire, would result in patches that are growing over time, clumped or coalescent25,26, whereas industrial clear-cutting may produce patches that have compact geometries and regular sizes and patterns8.

In addition to the causal agent, the local environment also plays a major role in shaping disturbance patches. Climate, weather and topography can strongly modulate the formation process of individual patches. For example, dry windy conditions may lead to stretched burn scars, while steep slopes may favour downslope flows and mass movements27. Forest properties, such as tree health and species composition, also influence the propagation of the disturbance process. For example, the speed of a fire spreading may vary depending on tree flammability traits28, while trees collectively subjected to a stress would be weakened and thus more vulnerable to disease outbreaks29. Finally, disturbance patches may simply reflect the shape and configuration of the original forest fragment, which is the case for degraded or land-use-fragmented forests30,31. While links between forest composition and disturbance agent suggest a link between disturbance patch structure and biomes, other factors suggest a more subtle variation of disturbance patch structure following the physical environment and local management norms.

The structure of disturbance patches has fundamental implications for forest recovery and the biodiversity supported. Complex or elongated patches increase edge effect, thus forest exposure to the altered microclimates inside the patch and permeability for other species32,33. Recolonization from the surrounding forest can be fast in smaller patches, while it may be slowed down in larger ones, owing to longer distances to seed sources and altered microclimatic conditions inside the patches4,34,35. Depending on forest types, these differences in conditions can be expected to influence the successional trajectory34,35. Clusters of small patches within a closed-canopy forest affect habitat quality by deteriorating the physical conditions required by shade-tolerant and forest interior specialists36,37, whereas with larger patches, they may create substantial barriers to habitat connectivity, impeding species movements, while also staging the way for fragmentation or degradation processes38. Identifying such structural characteristics in disturbances across the world opens the door for a better understanding of the biogeography of a broad range of species, communities and ecological processes, which is the very basis for constructing sustainable action plans for forests.

In this study, we used a Landsat-based tree cover loss data (Global Forest Change (GFC))10 and a selection of established and novel metrics describing patch structure in terms of magnitude, complexity and context (Table 1) to categorize and map major patterns in forest disturbances wall to wall across the world and over the period 2002–2014, when the tree cover loss data were the most consistent (Supplementary Fig. 3). We sought to (1) identify the most prominent disturbance patterns, (2) explore their distribution across biomes and potential linkages with underlying drivers and (3) address the effects of human activities on disturbance patterns, contrasting such effects across biomes. We define as forest disturbance any loss in tree cover identifiable by remote sensing at 30 m resolution, where the land is assumed to remain a forest. The definition focuses on disturbances as temporary events, excluding large land use changes (>300 m resolution pixels, that is, 9 ha, from European Space Agency Climate Change Initiative (ESA CCI) Land Cover data39), but including anthropogenic activities (that is, harvest and shifting agriculture). Disturbances occurring after the loss reported in the GFC data are ignored, as the vegetation cover is assumed not to contain mature trees (that is, only the spatial characteristics of the first disturbance event are considered). Our definition of forest includes open-canopy forests and woodlands where tree cover can be below 10%.

Table 1 Disturbance patch metrics selected to characterize the structure of disturbances (values rounded to two decimal places)
Full size table

Statistics of disturbance patch metrics at the global scale

Between 2002 and 2014, we identified 249,149,911 patches aligning with our definition of forest disturbances. The magnitude of disturbance patches, represented by mean area and count of years (Table 1), peaks in Sweden, Finland, western Russia, southeastern Asia, southeastern Amazon and northwestern and southeastern United States (Supplementary Fig. 4). However, patch areas remain small in most of the African continent. Complexity metrics, represented by mean shape and elongation indices, consistently indicate more complex structures in the aforementioned regions, along with central Africa. Context metrics, reflected by spatio-temporal clustering within a 5 km buffer, show that disturbance patches tend to be spatially clustered in Siberia, Africa, southeastern Asia and parts of the Amazon and Canada. Disturbances also appear less temporally clustered in Europe, western Russia, southeastern Asia, northwestern and eastern United States, eastern Amazon and central and western Africa than in other parts of the world.

Four disturbance patterns identified globally

Grouping of the 2002–2014 disturbance patches using k-means clustering40 revealed four distinct disturbance structures across the world (Table 2). The patterns identified correspond to (1) small-isolated patches, characterized by the smallest mean areas (mean = 0.20 ha, that is, ~2 pixels), and the lowest spatio-temporal clustering (mean = 88.69 concomitant patches within 5 km and accounting for 8% of the disturbance area during the period studied); (2) clustered patches, which are the most spatiotemporally clustered (mean = 386.16 concomitant patches within 5 km and 50% of the disturbance area over the period studied); (3) complex patches, which are of medium size (mean = 1.17 ha) and complex, with the most stretched shapes (mean = 0.66 for the elongation index); and (4) large-multiyear patches, which are the largest (mean = 9.40 ha), the most long-lasting (mean = 2.12 years) and also the most complex (mean = 1.86 in shape index).

Table 2 Statistics of the metrics by the disturbance patterns identified with unsupervised clustering (values rounded to two decimal places)
Full size table

There was a remarkable consistency across biomes in the relative composition of the different disturbance patterns in terms of counts (Fig. 1a). Only the Tundra biome diverged markedly, with a much higher relative frequency of clustered patches, reflecting the very sparse woody cover in this region. This corroborates and extends previous conclusions that found little difference in disturbance characteristics between boreal and temperate biomes17. However, there were very large differences in the total area affected by each disturbance pattern, with large-multiyear patterns dominating disturbed areas in the boreal and Mediterranean regions, whereas complex and large-multiyear patterns contributed roughly equally to areas affected in tropical and temperate biomes (Fig. 1b). Small-isolated patches are the most prevalent globally comprising between 56% and 70% of patches across all biomes (Fig. 2a). Overall, they account for 62.90% of the total number of patches, but collectively represent only 12.94% of the global forest area disturbed (Table 1 and Supplementary Figs. 8 and 9). They dominate by their total areas in the arctic regions, as well as in some mountain ranges, such as the Himalayas, the Scandinavian mountains and central Madagascar (Fig. 2a and Supplementary Fig. 9). They are also dominant in tropical forest areas, in parts of the Amazon, Africa and central Borneo Island.

Fig. 1: Proportion of disturbance patterns by biome based on count and area disturbed.
figure 1

a, In terms of counts, the prevalence of the patterns remains similar across most biomes, with the small-isolated pattern having the largest fraction. The complex pattern is the second most prevalent, except in the tundra biome where it is exceeded by the clustered pattern. The large-multiyear pattern is the least prevalent, except in mangroves, temperate broadleaf and mixed forests, and tropical moist broadleaf forests, where it exceeds the fraction of the clustered pattern. b, In terms of total area, substantial variability appears across biomes, although the largest fraction remains taken by the large-multiyear pattern. The only exception is in the tropical coniferous biome, where the complex pattern has a greater fraction.

Full size image
Fig. 2: Distribution of disturbance patterns based on total areas covered.
figure 2

a, Relative importance of disturbance patterns. To reflect the contribution of each patch type within a grid cell based on areal dominance, we used a cyan–magenta–yellow–key (black) colour scheme (CMYK), with the percentage of each patch type added as a percentage of colour in the CMYK space. We used black for large-multiyear, cyan for complex, magenta for clustered and yellow for small-isolated patterns. No forest mask is applied. b, Shannon diversity index in disturbance patterns based on cover percentage. No forest mask is applied. Most of Africa is characterized by a high diversity index compared with other regions.

Full size image

Clustered patches form a minority in terms of counts (8.55% of the total number of patches) and coverage (3.39% of the total area disturbed) (Table 1, Fig. 2a, and Supplementary Figs. 8 and 9). They are characteristically present along transitional zones with drylands, notably with the African Sahara, the deserts of central Australia, the grasslands of central North America and the tundra of northmost Canada and eastern Russia. Clustered patches also appear in the remote and intact forests of the Amazon. Intact forests are defined as large forest fragments of more than 50,000 ha, assumed to be minimally influenced by human activities41.

Complex patches are substantial both in terms of counts (22.58% of the total number of patches) and area (26.84% of the total area disturbed) (Table 1, Fig. 2a, and Supplementary Figs. 8 and 9). Like the small-isolated patches, they are ubiquitous, with a spatial dominance in human-influenced regions, notably in northern Europe, China, southeastern Asia, southeastern United States and southeastern Amazon.

Finally, large-multiyear patches, although relatively rare in number (5.96%), account for most of the disturbed areas globally (56.83%) (Table 1, Fig. 2a, and Supplementary Figs. 8 and 9). They represent typically 2.5–6.5% of the total number of patches in each biome and are prevalent in regions affected by large wildfires, notably in fire-prone boreal Canada and Siberia. They are also concentrated in regions subject to human activities, where they co-dominate with the complex patches, notably throughout southeastern Amazon, Southeast Asia and the United States, and central Chile.

Large variability within biomes

Although at the scale of biomes, the relative frequency of the different patch structures was consistent within most biomes, we found marked variability in their spatial distribution (Fig. 2a), as well as the diversity of the pattern types found (Fig. 2b). This variability showed clear spatial consistency, indicating that it is not simply a result of the 13 year period of not sufficiently sampling rare disturbance events. Some of the patterns in Fig. 2a are consistent with known areas of strong human influence. For instance, the large-multiyear patches dominating in Indonesia and Malaysia, which mark them out from the rest of the tropics, are consistent with the locations of oil palm and rubber plantations in these countries42,43. Similarly, the abundance of small-isolated patches in England is probably linked to the highly fragmented forest landscape limiting patch size44,45. The diversity of disturbance patterns, estimated with the Shannon diversity index46 based on cover area, presents marked hotspots across the world. This is particularly visible in many parts of Africa, including the western coasts of Madagascar (Fig. 2b), which suggests a more uneven distribution of disturbance patterns compared with other continents. This may be explained by the distribution of the clustered and large-multiyear patches (Supplementary Figs. 8 and 9) and the smaller areas and shorter perimeters observed in disturbances there (Supplementary Fig. 4). China, Portugal, central and northern Europe, Central America and eastern Brazil also emerge as areas of high diversity. Overall, the biome appears to be a poor indicator for disturbance patch characteristics. We speculate that the physical environment of the forest; the type, intensity and legacy of human influence; and the traits of the species found there (as filtered by their history) may instead be the driving patterns of the current biogeography of disturbance.

Linkages with potential underlying drivers

The association of disturbance patterns with existing global databases on environmental conditions and human activities provides some insights into the drivers of these patterns. In high latitudes and elevations, between the lower timberline and the tree line, our definition of forests is represented by scattered small groups of trees subsisting under the harsh conditions characterizing these environments. The small and isolated disturbance patterns prevailing in these regions may reflect the size of the isolated tree stands that can be found there (Supplementary Fig. 8). Small-isolated patterns also predominate in some intact closed-canopy forests in the Amazon where the effects of human activities and fragmentation are assumed to be minimal (Figs. 2a and 3, and Supplementary Fig. 12b). This underpins the rarity of large disturbances in these forests and highlights the importance of small-scale gap dynamics that drive natural canopy turnover there47,48. These gaps are punctual, caused by small-scale events (for example, blowdowns) that result in openings in the otherwise continuous canopy. It should be noted that small-isolated patches often correspond to single pixels from the original tree cover loss data. The commission and omission errors associated with these pixels49 suggest that there may be overestimations of the small-isolated patches in the temperate and boreal forests (almost the double of false negatives, as false positives represent 11.8% in the temperate biome and 12.0% in the boreal biome, while false negatives represent 6.1% in both biomes) and possible underestimations in the tropical biome (13.0% for false positives and 16.9 for false negatives). While single-pixel events could be partially caused by classification noise, their clear and coherent spatial variation, aligning with our knowledge of forest disturbances across regions, suggests that the signal may be more than just noise and that any such noise has a minimal influence on the overall patterns we identify.

Fig. 3: Proportion of disturbance patterns inside and outside intact forests by biome based on counts and areas.
figure 3

The total areas of disturbance patterns show that in the intact tropical moist broadleaf forests, the small-isolated and clustered patterns predominate, while large-multiyear patterns are minor.

Full size image

The dominance of clustered patches in large areas of the intact contiguous forests of the Amazon (Figs. 2a and 3, and Supplementary Fig. 12b) is consistent with the blowdown and drought-related mortality, which are known to be major agents of disturbance there50,51. Wind can damage multiple nearby locations at the same time, while drought impacts are strongly mediated by small-scale variations in water availability52, thus generating the clustered pattern, which has been suggested to be broadly associated with natural disturbances16,53,54. Clustered patches are also prominent throughout the ecotones with drylands, parts of the Mediterranean regions and in the boreal peatlands (Fig. 2a). They also show a substantial association with fire (Fig. 4). Forests in these regions are fragmented and often disturbed by large wildfires55. The clustered patch structure may reflect such fires spreading in sparsely wooded landscapes with many disconnected stands. The occurrence of clustered patches within the intact forests of the tropics (Figs. 2a and 3, and Supplementary Fig. 12) and their overall association with a higher forest landscape integrity index, an indicator of human influence on forest landscapes56 (Supplementary Fig. 11), suggest that these patterns are probably more commonly the results of natural processes, rather than anthropogenic activities.

Fig. 4: Proportions of dominant agents by forest biome for each disturbance pattern based on counts and areas.
figure 4

The ‘other’ category represents non-identified dominant drivers, which may include natural disturbances, such as biotic outbreaks and windthrow. The clustered pattern shows a distinct distribution among biomes, with a substantial association with fire in the boreal, Mediterranean and temperate conifer biomes, and an association with other drivers in the tropical coniferous biome. The large-multiyear pattern also shows a substantial association with fire and harvest, particularly in the boreal and Mediterranean biomes for the former and the tropical broadleaf biomes for the latter. Bars with values less than 1 are not labelled.

Full size image

Complex patches predominate in temperate and tropical regions where human activities are prevalent (Fig. 2a and Supplementary Fig. 13). This is particularly visible in areas where tree farming, urbanization and shifting agriculture are occurring (Fig. 4 and Supplementary Fig. 13). The link with human activities is also supported by the association with lower values in the forest integrity index (Supplementary Fig. 11). Although forestry, perhaps the most common anthropogenic disturbance, is often associated with simple geometries, such as a square or a rectangle, the high complexity measured may be the result of intricate and/or interconnected patches that complicate the shapes (for example, dendritic and fishbone patterns taken by deforestation in the Amazon; elongated rectangles of harvesting in Poland; checkerboard pattern of clear-cutting in Russia; linear features, such as roads and tracks carved within forests; and where the density of different activities may lead to intricate shapes).

Large-multiyear patches are mainly found in boreal and Mediterranean regions in association with fires and outside intact forests in the tropics, in areas dominated by harvest (Figs. 1, 2a, 3 and 4, and Supplementary Fig. 12). Notable regions include western North America, where harvest is mixed with fire and biotic outbreaks; southeast North America, where harvest is combined with hurricanes57; and tropical Southeast Asia and the southeastern Amazon where fire is used for forest clearing58. They may also be associated with regions where patch detection may be limited by cloud obstruction, generating an erroneous multiyear composition (for example, tropical cloud forests). Multiyear patches may also reflect natural long-term processes, possibly biotic spreading25,26 and cascading effects from wind or wildfire59.

Implications for forest ecosystems

Our global geography of disturbance patch structure provides insights into how forests are undergoing changes. While in some cases the observed patterns fit well to received knowledge (for example, the belt of large-multiyear patches in the boreal where large fires are known to be the major driver of disturbances)55,60, in others, the patterns do not align with our understanding of how unmanaged forests would behave. For instance, the small-isolated patches dominating in the remote forests of the Amazon (Fig. 2a) are consistent with the known structure of the lowland continental rainforest ecosystem, characterized by large continuous areas of mixed-age forests, with limited amounts of natural regrowth stands47,61,62. Also, the larger proportions of large-multiyear patches, compared with the intact forests of the African and Indomalayan realms (Supplementary Fig. 12b), are expected with the increasing vulnerability to wildfire and extreme weather events reported in the Amazon as a result of El Niño, climate stress and anthropogenic pressures58. However, in the Congo basin in Africa, the smaller rainforest fragments, dissected by road networks and encroaching farming, show an overall tendency towards large-multiyear patches (Fig. 2a), well above the typical small size of shifting agriculture60,63. While Amazonian and African rainforests differ in structure and composition64, the broad characteristics of these forests remain similar65 and the larger disturbance patches found imply that a fundamental change to the forest structure is underway. In tropical forests, small-isolated patches can positively influence biodiversity, as they contribute to increased light penetration in a moderate way and generate structural complexity within the environment, thus increasing habitat diversity and richness66, whereas large-multiyear patches are likely to clear large areas, changing both the types of tree species that will repopulate them, with long-term implications for food sources and habitats67,68. Broad-scale maps of tree species functional strategies are not yet available across the tropics, but we hypothesize that a comparison of our maps with species life-history strategies would reveal large regions where the distribution of early and late successional species across the landscape is not consistent with the prevailing disturbance patch structure.

There are several instances in which a climatically and pedologically similar region is bisected by a marked change in dominant patch structure, for instance, the borders between Finland and Russia (Fig. 2b and Supplementary Fig. 4). All these cases are characterized by a shift in dominant patch characteristics from a relatively uniform structure across the country to a more varied regime. Finland has much simpler patch shapes than the neighbouring region of Russia, and those patches also tend to be more clustered in time. Such changes along country borders are indicative of how forest management practices have fundamentally altered the disturbance regimes. Forests in both Sweden and Finland are very intensively managed69, and it is notable how the dominant cluster within the country is much more uniform than in their less intensively managed neighbours. Forests known to have little harvesting activity, such as the Amazon47 or where natural disturbances are still the dominant regime, such as the northern Canadian boreal forest12, tend to show much less uniform spatial patterns at the regional level. This has similarities with the process of biotic homogenization that has been reported as a result of transportation of species by humans70 and is anticipated as a result of climate change71,72. This suggests the hypothesis that regions with more uniform disturbance patch structures are both more biotically homogeneous and more inclined to biotic homogenization73.

Counter to this idea of regional homogenization, however, we also found more diversity in patches at the landscape scale (that is, within the 0.1° grid cells) in intensively managed areas, including not just Finland and Sweden, but also much of Europe, southwestern United States and southern China (Fig. 2a)20. The higher patch diversity reflects a superposition of relatively regular harvest interventions and natural disturbance regimes. This will ultimately have a diversifying effect on forest landscape structure, relative to the natural state, but whether the increased habitat variability will result in higher diversity of flora and fauna will depend on the intensity of management (for example, the use of exotic monocultures and removal of deadwood) and its spatial composition and the traits of the species. A review of recent empirical studies has shown that higher patch structural diversity results in greater species diversity at the landscape scale74.

Pervasive anthropogenic signature

This Analysis constitutes a comprehensive description of the characteristics of forest disturbances across the globe from a patch perspective. The patterns identified provide a new perspective on forest biogeography, revealing how the realized form of one of the key drivers of forest structure varies dramatically across the globe, emerging from the likely interplay of physical, biological and anthropogenic drivers. Four broad classes of disturbance patterns were identified integrating across patch areas, shapes, spatial arrangement and temporal continuity. These classes represented a similar fraction of disturbances across different forest biomes, but there was spatial variation in the dominant class within individual biomes. Variability within biomes emerges from both widespread anthropogenic modification of forests and natural variability within the broad classification of biomes, indicating that broad generalizations about disturbance patch structures by biome are probably inappropriate.

We found indications that human activities (harvest or small agriculture) may leave a consistent and uniform mark on landscapes, independently of the biome and regional climate, characterized by patches that are either complex or large and spread across multiple years. Such consistency would imply that human activities are leading towards a structural homogenization of the world’s forests, with potential consequences on forest functions and the diversity of species they host. However, the increases in patch diversity in heavily managed ecosystems suggest that even when humans have enacted dramatic structural changes to forests, the characteristics of the natural disturbance regime are not entirely wiped out. Our results also suggest that natural agents, such as storms and fire, may leave similar structural signatures across regions assuming they are under similar bioclimatic conditions, as reported across protected forests under temperate and boreal climates17,18.

The data we generated and combined for individual disturbance patches provide a rich dataset for further research, notably for quantifying disturbance regimes across regions, discriminating anthropogenic from natural processes and potentially elucidating the causal agents (for example, refs. 75,76). Such information on differences in disturbance characteristics between intact and non-intact forests can also provide a powerful reference for guiding the implementation of nature-based forest planning and management.

Methods

Tree cover loss data

We used the Landsat-derived GFC data version 1.6 to delineate tree cover loss patches between 2000 and 201810. This dataset captures annually any complete or near-complete loss of woody vegetation that is taller than 5 m (considered as trees) and within a 1 arcsec resolution pixel (around 30 m at the equator). The loss detected is relative to 2000, whether the tree cover was natural or planted. The GFC data were processed using median observations from high-quality Landsat imagery acquired during the growing season between 2000 and 2018 and a bagged decision tree algorithm10. Years of tree cover loss were disaggregated by pixel using a set of heuristics derived from the maximum annual decline in both percentage tree cover and minimum growing season normalized vegetation difference index (NDVI)10. Thus, pixel values represent the year when the loss occurred the first time since 2000. Disturbances that follow this first loss (for example, due to seasonal agents, such as wildfire) are ignored. No initial canopy density threshold was applied while mapping the change (that is, losses in open-canopy forests and other treed vegetation communities are included). Disturbances occurring late in the year, after acquisition of the annual Landsat data, may be assigned to the following year.

At the global scale, the overall accuracy reported for the first version of the tree cover loss product is 99.6 (based on 1,500 sample blocks, 120 m in length per side)10. The prevalence of errors varies by biome and is equal to 13% for false positives (commission errors) and 12% for false negatives (omission errors)49. However, no uncertainty layer is provided in the data published for pixelwise quality filtering. Also, the validation of the loss classification did not include results from Landsat 8 images.

Projection

To calculate patch shape metrics accurately, we projected the global tree cover loss rasters with the Universal Transverse Mercator (UTM) projection system. The latter consists of a group of projections that are parametrized by geographic locations. These UTM zones span every 6° of longitude and are further subdivided every 8° of latitude to form a UTM grid zone77. The UTM projection is conformal, allowing the conservation of shapes and thus the retrieval of pixel connectedness77. Area distortions are negligible within the margins of a UTM zone, but accumulate more in higher latitudes because of more tiling (Supplementary Fig. 1). Splitting the global data by UTM grid zones also allowed optimizing computation time by parallelizing data processing among tiles. The UTM grid zone dataset was downloaded from the ESRI website78 and consists of 1,201 tiles, among which 489 contained tree cover loss information (Supplementary Fig. 1). Tiling was performed in Google Earth Engine79. Exported tiles were reprojected by UTM grid zone at a resolution of 30 m using the ArcPy library from ArcGIS Desktop 10.7 (ref. 80) and Python 3.8 (ref. 81).

Pixel correction

To accurately retrieve the true shape and size of individual forest loss patches and at the same time identify patches that are growing over years, we applied, before patch delineation, a correction of the year assigned to the original tree cover loss pixels (Supplementary Fig. 2). The aim was to identify spatially and temporally contiguous pixels reflecting tree cover loss as a single composite patch. This allowed us to correct potential interannual splitting due to delayed detection resulting from the Landsat ETM+ Scan Line Corrector anomaly82, cloud obstruction, events occurring late in the year (for example, December) or residual loss around patches. Grouping these tree cover loss areas together also enabled us to capture those processes that are moving and operating over multiple years, which manifest in our data as patches that are gradually growing outwards83. These include, among others, biotic outbreaks, plantation expansion with annual harvesting of neighbouring forests or cascading events, such as wind damage followed by insect infestations. The algorithm applied used R 4.3.1 (ref. 84) and C++11 (ref. 85) to identify loss pixels that are in consecutive years and adjoining each other. It then reassigned them to the most frequent year using the mode statistic at each iteration. Where there were equal frequencies (ties), the year with the lowest value was retained.

Patch delineation

We used the corrected tree cover loss rasters to delineate patches globally by tile. The delineation was conducted in R using an eight-cell (queen) neighbourhood rule. It was performed regardless of canopy density percentage, as our objective was to capture the spatial imprint left by a disturbance on the treed landscape. The minimum patch size was approximately 0.09 ha (30 m pixel resolution).

In total, 344,893,162 tree cover loss patches were identified for the period 2001–2018, including land use change. Individual patches with extreme values were checked by visual interpretation using aerial imagery in Google Earth Pro 7.3 (refs. 86,87 and Supplementary Fig. 6). The patch that has the largest area (221,025.5 ha) and also the longest perimeter (26,869,920 m) is a wildfire that occurred in 2002 in boreal eastern Siberia. The most elongated patch (0.995) is a road that was not large enough to be excluded with the ESA CCI Land Cover data39. The patch with the largest number of surrounding patches (4,155) is also a wildfire that occurred in Mediterranean Australia in 2010. The patch spanning the longest period (8 years) corresponds to a slash-and-burn agricultural expansion in Cambodia.

Limitations of the patch correction and delineation approach

A limitation of the patch correction applied with the majority vote (that is, most frequent year) is that it caps the size of growing patches after the number of new pixels (that is, years) to add becomes a minority. In fact, the merging of subsequent patches is chronologically unidirectional starting from 2001 and ending up until the number of new pixels to merge becomes smaller than the pixels already merged, which breaks the sequence with other potential temporally consecutive pixels, which can make the count of years (growth) smaller than it should be. It does, however, avoid the case in which a large patch continues to grow indefinitely by a few pixels each year, which would not be consistent with the intention of using this temporal metric to identify patches that grow substantially over multiple years (Supplementary Fig. 2b). Such cases of long serial merging are, in any case, rare in our database, as most patches that have been merged (count of years > 1) represent 8.17% of the global database (28,190,664 over 344,893,162 polygons), with a mean of 2.14 for the count of years of these merged patches. Composite patches counting more than 3 years represent only 1% of the database (354,539 polygons).

The gridding of global tree cover loss rasters with the UTM grid zones artificially split disturbance patches located at the edges of the UTM tiles, which may increase the number of patches identified and reduce their calculated sizes, thus affecting derived patch metrics. Before deploying the patch delineation globally, we tested on a smaller region whether a correction across tiles (for example, moving window) was needed. We selected a region susceptible to this artefact: the highly disturbed North American boreal biome, where patches tend to be larger and in higher numbers and thus more likely to be found at the edges of tiles. We found that the number of patches that were broken apart by the tiling there represented only 0.025% (6,197 polygons crossing the margins, over 24,315,498 polygons). We considered this proportion negligible and thus applied no corrections for this artefact.

Metrics calculation

To quantify patch structure, we considered several candidate metrics measuring magnitude (size and growth), shape complexity and context (spatio-temporal clustering) (Supplementary Table 1). Corrected patches were first polygonized using the ArcPy library in ArcGIS Server 1.7.1 (ref. 88). The geodesic area, perimeter and radius of the minimum circumscribing circle were then calculated. Shape indices89 (that is, perimeter-to-area ratio, shape index, elongation index and fractal dimension index) were derived from these primary metrics. Patch growth was extracted during patch correction by counting the number of years composing the corrected patches. Patch context metrics were calculated within 5 km of the patch internal centroids using R and the UTM-projected rasters.

Isolating disturbances from deforestation

We used the ESA CCI Land Cover product39 to isolate disturbances (temporary loss patches) from land use change (persistent loss patches caused by anthropogenic activities). As the resolution of ESA land cover is 10 times coarser than the GFC product (300 m versus 30 m), we excluded only large forest conversions that are at a scale detectable by the ESA product and considered smaller events as disturbances (including shifting agriculture and selective logging, which are small and generally non-permanent). Areas of large forest conversion were identified as pixels, classified as forests in 2000 (cover classes 50, 60, 61, 62, 70, 71, 72, 80, 81, 82, 90, 100, 160 and 170), which were changed to non-forest-related, assumed human-modified, covers at the end of the period studied (cover classes 10, 11, 20, 30, 110, 130 and 190, corresponding mainly to croplands, pasturelands and built-up areas). Corrected tree cover loss patches were then labelled as land use change if 10% of their area overlaps with these large forest conversion pixels. We increased the sensitivity to overlapping because, considering the coarse resolution of the land cover data (9 ha), we preferred to exclude a maximum of land-use-related patches and minimize cases in which conversion and disturbances were mixed to not obscure our conclusions with deforestation patterns. Thus, our approach excludes cases in which land use directly causes neighbouring disturbances, such as fire from land clearing spreading into nearby forests. We further conducted a sensitivity analysis experimenting with different thresholds (10%, 25% and 50%) and comparing how many patches were classified as land use across different biomes. To check the case of a fire from human activities spreading into intact forests, we took a subset of large patches (>1,000 ha), whose majority area is within intact forests, but their portion outside fulfil the land use attribution condition. The number of patches classified as land use with the 10% threshold and extending inside intact forests was very small (1 for the boreal forests and 6 for tropical moist forests) (Supplementary Table 3). This supported our decision to keep the 10% threshold.

Ensuring spatio-temporal consistency

After applying the patch correction algorithm, we observed an abrupt increase in mean patch sizes for 2015 onwards across all biomes (Supplementary Fig. 3). The GFC tree cover loss time series data are not consistent over time beyond 2012, as they have been affected by an increase in detection sensitivity, marked after 2015, due to the combined effects of adjustments in the classification algorithm, improvements in the Landsat instruments and increased image richness90,91. The increase in sensitivity concerns partial and short-term changes in tree cover, including residual loss and delayed tree mortality around disturbance patches. As these are more prevalent in the new GFC tree cover loss dataset, this led to more pixels being grouped together by our algorithm as a single patch, leading in turn to abruptly larger patches after 2014. We therefore focused our analysis on the years between 2002 and 2014, during which the mean corrected patch area remained stable (Supplementary Fig. 3), and considered the patch metrics derived during this period as viable indicators. Regarding the correction applied, we decided to keep the areas added to the composite patches beyond 2002–2014, as we consider them reflecting the full extent of the disturbance processes.

We expect that the spatial variability in Landsat data richness13 would not affect our results as much as the temporal inconsistencies, as we are not using directly the year attribution of the loss in our analysis. In fact, we assume that any spatial bias would be minimized by the temporal flattening applied for delineating patches. In areas affected by a lack of observations due to cloud or snow obstruction, patches may be detected with a 1 or 2 year delay, but they should in most cases still be detected. If a patch is partially detected, the algorithm used to merge contiguous patches will allow us to retrieve its true and realized size and shape by merging it with its other parts detected in the following years. However, in areas where regrowth may be fast enough to cover the patch entirely or partially before it is detected in the following years, this would result in patches delineated with a smaller size, fractured, or in a smaller number, with single-pixel disturbances probably missed. This situation is most likely to occur in dense-canopy rainforests, particularly in the tropics, where regrowth may be fast.

This being said, we consider that overall, there are sufficient usable observations over the period studied to minimize the errors described above and expect these would not affect our conclusions at regional and biome scales. This is all the more supported by the various patch types found across rainforests in the pantropical regions (for example, small-isolated and clustered in the Amazon, small-isolated inside intact forests and large-multiyear outside intact forests in Borneo), which agree more with the literature describing disturbances in these regions (for example, blowdowns in the Amazon, oil palm plantations in Indonesia) (Fig. 2a). Also, considering the Congolian lowland forests in western Africa, one of the regions with the least usable observations13, we can retrieve similar patterns, with small-isolated and clustered patches inside the intact forests there and more complex multiyear patches outside the intact forests. We therefore conclude that our results and conclusions are robust to any variations in spatial consistency of disturbance classification errors.

Metrics selection and cluster analysis

We first selected the metrics that are meaningful for capturing the structural aspects considered: magnitude, complexity and clustering (Supplementary Table 1). We then assessed the correlation between these metrics to exclude those that are strongly multicollinear. This is needed to avoid conceptual redundancy between factors and consequent overweighting in the clustering, as well as to reduce dimensionality and thus data sparsity. We measured correlation using R and the Spearman method92, as some metrics were severely skewed and thus not fulfilling the normality assumption. We set 0.75 as the threshold for the mean absolute correlation coefficient beyond which the metrics are considered multicollinear. We chose a high value for the threshold, as patch metrics were derived from each other and thus tend to be correlated, although reflecting subtle differences in structural traits. The area, perimeter, shape and fractal dimension indices showed high correlations (Supplementary Fig. 7). We thus excluded the perimeter and fractal dimension metrics from the clustering. However, we kept the area metric despite its multicollinearity, as we consider it an essential descriptor of patch types. We also preferred the shape metric over the perimeter-to-area ratio, as it is corrected for patch size differences. This also resulted in a balanced representation of complexity with two metrics for each conceptual class of magnitude, shape and context (Table 1). We then standardized the observations for each metric by subtracting the mean of the metric and dividing the result by the standard deviation. This allowed us to have equivalent variance among the variables and thus avoid building clusters only with the metrics having the highest amount of variation.

To identify natural grouping of patch structures, we used R to apply unsupervised clustering with k-means40. This method partitions data into a specific number of clusters by minimizing distances between observations and the centroid of clusters to which they are assigned40. We selected this algorithm for its ease in interpretation and implementation on large datasets. We specified Euclidean distances for measuring dissimilarities and 100 as the maximum number of iterations. For selecting the optimal number of clusters for the grouping, we tested numbers between 1 and 10 on samples of 10,000 patches randomly selected by forest ecobiomes and 50,000 patches globally. We checked the goodness of the clustering using silhouette coefficients and the Dunn index with 100 bootstraps. The optimal number of clusters varied substantially depending on regions and the patterns that dominate there. We decided to use four clusters, as it resulted in a stable and interpretable grouping, associated with substantially different metric means at the global scale, while also maximizing cluster separability and compactness (for the global sample, the average silhouette width = 0.42).

Ancillary data calculation

To explore linkages with environmental characteristics, we associated each patch with the following datasets:

  1. 1.

    Ecoregions, ecobiomes and biomes: Ecoregions201793 (polygons)

  2. 2.

    Forest intactness: Intact Forest Landscapes 201641 (polygons)

  3. 3.

    Forest types in 2000: ESA CCI Land Cover39 (raster 300 m)

  4. 4.

    Fire: GFC Fire58 (raster 30 m)

  5. 5.

    Drivers: dominant agents60 (raster 10 km)

  6. 6.

    Forest landscape integrity index56 (raster 300 m)

  7. 7.

    Elevation: JAXA ALOS DEM94 (raster 30 m)

  8. 8.

    Countries: GADM95 (polygons)

Ancillary rasters with a resolution smaller than 1 km were split with the UTM grid zones, reprojected accordingly and snapped to the extent of the projected disturbance rasters, so as to align with the patch pixels. Zonal statistics were applied using ArcPy. Each patch was attributed either the mean (elevation and forest landscape integrity index) or the most frequent value (ecoregions, forest types) of the ancillary data. A cover percentage threshold was used for attributing fire (25% of the patch area) and forest intactness (50%). For ancillary rasters with a resolution coarser than 1 km (that is, dominant agents), patches were assigned the values extracted at the centroid of the patch using R.

Mapping

To visualize the distribution of patch structural patterns, we built grids of patch statistics for every 0.1° grid cell. Statistics were summarized depending on the patch internal centroid using R. We quantified the relative importance of patterns based on the total count of patches and the associated total disturbance area. We used the Shannon–Wiener diversity index to measure the diversity of patterns, accounting for the evenness of their abundance and emphasizing rare types. Forest total areas were calculated per 0.1° grid cell, accounting only for tree cover with a canopy density greater than 10% as measured for 2000 in the GFC product. The raster was produced in Google Earth Engine and used to normalize disturbance patch counts per grid cell. Map rendering was performed in ArcGIS Desktop.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.