Declining human pressure and opportunities for rewilding in the steppes of Eurasia

Large and ecologically functioning steppe complexes have been lost historically across the globe, but recent land‐use changes may allow the reversal of this trend in some regions. We aimed to develop and map indicators of changing human influence using satellite imagery and historical maps, and to use these indicators to identify areas for broad‐scale steppe rewilding.

Both locating candidate areas for grassland rewilding and measuring rewilding progress remain challenging. In part, this is because rewilding research has often focused on forests (Jepson, 2016), and adequate tools and datasets for steppe regions are often missing.
Furthermore, while the restoration potential of temperate grasslands has repeatedly been assessed, existing work has typically focused on small regions or individual sites within agricultural landscapes (Fuhlendorf et al., 2018). We are not aware of any assessment for rewilding opportunities at the broad scales needed to establish connected and self-regulating grassland complexes. Recently, considerable progress has been made in framing what such assessments could look like. Specifically, two fundamental dimensions need to be considered when identifying rewilding opportunities and progress: changes in human influence and changes in ecological integrity (Torres et al., 2018), where the latter can be framed to collectively capture changes in disturbance regimes, connectivity and trophic complexity (Perino et al., 2019). To our knowledge, no study has yet applied this framework to any grassland region of the world.
Identifying and mapping indicators capturing different aspects of rewilding can reveal priorities for conservation planning. For instance, protected areas in many grassland regions are typically sparse and isolated from each other (Saura, Bastin, Battistella, Mandrici, & Dubois, 2017;Saura et al., 2019). Passive rewilding might provide opportunities to enlarge protected areas, to expand protected area networks by adding new reserves or to establish corridors to restore and maintain connectivity between protected areas (Perino et al., 2019). Embedding protected areas in landscapes where human pressure is declining and rewilding is taking place is also important, as protected areas can contain source populations of conservation-dependent species (Wolf & Ripple, 2018). Such protected areas can serve as starting points for range recolonization where rewilding leads to increasing habitat availability and reduced human-induced mortality (e.g. due to hunting). It is therefore important to map indicators capturing key dimensions of rewilding and relate them to the spatial distribution of current protected areas and potential corridors linking these.
The Eurasian steppe is particularly interesting in the context of rewilding opportunities. This region, stretching from Eastern Europe to the Altai mountains, is situated nearly entirely within the former Soviet Union and contains the vast majority of Old World Steppe captures key components of rewilding and can help to devise strategies for fostering large, connected networks of protected areas in the steppe.

K E Y W O R D S
agricultural abandonment, ecological integrity, human pressure, Landsat, landscape connectivity, passive rewilding, steppe restoration (Wesche et al., 2016). Remnant populations of large grazers, such as the critically endangered saiga antelope (Saiga tatarica) or kulan (Equus hemionus kulan), still roam these steppes or have been recently reintroduced (Kock et al., 2018;Robinson, Milner-Gulland, & Alimaev, 2003). The region provides critical stopover habitat for Eurasia's migratory birds and hosts large populations of many species that are of high conservation concern in Western Europe Rounsevell, Fischer, Torre-Marin Rando, & Mader, 2018).
However, so far there has not been an assessment of the broadscale spatial patterns of declining human influence in these steppes that would allow the formulation of rewilding visions. Kazakhstan has an ambitious programme to expand its protected areas system considerably, starting in 2010 (Kamp et al., 2015), so there is now a "hot moment" (sensu Radeloff et al., 2013) for conservation in Kazakhstan. At the same time, protected areas in Kazakhstan are typically very large, understaffed and underfunded. Identifying places where human influence is declining or low might therefore help to establish a manageable network of protected areas in the country (sensu Pringle, 2017).
Our overarching goal was to develop and map rewilding opportunities, using the Landsat archives to map post-Soviet land-use change across large areas (Dara et al., 2018;Yin et al., 2018), as well as high-resolution Google Earth imagery together with historical maps to monitor additional indicators of changes in human influence in steppes. Focusing on northern Kazakhstan and the period 1990-2015, we asked: 1. What were the patterns of post-Soviet changes in cropland area, livestock density and human population density across the steppe? 2. Where are the steppe areas that have seen declining human influence, which might therefore undergo passive rewilding?
3. How has declining human influence affected the connectivity among protected areas in the region?

| Study area
Our study region comprises three provinces in northern Kazakhstan (Kostanay, Akmola and North Kazakhstan oblasts), covering 38 million ha (Figure 1). The region extends across three ecoregions, namely forest steppe, steppe and semi-desert (Olson et al., 2001).
The region has a rainfall gradient from 400 mm annually in the north to 200 mm in the south. The climate is continental with average temperatures of 22°C in July and −18°C in February (Afonin, Greene, Dzyubenko, & Frolov, 2008).
Across the ecotonal forest steppe, there is a mosaic of herb-rich meadows and forest patches composed of birch (Betula pendula) and Scots pine (Pinus silvestris). Historically, the overall ecological conditions across the steppe belt have been remarkably stable across the past 18,000 years (Tarasov et al., 1998(Tarasov et al., , 2000. During the last glacial maximum, steppe was already the dominant vegetation type across the wider region and steppes also occupied a much larger area in the European and southern Siberian parts of Eurasia, extending further north than today (Tarasov et al., 2000). From the early Holocene onwards, birch and pine spread northwards and formed the forest steppe ecotone found in northern Kazakhstan today (Tarasov, Jolly, & Kaplan, 1997). Generally, the distribution of steppe and forest patches is driven by soil factors and disturbance such as fire and F I G U R E 1 (a) Cropland dynamics in the study area, mapped from Landsat images. (b) Location of our study area in north-central Kazakhstan grazing. The steppe belt is a "brown black region," in which grazing and fire are the main consumers of biomass (Bond, 2005).
Historically, nomadic pastoralism was the main land use across the Eurasian steppe. The first signs of horse domestication date back to the Botai culture in northern Kazakhstan 5.500 BP. Nomadic pastoralism developed around 3.200 BP (Hanks, 2010), involving movements between the more productive steppes in the north that were grazed in summer and the southern, more arid steppe with dwarf shrub vegetation and little snow cover that were grazed in winter.
This form of pastoralism persisted until the 1930s, when nomads were largely forced into a sedentary lifestyle by the Soviet rulers . Later on, between 1954 and 1963, over 5 Mha of Kazakh steppes was converted to croplands during the Soviet "Virgin Lands Campaign" (Kraemer et al., 2015), and livestock numbers increased strongly again, primarily in form of large state farms (Yan et al., 2020). After 1991, as a result of institutional change, diminishing support for agriculture and large-scale human outmigration (Lesiv et al., 2018;Schierhorn et al., 2013), at least 48 million ha of cropland was abandoned across Russia and Kazakhstan alone (Lesiv et al., 2018). In Kazakhstan, grazing livestock numbers decreased by as much as 70% (Lioubimtseva & Henebry, 2012;Schierhorn et al., 2016), while remaining livestock were increasingly concentrated around larger settlements (Hankerson et al., 2019;Kamp et al., 2015).

| Mapping changes in cropland extent
To map changes in cropland extent, we generated Landsat image composites for the years ca. 1990 (i.e. the end of the Soviet era), ca. 2000 (first decade of the transition period, and the period when land-use intensity decreased heavily) and ca. 2015 (current situation, after a partial revival of the agricultural sector). Image composites are gap-and cloud-free mosaics based on Landsat images (Griffiths, Jakimow, & Hostert, 2018). For each of the three time steps, we calculated three composites centred on spring (Julian day 121), summer (day 180) and fall (day 260) to capture phenology differences that are important for mapping cropland-grassland dynamics (Baumann et al., 2011). We also calculated a set of spectral-temporal metrics for which we considered all available cloud-free observations for each year.
We gathered training data through on-screen digitization of high-resolution images in Google Earth, visual examination of the Landsat composites and land-use information collected in the field (see Dara et al. (2018) for details). We then classified our Landsat image composites using random forests, a nonparametric machine-learning technique (Breiman, 2001). Finally, we applied a minimum mapping unit of 10 Landsat pixels (equal to 0.9 ha) and validated the resulting land-cover map using 100 randomly sampled points per class, following best-practice protocols (Olofsson et al., 2014). Our land-cover change map had an overall accuracy of 86.3% (for details on the accuracy assessment, see Text-S1 and Table SI-1).

| Mapping changes in human population density and livestock distribution
We used human population and livestock numbers as proxies for overall human influence on steppes. To map human influence, we assessed changes in the extent and condition of settlements and livestock stations from the Soviet period until today. Livestock stations in the study area are outposts where livestock are concentrated in summer ("Letovki") or winter ("Zimovki"). These stations usually consist of up to three houses or tents ("yurts") for shepherd accommodation, stables and corrals. To assess changes in settlement and livestock station density, we digitized both for circa 1984 (representing infrastructure in the Soviet period) and circa 2012 (representing the current situation). For the Soviet period, we manually digitized settlements and livestock stations across the study region from georeferenced, declassified Soviet military topographic maps scaled 1:200,000 and labelled them as "fully intact" assuming that abandonment did not start prior to 1991, in-line with estimates of livestock numbers which declined only after 1990 in Kazakhstan . For the current situation, we used publicly available, high-resolution satellite images (2.5-m resolution or higher) in Google Earth and Bing Maps (for details see Koshkina et al., 2019), to determine the level of intactness of settlements and livestock stations (10% intact, 20% intact, etc.; see Text-S2 for further information).

| Changes in human influence across the steppe
Using our maps of cropland extent, grazing stations and settlements, we mapped changes in human influence (Carver, Comber, McMorran, & Nutter, 2012) from 1990 to 2015. We generated three layers with a common spatial resolution (300 m; 10 × 10 pixels in our Landsat-based land-cover map) that contained (a) the share of cropland per grid cell, (b) the distance to settlements and (c) the distance to livestock stations. We scaled the values from 0 to 1 such that higher values represented higher human influence (e.g. areas near active livestock stations and settlements). Next, we combined these three layers into a "human influence index," comparing two alternatives: (a) the product of the three layers (assuming overall pressure is the combined impact of these pressures) and (b) the average of the three layers (i.e. assuming additivity of pressures; see Appendix S1).
This index assumes that lower human influence is beneficial from a passive rewilding perspective. Last, we calculated changes in our human influence indices from 1990 to 2015, which captures the extent to which areas are might undergo passive rewilding-ranging from 0 (low) to 1 (high).

| Changes in landscape connectivity
To determine how changes in human influence impacted connectivity, we assessed landscape connectivity across our study region using circuit theory (McRae & Kavanagh, 2012). Circuit theory describes the movement of individuals through a landscape by considering all possible pathways between grid cells of a resistance surface (i.e. our maps of human influence). Each pathway can be interpreted as a current. Grid cells that are part of many pathways thus have a higher current density compared to grid cells that are part of fewer pathways. The cumulative current density map of all pathways can be interpreted as overall landscape connectivity (Koen, Bowman, Sadowski, & Walpole, 2014). We applied the CircuitScape algorithm , which calculates the current between nodes. The nature of the algorithm causes the highest current densities to be found around these nodes (Koen et al., 2014;McRae, Dickson, Keitt, & Shah, 2008). In the context of predicting functional connectivity across a larger region, this can result in biased results, particularly if distances between some nodes are small. We avoided this problem by randomly placing nodes within a buffer region outside our study area, and we followed Koen et al. (2014) and Leonard et al. (2017) to find the best combination of (a) the buffer width around our study region and (b) the number of nodes within the buffer, by subsequently increasing the number of points and buffer widths. For each combination, we calculated current density maps and compared them to previous maps (i.e. with less points and a smaller buffer) using Pearson's r.
We stopped increasing buffer width and number of nodes when our current density map did not change compared to the previous map, and we defined this as when r exceeded .98. The final buffer width was 50% of the study region extent, and the final number of nodes in that buffer was 60. Based on this parameter combination, we built current density maps for 1990 and 2015, and calculated the difference between the two maps (for more details on the connectivity analyses, see Text-S3).

| Rewilding effects on protected area connectivity
We compared changes in our human influence index to the network of protected areas and assessed the location of protected areas relative to areas of higher landscape connectivity. In addition, we evaluated how older (i.e. established prior to 1990) and newer (i.e.

| Changes in connectivity
The changes in human influence also resulted in marked changes in landscape connectivity. Landscape connectivity mostly increased in Kostanay, whereas in North Kazakhstan and Akmola such changes were not widespread (Figure 4). Relative to protected areas, decreasing human influence resulted in increased connectivity

| D ISCUSS I ON
The world's temperate grasslands have historically been transformed due to land-use change on a large scale. Rewilding large grassland complexes that are characterized by natural disturbances, high connectivity and trophic complexity, and that thus foster the interactions between fire and native grazers that have shaped grasslands for millennia, is a bold conservation vision (Fuhlendorf et al., 2018). Post-Soviet changes in land use may provide opportunities for turning such visions into reality across the Eurasian steppe, which contains some of the largest remaining stretches of temperate grasslands in the world. However, adequate spatial data for identifying where human pressure has declined were missing. It was hence unclear where passive rewilding might currently take place or where restoration interventions, such as reintroducing large native grazers, could take place.
Focusing on a 38 million ha region of the Eurasian steppe in Kazakhstan, we used a novel approach to map an aggregate measure of changes in human influence, based on changes in cropland extent, grazing pressure and human population density. Our study provides three main novel insights. First, our analyses highlight a massive decline in human influence following the collapse of the Soviet Union in 1991, with more than 6 million ha of cropland abandoned. In addition, we detected that 97% of livestock grazing stations and more than 90% of settlements were substantially reduced in size or completely dismantled. This major decline in human influence suggests substantial potential for restoration and conservation. Second, our analyses highlight that areas of ongoing passive rewilding have the potential to link existing protected areas. Protected areas in our study are sparse and isolated, and recent trends can help establish a protected area network that benefits a wider array of species, such as large ungulates (Figure 4b), and natural processes, such as grazing-vegetation-fire interactions, than are currently protected.
Finally, while our study highlights major conservation potential, the window of opportunity for implementing such broad scale protected area networks, and bold rewilding visions more broadly, may soon close. This is because recultivation of abandoned cropland is gaining momentum, Kazakhstan's population is growing , and foreign investment is increasingly directed at natural resources F I G U R E 3 (a) Human influence index 1990 and 2015 as well as changes therein for our study area (we here show human influence calculated as is the product of the three input layers). (b) Human influence index within the protected area network. A list with full names of the protected areas and their location is provided in the Appendix S1 across the Kazakh steppe (e.g. via China's new silk road initiative (Fallon, 2015)).
Cropland abandonment and outmigration of the rural human population happened across the former Soviet Union, but here we show that human influence declined particularly strongly in the steppes of Kazakhstan. Across the former Soviet Union, cropland abandonment was high in European Russia, Belarus and Ukraine (Prishchepov, Radeloff, Dubinin, & Alcantara, 2012;Schierhorn et al., 2013), with abandonment rates of over 40%, and the cropland abandonment rates we document for northern Kazakhstan were similarly high (over 40% in Kostanay). Importantly though, ours is the first study that quantifies the additional, massive decline in grazing pressure-for any steppe region in the former Soviet Unionwith an area footprint many times larger than that of cropland abandonment ( Figure 3). In the case of north-central Kazakhstan, the main drivers of these trends were large-scale human outmigration , the transition from state to market-driven economies (Rozelle & Swinnen, 2004), which made crop production unprofitable, as well as the collapse of state-owned farms. While agricultural sectors rebounded to some extent after 2000, many of the land-use changes that have happened since 1991 are likely to persist, for example because Soviet-era agriculture expanded onto marginal areas, or because infrastructure has been dismantled since 1991. Croplands have been partly recultivated, as we show here, but a redistribution of grazing stocks to abandoned regions has not yet been detected (Hankerson et al., 2019). Northern Kazakhstan should therefore be a priority region for active or passive rewilding with the goal to restore steppes.
Our analyses capture changes in human influence, one of two core dimensions along which to assess progress towards rewilding (Torres et al., 2018). While we did not directly measure all components of ecological integrity recently suggested as pivotal (Perino et al., 2019), three lines of evidence suggest that many of the areas highlighted in our maps are indeed undergoing passive rewilding.
First, regarding connectivity, our analyses highlight that declining human influence might have resulted in increased landscape connectivity among steppe patches, as well as among protected areas ( Figure 3). This increased connectivity should immediately benefit the movements of large grazers such as saiga antelope, which can now roam over larger areas with less human disturbance-a key factor influencing their distribution (Singh & Milner-Gulland, 2011).
In time, increasing connectivity of steppe habitat can also help grassland species that have suffered from conversion over long periods and might now rebound and expand their range towards  (Munteanu et al., 2020). Likewise, increased connectivity of steppe areas will allow plants and animals with low dispersal ability to recolonize regions where they are extinct due to historical steppe conversion, such as endemic tulip species (Tulipa spp., (Brinkert et al., 2016)).
Second, regarding natural disturbance regimes, it is now well documented that declining human pressure in the area has been accompanied by dramatic changes in fire regimes, with a general increase in fire frequency and severity (Dara et al., 2019), as elsewhere in the former Soviet Union (Dubinin, Luschekina, & Radeloff, 2011).
Third, regarding trophic complexity, large mammals throughout the former Soviet Union have rebounded from high poaching rates in the 1990s, including wild grazer populations (Bragina et al., 2015), and several trophic rewilding initiatives are now underway to bring native grazers back to the steppe areas they disappeared from historically (Kock et al., 2018). Nevertheless, the massive decline in domestic livestock suggests many steppes may now suffer from undergrazing (Hankerson et al., 2019), which might be one of the main drivers of intensifying fire regimes due to fuel accumulation (Dara et al., 2019). Higher fire frequency leads to a decrease in fire-sensitive species and an increase in grass, a potential feedback loop that could be broken through ramping up active rewilding (i.e. the restoration of wild grazer populations such as Kulan).
In addition to these changes along the three dimensions of ecological integrity relevant for rewilding as proposed by Perino et al. (2019), declining human influence has also affected a wide range of ecosystem processes. For instance, cropland abandonment has increased soil carbon pools Wertebach et al., 2017), and fire regimes have intensified substantially (Dara et al., 2019). Declining human influence has also markedly affected biodiversity, such as bird diversity which is recovering (Kamp et al., 2011), as well as plant community composition and species richness (Kämpf, Mathar, Kuzmin, Hölzel, & Kiehl, 2016) on abandoned croplands. Altogether, this suggests that our analyses indeed pinpoint areas where passive rewilding takes place, and our indices are useful for measuring progress towards more functional, self-regulating and complex ecosystems (Perino et al., 2019;Torres et al., 2018).
Our indices capture key dimensions of declining human impact and are widely applicable given the increasing availability of high-resolution satellite imagery, both current (e.g. Sentinel-II, Planet, imagery accessible in Google Earth or Bing) and historical (e.g. Landsat archives, aerial photographs, Corona imagery). Likewise, the historical maps we used are available across the entire Eurasian steppe, and similar maps are available elsewhere. Our work thus also underlines the value of making historical maps, here used to identify Soviet-era livestock stations, available in order to better understand historical human pressure. Similarly, historical aerial photographs or Corona imagery could be a powerful data source to track signs of human influence over time. It is important to note that our index represents a start but could be easily expanded to cover other aspects of human pressure (e.g. road infrastructure, land-use intensity, hunting pressure) once additional data become available. Likewise, our index could be integrated with a composite measure of ecological integrity, measuring for instance fire dynamics (e.g. Dara et al., 2019), steppe connectivity ( Figure 3) and the observed or modelled distribution of keystone species (e.g. saiga, Figure SI-8), in order to measure progress towards increasing ecological integrity (Torres et al., 2018).
Our analyses also highlight key areas currently likely undergoing passive rewilding that may represent target areas for extending the region's protected area network. Existing protected areas are far from each other, as many of them were formed primarily to protect stopover sites for migratory birds (Schweizer, Ayé, Kashkarov, & Roth, 2014). Most of them are also not strictly protected (though our analyses suggest low human influence inside them; Figure 3).
Expanding the existing protection area network seems particularly useful in the southern part of our study region, where the protection of relatively small areas would provide large benefits in terms of connectivity, while at the same time protecting critical saiga calving grounds ( Figure SI-8) (Singh, Grachev, Bekenov, & Milner-Gulland, 2010). Integrating our human influence indices and connectivity analyses with distribution data for species of conservation concern would allow the identification of those areas and corridors that would maximize benefit for biodiversity (e.g. through allowing existing populations to occupy historic home ranges), while restoring functional steppes.
However, the window of opportunity to establish such a protected area network may be closing. While human pressure has declined drastically across the region, our analyses show that recultivation of previously abandoned areas is occurring in parts of the study area. While cropland has still not reached the Soviet extent, recultivation trends are worrisome from a conservation perspective, as recultivated areas appear to coincide spatially with areas highlighted in our analyses as important connectors between protected areas, as well as with key areas of saiga ranges ( Figure SI-8). Reviving the agricultural sector, both in terms of higher crop production and an expansion of the livestock sector, is explicit goal of Kazakhstan's development of the agro-industrial sector (Ministry of Agriculture of the Republic of Kazakhstan, 2018). At the same time, Kazakhstan also is actively expanding its protected area network (Kamp et al., 2015).
Conservation and land-use planning that seek to balance conflicts of both goals in areas particularly valuable for rewilding are therefore needed. This study provides insights on where such areas may be located, for both identifying areas particularly useful for conservation and rewilding (e.g. areas with low human pressure and high connectivity), as well as to identify areas where agricultural development would harm conservation opportunities less (Figure 4b, right).
At a time when human pressure is increasing in most world regions, making use of rewilding opportunities as they emerge is critical. Grasslands are among the most imperiled biomes of the world (Fuhlendorf et al., 2018), and the substantially reduced human pressure in the Eurasian steppe after the breakdown of the Soviet Union provides major opportunities for broad-scale steppe restoration.
Our analyses highlight how a range of human influence indicators can be combined to provide a detailed and multidimensional picture of where and why human pressure declines, and where possible rewilding has been taking place, across large areas. This should provide a basis for conservation and land-use planning that makes use of opportunities to establish large, connected habitat complexes in the Eurasian steppe.

ACK N OWLED G EM ENTS
We gratefully acknowledge funding by the Volkswagen Foundation