Recent grain production boom in Russia in historical context

In recent years, Russia has established itself as the leading worldwide supplier of grain and continues to make ambitious plans for raising its grain production in the long term. Within the context of Russian agricultural history, the recent high growth in grain production is exceptional. This growth however is not fully replicated by the “weather-yield” crop models, which project only moderate yield increase in the twenty-first century and fail to predict the most recent record growth in grain yields. The difference between the projected climate-dependent yields and observations is especially high in two of the most important agricultural regions, the Central Black Earth and Northern Caucasus regions while the remaining agricultural zones show good agreement with the regression models. Similar differences were observed in the late 1960s, which we interpret in terms of the rapid changes in agricultural technology during the Union of Soviet Socialist Republic (USSR) agricultural reforms followed by periods of reversal. We also interpret the current period of high differentiation between weather-yield model results and collected yields as evidence of a higher than usual contribution of agricultural reforms in yield improvements, which, however, primarily benefit the large-scale producers located in the most productive areas of Russia.


Introduction
The explosive growth of Russia's grain exports in the twenty-first century is exceptional in the country's history. Between 1992 (first year grain export information was reported to the Food and Agricultural Organization [FAO]) and 1996, Russia exported only 0.9 million tons of grains per annum on average (FAOSTAT 2020). From 2000From to 2004 This article belongs to the topical collection "Climate Change in Russia -history, science and politics in global perspectives", edited by Benjamin Beuerle, Katja Doose, and Marianna Poberezhskaya factor missing from most studies is state policies toward agriculture that can affect producers' incentives, economic environment in which they operate, and social conditions defining rural support for those policies (Wegren 1998). It has been argued that the strong government and weak societal institutions of Russia has made the impact of these centralized policies on agricultural practices somewhat unique in their importance (Wegren 1998). The diversity of the contradictory goals of the centralized policies, such as advancement of production and advancement of society (such as "liquidation of kulaks," that is, the most efficient private farmers) has led to sharp changes in agricultural policies along with the shifts in "Party line," which in turn exercised either a positive or negative impact on food production (Wegren 1998) and manifested itself in the periods of fast technological progress in agriculture or years of stagnation (Kirilenko and Dronin 2005). This historical context is discussed in Supplement 1. 2 In contrast to the long-and short-term characteristic time of the abovementioned factors, these policies tended to act at a decadal mediumterm characteristic time. The effects of the long-term (for simplicity, further referred to as "technological"), mid-term ("policy"), and short-term ("climate variability and weather") factors overlap. For example, few episodes of fast growth of agricultural production in the twentieth century occurred when pro-farmer policies, such as Kosygin's reforms in late 1960s, coincided with favorable weather.
Recent (2003)(2004)(2005)(2006)(2007)(2008)(2009)(2010)(2011)(2012)(2013)(2014)(2015)(2016)(2017) rates in the annual growth of grain production (2.6%) is almost double that during the previous periods of agricultural expansion. Thus, the goal of this paper was to investigate the sources that are responsible for the present grain production boom. Several specific objectives are highlighted below: 1. To demonstrate that the current climate is only marginally beneficial for grain production in Russia using the weather-yield regression models trained on historical agricultural statistics 2. To show that the difference between predicted and actual changes in yield can be explained with structural changes brought by the modern agricultural policy of Russia 3. To identify the regions which are winners and losers of these policies In our study, we concentrated on wheat, which constitutes 66% of Russian production of cereals, followed by barley and maize (15% and 10%, respectively). Wheat also constitutes 80% of grain exports (by weight: FAOSTAT 2020). "Section 2" gives a brief literature review. "Section 3" introduces data and methods in this study. Model results are provided in "Section 4" and discussed in "Section 5" while the overall findings are summarized in "Section 6."

Literature review
In this section, we provide a brief literature review of statistical weather-yield models; for a more thorough discussion in a historical context, see Supplement 1. The early history of the agricultural meteorology in Russia was thoroughly described in our earlier publication (Dronin and Kirilenko 2013). These studies dating back to the end of nineteenth century found a strong correlation between yields and weather, especially in late spring and early summer (Brounov 1913;Alsberg and Griffing 1928) and proposed splitting the observed dynamics of yields into the long-term linear part explained by progress in technological and management practices and deviations from the trend explained by weather (Obukhov 1927).
Early scientific studies exploring the possibilities to predict yield based on the weather did not sit well with the ideological doctrine of central planning as the "Comrade yield [should be] the object of the planned action of the productive forces of the Socialist state" (cit. Wheatcroft 1977: 12). The central planning and new collective forms of work organization brought a series of inconsistent reforms in agriculture starting with compulsory grain procurement ("prodrazverstka"), replaced with a free market "new economic policy" (NEP), which was in turn replaced with discriminatory market regulations (a practice of "price scissors," which artificially inflated prices for industrial goods and deflated prices for agricultural products during that period). Later, a collectivization campaign with a partial return of grain procurement and a policy of "liquidation" of the most successful farmers ("kulaks") was undertaken. Notably, prior to Russian Revolution (1900)(1901)(1902)(1903)(1904)(1905)(1906)(1907)(1908)(1909)(1910)(1911)(1912)(1913)(1914)(1915)(1916)(1917), the correlation between reported and climate-driven yields was high (R = 0.91; p < 0.01); however, it had been reduced to R = 0.37 and p > 0. 05 from 1917to 1928. Thus, Wheatcroft (1977 fitted the de-trended yield to monthly rainfall and mean surface air temperature and explained model residuals in terms of political factors impacting agricultural production. The disruptive and transformational implementation of new agricultural policies was also common in the decades after Stalin's death. Prominent examples include the "Virgin Land Campaign" (1954)(1955)(1956)(1957)(1958)(1959)(1960)(1961)(1962)(1963)(1964), Kosygin's agricultural intensification reform (1965-1975), Food Program (1982 followed by Perestroika (1985Perestroika ( -1991, liberalization of economics including agriculture (1991)(1992)(1993)(1994)(1995)(1996)(1997)(1998)(1999)(2000), and the current state support of agriculture targeting food independency. Models allow us to trace success or failure of these reforms in reported yields; for example, favorable climate and intensification policies equally contributed to the 20% gain in grain yields during Kosygin's reforms while a 15% yield reduction in the 1990s could mostly be attributed to an economic crisis due to unsuccessful liberalization in agriculture (Dronin and Kirilenko 2013). This interpretation of the residuals of yieldweather model in terms of the effects of agricultural policies is not limited to Russia. For instance, Chand and Raju (2009) estimated variability in agricultural and food production in Indian provinces resulting from the adoption of new technology. The authors analyzed the residual values from de-trended yield time series and found that the residuals' variance had decreased, which was interpreted as an effect resulting from green revolution policies. Similar methods for estimating the patterns of yield variability have been employed by many authors for studies on different regions (see Naylor et al. 1997).

Data and methods
Following the conventional approach (Obukhov 1927;Wheatcroft 1977;Lobell et al. 2011;Shi et al. 2013;Dronin and Kirilenko 2013;Chavas and Nauges 2020), the starting point in our analysis was de-trending the yield time series to remove long-term improvements in crop production attributed to slow and gradual changes in technology. Multiple methods for removing the trend from the yield data can be found (Shi et al. 2013); however, Lobell and Ortiz-Monasterio (2007) demonstrated that different de-trending techniques returned similar results. Hence, we applied the simplest one with a linear trend. The residuals were then used as a dependent variable in a multiple regression model with agro-climatic variables, such as temperature and precipitation, which were used as predictors (Nicholls 1997;Lobell and Ortiz-Monasterio 2007;Matiu et al. 2017). For example, Lobell and Ortiz-Monasterio (2007) explained variability in crop yields based on minimum and maximum temperatures and precipitation over the growing period. More recent models frequently use machine learning methods, such as random forests (Hoffman et al. 2018), artificial neural networks (Crane-Droesch 2018), and Bayesian inference (Shirley et al. 2020). While these techniques could improve model fit (Hoffman et al. 2018), the improvements are made at the cost of reduced interpretability shared by black-box models. 3 A common and difficult to deal with issue in climate-yield models is collinearity of climate variables (Lobell and Burke 2009;Shi et al. 2013). To build up result robustness, we used three models: (1) multiple linear ordinary least squares regression (MLR), (2) lasso regression (LR), and (3) ridge regression (RR). Either of the latter two models cause a reduction in the pitfall of variable collinearity while preventing model overfitting. In comparison, lasso regression tends to set more covariates to zero as compared to ridge regression. In addition, a comparison of model projections helps to estimate the robustness of results. Incidentally, all three models have the same form: in which x i represents climate variables (monthly or quarterly temperature and precipitation), and y represents observed yield. The MLR model followed the earlier weather-yield models discussed in the previous section for finding the values of the best fit parameters, i , using the ordinary least squares (OLS) method. Specifically, the OLS finds model parameters that minimize the cost function, Σ(y j −ŷ i ) 2 , in which y j are observations, and ŷ i are respective model predictions. To address overfitting and multicollinearity problems in the model, we applied the variable selection by adhering to the following protocol. First, the model was fitted to all climate variables. Five variables with the best explanatory power for all administrative units were then selected: the temperature for months of June and July, second-quarter precipitation, and the precipitation for months of July and December: in which ŷ represents estimated climatic yield, t i represents monthly temperature, p i is monthly precipitation, i is the number of the month, and a j indicates model parameters.
The MLR model still suffered from the uniformity in the variable selection, which did not reflect the multitude of climates in the country and resulted in poor model performance in some regions, such as the Far Eastern section of the country. To address this problem, we used more robust alternatives, namely lasso and ridge regressions. Both models modify the OLS cost function targeting reduction in model complexity. Specifically, the OLS cost function, Σ(y j −ŷ i ) 2 , does not discriminate against the number of independent variables and leads to an overfitting problem. That is, the model may demonstrate a very good fit on the training data but shows a poor fit when new data not used in model training are introduced. Both the LR and RR solve this problem by changing the cost function through introduction of a penalty to minimize the coefficients at the independent variable, β.
For instance, in the RR, the cost function is of the form, Note that when the parameter = 0 , both LR and RR become the familiar OLS model. The advantage of LR and RR is that both models automatically select the most significant variables for the model either by penalizing or excluding the least significant variables. A detailed discussion of the LR and RR is provided by Tibshirani (1996). The selection of the variable in the LR and RR models was individual for each administrative unit. The most frequently selected variables reflecting at least half of the administrative units were the growing period precipitation (April to July with positive coefficients), harvesting period precipitation (August to September with negative coefficients), and February precipitation (with negative coefficients). The time frame of the model was a 62-year period from 1958 to 2019 that encompassed multiple attempts to improve the agricultural sector of the country. For consistency and to accommodate for changes in administrative units in the Russian Federation, the model adopted the upper-level administrative divisions of the country at the time of Declaration of State Sovereignty of Russian Soviet Federated Socialist Republic (June 12, 1990). Consequently, the missing data points for yields in newly introduced administrative units were treated in the model as missing data and were excluded. For climate variables, we followed the family of statistical models introduced by Lobell (see Lobell and Ortiz-Monasterio 2007) with temperature and precipitation as the variables. Yield statistics for provinces are presented in Soviet and Russian official statistical reports related to corresponding years.
The mean monthly air temperature was acquired from the temperature product GISS GISTEMP v4 SBBX.Tsurf250 from the National Aeronautics Space Administration (NASA) Goddard Institute for Space Studies (GISS), gridded at 1 × 1 degrees of geographical latitude and longitude with application of a 250-km smoothing filter (Hansen and Lebedeff 1987; Lenssen et al. 2019). For the monthly precipitation from 1958 to 2016, we used the Full Data Monthly Product V.2018 (V8) from Global Precipitation Climatology Centre ([GPCC], Schneider et al. 2018a, b). The recent years missing from the fully vetted product (2017-2019) were acquired from the "First Guess" product (Ziese et al. 2011;Schamm et al. 2014). Precipitation amounts from both products were compared for one overlapping year (2016) to ensure data compatibility. All products were interpolated into a standard grid of 0.5 × 0.5 degrees of geographical latitude and longitude.
Following established practice (see Lobell and Ortiz-Monasterio 2007), the gridded parameters of climate variables were unified at the level of administrative units of the country by using their respective weighted means. The weights represented the 1992 area taken by agriculture in each cell of the grid (Ramankutty and Foley 1998) to match the reference year for administrative division of the country. Twenty-eight administrative units with little or no grain production or with ten or more missing years of data were excluded from consideration resulting in 59 remaining units for further analysis. Note that the study excludes the period prior to late 1958 when grain production tended to be misreported (for estimates, see e.g., Wheatcroft and Davies 1994). One limitation of our study is related to the robustness of the linear regression. Specifically, a median based estimator, such as Theil-Sen, would be advisable if the normality or homoscedasticity assumption had been broken. However, severe problems were not found with either estimator. Normality assumption is broken in five out of 42 provinces as demonstrated by the Shapiro-Wilk test at a p = 0.05 level; the provinces exhibiting this problem have insignificant wheat production. Homoscedasticity assumption is broken in one unit as evidenced by the Breusch-Pagan test at a p = 0.05 level. Based on these results, our decision was to stay with OLS approach in view of previous studies.

Results
The model residuals were averaged across all administrative units and are presented in Fig. 1. Recall that the residuals represent yield variability that is unexplained by the longterm technology trend and climate; the residuals were attributed to the changes in agricultural policies. The results clearly show distinct periods in deviation between the predicted and actual yields, which are consistent over all three models (MLR, RR, and LR). The first period of high yields encompassed the mid-1960s to mid-1970s, which were attributed to the success of Kosygin's agricultural reforms, including simplified credits for collective farms, tax reductions, irrigation, and reduced central management. The initial boost from those initiatives eventually regressed; one of the drivers of this regression was the return of rigid and poorly coordinated central management, publicly known at the period as "vedomstvennost," loosely translated as "multiplicity of controlling departments." The second period of high yields spanned from the mid-1980s to the early 1990s following the 1982 "Food program," probably initiated by Gorbachev, who at that time was overseeing agriculture at the Political Bureau (Dronin and Kirilenko 2013). The final period started in late 2000s and is ongoing.
Notably, starting from circa 2010, the simulated yield based on climate variables alone clearly deviated from the reported yields in a manner similar to the previous periods of agricultural reforms (Fig. 2). Interestingly, this period was also characterized by reduced precipitation in the main producing areas (Fig. 2B). The first indication of divergence could be observed in 2009 when actual grain production was well out of our simulation. However, the disastrous drought of 2010 that hit the entire European part of Russia masks the beginning of the divergence. In retrospect, the positive gap between the projected climatic and observed yields reached its historic maximum in the 2010s and 2020 (Table 1), making the past decade exceptionally productive for Russian agriculture.
The variety of climates in grain producing areas of the country affected the dynamics of yield with the possibility of some areas affected by unfavorable weather while the others exhibited beneficial agrometeorological conditions. We attempted a cluster analysis of model residuals aiming at segmentation of the agricultural areas of the country. A grouping algorithm was run for the number of clusters k = 2 … 15, and then, the optimal number of clusters was determined based on Calinski-Harabasz pseudo F-statistic. For all three models (MLR, RR, and LR), the optimal number of groups was found to be k = 2. Notably, the obtained clusters were nearly contiguous, clearly following the Tobler's (1970) First Law of Geography "everything is related to everything else, but near things are more related than distant things," which provided additional support for the validity of the yield model. Group 1 is primarily located in the fertile steppe and forest-steppe zones while group 2 is mostly situated in forest zone. All three models returned similar clustering with some differences at the edges of the clusters. The proportion agreements were 0.86, 0.92, and 0.95 for the MLR-LR, LR-RR, and MLR-RR comparisons, respectively. Figure 3 illustrates the grouping for the MLR, RR, and LR models.
The residual yield change differs between the groups, especially during the period following the USSR breakup. In the 1980s, the residual yields in groups 1 and 2 changed similarly. In response to crisis in agriculture in the 1990s, the marginal productivity lands were removed from crop production. This reduction worked at a different rate in group 1 and 2 areas. In the more productive group 1 area, the reduction was small; the area under crops was reduced by 20% by mid-1990s and remained relatively constant until the mid-2000s after which it showed a rebound. In the less productive group 2, the reduction was steep with 40% of crop land lost by year 2000 with no sign of recovery (Nefedova 2019). Therefore, removal of land from crop production mitigated the loss of productivity in group 2 at a much higher rate as compared to  Fig. 2 A, B Five-year running mean for the mean annual temperature (C) and annual precipitation (mm) for the wheat producing areas of Russia (A) and for the top five wheat producing areas in the European part of Russia (Krasnodar, Rostov, Stavropol, Voronezh, and Kursk) (B). C Observed and projected yield with projections (the mean over the MLR, LR, and RR models) based on climate plus long-term trend group 1 regions, resulting in a higher drop of residual yields in group 1. On the opposite hand, with restoration of agriculture in the 2010s, residual crop yields in group 1 regions grew much faster than in group 2.  In the group 1 cluster, the actual yields significantly exeeds predictions based on climate alone despite lower than normal precipitation levels. Group 2 regions (blue) were less productive and showed little divergence between the actual and climatic yields

Discussion
The climate in the main agricultural regions of Russia is changing. All eight provinces of the Central Black Earth and three provinces of Northern Caucasus regions have reported warmer temperatures with small changes in precipitation resulting in a drier condition in the vegetation period. For example, between the 1960s and 2010s in Voronezh Oblast (Central Black Earth region), the mean temperature of the April to September growing period increased by 1.8 °C. While the total amount of precipitation did not change, the occurrences of heavy rains increased, leading to crop damage (Gordeev and Turusov 2015). These heavy rains are usually followed by hot and dry weather with dry winds ("sukhovey"), leading to high soil evaporation (Gordeev and Turusov 2015) and depletion of water resources. In summer of 2020 for the first time since its establishment in 1892, the Dokuchaev monitoring well in Kamennaja Steppe went dry as the groundwater level fell below 8 m (RIA Voronezh 2020). In Stavropol krai (North Caucasus region), an increase in frequency of very hot days with temperatures exceeding 40 °C were observed at 12 out of 16 meteorological stations. Over the 18 last years, eight catastrophic rainfalls that exceeded 100 mm out of 18 over the entire observation period occurred (Vliyanie 2019). During the same timeframe, three large, prolonged droughts were observed. Local agronomists have called for shifting crop selection from frost-resistant to drought-tolerant cultivars.
As far as we know, recent success of Russian agriculture is also challenging the results of all contemporary weather-yield models trained on historical yield data, which have failed to replicate the fast increase in grain yields in the best agricultural areas of Russia. Our earlier grain production estimates (Lioubimtseva et al. 2015) as well as the climatic yield reported in this article (Fig. 2) fell considerably below the observed yield. Similarly, a model by Belyaeva and Bokusheva (2018) estimated that each additional heat degree day over the base temperature of 25 °C causes a reduction in the yield of winter wheat by 0.8%, spring barley by 1%, and spring wheat by 1.44%. This model also failed to explain the most recent grain production boom (Bokusheva, personal communication). The model by Sirotenko and Pavlova (2012) found marginal growth of weather explained yields 4 over the 1975 to 2006 period at a rate of 0.4% per decade in the Central economic region to 2.8% per decade in the Volga region. Finally, over a longer period of time, Lobell et al. (2011) estimated that climate trends have caused a decline in Russian wheat yields by 3.9% to 6.5% per decade during the period from 1980 to 2008, while in fact the trend over that period was close to zero.
In contrast to the moderate yield predictions based on the weather-yield models, experts point out the huge untapped agricultural potential in Russia. Assuming that similarity of agro-climatic conditions in Russia and Canada indicates potential to collect similar yields, an introduction of better technologies, management, and cultivars should allow Russian agriculture industry to increase average yields by 65% (EBRD-FAO 2008). This projection is in agreement with a much earlier study by Sirotenko et al. (1997), which estimated that Russian grain production could be increased by as much as 120% (relative to 1986-1990) if the soil fertility had been improved. Similarly, Deppermann et al. (2018) utilized the Environmental Policy Integrated Model EPIC-IIASA model and found a potential increase in cereal production in 2030 in Russia by 70% (up to 3.0 to 3.2 t/ha) from the basic 2000-2010 level in "the strongest intensification" agricultural scenario; notably, only 9% of the reported additional production was due to recultivation, whereas 91% was due to better application of fertilizer, pesticides, and other technological improvements. Other experts also reported a potential increase in grain production with improved technologies and recultivation of lands abandoned during the free market reforms of the 1990s (Deppermann et al. 2018;Meyfroidt et al. 2016;Schierhorn et al. 2014).
The Russian government has been supporting the country's agriculture after the prolonged period of frequently unsuccessful experiments with free market reforms in the agricultural sector of the economy. The goal of improving food independence (food sovereignty) of the country was set as early as in 2000 (Wegren 2002;Spoor et al. 2013) when it was exemplified as a primary task in the "Main Directions of the Agricultural Food Policy of the Government of the Russian Federation for 2001 to 2010" (Osnovnye .... 2001). The food security doctrine was adopted in 2005 and aimed at achieving national self-sufficiency by producing locally 95% consumed grain, 80% sugar, 80% vegetable oil, 85% meat, 90% milk, 80% fish, and 95% potatoes (Burkirbaeva et al. 2020). The main tools of this policy are state grain purchases, commodity interventions limiting volatility of prices, and grain export restrictions. The direct support to agricultural actors includes subsidies for lowinterest credits, seeds, insurance, and others, investment grants, direct payments per hectare and per kg of milk; grants to starting farmers, subsidies for melioration and land recultivation, and others (Uzun et al. 2019). A thorough discussion of these policies was published by .
While some experts attribute the recovery of Russian agriculture to these policies (Uzun 2004;Serova 2007), these policies are clearly insufficient. Most grain production in Russia still relies on low-cost agrotechnologies with a minimum use of agrochemicals (USDA 2017). The fertilizer application especially remains low (22 kg/ha on average vs. 134 kg/ ha in the USA and 199 kg/ha in Germany (Dyatlovskaya 2018)) due to high prices, volatility of returns in agriculture, and insufficient insurance systems thus limiting producers' capabilities to increase inputs in order to avoid financial losses (Bobojonov et al 2014). Up to 20% of seeds (except for corn) are low-quality reducing the harvest by up to 3 million tons (Samofalova 2016) as an average farmer household cannot afford to purchase imported seeds while the domestic seed production is undeveloped (USDA 2017). The utilization rates for tractors and harvesters are double or triple the norm (Alabushev 2010) leading to shortage of agricultural machinery. With the increase in agricultural production and exports, the shipping and storage infrastructure has become over-stressed, especially in Siberian regions due to their remoteness from the main markets; inadequate storage leads to a loss of at least 10% of harvest. That loss of storage discourages local businesses from investing in the grain producing sector (USDA 2010;Liefert et al 2013). Only recently has the government started modernization and expansion of storage and transportation capacities (Wegren 2018). Aggravating the problem, the federal support for agriculture has been shrinking in recent ("boom") years (IKAR 2017).
A direct yet unexpected (Davydova and Franks 2015) result of the market reforms in Russia was the emergence of large corporate agricultural companies, agroholdings, in the 2000s. Originally, reformers of the agriculture primarily aimed to develop the family-based private farming sector. Instead, the agroholdings appeared in major agricultural regions of the country by absorbing (often through land grabbing) former collective farms in addition to individual farms (Davydova and Franks 2015). At present, Russia, Ukraine, and Kazakhstan have the highest level of concentration of cultivated land even by world standards (Deininger and Byerlee 2011).
This increase in concentration of agriculture seems to be determined by the way the agricultural subsidies are distributed. While the primary interest of current state interventions into agriculture is production increase with the strategic goal of reaching food independency, the factual support varies widely among the regions. One reason for that difference are the shared responsibilities between the federal and regional governments, which are calculated and negotiated by lobbyists in a non-transparent process (Kvartiuk and Herzfeld 2021). Large enterprises, such as agroholdings, have more opportunities to influence the distribution of subsidies among the regions. In turn, on numerous occasions the Federal government has expressed the priority support to large-scale agricultural enterprises, positioning them as the "locomotive" for agrarian development (Wegren 2021). Kvartiuk and Herzfeld (2021) argue that another important factor is that the federal government is using the agricultural subsidies to cement the dominance of the current elites in regions in which it has less support via mobilization of rural voters.
As a result, the access to state support, credits, infrastructure, and markets differs among the agrarian actors and regions. The smallholders, in particular, suffer from a lack of access to new technologies, poor connection to retail, food processors, lack of transportation, labor shortage, access to credit, and others (Wegren 2018) leading to a decline from 57% in 1997 to 35% in 2016 in their share of food production (Wegren 2018). On the other part of size spectrum, large agroholdings benefit from the access to investments, federal support, and political connections, steadily increasing their share in production (Wegren 2018). Notably, one of the indicators used in evaluation of regional governments is the size of investments in agriculture, promoting priority support of bringing new investments agroholdings not only at the federal, but also at a regional level .
Inadequate support of the agricultural sector together with prioritization of large enterprises has also caused acceleration of the geographical divergence in agricultural development. In addition, the federal funds are processed and augmented by the regional governments prior to reaching individual farmers and enterprises, resulting in regional variation of subsidies (Uzun et al. 2019). As a result, we found that the current boom in grain production is mostly driven by a compact group of provinces in the Central Black Earth and Northern Caucasus regions (Fig. 3). This area is also the one in which the difference between the weather-yield model and the observed yield in the recent years is the greatest (Fig. 4). This area boasts prime soils and relatively mild agricultural climate, which however, has an elevated risk for the occurrence of droughts. Climate change projections show further increase in drought frequency (Dronin and Kirilenko 2011b). This area also has the highest in the percentage of croplands managed by the large-scale business groups frequently operating over one million hectares of land in Russia. Accordingly, in the fertile Central Black Earth, agroholdings operate 45% of the arable lands, whereas in Northern Caucasus, this share is 21%, in the Volga region 17%, in the South Ural 9%, and in Western Siberia only 7% (Rylko 2011;Grouiez 2012). The percentage of land operated by agroholdings generally decreases along with yields. Between the top 100 grain producers in Russia, 88 are in the Central Black Earth and Northern Caucasus regions, two in the Volga region, two in Tatarstan, and four in Siberia (VIAPI 2009).
Few experts have reviewed the economic efficiency and productivity of Russian agroholdings (Visser et al. 2017) as they are not considered separately in statistical data by Russia's statistical agency RosStat . 5 According to Uzun et al. (2012), agroholdings differ significantly in terms of profitability (as percentage of costs) from 26% on farms operated by foreign owners, which was more than twice the average for all other corporate farms in Russia, to − 12.5% in municipal-owned agroholdings. Anyway, sparse data show that the fertilizer use in agroholdings is 260% higher than in other agricultural companies (Uzun et al. 2012). In 2009, the average grain yield in Russia was 1.79 tons per ha when compared with 3.56 tons per ha in the top 100 largest grain producers (VIAPI 2009). Therefore, it seems that in times of limited monetary and logistical support from the state, those large-scale operators are the main drivers of productivity and efficiency in the grain sector since they will gain the most advantage from state support, attract investments, obtain better seeds, purchase fertilizers, improve infrastructure and storage capacity, and increase grain exports due to their proximity to seaports (Liefert et al 2013).
One of the unintended consequences of the regional diversions of agriculture is a persistent abandonment of millions of hectares of marginal crop lands (Ioffe et al 2012), especially pronounced in group 2 regions. The Russian government strongly intends to bring this land back to cultivation if not for wheat but coarse grains to support an ambitious plan of self-sufficiency in meat production (Visser et al. 2014). Opportunities for increasing ecosystem services, such as carbon sequestration on abandoned lands and increasing habitats for umbrella species, are not officially regarded as options for development of marginal areas (Meyfroidt et al 2016;Kurganova et al 2015;Schierhorn et al 2013). Meanwhile land abandonment combined with warmer snowmelt season reduced soil degradation (Litvin et al. 2017) as croplands of Ukraine, Belarus, and European Russia switched from a small CO 2 source of 10 g C m −2 yearr −1 , to a 47 g C m −2 year −1 sink (Vuichard et al. 2008).
The environmental impact of agroholdings is assumed to be higher than in other categories of agricultural enterprises because of their greater specialization and mono-culture types of crop production, higher uses of mineral fertilizers (Gagalyuk and Schaft 2016), preferences for the "block" cultivation in which neighboring fields are farmed as a single block (Davydova and Franks 2015), short high-intensity land lease terms (Levkivska and Levkovych 2017), and others. In response to public criticism, some agroholdings with international stakeholders are voluntarily adopting the Corporate Social Responsibility (CSR) framework, which focuses on the ecological and social dimension of industrial farming. However, this process is still in its infancy; for example, in Krasnodar region only seven of the 20 largest farms mention CSR activities on their websites (Visser et al. 2014).
Other provinces of Russia tend to have even lower low rates of CSR activities (Davydova and Franks 2015). The major motivation to adopt the CSR concept by Russian agroholdings is to maintain good relations with local authorities and ensure access to land by development of rural social and physical infrastructure (Davydova and Franks 2015) rather than clearly defined state policies. Gagalyuk and Schaft (2016) note that the primary environmentally friendly activities are technological innovations (no-till, drip irrigation, mechanical weed control, and others) rather than a switch to alternative energy sources, animal welfare programs, addressing biodiversity loss, and similar trends. The apparent regional divergences in Russian agricultural sector have created uncertainty in projections of the impact of future climate change on grain production. The global circulation model (GCM) projections for the mid-twenty-first century show a somewhat poorer climate in most of today's principal agricultural lands affected by droughts. In the main agricultural areas of Russia, the majority of GCMs project a small increase or decrease in precipitation with increasing temperatures thus elevating risks of droughts (Alcamo et al. 2007;Bobylev et al. 2012;Cook et al. 2020;Safonov and Safonova 2013;Monier et al. 2017;Pavlova et al. 2019;Ukkola et al 2020). Following this trend of increasing moisture deficit in the main grain producing belt, the Russian hydrometeorological service has estimated grain production to fall by 10% to 20% by 2035 relative to the end of the twentieth century (Roshydromet 2014). Note that climate change benefits grain production in the secondary grain production areas in East Siberia in which the climate becomes warmer and milder. Some weather-yield models project that the negative impact of climate change will be compensated with a significant increase in grain production driven by a warmer and longer growing season in the currently least productive northern regions of the forest zone (Pegov et al., 2000). This projection, however, is somewhat curtailed as other natural (such as land availability and fertility) and social (such as sparse population and lack of infrastructure) factors should be considered (Alcamo et al. 2007). Still, no signs of recovery of sown areas in central and northern regions of European Russia are observed (Nefedova 2019).

Conclusion
Russian grain production is booming. Recently, Russia became the third largest country in grain exports and top in wheat export. This success has mainly been achieved through improved yields in two regions: (1) The Central Black Earth and (2) North Caucasus regions. Meanwhile, climate favorability for agriculture in those regions is not improving but rather bringing new problems for farmers. In the boom period (2012-2020), seven large droughts were reported to affect 4.4 million ha of agricultural lands and bring losses at 10.2 billion rubles on average (Wegren and Nilssen 2022). Hence, we showed that both regions have significant differences between the yield projected by weather-yield models and actual yield amounts. We explain this difference as a substantial increase in the importance of non-climatic factors in recent dynamics of grain production in these regions.
Most experts suggest that the observed agricultural boom in both regions reflects large structural improvements including subsidies, state control over price volatility, modernization, and expansion of the physical infrastructure (Uzun 2004;Serova 2007). These improvements are mainly recuperations from the long-term deficiencies in the agricultural policies, first suggested in research dating back to 1990s. However, our model shows that there is a compact group of regions with much higher yields than expected from simulation. These regions with the best soils and access to seaports are the main beneficiaries of these policies, which are reflected in the development of large agroholdings leading the agricultural sector. Although not recognized in the official statistics, agroholdings attract a considerable portion of state financial support and play a crucial role in unlocking the untapped agricultural potential in Russia. At the same time, smaller operators and less advantageous regions are trailing with respect to this major agricultural improvement. The forest zone of Russia in which vast tracts of agricultural land, once heavily subsidized collective farms under Soviet rule, still lie uncultivated with 20-year-old birch groves covering the old fields, which ironically increase their importance as carbon sink. Therefore, the role of state support of agriculture differs among the regions. Accessibility to state support and credits, development of infrastructure, and proximity to major markets are drivers of increasing divergence of Russian regions in terms of agriculture performance, which is an unintended result of the state agricultural policy that primary focuses on food security issue.
In the future, GCM projections suggest deterioration of the agricultural climate in the main grain producing areas, mainly due to the increase in water deficit. 6 The rapid increase in agricultural production apparently runs against this dynamic, suggesting significant potential for adaptation to climate change in the agricultural sector, including a shift to drought-tolerant cultivars and crops, such as corn (including winter corn) and sunflower, a shift in sowing time, expansion of the area under winter crops and thermophilic spring crops (Dronin and Kirilenko 2011a), and many other changes. The resilience of the agricultural sector is however limited by the apparent difference in adaptation capacity among the large producers located in the most productive areas close to population centers and seaports.