Abstract
During the Soviet Virgin Lands Campaign, approximately 23 million hectares (Mha) of Eurasian steppe grassland were converted into cropland in Northern Kazakhstan from 1954 to 1963. As a result Kazakhstan became an important breadbasket of the former Soviet Union. However, the collapse of the Soviet Union in 1991 triggered widespread agricultural abandonment, and much cropland reverted to grasslands. Our goal in this study was to reconstruct and analyze agricultural land-cover change since the eve of the Virgin Lands Campaign, from 1953 to 2010 in Kostanay Province, a region that is representative of Northern Kazakhstan. Further, we assessed the potential of currently idle cropland for re-cultivation. We reconstructed the cropland extent before and after the Virgin Lands Campaign using archival maps, and we mapped the agricultural land cover in the late Soviet and post-Soviet period using multi-seasonal Landsat TM/ETM+ images from circa 1990, 2000 and 2010. Cropland extent peaked at approximately 3.1 Mha in our study area in 1990, 38% of which had been converted from grasslands from 1954 to 1961. After the collapse of the Soviet Union, 45% of the Soviet cropland was abandoned and had reverted to grassland by 2000. After 2000, cropland contraction and re-cultivation were balanced. Using spatial logistic regressions we found that cropland expansion during the Virgin Lands Campaign was significantly associated with favorable agro-environmental conditions. In contrast, cropland expansion after the Campaign until 1990, as well as cropland contraction after 1990, occurred mainly in areas that were less favorable for agriculture. Cropland re-cultivation after 2000 was occurring on lands with relatively favorable agro-environmental conditions in comparison to remaining idle croplands, albeit with much lower agro-environmental endowment compared to stable croplands from 1990 to 2010. In sum, we found that cropland production potentials of the currently uncultivated areas are much lower than commonly believed, and further cropland expansion is only possible at the expense of marginal lands. Our results suggest if increasing production is a goal, improving crop yields in currently cultivated lands should be a focus, whereas extensive livestock grazing as well as the conservation of non-provisioning ecosystem services and biodiversity should be priority on more marginal lands.
Export citation and abstract BibTeX RIS
Content from this work may be used under the terms of the Creative Commons Attribution 3.0 licence. Any further distribution of this work must maintain attribution to the author(s) and the title of the work, journal citation and DOI.
1. Introduction
Global agricultural production will need to supply substantially more feed, food, and bioenergy in the coming decades, but crop yields are stagnating, soil degradation is widespread (Godfray et al 2010, FAO 2011, Foley et al 2011), and available fertile lands for further cropland expansion are becoming increasingly scarce (Ramankutty et al 2002, Fischer et al 2011, Lambin and Meyfroidt 2011). Past agricultural expansion has been particularly widespread in the steppe and savanna biomes, due to abundant fertile soils and the low costs of land conversion (Ramankutty et al 2006, Ellis et al 2010, Müller et al 2015). Approximately 400 million hectares (Mha) of these biomes were converted in the 20th century alone (Ramankutty and Foley 1999, Goldewijk 2001).
The former Soviet Union is a prime example of a region with rapid agricultural expansion into the steppes. After the Second World War, due to an acute grain supply shortage, the general secretary of the Communist party of the USSR, Nikita Khrushchev, initiated a unique cropland expansion project called the Virgin Lands Campaign (McCauley 1976, Wein 1980). From 1954 to 1963, approximately 45 Mha of the Eurasian steppe grasslands, an area nearly the size of Spain, were converted into cropland, roughly half each in the Russian and Kazakh territory of the USSR (McCauley 1976). This rapid land conversion boosted Soviet agricultural production, particularly of wheat, but also resulted in negative environmental and socio-economic outcomes, such as massive soil degradation that reduced soil organic matter, and promoted widespread salinization and dust storms (Hahn 1964, Amerguzhin 2003, Funakawa et al 2007, Josephson et al 2013).
Recently, cropland abandonment has become a common phenomenon in many parts of northern Eurasia, including the steppe region of Kazakhstan in Central Asia (Ioffe et al 2004, Wright et al 2012, Schierhorn et al 2013). Cropland abandonment in Kazakhstan reflects the socio- political and structural changes in agriculture that followed the breakup of the Soviet Union, and the subsequent transition from a state-commanded to a market-driven economy ('transition' hereafter) (Smith 1999, Lioubimtseva 2010, Prishchepov et al 2013). From 1990 to 2000, Kazakhstan's grain production dropped from 23.4 to 10.7 million tons, and livestock numbers declined from 48.6 to 14.5 million heads (World Bank 2004, ASK 2014). As a result, nearly 19 Mha of areas cultivated with grain and fodder crops in 1990 were withdrawn from cropland production by 2000, which is a decline of 54% (ASK 2003).
After 2000, Kazakhstan's crop and livestock production started to recover, largely in response to policy reforms and increasing governmental support for agriculture (Dudwick et al 2007, Belaya and Mykhaylenko 2010, OECD 2013). As a result, idle cropland was partially re-cultivated and livestock numbers increased. Yet, 14 Mha of Kazakhstan's rain-fed croplands cultivated in 1990 remained idle in 2010 (ASK 2014). Some of these abandoned croplands may represent a valuable land resource for future cropland expansion (Liefert et al 2010, FAO 2011, Lambin et al 2013).
Satellite remote sensing is a powerful tool for mapping trajectories and patterns of agricultural land-cover change, including agricultural land abandonment in the post-socialist countries (Prishchepov et al 2012a, Alcantara et al 2013, de Beurs and Ioffe 2013, Kuemmerle et al 2013). However, prior studies of post-socialist agricultural land-cover change focused largely on Eastern and Central Europe, and only a few studies have examined Central Asia (e.g., Klein et al 2012, Chen et al 2013, Dubovyk et al 2013). Existing evidence about land-use and land-cover change in Central Asia mainly stems from coarse-scale satellite imagery, showing biomass and land-surface phenology dynamics, which are likely related to post-Soviet agricultural change (de Beurs and Henebry 2004, Propastin et al 2008, de Beurs et al 2009, Zhou et al 2015). Unfortunately, the resolution of coarse-scale satellite imagery can limit understanding of the spatial patterns of land change.
Multispectral Landsat TM/ETM+ satellite imagery may be well suited to map the dynamics of land cover in the steppe region of Central Asia. Landsat images have been previously used to map broad land-cover classes, such as cropland, grassland, shrubs and trees, in sparse vegetation environments (Guerschman et al 2003, Estes et al 2012, Müller et al 2015, Senf et al 2015), and also to monitor land-cover change in northern Kazakhstan (Terekhov 2010). Moreover, the Landsat TM/ETM+ resolution (30 m) suits well to the average size of agricultural fields in our study area, which is approximately 4 km2 (Kazakh Space Research Institute 2013b). Finally, cost-free access to historical archives with multi-date Landsat TM/ETM+ satellite images that date to the mid-1980s allows the reconstruction of agricultural land-cover change back to the Soviet period.
Our major goal was to assess available idle croplands and their suitability for future cropland expansion in northern Kazakhstan by analyzing long-term agricultural land-cover change, and the biophysical characteristics of idle cropland. Detailed maps were available from the peak time of the Virgin Lands Campaign (Moscow State University 1964), which allow us, in combination with Landsat TM/ETM+ satellite images, to reconstruct the trajectories of agricultural land-cover change from shortly before the beginning of the Virgin Lands Campaign in 1953 to 2010.
Our first objective was to reveal the rates and patterns of agricultural land-cover change from 1953 until 2010. Our second objective was to assess whether statistical relationships exist between observed patterns of agricultural land-cover change and biophysical conditions (namely, elevation, soil types and degree of aridity). Finally, we assessed suitability for idle cropland re-cultivation and contrasted them with ongoing increase of livestock numbers in order to assess the remaining idle croplands under different agro-environmental conditions and the competition of land-use.
2. Materials and methods
2.1. Study area
We chose Kostanay Province (oblast) as our study area because it is representative of the northern Kazakh steppe region, the core area for rain-fed crop production in Kazakhstan (figure 1, table A1). Within Kostanay Province, we analyzed two Landsat footprints that cover approximately 5.8 Mha, or 30% of the province area (figure 1). The elevation in our study area ranges from 90 to 300 m, with an increase in elevation toward the south. The climate of Kostanay Province is dry-continental, with a mean annual precipitation ranging from 400 mm in the north to 200 mm in the south (Afonin et al 2008). The potential annual evaporation ranges from 600 to 700 mm (Hahn 1964, Wein 1980) and frequently results in drought conditions, especially in the south (Iijima et al 2008). Severe droughts and dry storms (Sukhovey) occur every three to four years. The average number of frost-free days in Kostanay Province is 131, and decreases toward the north (Afonin et al 2008). Croplands are prevalent in the north, while grasslands dominate in the south. Numerous seasonal lakes and wetlands, which are mostly drainless and salty, are scattered throughout the area (DIK and Feorija 2011).
The semi-arid climate also produces a high mineral content in the soils and increases the risk of salinization, particularly in lowlands (Hahn 1964, Florinsky et al 2000). The most common and most fertile soil type in our study area is Chernozem (black earth), followed by Kastanozem (chestnut soils). The least fertile soils are the wet and salty Solonetz. Overall, the environmental conditions in the study area are not ideal for agricultural production (Lioubimtseva and Henebry 2012, Pavlova et al 2014). In particular, recurring droughts cause yield shortfalls, and average wheat yields from 2000 to 2010 were 1 ton ha–1, with high inter-annual fluctuations (ASK 2003, 2014).
Figure 1. Overview of Kostanay Province, Northern Kazakhstan, study area (Landsat footprints) and test sites for collection of validation data.
Download figure:
Standard image High-resolution imageThe steppes of northern Kazakhstan were traditionally used as pastures by nomadic herders (Olcott 1995, Robinson et al 2003). Initial efforts to plough the steppes date back to the late 19th and early 20th century. However, official statistics suggest, the main expansion of cropland area occurred during the Virgin Lands Campaign (the Campaign hereafter) in the 1950s and 1960s, when croplands in Kostanay Province increased from 1.0 to 6.4 Mha (ASK 2003). After 1963, cropland expansion slowed, and croplands reached a maximum extent (8.5 Mha) in the early 1980s. Toward the end of the Soviet era, the cropland area in Kostanay Province started to decline, and after the dissolution of the Soviet Union in 1991, the cropland area decreased from 6.8 to 3.1 Mha until 1999, and has only slightly rebounded since 2000 (KDS 2011, 2013) (figure 2). Similarly, livestock numbers decreased from 3.0 to 0.7 million head between 1990 and 1999, with a moderate recovery since 2000 (KDS 2011, 2013) (figure 2).
Figure 2. Agricultural development in Kostanay Province, 1990–2012 (left axis: cropland area, right axis: number of livestock), source: ASK 2003, KDS 2011, 2013.
Download figure:
Standard image High-resolution imageIn 2010, the agricultural sector of Kazakhstan comprised three main types of producers: corporate enterprises, which are successors of Soviet collective farms (kolkhoz) or state farms (sovkhoz), registered and commercially-oriented individual farms, and small, partly subsistence-oriented household farms (Dudwick et al 2007, OECD 2013, Petrick et al 2013). The large-scale corporate enterprises constitute the backbone of the Kazakh crop production, controlling about three fourths of the total cropland in northern Kazakhstan (ASK 2014). The main crop grown in Kostanay Province is spring wheat, which accounts for over 90% of the total crop production (KDS 2012). After 1991, livestock production shifted largely from corporate farms to individual and household farms (OECD 2013). In 2010, household farms owned 83% of the livestock in Kostanay Province (KDS 2012). From 1990 to 2000, Kostanay Province experienced a drastic population decline, from 1.24 million to 995,000 inhabitants (KDS 2011), which coincided with socio-economic and agricultural decline. Population totaled 881,000 inhabitants in 2010 (KDS 2011).
2.2. Archival land use data and satellite imagery
To estimate the cropland extent during the 1950s and 1960s, we digitized a 1:3,000,000 map from the Atlas of Virgin Territory (Moscow State University 1964) (for a generalized subset of the scanned map, see figure S1). This map contains information about cropland areas by 1953, the year before the start of the Campaign, and by 1961, which represented the peak of cropland expansion during the Campaign. The atlas was produced based on the results of two in-depth field campaigns running from 1953 to 1961, conducted by the Geography Department of Moscow State University in Northern Kazakhstan, which aimed to scientifically support the cropland allocation during the Campaign. Detailed cropland expansion plans, updated land-use maps of newly established kolhozes and sovkhozes in Kazakhstan and information from the local cadaster offices were also used to complete the Atlas (Moscow State University 1964). We scanned the cropland expansion map from the atlas, georeferenced it and digitized (vectorized) two thematic classes: the cropland extent by 1953 ('pre-Campaign cropland' hereafter), and the grassland converted into cropland during the peak of the Campaign, from 1954 to 1961 ('Campaign cropland' hereafter).
To assess agricultural land-cover change from 1990 to 2010, we analyzed 30 m resolution Landsat TM/ETM+ satellite imagery from two Landsat footprints (WRS-2 path/row 160/23 and 160/24, figure 1). To account for crop phenology (Prishchepov et al 2012b), we acquired multi-seasonal images from the U.S. Geological Survey (USGS) circa 1990, 2000 and 2010. We then selected suitable multi-seasonal satellite images based on cropping schedules (table 1, figure S2). In total, we analyzed 11 images for footprint 160/23 and 12 images for footprint 160/24 (table 1).
Table 1. Landsat TM/ETM+ images used for land-cover change analysis with acquisition date and cloud contamination.
| WRS-2 path/row | ||||
|---|---|---|---|---|
| 160/23 | 160/24 | |||
| Time step | Date | Clouds (%) | Date | Clouds (%) |
| Circa 1990 | 1986-06-22 | 8 | 1986-06-06 | 13 |
| 1986-10-28 | 7 | 1986-06-22 | 2 | |
| 1987-09-13 | 2 | 1986-07-24 | 5 | |
| 1988-10-09 | 0 | 1987-09-13 | 1 | |
| 1988-10-09 | 0 | |||
| Circa 2000 | 2000-09-08 | 0 | 2000-09-08 | 19 |
| 2000-11-11 | 0 | 2000-11-11 | 0 | |
| 2001-06-07 | 48 | 2001-06-07 | 4 | |
| 2001-07-25 | 22 | |||
| Circa 2010 | 2009-06-05 | 0 | 2009-06-05 | 0 |
| 2009-09-09 | 5 | 2009-06-21 | 2 | |
| 2009-10-11 | 0 | 2009-09-09 | 0 | |
| 2009-10-11 | 0 | |||
For our satellite data analysis, we acquired ortho-rectified Landsat satellite images (USGS 2013). Due to the flat terrain in our study area, additional topographic correction was not necessary as we used systematically terrain-corrected Landsat images (Level 1T product). We also did not apply an additional atmospheric correction because we used a classification change detection approach based on all (stacked) satellite images (Coppin et al 2004, Jensen 2005, Chen et al 2012). Since the training dataset is derived from this multi-temporal composite (stack) (Jensen 2005, Prishchepov et al 2012b), an atmospheric correction would have no effect on such an image classification (Song et al 2001). For each image date, 30 m resolution spectral bands 1–5 and 7 of the TM/ETM+ sensors were stacked to one multi-spectral image. Clouds and associated shadows were masked using 'Fmask' (Zhu and Woodcock 2012).
2.3. Classification of satellite images
Our classification scheme focused on agricultural land-cover change for 1990, 2000 and 2010 (table 2). We defined 'cropland' as agricultural land that was ploughed prior to sowing in either spring or fall or that was kept temporarily fallow as a part of a crop rotation. Initial tests to separate grasslands with different management regimes and degrees of degradation indicated high spectral collinearity among these classes. Therefore, we mapped all grassland types as one single class ('grassland') that included pastures, hay cutting areas and natural grasslands. In total, we mapped six classes of cropland and grassland, and three non-agricultural classes (forest, wetland and 'other', which included water bodies, bare soil and impervious surfaces) (table 2).
Table 2. Classification scheme with number of training pixels per footprint (WRS-2 path/row) and validation pixels for entire map.
| No. of training pixels | ||||
|---|---|---|---|---|
| Class name | Acronym | 160/23 | 160/24 | No. of validation pixels |
| Stable cropland in 1990, 2000, 2010 | C-C-C | 488 | 440 | 85 |
| Stable grassland in 1990, 2000, 2010 | G-G-G | 212 | 521 | 135 |
| Cropland in 1990, grassland in 2000, 2010 | C-G-G | 217 | 519 | 133 |
| Cropland in 1990, 2000, grassland in 2010 | C-C-G | 120 | 211 | 52 |
| Cropland in 1990, grassland in 2000, cropland in 2010 | C-G-C | 152 | 231 | 56 |
| Grassland in 1990, 2000, cropland in 2010 | G-G-C | 36 | Not observed | 13 |
| Forest | F | 399 | 557 | 58 |
| Wetland | W | 348 | 383 | 45 |
| Other (water, bare soil, impervious surfaces) | Other | 162 | 269 | 28 |
Utilized multi-spectral images (table 1, six spectral bands each) were assembled into multi-temporal image composites, which comprised a total of 66 and 72 bands for footprint 160/23 and footprint 160/24, respectively. This depth of spectral information of each pixel over all time steps helps the classifier to group pixels into thematic classes. For the classification, we utilized semi- automatic non-parametric support vector machines (SVM) (Vapnik 1995, Huang et al 2002), which have been successfully used to map post-socialist agricultural land-cover change elsewhere (e.g., Kuemmerle et al 2008, Prishchepov et al 2012b). We used the SVM implementation in the EnMAP-Box (version 1.4) (Rabe et al 2010) based on LIBSVM (Chang and Lin 2011), which iteratively selects optimal SVM parameters (for details, please refer to text S1 available at stacks.iop.org/ERL/10/054012/mmedia).
2.4. Collection of training data
To facilitate the collection of training and validation data, we stratified the multi-temporal image stacks into 50 classes using iterative automatic clustering (ISODATA, k-means), which we grouped into nine thematic classes to separate all trajectories of agricultural land-cover change and predominant non-agricultural land-cover classes (table 2). Based on this stratification, we sampled between 2000 and 3000 random training points (single pixels) per Landsat TM/ETM+ footprint, with a minimum distance of 500 m between points to reduce potential spatial autocorrelation (table 2). We assigned a thematic class to each training point using expert-based visual interpretation of the dense stacks of Landsat images available from the USGS archive along with very high- resolution images (e.g., QuickBird™ and WorldView™ images) from Bing™ and GoogleEarth™ online mapping services (table S1) and from digitized agricultural plots from the 1980s (Kazakh Space Research Institute 2013a). We selected at least 200 training points in classes that were common and at least 100 in rare classes (table 2). The only exception was the new cultivation class after 2000 (G-G-C), for which we could allocate only 36 training points (table 2). Using these points, we trained the SVM to classify the multi-temporal Landsat stacks into our land-cover change classes. Finally, we applied a 3 × 3 majority filter to reduce the salt-and-pepper effect in the initial classifications.
2.5. Collection of reference data and accuracy assessment
For our accuracy assessment of the land-cover change map, we collected validation data, using a stratified clustered random sample (Edwards et al 1998, Prishchepov et al 2012b). Points were then labeled during a field campaign in 2013, and using very high-resolution QuickBird™ and WorldView™ images available via Bing™ and GoogleEarth™ (table S1). We selected six 20 × 20 km blocks in our study area, which matched footprints of the high-resolution satellite images (figure 1). Then, we digitized all roads and randomly selected points that were at least 60 meters and at most 360 meters (i.e., 12 Landsat pixels) away from these roads (to ensure accessibility in the field), while maintaining a minimum distance of 500 meters between validation and training points. During the field campaign, we located validation points and recorded additional points with a non-differential GPS.
Field surveying, however, was sometimes difficult due to poor road conditions. Thus, we had to assign the land-cover classes for non-visited points manually based on their spectral similarity to known, non-random points nearby. To reconstruct land use for 1990 and 2000, we interviewed local farmers and officials of farm enterprises using semi-structured questionnaires, and we obtained information such as detailed land-use plans for each plot. To validate the 'forest', 'wetland' and 'other' classes, we used high-resolution images and spectral profiles for validation points derived from dense stacks of Landsat images (Baumann et al 2012, Griffiths et al 2013).
To assess the map accuracy, we calculated contingency matrices, producer's and user's accuracies and the overall accuracy (Congalton et al 1983, Olofsson et al 2013). We derived area estimates, which we adjusted for possible sampling bias, and we calculated 95% confidence intervals around these estimates (Olofsson et al 2013).
2.6. Combination of archival and satellite-derived maps and comparison of land-cover change with biophysical conditions
The combination of the satellite-derived map and the digitized Virgin Lands Campaign atlas for 1953 to 1961 allowed us to identify cropland expansion after the peak of the Campaign from 1962 to 1990 ('post-Campaign cropland' hereafter), as well as the abandonment of pre- Campaign cropland and Campaign cropland from 1962 to 1990. We also assessed the biophysical characteristics of each agricultural land-cover change class, such as elevation (CGIAR–CSI 2014), Selyaninov's hydrothermal coefficient (HTC) (Selyaninov 1928, Afonin et al 2008) and soil quality (COGC 1976) (figure 3). The HTC represents the ratio of total precipitation and average daily air temperature during the growing season (days with an average temperature >5 °C). Areas with insufficient humidity are defined by values below one (Gathara et al 2006, Voropay et al 2011). We digitized the soil types from the 1:2 500 000 Soil Map of Kazakhstan (COGC 1976), grouped them into three classes and ranked them according to the suitability for crop production into pure Chernozem and Kastanozem (highest crop production suitability, and rank 1), Chernozem Solonetz and Kastanozem Solonetz (medium crop production suitability, and rank 2) and Solonetz and meadow soils (lowest crop production suitability, and rank 3). For further spatial analysis, all data were rasterized, including the Virgin Lands Campaign map, resampled to a 30 m resolution to match the satellite-derived agricultural land-cover-change map, and analyzed using GRASS GIS (GRASS Development Team 2014), where we summarized each land-cover change class and its association with elevation, HTC and soil rank.
Figure 3. Violin plots for selected biophysical parameters for major agricultural land-cover change classes 1953–2010. Note: the violin plots contain boxplots with median and interquartile range (Hintze and Nelson 1998). Acronyms for land-cover classes introduced in table 2. The G-G-C class was not assessed due to rare occurrence.
Download figure:
Standard image High-resolution image2.7. Biophysical determinants of agricultural land-cover change and suitability for cropland expansion
In addition to summary statistics, we also analyzed the relationships between the spatial patterns of the agricultural land-cover change from 1953 to 2010 and the biophysical variables (figure 3). We expected that cropland expansion would be more common in areas with better soils, higher HTC and at higher elevation, while abandonment would dominate marginal plots with lower HTC, lower soil quality and at lower elevations. Similarly, we expected re-cultivation of abandoned plots would be most common in areas with better agro-environmental conditions. However, we were unable to analyze other socio-economic variables because no fine-scale data were available for the study period.
For our statistical analysis, we conducted several binary logistic regressions separately for two periods (1953–1990 and 1990–2010) (table 3). For instance, for the period 1953–1990, 'pre-Campaign cropland' was coded as '1', while the 'stable grassland in 1953, 1961 and 1990' class was coded as '0' (table 3, model 1–3). For the period 1990–2010, we set the agricultural land-cover changes equal to one and stable cropland (C-C-C) to zero (table 3, model 4–6). Finally, we constructed one additional model to map suitability for potential cropland expansion at the expense of abandoned croplands (table 3, model 7).
Table 3. Logistic regression modelsa.
| Coding | ||
|---|---|---|
| Model | '1' (presence) | '0' (absence) |
| 1 | 'pre-Campaign cropland' | 'Stable grassland in 1953, 1961, 1990' |
| 2 | 'Campaign cropland' | 'Stable grassland in 1953, 1961, 1990' |
| 3 | 'post-Campaign cropland' | 'Stable grassland in 1953, 1961, 1990' |
| 4 | C-G-G | C-C-C |
| 5 | C-G-C | C-C-C |
| 6 | C-C-G | C-C-C |
| 7 | C-G-C | G-G-G + C-G-G + C-C-G |
aAcronyms for land-cover classes in model 4–7 introduced in table 2.
We selected observations with a distance of at least 1 km among them, as this helped us to reduce spatial autocorrelation of Moran's I to 0.1–0.2, which we measured in ENVI™ package for five randomly selected 10 × 10 km blocks. All bivariate Pearson's correlation coefficients between predictors were below 0.5, indicating little multi-collinearity. We assessed the goodness-of-fit using the log-likelihood, the deviance for the residuals of the null and fitted models and the area under the receiver operating characteristics curve (AUC; Pontius and Schneider 2001). We checked robustness of models to balanced and unbalanced sampling, as often numbers of absence (0s) were more frequent than numbers of presence (1s). For model interpretation, we used odds ratios. Finally, using the statistically significant coefficients (p < 0.05) from model 7 (table 3), we calculated the relative suitability for cropland expansion at the expense of formerly abandoned croplands. For comparison, we also summarized recent trends of livestock increase in our study area using detailed population statistics available at the settlement level (ASK 2001b, 2011b) as well as district-level livestock dynamics for Kostanay Province (ASK 2001a, 2011a, KDS 2001 and 2012). For details on our methodology to disaggregate the livestock data please refer to text S2. We used R software for all statistical analysis (R Development Core Team 2011).
3. Results
3.1. Land-cover change from 1953 to 2010
Our classifications had an area adjusted overall accuracy of 78% (table 4). Among the agricultural land-cover classes, the 'stable cropland in 1990, 2000, 2010' class (C-C-C) had the highest user's accuracy (84%) and area adjusted producer's accuracy (95%), followed by 'stable grassland in 1990, 2000, 2010' (G-G-G), with 77% and 86%, respectively (table 4). Among the change classes, the user's and area adjusted producer's accuracies varied from 68% to 79%, though with lower producer's accuracies for 'cropland in 1990, 2000, grassland in 2010' (C-C-G, 49%) and 'grassland in 1990, 2000, cropland in 2010' (G-G-C, 35%).
Table 4. Confusion matrix with user´s (UA), producer´s (PA) and area adjusted producer´s accuracy (aPA), and conditional Kappa for the satellite-derived change mapa.
| Reference | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Classification | Other | Water | Forest | G-G-G | C-C-C | C-G-G | C-G-C | C-C-G | G-G-C | Total | UA (%) | Kappa |
| Other | 24 | 2 | 2 | 2 | 1 | 1 | 32 | 75.0 | 72.9 | |||
| Water | 3 | 28 | 5 | 4 | 1 | 3 | 1 | 1 | 1 | 47 | 59.6 | 54.0 |
| Forest | 46 | 1 | 47 | 97.9 | 93.9 | |||||||
| G-G-G | 13 | 6 | 105 | 9 | 1 | 3 | 137 | 76.6 | 70.4 | |||
| C-C-C | 1 | 1 | 73 | 3 | 3 | 5 | 1 | 87 | 83.9 | 78.3 | ||
| C-G-G | 2 | 1 | 15 | 1 | 103 | 4 | 4 | 130 | 79.2 | 74.2 | ||
| C-G-C | 5 | 4 | 46 | 3 | 58 | 79.3 | 77.2 | |||||
| C-C-G | 6 | 2 | 10 | 38 | 56 | 67.9 | 64.0 | |||||
| G-G-C | 1 | 1 | 1 | 8 | 11 | 72.7 | 79.0 | |||||
| Total | 28 | 45 | 58 | 135 | 85 | 133 | 56 | 52 | 13 | 605 | ||
| PA (%) | 85.7 | 62.2 | 79.3 | 77.8 | 85.9 | 77.4 | 82.1 | 73.1 | 61.5 | 77.9 | ||
| aPA (%) | 70.7 | 63.1 | 55.6 | 86.0 | 94.9 | 73.4 | 70.3 | 48.8 | 35.4 | 77.8 | ||
aAcronyms for land-cover classes introduced in table 2.
We observed high rates of cropland expansion since the beginning of the Campaign, and substantial cropland contraction after 1990, followed by a minor rebound of cropland area after 2000 (figure 4). There was an almost seven-fold increase in cropland from 1953 to 1990 (figures 4 and 5(A)). The cropland area was 465,000 ha in 1953, then expanded by 1.57 Mha through 1961 and by another 1.08 Mha through 1990 (figures 4 and 5(A)). At the same time, 120,000 ha of pre-Campaign cropland (by 1953) and 380,000 ha of Campaign cropland (ploughed 1954–1961) were already converted back to grassland by 1990 (figure 4).
Figure 4. Changes in cropland area from 1953 to 2010 by period and with proportions of first documented conversion of grassland into cropland.
Download figure:
Standard image High-resolution image
Figure 5. Agricultural land-cover change in the study area. (A) Cropland in 1990 by period of first documented conversion of grassland into cropland. (B) Classified land-cover map for post-Soviet period (1990–2000–2010). Acronyms for land-cover classes introduced in table 2.
Download figure:
Standard image High-resolution imageThe highest rates of cropland abandonment occurred between 1990 and 2000 (figures 4 and 5(B)). The cropland area was 3.12 ± 0.45 Mha (54 ± 7.7% of the study area) in 1990, but it decreased by 1.41 ± 0.21 Mha (∼45%, class C-G-G and C-G-C) by 2000 (figures 4 and 5(B)). By 2010, 373,000 ± 84,000 ha (∼26%) of prior abandoned cropland (i.e., from 1990 to 2000) was ploughed again (class C-G-C, 6.4 ± 1.5% of the study area), while another 341,000 ± 102,000 ha of earlier cultivated cropland was abandoned (class C-C-G, 5.9 ± 1.8% of the study area) (figures 4, 5(B) and 6). We also observed a minor expansion of cropland from 2000 to 2010 at the expense of virgin grasslands (class G-G-C: 72,000 ± 48,000 ha, 1.2 ± 0.8% of the study area, ∼2% of the 2010 cropland) (figures 4, 5(B) and 6). In total, cultivated cropland comprised 1.81 ± 0.26 Mha by 2010 (class C-C-C, C-G-C and G-G-C) or approximately 58% of the cropland extent of 1990 (figures 4, 5(B) and 6).
Figure 6. Error-adjusted area estimates for mapped agricultural land-cover classes of post-Soviet period (error bars indicate the error margins with 95% confidence interval). Acronyms for land-cover classes introduced in table 2.
Download figure:
Standard image High-resolution imageFrom 1990 to 2000, 64% of cropland abandonment (C-G-G) occurred in areas cultivated after the peak of the Campaign (1962–1990) (figures 4 and 5). This post-Campaign cropland alone decreased by 57% (from 1.59 to 0.91 Mha) during the first decade of transition. By 2010, 58% (0.93 Mha) of total post-Campaign croplands were still grasslands (part of C-G-G and C-C-G), despite efforts to re-cultivate grasslands starting in 2000. In contrast, the cropland that was ploughed during the peak of the Campaign (1954–1961) decreased by only 29% (from 1.19 to 0.84 Mha) from 1990 to 2000 (part of C-G-G and C-G-C), with rare re-cultivation events until 2010 (38,000 ha) (figures 4, 5 and 6). Interestingly, about 44% (203,000 ha) of the oldest pre-Campaign croplands were still cultivated in 2010 (part of class C-C-C or C-G-C) (figures 4, 5 and 6).
3.2. Comparison of land-cover changes with biophysical conditions
From 1953 to 1990, the cropland area gradually expanded toward the southern edge of our study area (i.e., the vicinity of Burevestnik settlement, which was established in 1965) (figure 5(A)). The Landsat satellite images revealed that stable cropland (C-C-C) was dominant in the north, whereas stable grassland (G-G-G) and early cropland abandonment and reversion to grassland (C-G-G) predominantly occurred in the central and the southern parts of our study area (figure 5(B)).
Descriptive statistics showed both Campaign and post-Campaign croplands were more common at higher elevations (a mean of 197 m for both classes) compared to pre-Campaign cropland (a mean of 181 m) (figure 3). Pre-Campaign and Campaign croplands ploughed by 1990 were located at slightly higher elevations compared to croplands that were abandoned prior to 1990. The post-Campaign croplands had a lower HTC (0.68 mean) compared to pre-Campaign cropland and Campaign cropland (means of 0.72 and 0.73, respectively) (figure 3). Pre-Campaign cropland had the lowest variation in HTC compared to the Campaign and post-Campaign cropland, which occurred largely in very dry areas with an HTC of approximately 0.5 (figure 3).
From 1990 to 2010, stable cropland (C-C-C) occurred on average at higher elevations (203 m mean) than all other agricultural classes (figure 3). Among the change classes, cropland re-cultivation (C-G-C) tended to occur at higher elevations than the other classes. Stable grassland (G-G-G) existed at lower elevations and in areas with distinct lower HTC values than all the classes that were croplands in 1990 (figure 3). Furthermore, we found that stable cropland (C-C-C) and recent conversion of grasslands for crop production (G-G-C) had the best hydrothermal conditions, with a mean and median HTC above 0.7 (figure 3).
In terms of soil types, pre-Campaign cropland primarily occurred on soils with the highest or medium suitability for crop production (43% of pure Chernozem or Kastanozem, and 50% of Chernozem or Kastanozem Solonetz) (figures 3 and 7(B)). Cropland expansion during the Campaign largely occurred on the most suitable soils (by ∼58%), while nearly 60% of the expansion after the Campaign, from 1962 to 1990, were found on soils of lower suitability (figure 7(B)). Moreover, 19% of post-Campaign cropland expansion occurred on the least suitable soils. In the post-Soviet period, 69% of the stable cropland (C-C-C), and 45% of the abandoned cropland re-cultivation (C-G-C), occurred on the most suitable soils, whereas 76% of the early reversion of cropland to grassland, and 71% of the late reversion, occurred on soils with medium or lowest suitability (figure 7(A)).
Figure 7. (A) Distribution of soil classes per agricultural land-cover class for post-Soviet period. (B) Distribution of soil classes for 1990´s cropland by period of first documented conversion into cropland (soil rank no. 1 = highest suitability for crop production: Chernozem and Kastanozem; soil rank no. 2 = medium suitability: Chernozem and Kastanozem Solonetz; Soil rank no. 3 = lowest suitability: Solonetz and meadow soils). Acronyms for land-cover classes introduced in table 2.
Download figure:
Standard image High-resolution image3.3. Biophysical determinants of land-cover change and suitability for cropland expansion
We found a statistically significant association of land-cover changes with elevation, soil types and Selyaninov's HTC (table 5). Our models were robust to equal and unequal sampling of '0s' and '1s' with very marginal change of the coefficients, and models largely confirmed the results in section 3.2. For instance, an increase of soil rank by one unit increased the chances of land to be converted into cropland by 62% in the pre-Campaign period (model 1, table 5) and by 61% during the Campaign (model 2, table 5), while the chance of conversion was only 38% in the post-Campaign period (model 3, table 5). The higher odds ratio for the pre-Campaign and Campaign periods, compared to the post-Campaign cropland expansion model, suggest that cropland expansion during these periods primarily occurred on lands with better agro-environmental conditions relative to post-Campaign cropland expansion. Similarly, after 1990 a decrease of soil rank by one unit increased the likelihood of abandonment until 2000 (C-G-G) by a factor of three (model 4, table 5). After 2000, a decrease of soil rank by one unit increased the likelihood of abandonment until 2010 (C-C-G) by a factor of 2.4 (model 6, table 5).
Table 5. Odds ratiosa for agricultural land-cover change models from 1953 to 1990 and from 1990 to 2010.
| 1953–1990 | 1990–2010b | ||||||
|---|---|---|---|---|---|---|---|
| '1' (presence) | Model 1: pre-Campaign cropland | Model 2: Campaign cropland expansion | Model 3: post-Campaign cropland expansion | Model 4: C-G-G | Model 5: C-G-C | Model 6: C-C-G | Model 7: C-G-C |
| Variable/'0' (absence) | Stable grassland in 1953, 1961, 1990 | C-C-C | G-G-G + C-G-G + C-C-G | ||||
| Elevation (100 m) | 1.041 | 1.170 | 1.108 | 0.887 | 0.932 | 0.930 | 1.090 |
| Soil type (rank) | 0.377 | 0.393 | 0.621 | 3.072 | 1.761 | 2.451 | 0.561 |
| Selyaninov's HTC (unit * 100) | 3.435 | 4.984 | 2.185 | 0.398 | 0.543 | 0.634 | 1.635 |
| Number of presence observations ('1') | 3095 | 10 495 | 10 236 | 6387 | 2192 | 1632 | 2192 |
| Number of absence observations ('0') | 9929 | 9929 | 9929 | 10 320 | 10 320 | 10 320 | 18 151 |
| Adjusted R2 | 0.371 | 0.519 | 0.221 | 0.319 | 0.103 | 0.136 | 0.093 |
| AUC | 0.85 | 0.875 | 0.746 | 0.796 | 0.68 | 0.729 | 0.686 |
aAll odds ratios are statistically significant at p < 0.001. bAcronyms for land-cover classes introduced in table 2.
Agricultural land abandonment primarily took place on marginal lands from 1990 to 2000. Once socio-economic conditions changed, re-cultivation took place at the expense of these marginal lands. The likelihood to observe re-cultivation of such plots by 2010 on soils with low suitability (rank) was high (76%), albeit much lower compared to the likelihood of agricultural abandonment from 1990 to 2000 and from 2000 to 2010. This suggests, recent re-cultivation efforts focused on the best soils available, and least suitable soils for crop production remained abandoned.
We used the results from the logistic regression model (model 7, table 5) and assessed the suitability for future re-cultivation of currently abandoned croplands (figure B1 (1)). The comparison of the likelihood for re-cultivation with ongoing livestock expansion (figures B1 (2) and (3)) revealed that only few idle cropland plots with good agro-environmental characteristics remained available for cultivation due to competition of land use. Often, expansion of livestock grazing and associated provision of hay as fodder for subsistence farms was taking place on abandoned croplands.
4. Discussion
We conducted the first assessment of agricultural land-cover change in the northern Kazakh grain region during sixty years of Soviet and post-Soviet cropland development. Our results revealed high spatial and temporal dynamics in cropland extent from 1953 to 2010. Because of the importance of wheat procurement in the post-WWII Soviet Union, croplands substantially expanded in our study area, from only 8% of the total area in 1953 to 54% in 1990. At the same time, our results showed that the conversion of virgin steppe into croplands during the peak of the Virgin Lands Campaign (1954–1961) was concentrated in areas most suitable for agriculture. After 1961, cropland cultivation gradually expanded southward, irrespective of lower agricultural suitability there (figures 5(A) and 7(B)). Expansion in these marginal areas often led to rapid land degradation and subsequent cropland abandonment after only a few years of cultivation (Geipel 1964, OECD 2013). Thus, our results show that the majority of highly suitable lands had already been converted to cropland by 1961 in the course of the Campaign (Wein 1980).
The dissolution of the Soviet Union and the transition to a market economy were the underlying causes of the drastic decline in agricultural land use in Kazakhstan in the 1990s (World Bank 2004, Lioubimtseva 2010). The loss of guaranteed markets, disintegration of value chain supplies and deteriorating price relationships between inputs and outputs during the transition, promoted the decline in agricultural production and the agricultural land abandonment in post-Soviet times (Smith 1999, Ioffe et al 2004, Prishchepov et al 2013). In northern Kazakhstan, particularly, the decline of livestock and associated fodder crop production after 1990 notably contributed to the abandonment of cultivated croplands (Dudwick et al 2007, Suleimenov and Oram 2000), as confirmed by our interviews in the field.
Overall, the proportion of cropland in our study area decreased from 54% in 1990 to 30% in 2000, a relative reduction of 45%. Interestingly, most of the cropland that reverted to grassland after 1990 was only marginally suitable for agriculture and was, for the first time, ploughed only after the peak of the Campaign. In general, we observed higher rates of cropland abandonment from 1990 to 2000 compared to the subsequent decade, similar to other regions in post-Soviet Eastern Europe (Baumann et al 2011, Prishchepov et al 2012a, Griffiths et al 2013). Economic changes, such as economic adjustment toward open-market conditions, and the increase of world wheat prices after 2000 (FAO 2014) fostered re-cultivation of 6% of the idle cropland in our study region from 2000 to 2010. Similar re-cultivation rates of abandoned lands have also been observed in Romania and post-Soviet Ukraine (Griffiths et al 2013). Together, these findings suggest that comparable underlying drivers of land-cover change operated across former socialist countries after 1990.
Cropland abandonment and reversion to grasslands continues in our study area, partly due to incomplete land reforms, termination of agricultural production by bankrupt enterprises and ongoing structural change in the agricultural sector (OECD 2013, Petrick et al 2013, Glauben et al 2014). Although the overall amounts of re-cultivation and abandonment from 2000 to 2010 were almost equal, the two processes resulted in distinct spatial patterns of agricultural land-cover change because cropland abandonment primarily affected marginal areas (table 4, figure 5(B)). Conversely, the re-cultivation from 2000 to 2010 primarily occurred on lands with relatively favorable agro-environmental conditions compared to remaining idle croplands, albeit with much lower agro-environmental endowment in contrast to stable croplands from 1990 to 2010 (table 5, figure 7(A)).
Overall, we observed that 80%, or approximately 1.2 Mha, of the previously used cropland on the best soils was still under cultivation in 2010 (figure 7(A)). In contrast to optimistic expectations about untapped agricultural potentials on the abandoned agricultural lands in Kazakhstan (Liefert et al 2010, FAO 2011, Lambin et al 2013), our analysis of land-cover change and the increase of livestock density since 2000 (figures B1 (2) and (3)) showed that not much suitable land remains for future cropland expansion. In our study area, just about approximately 300,000 ha out of 1.7 Mha of idle croplands in 2010 have a high suitability for agricultural production. Any further cropland expansion would only be possible in marginal lands that originally were ploughed during and after the Campaign, but were quickly abandoned, especially after 1990. Moreover, strong competition for available idle croplands is expected from other land uses, for instance, due to implementation of governmental programs on agricultural diversification, and to support livestock production, including the development of the livestock fodder base (OECD 2013).
The fit of our model to map suitability for re-cultivation is adequate (table 5, model 7, AUC 0.68), given that we only used biophysical parameters as explanatory variables. Accounting for socio-economic parameters (e.g., proximity to grain processing facilities and markets) would have potentially improved the model fit. To proxy the spatial pattern of livestock dynamics (figures B1 (2) and (3)), we used data for subsistence livestock that we disaggregated by the aid of population census data (text S2). Maps of grassland productivity would also potentially corroborate the representation of livestock density, and thus the analysis of competition between crop and livestock production. However, such data were not available for our study. Nevertheless, our results support the recent findings about limited potential for cropland expansion on abandoned agricultural lands in post-Soviet Russia, where cropland abandonment is widespread (Schierhorn et al 2014). Ultimately, our results underscore the need to increase crop yields in existing fields and to use land that is currently uncultivated for extensive livestock grazing and the preservation of non-provisioning ecosystem services, such as carbon sequestration, as well as for biodiversity conservation (Kamp et al 2011, Kurganova et al 2014, Schierhorn et al 2013).
The remote sensing classifications had high accuracies for the stable land-cover classes, such as stable croplands and grasslands, whereas accuracies for the change classes were somewhat lower, despite the selection of multi-seasonal images, and the use of a non-parametric machine- learning classifier. Difficulties experienced in separating the change classes can be mainly attributed to the spectral similarities of more or less intensively managed and unmanaged grasslands in dry environments, and the limited spectral and temporal resolutions of the Landsat TM/ETM+ imagery (Klein et al 2012, Gong et al 2013). This is why we mapped broad agricultural land-cover change classes and complemented with spatial modeling of livestock density (text S2, figures B2 and B3).
Our cropland area estimates, using remote sensing data, were fairly close to the late Soviet official statistics (1990), and recent agricultural statistics (2010), but differed from official reports during the transition period (2000) (table C1). Given that official estimates lie well outside the error margins we calculated for our cropland area estimates, the possible difference between our and the official estimates can be largely attributed to uncertainties in official statistics, which were common during the early transition period across post-Soviet countries (Rumer and Zhukov 1998, Ioffe et al 2004). This accentuates the importance of monitoring land-cover change with remote sensing data, when official statistics remain uncertain.
Our remote sensing approach is well suited to map land-cover change in grassland and savanna ecosystems, and this approach could be used throughout the steppes of Central Asia. Furthermore, the increase in image availability after the launch of Landsat 8 in 2013 and, potentially, of Sentinel-2 in 2015, may offer new research opportunities at higher temporal and spatial resolutions, especially when used with imagery of higher temporal frequencies, such as MODIS.
5. Conclusion
Abandoned croplands may offer opportunities for cropland expansion, and re-cultivation of currently unused cropland may contribute to regional and global food security. Our study area, which is representative of the northern Kazakh rain-fed grain region, encompasses a substantial amount of abandoned cropland. However, we found that the potential for cropland expansion is limited in northern Kazakhstan, because the remaining idle croplands are mainly located in areas that are little suitable for crop production. In addition, given that recent agricultural policies in Kazakhstan are targeting an increase in livestock production and the diversification of crop production, increasing competition between different land uses can be expected. Increasing grain production may thus be more easily achieved by improving yields on existing croplands, rather than through additional cropland expansion. The remaining abandoned croplands in less suitable areas of northern Kazakhstan are likely more apt for grazing, and for conserving biodiversity and non-provisioning ecosystem services.
Acknowledgments
We gratefully acknowledge the support of the GERUKA project, which is funded by the German Federal Ministry of Food and Agriculture (BMEL), the Federal Office for Agriculture and Food (BLE) and the EPIKUR project, which is funded by the Leibniz Association's 'Joint Initiative for Research and Innovation' ('Pakt für Forschung und Innovation'). We also acknowledge the financial support of the German Ministry for Research and Education (BMBF, project KULUNDA), the NASA Land-Cover/Land-Use Change Program (LCLUC), and the Volkswagen Foundation (project BALTRAK). We thank our colleagues from the Analytical Centre of the Economic Policy in the Agricultural Sector (ACEPAS) in Astana, and Dauren and Rakhim Oshakbaev for their support in the field. We are grateful to Martin Petrick, Florian Schierhorn, Friedrich Koch and Brett Hankerson from IAMO for constructive discussions, and also to Achim Zetek and Alexander Mizgirev for their technical assistance. Two anonymous reviewers and an anonymous editorial board member provided valuable comments that greatly improved our manuscript.
Appendix A
Table A1. Comparison of Northern Kazakhstana and Kostanay Province using selected socio-economic and agro-environmental indicators.
| Group | Indicator | Northern Kazakhstana | Kostanay province |
|---|---|---|---|
| Socio-economicb | Cropland area (Mha) (share in the total area) | 15.4 (26.8%) | 5.0 (24.9%) |
| Area under grain crops (Mha) (share in the cropland area) | 13.2 (85.6%) | 4.4 (86.9%) | |
| Share of agricultural sector in the total gross regional product | 16.1% | 20.2% | |
| Share of crop production in the gross output of the agricultural sector | 65.3% | 73.3% | |
| Agro-environmentalc | Annual frost free days, mean (min, max) | 118.6 (89, 159) | 130.8 (109, 159) |
| HTC, mean (min, max) | 0.73 (0.28, 1.15) | 0.62 (0.28, 1.15) | |
| Share of pure Chernozem and Kastanozem over all soils | 45.0% | 41.8% |
aNorthern Kazakhstan = provinces Akmola, Kostanay, North Kazakhstan, Pavlodar. b2009–2011 average, source: ASK (2014). cSource: Afonin et al (2008).
Appendix B
Figure B1. (1) Tertiles of suitability of abandoned cropland re-cultivation by 2009 using results from spatially-explicit logistic regression (model 7, table 5). (2) and (3): livestock density for 1999 and 2009. For computation of livestock density, please refer to text S2.
Download figure:
Standard image High-resolution image
Appendix C
Table C1. Comparison of cropland area in Altynsarin and Auliekol districtsa by map estimates and official statistics for time steps 1990, 2000, 2010.
| District | Year | Remote-sensing based land-cover change mapb (thousand ha) | Official statisticsc (thousand ha) |
|---|---|---|---|
| Altynsarin | 1990 | 328.2 (±25.3) | 290.5 |
| 2000 | 199.9 (±8.0) | 161.5 | |
| 2010 | 239.6 (±10.8) | 225.1 | |
| Auliekol | 1990 | 568.5 (±43.8) | 505.6 |
| 2000 | 264.3 (±10.6) | 196.5 | |
| 2010 | 251.4 (±11.4) | 293.4 |
aThese are the only two districts completely lying within our study area. For location of districts in the study area please refer to figure S3. bNumbers in brackets stand for error margins with 95% confidence interval based on confusion matrix (see table 4). cSource: KDS (2001), (2012).