Mapping and assessing crop diversity in the irrigated Fergana Valley, Uzbekistan

Applied Geography 86 (2017) 102e117

Contents lists available at ScienceDirect

Applied Geography journal homepage: www.elsevier.com/locate/apgeog

Mapping and assessing crop diversity in the irrigated Fergana Valley, Uzbekistan € w b, John P.A. Lamers c Christopher Conrad a, *, Fabian Lo a

Remote Sensing Unit at the Institute of Geography and Geology, University of Würzburg, Oswald-Külpe-Weg 86, 97074 Würzburg, Germany MapTailor Geospatial Consulting, Bonn, Germany c ZEF (Center for Development Research), University of Bonn, Walter-Flex Str. 3, 53113 Bonn, Germany b

a r t i c l e i n f o

a b s t r a c t

Article history: Received 10 February 2016 Received in revised form 25 May 2017 Accepted 10 June 2017

Crop diversity (e.g. the number of agricultural crop types and the level of evenness in area distribution) in the agricultural systems of arid Central Asia has recently been increased mainly to achieve food security of the rural population, however, not throughout the irrigation system. Site-specific factors that promote or hamper crop diversification after the dissolvent of the Soviet Union have hardly been assessed yet. While tapping the potential of remote sensing, the objective was to map and explain spatial patterns of current crop diversity by the example of the irrigated agricultural landscapes of the Fergana Valley, Uzbekistan. Multi-temporal Landsat and RapidEye satellite data formed the basis for creating annual and multi-annual crop maps for 2010e2012 while using supervised classifications. Applying the Simpson index of diversity (SID) to circular buffers with radii of 1.5 and 5 km elucidated the spatial distribution of crop diversity at both the local and landscape spatial scales. A variable importance analysis, rooted in the conditional forest algorithm, investigated potential environmental and socioeconomic drivers of the spatial patterns of crop diversity. Overall accuracy of the annual crop maps ranged from 0.84 to 0.86 whilst the SID varied between 0.1 and 0.85. The findings confirmed the existence of areas under monocultures as well as of crop diverse patches. Higher crop diversity occurred in the more distal parts of the irrigation system and sparsely settled areas, especially due to orchards. In contrast, in water-secure and densely settled areas, cotton-wheat rotations dominated due to the state interventions in crop cultivation. Distances to irrigation infrastructure, settlements and the road network influenced crop diversity the most. Spatial explicit information on crop diversity per se has the potential to support policymaking and spatial planning towards crop diversification. Driver analysis as exemplified at the study region in Uzbekistan can help reaching the declared policy to increase crop diversity throughout the country and even beyond. © 2017 Elsevier Ltd. All rights reserved.

Keywords: Crop diversity Crop rotations Multi-sensor mapping Random forest Conditional variable importance Conditional inference trees Aral Sea Basin Fergana Valley

1. Introduction Crop diversification positively influences ecosystem services and hence agricultural production within agricultural landscapes since such practices can contribute to preventing soil degradation, maintaining soil fertility and soil health, or reducing soil erosion (Bullock, 1992; Dick, 1992; Naeem, Thompson, Lawler, Lawton, & Woodfin, 1994; Thrupp, 2000). Crop diversity is known also to decrease pest propagation and harvest damage (Matson, Parton,

* Corresponding author. E-mail address: [email protected] (C. Conrad). http://dx.doi.org/10.1016/j.apgeog.2017.06.016 0143-6228/© 2017 Elsevier Ltd. All rights reserved.

Power, & Swift, 1997; Tilman, Cassman, Matson, Naylor, & Polasky, 2002). Growing a mix of different crops can therefore be of paramount relevance for livelihood security (Smale & King, 2005). Consequently, the FAO sees an increasing land use diversity, i.e. the cultivation of wide ranges of annual and perennial plant species such as fruit trees, shrubs, pastures, and crops, as a one approach for improving resilience of agricultural ecosystems (FAO, 2011). This is in particular important in the agrarian landscapes of the Aral Sea Basin in Central Asia (CA) that have inherited a cotton-dominated farming system following seven decades of Soviet reign (1924e1991). The current practice of cotton monoculture has been considered a key culprit to the wide-spread and on-going soil degradation (e.g., Giese, Bahro, & Betke, 1998;

C. Conrad et al. / Applied Geography 86 (2017) 102e117

Orlovsky, Glantz, & Orlovsky, 2001). Given the potential of crop diversification practices such as an increase of crop types, crop rotations, and of the area share of minor crops they can be regarded as an important contribution to reaching more sustainable agricultural production systems in CA. Concerns about ensuring national food security and sustaining agricultural production stimulated Uzbekistan in CA to diversify crop type composition on its territory (ca. 4.1 Million ha (Mha) of irrigated land). To reduce its dependency from wheat imports, Uzbekistan started to promote irrigated winter wheat production after gaining independence in 1991 (Abdullaev, De Fraiture, Giordano, Yakubov, & Rasulov, 2009). As a consequence, in virtually a decade, cotton-wheat rotation systems covered already 70% of the arable area (Bobojonov et al., 2013). Recently, the national administration extended their crop diversity perspectives by promoting in addition the cultivation of vegetables, vineyards, and fruit trees where possible (Uzbekistan News, 2016) also because recent evidence underlined the potential of tree plantations to increase economic stability on highly unproductive areas in arid CA (Lamers and Bobojonov, 2008). Despite these efforts, spatial sectors of cotton mono-cropping patterns remained underling indirectly the complexity of increasing the current crop mix at the local scale such as suggested for West-Uzbekistan (Bobojonov € w, & Martius, 2016). et al., 2013; Conrad, Lamers, Ibragimov, Lo When comparing Uzbekistan with other successor states of the Soviet Union, Bobojonov et al. (2013) identified numerous factors that are likely to hamper a further development of crop diversification in Uzbekistan even in case state orders would be eased. These factors include water demand, processing knowledge and technology, as well as market access. Thus, the diversification of crops remains a pressing subject in Uzbekistan certainly in the light of the predicted increase of variability of the climate and in turn water availability due to human and natural impacts (Siegfried et al., 2012). Crop diversity used to be analysed from different perspectives. For instance, ecologists investigated predominantly the effects of crop diversity on soil development (Russell, 2002) or biodiversity (e.g., Duro et al., 2014; Palmu, Ekroos, Hanson, Smith, & Hedlund, 2014) and frequently identified crop diversity as being important for improving ecological conditions. Others explained crop diversity as a result of farmer's decisions in complex causal chains (e.g., Bobojonov et al., 2013; Sichoongwe, Mapemba, Tembo, & Ng’ong’ola, 2014; Singh, Kumar, & Singh, 2006). The results of these studies underscored that such causal chains are linked as well to political, economic, societal, cultural, and environmental or biophysical factors ranging from the world market over higher administrative decisions to local potentials and constraints (e.g., Bowman & Zilberman, 2013). Most studies attempting to explain crop diversity used statistical census data on higher administrative aggregation levels (e.g., Bobojonov et al., 2013; Rahman & Kazal, 2015; Sichoongwe et al., 2014; Singh et al., 2006). Depending on the specific purpose or scale, such statistical data may be sufficient, but when the subject is on understanding the variability of crop diversity, aggregated crop statistics are less suitable. Furthermore, when realizing the recurrently remarked incompleteness and inconsistencies of national agricultural databases once aggregated (e.g., Oberkircher et al., 2012; Forkuor et al., 2014), spatially explicit information about cropping patterns is indispensable when for instance aiming at the implementation of site-specific improvements e.g. for soil and crop conservation, soil degradation mitigation, or crop diversification. Satellite remote sensing has already proved to be a highly suitable tool for mapping the spatial distribution of crops (e.g., Wardlow, Egbert, & Kastens, 2007; Zhong, Wang, & Wu, 2015).

103

With the availability of multi-sensor data, crop distribution can be mapped with a much higher accuracy at field level than before, irrespective if in heterogeneous agricultural landscapes or over extensive study regions (e.g., De Wit and Clevers, 2004; Forkuor, €w, Duveiller, Conrad, Thiel, Ullmann, & Zoungrana, 2014; Lo Conrad, & Michel, 2015). Landscape metrics and indices have been frequently applied to such remotely sensed maps, e.g. for analysing spatial patterns as well as the changes thereof (Fahrig et al., 2011). Nevertheless, only few studies are available how such maps were used to process and further analyse information on biodiversity at the landscape scale. Pasher et al. (2013) suggested a scheme for optimizing landscape selection in ecological studies e.g. on biodiversity within agricultural landscapes and included a crop diversity index and field sizes derived from Landsat satellite data to describe landscape heterogeneity. Duro et al. (2014) utilized crop maps derived from Landsat to assess field sizes and crop diversity within a set of explanatory variables to predict the diversity of birds, butterflies, and plants. None of these studies targeted on the explanation of crop diversity within a landscape. Hence, at the example of the Fergana Valley in Uzbekistan and with the aim of tapping the potential of remotely sensed crop maps, the objective was to disclose not only the diversity of crop types and crop rotations in the existing production systems per se, but also the potential environmental and socio-economic drivers of current diversity patterns and its local variability. Diversity patterns were analysed by classifying multi-temporal Landsat and RapidEye data sets from 2010 to 2012. Afterwards, the resulting diversity patterns were regressed against a set of environmental and socio-economic indicators using random forest (RF) regression based on conditional inference trees. It is hypothesised that site-specific information supports decision-makers to phrase and implement means and measures urgently required to reducing a series of environmental and economic risks (Wehrheim & Martius, 2008). 2. Study area The Fergana Valley is located between two CA mountain ranges, the Alay and the Tien Shan. Cold winters and hot summers characterize the climate. Annual, predominately winter precipitation can be as low as 100e200 mm (Umarov, Kenjabaev, Stulina, & Dukhovny, 2010). Due to the arid climate, agricultural production in the valley depends on irrigation. Three out of the five countries in CA, Uzbekistan, Tajikistan, and Kyrgyzstan, share the territory of the Fergana Valley (Fig. 1a). With 90 inhabitants per km2 it is one of the most densely populated parts of the entire region (Reddy, Muhammedjanov, Jumaboev, & Eshmuratov, 2012). Irrigated agriculture is the leading economic activity as evidenced by the ca. 1.653 Mha of land annually supplied with irrigation water (SIC-ICWC, 2014). Main crops are cotton and winter-wheat, followed by orchards, and rice. Minor crops comprise sunflowers, watermelons, alfalfa, maize, sorghum, and vegetables (Conrad, Dech, Hafeez, Lamers, & Tischbein, 2013). Typically, the winter wheat season begins in autumn while maturity is expected in MayeJune and harvest late June and at the onset of July. The summer cropping season usually spans the period from mid-April till the end of October. To optimally utilize all available data sets (section 3.1.) this study concentrated on an area of 377,278 ha, covering 58,755 fields, in the inner part of the valley (Fig. 1). The field boundaries were derived based on image segmentation as previously reported (Conrad et al., 2013). Most of the fields are located in the three provinces Fergana, Namangan (Naryn and Syr Darya parts), and Andijan of Uzbekistan (Fig. 1a).

104

C. Conrad et al. / Applied Geography 86 (2017) 102e117

Fig. 1. a) Location of the study area in Central Asia (small map, UZB: Uzbekistan, KAZ: Kazakhstan, KZG: Kyrgyzstan, TKM: Turkmenistan, TJK: Tajikistan) and an overview of the investigated field objects (green colours, main map) within the different provinces (red), the location of the field samples collected in 2010, 2011, and 2012, are highlighted; b) zoom to the study area (red rectangle in figure part a) showing the RapidEye data (recorded on 19.5.2010) based on which the field objects had been extracted. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

well as all the Landsat data sets were co-registered to geometrically ground validated RapidEye data of 2010 (Conrad et al., 2013) using a 2nd order polynomial transformation in the AutoSync module of ERDAS Imagine.

3. Materials and methods 3.1. Data and pre-processing 3.1.1. Remote sensing data The acquisition dates of the RapidEye (RE) level 1B and Landsat 5 TM (LS) level 1G (WRS-2 path-row: 152-32) data sets are summarized in Table 1. One Landsat image, or at least two spatially adjacent RapidEye images were necessary to envelop the complete study area at a one time-step. The time lag between the two RapidEye data acquisitions for the same time step never exceeded seven days. During the field-based classification (section 3.2) the first RapidEye acquisition entirely covering one field polygon was selected in case of spatially overlapping data. Only data free of clouds over the study area, i.e. the cultivated part of the Fergana Valley, were selected. All data sets were atmospherically corrected using the MODTRAN radiative transfer code via the ATCOR-2 module (Richter, 2011). The RapidEye data of 2011 and 2012 as

3.1.2. Field data In 2010e2012, field samples were annually collected of the classes ‘cotton’, ‘rice’, ‘wheat-other’, ‘other’, ‘fallow’, ‘orchards’, and ‘water bodies’ (Table 2). The ‘wheat-other’ class describes an intraannual sequence of winter wheat followed by any second crop (usually rice, sorghum, or maize). The class ‘other’ summarizes all minor crops (like sunflower), meaning those crops, for which only few training samples were available. Fallow land refers to cropland temporarily not used, i.e. neither tilled nor sown in one or more years. The class ‘Orchards’ includes perennial woody vegetation such as fruits, but also other tree plantations or vineyards whilst the class ‘water bodies’ refers to a minor amount of fish ponds. The originally planned pure random scheme had to be adjusted due to

Table 1 Acquisition dates of the RapidEye (RE) level 1B and Landsat 5 TM (LS) level 1G (WRS-2 path-row: 152-32) data sets; Note: two RapidEye acquisitions were necessary for covering the study region which in most cases were acquired on two different dates. No of acquisitions

2010

2011

2012

1 2 3 4 5 6 7

14./19. May (RE) 13./15. June (RE) 2. July (LS) 4. September (LS) 6. October (LS)

2. May (LS) 13./20. May (RE) 3. June (LS) 23./29./31. July (RE) 7. August (RE) 22. August (LS) 7. September (LS)

3./4. April (RE) 21./23. May (RE) 30. May/1. June (RE) 17./29. June (RE) 2./5. July (RE) 1./3. August (RE)

C. Conrad et al. / Applied Geography 86 (2017) 102e117 Table 2 Number of field samples collected for each of the seven land use classes for the years 2010, 2011, and 2012. Class

Number of field samples 2010

2011

2012

Cotton Wheat-other Orchardsa Fallow Rice Other Water bodies (fishponds)

506 451 224 50 57 130 7

160 366 114 16 95 20 14

262 190 37 15 29 56 13

Total number

1425

785

602

a

Defined as perennial crops such as fruit, other tree plantations, and vineyards.

limited field access caused by bad roads, broken bridges, or missing permissions to visit the areas planned. All sample locations are shown in Fig. 1a. To capture differences in intra-annual ‘wheat-other’ rotations, field samples were collected after the wheat harvests in June and July 2010e2012. A common GPS device and maps showing the agricultural fields (plotted screenshots from Google Earth) were used for sampling. The collected field samples served the creation (training and validation) of the annual crop maps. In addition, 27 samples comprised information about three subsequent years (2010e2012) and were hence used for assessing the plausibility of the multi-annual map (crop rotations). They included the following rotations (the first, second and third class refers to 2010, 2011, 2012, respectively; the frequency is given in brackets): ‘cotton þ cotton þ cotton’ (1), ‘rice þ rice þ rice’ (1), ‘other þ other þ other’ (1), ‘orchard þ orchard þ orchard’ (3), ‘other þ wheat-other þ wheat-other’ (1), ‘rice þ rice þ cotton’ (1). Eleven ‘multi-annual’ samples showed two years of cotton and one year of wheat in rotation with other crops irrespective of the order. Eight samples consisted of two years of wheat and one year of cotton cultivation. 3.2. Crop mapping 3.2.1. Mapping algorithm Random forests (RF) are ensembles of single decision tree (DT) classifiers (Breiman, 1999), where a majority vote combines independent class decisions of the single trees. The RF results are highly accurate even when based on numerous input features such as multi-temporal satellite data (Rodriguez-Galiano, Chica-Olmo, Abarca-Hernandez, Atkinson, & Jeganathan, 2012). Within RF, bootstrapping is applied, i.e. each tree utilizes a subset of samples, and random selection of a limited number of features for node splitting. The randomForest package (Liaw & Wiener, 2002) in the R programming environment was used for classification. In accordance with previous recommendations (e.g., Breiman & Cutler, 2007), the number of DT in the RF ensemble was set to 500 and the number of features to split the nodes was set to half of the features within the corresponding input dataset. The input features for classification included the mean and standard deviation, which were calculated from spectral bands (i.e. five of RapidEye and six of Landsat TM) plus numerous vegetation indices described earlier € w, Michel, Dech, & Conrad, 2013) per agricultural field and per (Lo time step. This approach is referred to as object-based image classification (Blaschke, 2010). Each year was classified separately. Next, the resulting annual crop maps were concatenated to a three-year (multi-annual) crop map. It was assumed that orchards and water bodies remained

105

unaltered throughout 2010e2012, which was based on a crosscheck using the time slider in Google Earth (GE). Hence, any changes in temporally stable land use types in the annual crop maps most likely occurred due to classification confusion only. Thus, a majority vote was applied for labelling those fields for at least two years accordingly (in both, the annual and multi-annual crop maps). In case a field was classified as ‘orchard’ or ‘water body’ in only one out of the three annual crop maps, the label in the respective annual crop map and the multi-annual crop map changed to the class ‘other’. 3.2.2. Validation of the classification The classification accuracy for annual crop maps relied upon out-of-bag (OOB) samples. Those OOB samples were not drawn during the bootstrapping in RF (usually 1/3 of all samples, Breiman, 2002). Error matrices based on OOB samples were analysed in accordance to Congalton (1991), i.e. to derive the overall accuracy, and class-wise user's and producer's accuracies. For assessing the plausibility of the multi-annual crop map, i.e. the three-year crop sequences for each field, the 27 multi-year observations (section 3.1.2) were analysed. For assessing the impact of the class refinements for orchards in the multi-annual and annual crop maps, 100 independent samples of orchards were additionally collected from GE using the time slider function (producer's accuracy). The user's accuracy of orchards in multi-annual crop maps was analysed by inspecting 100 randomly selected orchard fields of the classification result in GE. 3.3. Crop diversity assessment The Simpson Index of diversity (SID, Simpson, 1949) is a widely used ecological indicator reflecting the probability of the next observed plant or animal being another species (Hurlbert, 1971). It indicates hence the richness and the evenness e.g. of species within a certain area (Magurran, 2004) and thus is a proxy for the spatial diversity of crop species at different scales, i.e. the occurrence and area (abundance) of the different crop types measured for a plot, an ecosystem or a region. Since the SID has served successfully as a measure for crop diversity in numerous studies (Palmu et al., 2014; Russell, 2002; Singh et al., 2006) including Uzbekistan (Bobojonov et al., 2013) it was preferred for estimating crop diversity. The SID was calculated according to:

PM SID ¼ 1 

 1Þ NðN  1Þ

m¼1 nðn

(1)

where n represents the area of one class m, M is the number of classes, and N refers to the area under observation. Values close to 1 point at a more heterogeneous and diverse cropping pattern, a value of 0 indicates in contrast a situation of monoculture. Practically, the SID was calculated by intersecting two layers of vector information: the field objects and the circular buffers using the function “isectpolyrst” as provided in the geospatial modelling environment (GME; http://www.spatialecology.com/gme/). From a geographical point of view, two aggregation levels were introduced. Two circular buffer patches around the field centres were delineated to investigate the crop diversity at both the local and landscape scales. The fact that the contour length of the field boundaries in the study region averaged approx. at 500e600 m suggested a buffer radius of 1.5 km around the field centres to be sufficient for observing the direct neighbourhood of the fields (local scale). However, to cover the neighbourhood of the smallest water administration levels in the region (water user associations with ca. 2.500 ha), an aggregation within a 5 km circular buffer (ca. 7854 ha) was considered to reflect the landscape scale. Testing

106

C. Conrad et al. / Applied Geography 86 (2017) 102e117

Table 3 Response variables for modelling spatial diversity patterns by using two radii (1.5 and 5 km circular zones). Variable

Description

Name

Y1 Y2 Y3 Y4 Y5 Y6 Y7 Y8

Crop diversity 2010 - 1.5 km Crop diversity 2011 - 1.5 km Crop diversity 2012 - 1.5 km Crop diversity 2010 - 5 km Crop diversity 2011 - 5 km Crop diversity 2012 - 5 km 3-year crop rotation diversity (2010e2012) - 1.5 km 3-year crop rotation diversity (2010e2012) - 5 km

SID_10_1.5 SID_11_1.5 SID_12_1.5 SID_10_5 SID_11_5 SID_12_5 SID_3y_1.5 SID_3y_5

additional circular buffer sizes (i.e., a wider radii around field centres) analysis raised some conflicts that obstructed collecting a sufficient amount of samples for the envisaged driver analyses (section 3.4.3). While excluding the class “water bodies”, the SID was calculated for the three annual crop maps. Intercropping i.e. subtree cultivation (e.g. maize, wheat and other crops) and private gardening (e.g. salads, onions, other vegetables), characterizes orchards in the Fergana Valley and suggests a higher crop diversity than e.g. on a cotton field. Hence, we artificially increased the crop diversity when orchards occurred in a patch or zone by replacing the area of these orchards with two classes covering half of the orchard area each. Afterwards, the SID was calculated for the 3-year crop rotations. The eight resulting variables (Table 3) served next as response variables in the follow-up analysis aiming at modelling spatial diversity patterns.

3.4. Statistical analysis of crop diversity 3.4.1. Set of indicators In accordance with the methodology reported in studies explaining crop diversity (e.g., Rahman & Kazal, 2015; Sichoongwe et al., 2014; Singh et al., 2006) the drivers of crop diversity in the Fergana Valley were detected through statistical analyses of indicators for crop diversity. The following drivers of crop diversity in Uzbekistan had previously been reported and these were linked to indicator variables as shown in Table 4: 1. Water demand and water availability play a major role in crop production (Awan, Tischbein, Conrad, Martius, & Hafeez, 2011; Conrad et al., 2013a; Horst, Shamutalov & Pereira, 2005; Martius, Rudenko, Lamers, & Vlek, 2012; Reddy et al., 2012; Tischbein et al., 2013), whilst crop type, irrigation strategy, and groundwater (level and quality) can be used as proxies for crop water demand at the field level. Water availability is limited by (institutionally regulated) water deliveries extracted from the only river (Amu Darya) as well as the capacity and quality of the canals used for bringing the extracted water to the irrigation areas/fields. Within the irrigation system, water availability depends also on the location (e.g. upstream-downstream, head or tail end) of a field within the system (Conrad et al., 2013b). Increased fluctuations in topography request increased irrigation efforts, which include higher demands for electricity to supply pumping stations, and hence, financial input. Consequently, elevation, slope, distance to river intake point, distance

Table 4 Indicators (predictor variables) including their description and preparation steps. The finally selected variables are numbered (X1-X8). Unselected indicators remained unnamed. Predictor variable

Name

Indication for

a) Indicators based on environmental variables Elevation (m a.s.l.) X1 a) land suitability for irrigation, high elevation suggests increased energy demands for pumping irrigation water to the fields b) precipitation: water demand decrease with increased elevation Slope (%) land suitability for irrigation, higher slopes restrict the production of some crops (increase irrigation efforts) Soil type X2 a) varying irrigation water demand (irrigation norms) b) varying crop (-rotation) recommendation, and c) other management demands (fertilizer application, etc.) d) risk of land degradation

b) Indicators based on socioeconomic/infrastructure variables Distance to river X3 access to irrigation water: long distances imply reduced amounts of intake point irrigation water and an increased risk of irrigation water losses during transport (due to losses or upstream water use, e.g. upstream- downstream disparities) Distance to canals X4 access to irrigation water, the same as X3 but at a higher level within the water distribution system Canal density X5 access to irrigation water: short-distance access to irrigation water enables the implementation of high crop intensity/diversity, but simultaneously monocultures of high-water demanding and/or profitable crops (e.g. cotton) Distance to X6 a) access to fields, it is assumed that on fields in the vicinity of Settlements settlements crops other than cotton or wheat are cultivated, which is a proxy for crop diversity that in turn necessitate intensive cultivation practices b) access to markets/processing factories Distance to roads X7 access to fields, near infrastructure is assumed to reduce management and transportation costs and increase crop diversity Road density X8 farmer's access to fields and opportunities for transportation to locations of processing or marketing

Source and preparation steps ASTER global Digital Elevation Model (DEM), (http://asterweb.jpl. nasa.gov/gdem.asp) Derived from ASTER global DEM using the slope function in the 3D analyst (ArcGIS) Zones extracted from the soil map of Uzbekistan from 1960 (Genusov et al., 1960): Mainly three out of the nine major soil categoriesa occurred in the fields under irrigation and were thus included: meadow march soil on alluvial, solonchak on alluvial deposits, meadow-saz soil of sierozem zone. One summary class of all other minor soil types was added Canal and river intake points (intersections between canals and rivers from OSMb waterways) were combined to cost-weighted distance layers. The Euklidean distance was calculated between the circular buffer zone center and the next canal from OSMb Polylines of the OSMb waterways layer were analysed within a radius of the 5 km buffer zone using the density function of the Spatial Analyst (ArcGIS) Road and settlement layers were combined to cost-weighted distance layers; Road hierarchy was considered. The settlement boundaries resulted from on-screen digitization using RapidEye € w et al., 2015b) data and Google Earth (GE) information of 2014 (Lo The Euklidean distance was calculated between the buffer zone center and the neighboured road from OSMb Polylines of the OSMb road layer were analysed within a radius of 2000 m using the density function of the Spatial Analyst (ArcGIS)

a Sandy cambic solonchak (1), takyric soils and takyrs (takyric salic solonetz (2), sandy desert soils (3), meadow soil on alluvial (4), meadow march soil on alluvial (5), solonchak on alluvial deposits (6), light sierozem (7), typical sierozem (8), meadow-saz soil of sierozem (9). b OSM: Open Street Map (© OpenStreetMap contributors).

C. Conrad et al. / Applied Geography 86 (2017) 102e117

to canals, and canal density were included as potential drivers into the analyses (Table 4). 2. Precipitation in the Fergana Valley increases with altitude, e.g. from the inner valley to the mountain foothills, making elevation thus to another indicator of water supply. Compared to water availability per se, other climate factors such as temperature or air humidity play a negligible role in the study area. The indicator lists therefore do not comprise climate information. 3. National recommendations for crop rotations are issued for the two state order crops cotton and winter wheat, although they hardly vary with soil types/conditions (Khalikov and Tillaev, 2006). Yet, poor soil conditions, reflecting the different stages of land degradation caused e.g. by soil salinity, strongly influence the growth and development of these crops (Dubovyk et al., 2013; Lamers and Bobojonov, 2008) and hence the most degraded fields can be excluded from state order areas. Thus, soil types were included in the list of indicators (Table 4). 4. Physical access to markets (or processing sites of commodities) and road infrastructure (e.g. an indicator for transportation costs) are important variables influencing crop diversification in the agricultural landscapes of, for example, Africa (Abdalla, €user, Bauer, & Elamin, 2013; Kankwamba, Mapila, & Leonha

107

Pauw, 2013), which has not been confirmed yet by empirical evidence for Uzbekistan, although this has often been indicated as a potential driver of commodity prices (Clement, Bhaduri, & Djanibekov, 2014). Depending on the geographical region, crop diversity can increase or decrease with an increasing distance from markets (Sichoongwe et al., 2014). In addition, population density has been previously accounted for as an indicator for crop diversity (Singh et al., 2006). Moreover, the low degree of agricultural mechanization that has been reported throughout Uzbekistan after the collapse of the Soviet Union (Spoor, 2007) reduces the accessibility of remote fields and hence influence crop diversity. We summarized these variables potentially for Uzbekistan as the distance to settlements (Table 4), assuming that neighboured settlements indicate both, residences of farmers and market or processing places. The density of the road network was added to indicate transportation costs and field accessibility for crop management.

3.4.2. Regression algorithm To account for correlations among the indicator variables (Table 4) and due to the presence of continuous and categorical

Fig. 2. Maps of indicator variables included in the modelling of crop diversity in the Fergana Valley. The sequence of the eight variables from upper left to lower right are: Elevation, soil zones, cost-weighted distance to water intake points (white triangles), distance to canal (grey lines), canal density (5 km), cost-weighted distance to settlements, distance to roads, road density (5 km, overlaid with settlements, white). Red lines demarcate the administrative provinces in the valley. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

108

C. Conrad et al. / Applied Geography 86 (2017) 102e117

Fig. 3. Example of a randomly selected circular buffers, a) 393 samples from the 1.5 km circular patches (used for ten model runs), b) 73 samples from the 5 km circular buffer (used for ten model runs).

indicator variables (Strobl, Hothorn, & Zeileis, 2009), the RF regression (RFR) was selected, which is based on conditional inference trees for modelling. In contrast to the original RF model, the identification of splits relies on the statistical correspondence between the predictor (independent) variables and the depending response variable (Hothorn, Hornik, & Zeileis, 2006). The variables of high relevance for the machine-learning model were identified through the use of the conditional variable importance (CVI). The CVI differs from unconditional measures (Louppe, Wehenkel, Sutera, & Geurts, 2013) since the former only permutes values of variable j within groups of observations and preserves the correlation structure between one input variable j and the other input variables (Strobl et al., 2008). Here, the ‘cforest’ and ‘varimp’ were used and implemented in the R package ‘party’ (Strobl et al., 2009) for the estimation of RFR and CVI. Tests returned stabilized results when setting the number of trees to 500 and the number of variables tested in each node to 5. All other parameters of ‘cforest_control’ and ‘ctree_control’ remained thus unchanged.

implemented in ArcGIS 10.2.1. This procedure allows for setting a minimum distance between two samples. Here, the minimum distance between the centres of selected circular buffer zones was set to 3 km (7.5 km) for the 1.5 km (5 km) buffer zone. This step avoided in addition the selection of similar and highly overlapping buffer zones (Pasher et al., 2013). The routine returned ca. 400 (ca. 75) buffer zones for the 1.5 km (5 km) buffer zone data sets for each out of the ten sample sets. Fig. 3 shows exemplarily one random sample for both circular buffers. Model performance was analysed by predicting the SID for all circular buffers of the study area and by comparing the results with the remotely sensed SID data. For quality assessment, correlation analysis after Pearson was conducted and the root mean square error (RMSE) was calculated (Richter, Hank, Atzberger, Locherer, & Mauser, 2011). 4. Results 4.1. Crop mapping: accuracy assessment

3.4.3. Model implementation The ASTER DEM showed steep slopes of more than 20% in many places of the usually flat inner part of the valley, where furrow irrigation for cotton production used to be practiced and which necessitates only reduced slopes (Reddy, Jumaboev, Matyakubov, & Eshmuratov, 2013). Thus, the values of the ASTER-DEM were considered as unrealistic and thus erroneously, possibly due to processing errors of the DEM product, for estimating slopes and this parameter was omitted from modelling leaving eight indicator variables that were considered for crop diversity modelling (Fig. 2). To analyse areas of comparable structure as suggested previously (Pasher et al., 2013), special attention was paid to the central part of the Fergana Valley (~400e520m a.s.l.) and the agricultural landscapes in the foothills were excluded. The remaining area, highlighted by the black delineation in Fig. 3, covers thus the inner valley. The condition ‘minimum percentage of agricultural area allowed within a 1.5 km and 5 km circular buffer’, was set to 50% and 30% (to achieve comparable SIDs), respectively. Altogether 33,041 fields remained for modelling. The eight predictor variables were spatially aggregated within the two assumed buffer zones (Table 4, variables X1 e X8). The average information was selected for the continuous variables and the majority information for the categorical variables. The RFR was run in total ten times for each of the ten different sample data sets that consisted of randomly selected circular zones. The ten sample sets were individually drawn by the routine “create random points”

The overall accuracies of the three annual crop maps were 0.86 (2010), 0.84 (2011), and 0.85 (2012). Class-wise accuracies showed similar values among the years, although their level varied among classes (Table 5). User's and producer's accuracy for both ‘cotton’ and ‘wheat-other’ classes were above 0.82. In all annual crop maps, confusion mainly occurred among the classes ‘other’, ‘orchard’, and ‘wheat-other’, mainly due to the fact that all these crop classes can show high vegetation signal in both, winter and summer seasons, in the remote sensing data. This confusion among these three classes meaningfully reduced the user's accuracy for ‘orchards’ Table 5 Class-wise accuracy assessment for the annual crop maps elaborated for 2010, 2011 and 2012 in the study area. Classes

2010

2011

2012

User's Producer's User's Producer's User's Producer's Cotton Rice Wheat-other Orchard Othera Fallow Water

0.97 0.84 0.97 0.89 0.42 0.74 0.86

Overall accuracy 0.86 a

0.89 0.89 0.93 0.71 0.75 0.84 0.86

0.90 0.85 0.86 0.68 0.50 0.87 0.86 0.84

0.94 0.80 0.89 0.65 0.30 0.81 0.86

0.92 0.78 0.82 0.56 0.69 0.93 1.00

0.98 0.72 0.86 0.65 0.36 0.87 0.92

0.85

Dominated by the crops sunflower, watermelon, alfalfa, maize, sorghum, and a series of vegetables.

C. Conrad et al. / Applied Geography 86 (2017) 102e117 Table 6 Agreement of the class ’orchard‘ in the three annual crop maps (no. of fields). 2010

2011

2012

No. of fields classified as ’orchard’

9217

10,906

8866

Agreement over two years (no. of fields) 2010 þ 2011 2011 þ 2012 2010 þ 2012

6275 7006 6262

Majority vote (2010e2012)a

7005

Agreement over three years (no. of fields)

6113

a

Extracted from the multi-annual crop map.

(0.56e0.89) and the producer's accuracy for ‘other crops’ (0.3e0.75). Similar patterns of confusion occurred found between ‘fallow’ and ‘other’, particularly in 2010 and 2012. The number of identified orchards varied over the three years (Table 6). An overlay of the annual crop maps returned 6113 fields as orchard in all 3 years. Applying the majority vote resulted in 7005 orchards in the multi-annual crop map. About 94% of the orchards of the adjusted, multi-annual crop map corresponded with the independent orchard samples from GE. For the individual years 2010, 2011, and 2012, the revisions improved the correspondence of the class ‘orchard’ to 93%, 81%, and 81%, respectively.

109

A comparison with the class-wide accuracy assessment (Table 5) underlined an increase of producer's accuracy for orchards in the annual maps. In addition, the user's accuracy for orchards improved as indicated by 75 out of the 100 orchard samples of the classification confirmed in GE. In addition, given that only 3 out of the 27, 3-year samples mismatched with the multi-annual crop map, multi-annual rotation map could be assessed to be plausible. 4.2. Spatial distribution of crops and crop rotations In all study years, irrigated cotton and wheat production dominated in the inner valley (Fig. 4aec). Orchards were scattered throughout the region, albeit dominating the moderate slopes of the geomorphological structures in the S and E, e.g. the alluvial fans at the foothills of the Alay Mountains (S). A concentration of rice fields occurred along the major canals in the NE-SW direction and, at least in 2010 and 2012, in the south of the Namangan-Syr Darya province. Fallow parcels and fishponds were concentrated in the center of the valley. The area shares and the number of fields detected for all classes and for all three study years virtually resembled each other (Table 7). About 270 different rotation classes occurred in the multiannual map. Nevertheless, visualizing all field objects where only ‘wheat-other’ and ’cotton’ were mapped for 2010e2012 exhibited the dominance of these two crops in the region. Mono-sequences

Fig. 4. Annual crop maps (a: 2010, b: 2011, c: 2012) and summarized multi-annual crop map (d), showing diverse rotations (i.e. three different crop classes in the annual crop maps 2010e2012), areas grown permanent with wheat or cotton, and mono-cropping areas (2010e2012) of rice, cotton, and wheat.

110

C. Conrad et al. / Applied Geography 86 (2017) 102e117

Table 7 Land use statistics according to land use classes as derived from the three annual crop maps (2010e2012) in the Fergana Valley. Crop

Cotton Rice Wheat-other Orchard Othera Fallow Water a

Number of objects

Objects (%)

Area (ha)

Area (%)

2010

2011

2012

2010

2011

2012

2010

2011

2012

2010

2011

2012

23,578 4601 16,351 7005 2419 4647 154

23,059 1416 19,072 7005 3832 4242 129

25,067 4817 15,628 7005 808 5183 247

40.1 7.8 27.8 11.9 4.1 7.9 0.3

39.2 2.4 32.5 11.9 6.5 7.2 0.2

42.7 8.2 26.6 11.9 1.4 8.8 0.4

157,435 21,347 132,913 32,760 11,840 19,338 1645

158,319 6625 135,660 32,760 20,603 21,655 1657

167,249 25,217 118,446 32,760 5012 26,418 2178

41.7 5.7 35.2 8.7 3.1 5.1 0.4

42.0 1.8 36.0 8.7 5.5 5.7 0.4

44.3 6.7 31.4 8.7 1.3 7.0 0.6

Dominated by the cropssunflowers, water melons, alfalfa, maize, sorghum, and vegetables.

or diversely cropped fields played a minor role. Field-based statistics of the multi-annual crop map showed a three-year lasting mono-sequence of several single crops such as cotton, rice, and wheat on 5611 fields (9.56% of the cropland area). Cotton dominated these mono-sequences (6.30%), followed by wheat (2.51%), and rice (0.72%). On the contrary, 7592 field objects (10.52%) showed diversified sequences at field level, i.e. three different cultivation classes within the observed time span. Two different crops in 2010e2012 were cultivated on 38,082 fields (70.28%).

the orchard-class dominated in the SE and displayed a comparatively low SID due to the temporal stability of this class. The E-W gradient showing an increasing SID in the annual crop maps, remained visible in the multi-annual map, albeit less pronounced. In particular the district Andijan (E) depicted higher SIDs than in the annual analysis pointing at planned rotations between cotton and wheat.

4.3. Crop diversity maps

Mean and standard deviations (SD) of the coefficient of determination (R2) and root mean square error (RMSE) were received from 100 RFR models applied to each of the ten sample sets (Table 9). Based on the R2 values, the eight predictors could explain between 37.6% and 50.8% of the SID variance. The lowest predictive power was observed for 2011, the highest for 2012. Model performance of the annual SID-maps for the 1.5 km circular buffers exceeded that for the 5 km buffer zone. The opposite was true for the SID-maps of the three-year rotations. Average CVI shows that the relevance of predictors for SID varied among the years and buffer sizes (Fig. 7). When assuming the 1.5 km radius (Fig. 7 aec, g), the distance to the river intake points was most important, followed either by the distance to settlements (2010, 2010e2012) or soil conditions (2011, 2012). Elevation played a major role as well. When analysing the 5 km radius (Fig. 7 eef, h), distance to settlements was a major predictor in 2010 and 2011. In 2012, the leading position became the distance to the water intake point at the river. Elevation, road density and the distance to settlements exhibited an outstanding importance in the 5 km zones as well (2010e2012). Distance to canals influenced the SID distribution only slightly. Distances to roads were found negligible for both circular buffers. The influence of soil conditions was generally higher in the 1.5 km patches than in the 5 km zones. Variable interactions among the predictor variables are shown in the node sequences of the conditional inference trees (CITs). Fig. 8 exemplarily depicts one CIT explaining the SID of the multiannual crop map in the 1.5 km patches. The nodes comprise the predictors and the thresholds. Final nodes (leaves) are highlighted with their SID levels (y-axis in the box plots of the leaves). In this example, the SID values mainly decrease with road density (root node 1), leading to increased SID distributions in the left leaves of the tree, but not in all cases. The SID distribution of those patches located in higher elevations (node 2) and in less densely settled areas (node 3) also show comparatively low SID.

The SID levels in the circular buffer zones of the annual crop maps depended on the observation years and the circular buffers assumed. Descriptive statistics (Table 8), such as minima, maxima, ranges, means, standard deviations (SD), and the medians of the annual and multi-annual maps (rotations) indicated a more compact distribution with higher means and smaller SD of SID in 5 km zones in comparison to 1.5 km patches. The SID values of the multi-annual map ranged at higher levels than those of the annual maps in both circular buffer zones. Spatial patterns of crop diversity expressed by the SID (Fig. 5) showed that within the 1.5 km radius (Fig. 5aec), more detailed spatial structures became visible in contrast to the 5 km radius (Fig. 5def). At this landscape scale (5 km radius), statistical effects balance out the variability of SID. Irrespectively of a single year, the SID increased from E to W. In comparison to the provinces Namangan and in particular Fergana (W), low crop diversity characterized the densely settled province Andijan (E), where cotton and wheat nearly exclusively covered the irrigated areas. Interannual variations of the SID were observable in the SW part of the map (province Fergana). At both scales of aggregation (1.5 km and 5 km), a decreased SID of the multi-annual crop map (three-year rotations) characterized the S and N locations of the study area (Fig. 6). Especially

Table 8 Distributions of the Simpson Index of Diversity (SID) according to the three study years and the multi-annual crop map within the 1.5 km and 5 km radii around the field centres. Year

Radius

Min

Based on the annual crop maps 2010 1.5 km 0.18 2011 0.23 2012 0.10 2010 5 km 0.53 2011 0.56 2012 0.53 For the composed multi-annual 2010e2012 1.5 km 0.06 2010e2012 5 km 0.45

Max

Range

0.86 0.68 0.84 0.61 0.85 0.75 0.85 0.32 0.84 0.28 0.84 0.31 crop map 0.97 0.91 0.97 0.53

Mean

SD

Median

0.65 0.66 0.67 0.69 0.71 0.71

0.09 0.08 0.09 0.07 0.05 0.06

0.65 0.67 0.68 0.69 0.71 0.72

0.84 0.89

0.07 0.06

0.86 0.90

4.4. Statistical analysis of crop diversity

5. Discussion The identification of spatial patterns of crop diversity in the Fergana Valley, Uzbekistan, was explored through multi-sensor

C. Conrad et al. / Applied Geography 86 (2017) 102e117

111

Fig. 5. Crop diversity maps for three years (2010e2012) in the Fergana Valley. Each field is attributed by the Simpson Index of Diversity (SID) values received for the zone around the center of the field. The left and the right parts refer to the 1.5 km (e.g. 5a,b and c) and 5 km (5d, e, f) radius zones.

remote sensing. Despite varying image acquisition dates during the study period (2010e2012), e.g. no data of September or October in 2012 (Table 1), comparable classification accuracies resulted for the three annual crop maps. In a systematic analysis of €w et al. (2013) temporal windows suitable for classification, Lo showed that in the Fergana Valley at least one data set of the late winter crop season (April-Mai) and one of the middle summer crop season (JulyeSeptember) are most important for successful

classification. On the one hand this explains that late season images, which in this study were available for 2010 and 2011, had lower impact on the classification accuracy. On the other hand, €w et al. (2015a) showed among others for the Fergana Valley Lo that in the CA irrigation systems from a certain number of data sets onwards distinct temporal windows are not necessary for a successful field-based classification, i.e. to achieve overall accuracies of approximately 0.85.

112

C. Conrad et al. / Applied Geography 86 (2017) 102e117

Fig. 6. Diversity maps of crop rotations over time (2010e2012) in the Fergana Valley. Each field is attributed by the Simpson Index of Diversity (SID) value of the circular buffer around the center of the field. The left and the right maps refer to the 1.5 km (6a) and 5 km (6b) buffers, respectively.

Table 9 Model performances, i.e. mean and standard deviations (SD) of R2 and RMSE, as a result of 100 RFR model runs applied to each of the ten sample sets collected in the Fergana Valley. Response variables (SID) 1.5 km

R2 SD(R2) RMSE SD(RMSE)

5 km

1.5 km

5 km

2010

2011

2012

2010

2011

2012

2010e2012

2010e2012

0.5080 0.0050 0.0620 0.0004

0.4080 0.0050 0.0635 0.0003

0.4960 0.0040 0.0692 0.0006

0.4220 0.0040 0.0511 0.0002

0.3760 0.0050 0.0415 0.0002

0.3890 0.0040 0.0491 0.0002

0.4130 0.0080 0.0584 0.0008

0.4650 0.0090 0.0426 0.0005

Multi-annual analysis showed some variations in the importance of the explanatory variables among the years and the two circular buffers (Fig. 7). Classification uncertainty that biases subsequent (Olofsson, Foody, Stehman, & Woodcock, 2013) or overfitting of the conditional forest to the training data may be a technical explanations for these deviations. But also crop distribution altered by years such as indicated by the drop in rice area in 2011 (Fig. 4). The latter consequently caused the inter-annual spatial variations of SID patterns (visible in Fig. 5; especially at the 1.5 km scale), which in turn allows for explaining the identification of different drivers. Changing land use patterns can be a result from variations in climatic conditions and water supply. Also, the implementation of governmental control on farmer's cropping pattern can lead to a decline of rice area in Uzbekistan in some years (Veldwisch, 2008). The differences in variable importance which were observed between both radii when modelling the SID (Fig. 7) could have been caused by overfitting, but also simply by the fact that the explanatory information (potential drivers) were averaged within the respective circular buffer. As a consequence, within the 5 km circular zones average distances to the next canal intake from the river varied in comparison to the 1.5 km buffer patches. However, despite these uncertainties major drivers were found to be similar, others continuously played a marginal role (e.g. distance to canal, distance to road) or varied with the scale of the analysis (distance to settlements, distance to intake).

For the study period (2010e2012) crop diversity in accordance with both governmental rules (for cotton and wheat) and cultivation recommendations (Abdullaev et al., 2009; SIC-ICWC, 2014) were widely implemented in the Uzbek part of the Fergana Valley. Cotton mono-sequences, recurrently mentioned as characteristic for the Uzbek irrigated landscape (e.g., International Crisis Group, 2005) occurred seldom, at least not during 2010e2012. On the other hand, more accurate assessments for instance to identify locations of long-term mono-sequences, however, require longer observation periods as previously shown (Conrad et al., 2016). The SID (Table 8) for the entire study region (0.72e0.73) exceeded that for whole Uzbekistan (0.68) estimated in 2008 albeit based on national statistical data sources only (Bobojonov et al., 2013). The relatively high share of orchards, at least compared to the rest of the country, is very likely a major cause of the high overall SID in the Fergana Valley. This is substantiated in particular by the findings in the Andijan oblast in E-Fergana, which is dominated by the typical cotton-wheat rotation. Yet, the SID levels of 0.65 are comparable to that reported for the Khorezm oblast in WUzbekistan (in 2008), where the same state order crops cotton and wheat dominated the crop portfolio (Bobojonov et al., 2013). The temporal difference between the similar results of both studies may not easily blur this comparison due to the strong control of the national government on the agricultural practices throughout the country (Abdullaev et al., 2009).

C. Conrad et al. / Applied Geography 86 (2017) 102e117

113

Fig. 7. Normalized conditional variable importance (CVI) of eight predictor variables; blue bars indicate the average of 100 model runs (10 samples with each 10 repetitions), a)-h) refer to the different models (years and circular buffer zones around the field centres). (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Locations within the irrigation system of the Fergana Valley that depicted very low SIDs can be prioritized for planning purposes when aiming at increasing crop diversification and diverse rotation systems. The introduction of alternative land use (e.g., vegetables and tree plantations), i.e. increase of crop diversity, can contribute to the reduction of water demands for irrigation (Tischbein et al., 2013), which in turn permits for decreasing groundwater levels and consequently secondary soil salinization

(Ibrakhimov, Martius, Lamers, & Tischbein, 2011). In Uzbekistan, the latter is frequently recognized as a strong indicator for a dysfunctional and unproductive irrigation system (Abdullaev et al., 2009). The presently elaborated maps could in addition be combined with economic models as previously suggested (Bobojonov et al., 2013) when aiming at selecting the most profitable alternative crops and locations for these to increase agro-biodiversity.

114

C. Conrad et al. / Applied Geography 86 (2017) 102e117

Fig. 8. Example of a conditional inference tree (CIT) from the RFR model 2010e2012. Important variables are frequently selected in the trees and show a clear impact on the accuracy of the tree when its predictor values appear within a node are permuted among each other (Strobl, Boulesteix, Kneib, Augustin, & Zeileis, 2008).

The identification of drivers of crop diversity patterns in the Fergana Valley showed low to moderate coefficients of determination only. It is quite likely that some unexplained variance can be attributed to indicators such as soil salinity (section 3.4.1), or soil suitability for irrigation in general, which is an important indicator used by the administration of Uzbekistan to assign the location for the two strategic crops cotton and wheat (Rudenko, Grote, Lamers & Martius, 2008). Also economic-based indicators at farm-level could have increased the predictive power of the model as e.g. found by Kankwamba et al. (2013) or Abdalla et al. (2013). Yet due to a limited data availability such potential key drivers could not be assessed, at least not without compromising on accuracy. Increased crop diversity was observed in 1.5 km circular patches of lower water availability, e.g. distal parts of the irrigation system, and at higher elevations necessitating more intensive irrigation efforts. In Malawi, water availability, due to a predominating rainfed agriculture, was also recognized as an important contributor to crop diversity (Kankwamba et al., 2013). The present crop maps showed that in remote locations less water demanding orchards occurred, which may be taken as a proxy that tree plantations are implemented for coping with climate-driven (water) variability (Lin, 2011). On the other hand, in the multi-annual derived SID maps, crop diversity decreased with elevation (as a most important predictor variable), because the occurrence of the orchards reduced the overarching number of rotations. Hence, the estimation of SIDs in areas dominated by perennial crops demand more long-term data before valid conclusions can be drawn. The assessment of the conditional inferences trees (CIT) showed that crop diversity increases with distance to the settlements whilst road density was a negative correlator to crop diversity. This is to some extend astonishing because crop diversity has recurrently been positively associated with short distances to markets and inherent reduced transportation costs (e.g., Sichoongwe et al., 2014). On the other hand, Singh et al. (2006) concluded in the vicinity of markets and transportation networks for India mechanized, high-income farming as an important factor reducing crop diversity. It seems that the present

findings in the Fergana Valley have been impacted by various, uncontrolled factors. For instance, a variety of crop types is hidden in the classes ‘other’ and ‘wheat-other’, which in turn comprises the intra-annual crop cycles of wheat with e.g. sunflowers, tobacco, or maize. However concurrently, the state-order crops cotton and wheat play an important role as is reflected as well in the appearance of less diverse cotton-wheat rotations (and particularly) in the most densely parts of the valley, where an increased crop diversity could have been expected. These findings together with those of Bobojonov et al. (2013) for W-Uzbekistan mirrors still the impact of the state order on crop diversification in the country and despite the declared strategy for increasing crop diversification in the past years. One recurrently mentioned prerequisite for flexible assessments of cropping patterns is the existence of field-based cropping archives (De Wit and Clevers, 2004), which under data-scarce conditions can be complemented by data obtained through other sources such as different optical systems (e.g., Landsat and RapidEye). Disentangling furthermore the ‘minor’ crops would allow for unravelling consequently the class ‘other’. This combined with different subtree cultivation patterns has a high potential to improve crop maps, which would be beneficial not only for spatial crop diversity assessments. The potential of high-resolution RapidEye data for distinguishing such classes has recently been demonstrated (Conrad et al., 2014) although obviously this procedure requires sampling strategies that take into account minor crop classes as well. In addition, long-term field sampling techniques certainly will improve the quantitative assessment accuracy of the multi-annual archives (crop rotation maps). While the accuracy of the elaborated maps was acceptable high, the aforementioned factors point nevertheless to several limitations. Besides more ground reference samples, open and free satellite archives of for instance the Copernicus Sentinel satellites offer positive prospects for more accurate classifications. The addition of radar images (e.g., Sentinel-1) could improve further the class discrimination as was previously argued (Stefanski et al., 2014; Waske & van der Linden, 2008).

C. Conrad et al. / Applied Geography 86 (2017) 102e117

6. Conclusions In particular the combination of Landsat and RapidEye data for 2010e2012 enabled the extraction of multi-annual field information. The improvement of accuracies for typical classes such as orchards, which resulted from overlaying multiple annual maps and applying a majority vote for the perennial class, suggests nevertheless additional room for improving land use maps in general by this approach. The Simpson index of diversity (SID) applied to crop maps within circular buffers representing the direct neighbourhood of fields and the landscape scale (1.5 km and 5 km) in Fergana Valley allowed for recognizing the potentially still, ill-managed patches (e.g. monocultures and fields under mono-cropping) within larger zones of homogenous crop diversity. However, identifying suitable buffer sizes may vary when the approach is transferred to other study regions. The flexibility of aggregation of SID when having per-field crop information also enables the delineation of crop diversity e.g. for administrative planning units, where tabular information is currently unavailable. This step can provide a more detailed scale in (economic) crop diversity studies, which mainly rely on national or provincial (tabular) information. Moreover, the use of SID and crop maps can support the selection of representative sites when aiming e.g. at investigating farm-level diversity. The application of the SID gave insights in the spatial cropping patterns in the Fergana Valley. Crop diversity varied throughout the irrigated farmland and, in contrast to recurrent statements, area-wide monocultures had not been detected, at least not for the study period. This confirms hence to a certain extend the declared orientation of Uzbekistan towards a higher crop diversification following their independence from the Soviet Union. Crop rotations (mainly comprising cotton and winter wheat) were widely implemented in the Fergana Valley. The comparatively high share of orchards and SID values exceeding national levels of crop diversity indicated that crop diversification is of interest to farmers and the administration alike, at least in some zones of the study area. Explorative statistical modelling of the SID showed that areas in the inner valley, typified by a high vicinity to settlements and a dense road network, indicated low crop diversity, and concurrently pointed at zones where cotton-wheat rotations dominated the crop portfolio. These modelling results support thus previous reviews on the impact of the state order on crop diversification in the country. However, the interpretation of the modelling results illustrated that complex causal chains linked by numerous factors at different scales influenced crop diversity patterns in the landscape. This challenges even more the need for future predictions of crop diversity not only in Fergana Valley, but also throughout Uzbekistan. Despite some data-oriented and methodological challenges, this study showed one possibility of using multi-sensor remote sensing data for establishing and maintaining multi-annual land use data sets at the field level. Such information enables consequently more detailed analyses of actual cropping situations in different landscapes and can thus contribute to economic and ecological studies on crop diversity. The driver analysis can help reaching the declared policy to increase crop diversity throughout the country but also beyond the study region. The spatial explicit information per se may also improves policymaking and spatial planning towards more crop diversification in agrarian-based countries worldwide. Acknowledgements This study was conducted within the CAWa project (‘Water in

115

Central Asia’) funded by the German Federal Foreign Office (funding no. AA7090002). The authors gratefully thank the Blackbridge Company and the German Aerospace Center (DLR) for providing data from the RapidEye Science Archive (RESA). We particularly thank Galina Stulina, Mirzahayot Ibrakhimov, Sylvia Seissiger, and Maren Rahmann for their assistance during the field campaigns and Sylvia Seissiger and Gunther Schorcht for supporting the data processing and classifications.

References Abdalla, S., Leonh€ auser, I., Bauer, S., & Elamin, E. (2013). Factors influencing crop diversity in dry land sector of Sudan. Sky Journal of Agricultural Research, 2(7), 88e97. Abdullaev, I., De Fraiture, C., Giordano, M., Yakubov, M., & Rasulov, A. (2009). Agricultural water use and trade in Uzbekistan: Situation and potential impacts of market liberalization. International Journal of Water Resources Development, 25(1), 47e63. http://doi.org/10.1080/07900620802517533. Awan, U. K., Tischbein, B., Conrad, C., Martius, C., & Hafeez, M. (2011). Remote sensing and hydrological measurements for irrigation performance assessments in a water user association in the lower Amu Darya river basin. Water Resources Management, 25(10), 2467e2485. http://doi.org/10.1007/s11269-0119821-2. Blaschke, T. (2010). Object based image analysis for remote sensing. ISPRS Journal of Photogrammetry and Remote Sensing, 65, 2e16. http://dx.doi.org/10.1016/ j.isprsjprs.2009.06.004. Bobojonov, I., Lamers, J. P. A., Bekchanov, M., Djanibekov, N., Franz-Vasdeki, J., Ruzimov, J., & Martius, C. (2013). Options and constraints for crop diversification: A case study in sustainable agriculture in Uzbekistan. Agroecology and Sustainable Food Systems, 37(7), 788e811. http://doi.org/10.1080/21683565. 2013.775539. Bowman, M. S., & Zilberman, D. (2013). Economic factors affecting diversified farming systems. Ecology and Society, 18(1). http://doi.org/10.5751/es-05574180133. Breiman, L. (1999). Random forest. Machine Learning, 45(5), 1e35. http://doi.org/10. 1023/A:1010933404324. Breiman, L. (2002). Manual on setting up, using, and understanding random forests v3. 1. Statistics Department University of California Berkeley. Retrieved from: http://oz.berkeley.edu/users/breiman/Using_random_forests_V3.1.pdf (last call: 14 April 2017). Breiman, L., & Cutler, A. (2007). Random forests-classification description. Berkeley: Department of Statistics. Bullock, D. (1992). Crop rotation. Critical Reviews in Plant Sciences, 11(4), 309e326. Retrieved from: http://www.tandfonline.com/doi/abs/10.1080/ 07352689209382349. Clement, Y. M., Bhaduri, A., & Djanibekov, N. (2014). Food price fluctuations in Uzbekistan: Evidence from local markets in 2002-2010. In J. P. A. Lamers, A. Khamzina, I. Rudenko, & P. L. G. Vlek (Eds.), Restructuring land allocation, water use and agricultural value chains. Technologies, policies and practices for the lower Amudarya region (pp. 275e294). Goettingen: V & R Unipress. Bonn University Press. Congalton, R. (1991). A review of assessing the accuracy of classifications of remotely sensed data. Remote Sensing of Environment, 37(1), 35e46. http://doi. org/10.1016/0034-4257(91)90048-B. €w, F., et al. (2014). Derivation Conrad, C., Dech, S., Dubovyk, O., Fritsch, S., Klein, D., Lo of temporal windows for accurate crop discrimination in heterogeneous croplands of Uzbekistan using multitemporal RapidEye images. Computers and Electronics in Agriculture, 103, 63e74. Conrad, C., Dech, S. W., Hafeez, M., Lamers, J. P. A., & Tischbein, B. (2013). Remote sensing and hydrological measurement based irrigation performance assessments in the upper Amu Darya Delta, Central Asia. Physics and Chemistry of the Earth, 61e62, 52e62. €w, F., & Martius, C. (2016). Analysing Conrad, C., Lamers, J. P. A., Ibragimov, N., Lo irrigated crop rotation patterns in arid Uzbekistan by the means of remote sensing: A case study on post-Soviet agricultural land use. Journal of Arid Environments, 124, 150e159. http://doi.org/10.1016/j.jaridenv.2015.08.008. Conrad, C., Rahmann, M., Machwitz, M., Stulina, G., Paeth, H., & Dech, S. (2013). Satellite based calculation of spatially distributed crop water requirements for cotton and wheat cultivation in Fergana Valley, Uzbekistan. Global and Planetary Change, 110, 88e98. De Wit, A. J. W., & Clevers, J. (2004). Efficiency and accuracy of per-field classification for operational crop mapping. International Journal of Remote Sensing, 20, 4091e4112. Dick, R. P. (1992). A review: Long-term effects of agricultural systems on soil biochemical and microbial parameters. Agriculture, Ecosystems & Environment, 40(1e4), 25e36. Duro, D. C., Girard, J., King, D. J., Fahrig, L., Mitchell, S., Lindsay, K., et al. (2014). Predicting species diversity in agricultural environments using Landsat TM imagery. Remote Sensing of Environment, 144, 214e225. http://doi.org/10.1016/j. rse.2014.01.001. Dubovyk, O., Menz, G., Conrad, C., Kan, E., Machwitz, M., & Khamzina, A. (2013).

116

C. Conrad et al. / Applied Geography 86 (2017) 102e117

Spatio-temporal analyses of cropland degradation in the irrigated lowlands of Uzbekistan using remote-sensing and logistic regression modeling. Environmental Monitoring and Assessment, 185(6), 4775e4790. http://dx.doi.org/ 10.1007/s10661-012-2904-6. Fahrig, L., Baudry, J., Brotons, L., Burel, F. G., Crist, T. O., Fuller, R. J., et al. (2011). Functional landscape heterogeneity and animal biodiversity in agricultural landscapes. Ecology Letters, 14(2), 101e112. http://doi.org/10.1111/j.1461-0248. 2010.01559.x. FAO. (2011). Save and Grow. A policymaker's guide to the sustainable intensification of smallholder crop production. http://doi.org/10.1787/9789264162600-en. Forkuor, G., Conrad, C., Thiel, M., Ullmann, T., & Zoungrana, E. (2014). Integration of optical and synthetic aperture radar imagery for improving crop mapping in northwestern Benin, West Africa. Remote Sensing, 6(7), 6472e6499. http://doi. org/10.3390/rs6076472. Genusov, A. Z., Gorbunov, B. V., & Kimberg, N. V. (1960). Soil map of Uzbekistan. General Directorate of Surveying and Cartography of the Soviet Ministry, GUGK. SSSR 1,500,000. €rungen in Trockengebieten Giese, E., Bahro, G., & Betke, D. (1998). Umweltzersto Zentralasiens (West-und Ost-Turkestan): Ursachen, Auswirkungen, Maßnahmen. Stuttgart. Horst, M. G., Shamutalov, S. S.., & Pereira, L. S. (2005). Field assessment of the water saving potential with furrow irrigation in Fergana, Aral Sea basin. Agricultural Water Management, 77, 210e231. http://doi.org/10.1016/j.agwat.2004.09.041. Hothorn, T., Hornik, K., & Zeileis, A. (2006). Unbiased recursive partitioning: A conditional inference framework. Journal of Computational and Graphical Statistics, 15(July), 651e674. http://doi.org/10.1198/106186006X133933. Hurlbert, S. H. (1971). The nonconcept of species diversity: A critique and alternative paramaters. Ecology, 52(4), 577e586. http://doi.org/10.2307/1934145. Ibrakhimov, M., Martius, C., Lamers, J. P. a, & Tischbein, B. (2011). The dynamics of groundwater table and salinity over 17 years in Khorezm. Agricultural Water Management, 101(1), 52e61. http://doi.org/10.1016/j.agwat.2011.09.002. International Crisis Group. (2005). The curse of Cotton: Central Asia's destructive monoculture. In Asia report (Vol. 93). Retrieved from: http://www.crisisgroup. org/en/regions/asia/central-asia/093-the-curse-of-cotton-central-asiasdestructivemonoculture (last call: 14 April 2017). Kankwamba, H., Mapila, M., & Pauw, K. (2013). Determinants and spatiotemporal dimensions of crop diversification in Malawi, (December). http://doi.org/10. 13140/2.1.2455.2964. Khalikov, B., & Tillaev, R. S. (2010). Practical recommendations on crop rotations in Uzbekistan. Uzbekistan (in Uzbek): Cotton Research Institute. Lamers, J. P. A., & Bobojonov, I. (2008). Financial analysis of small-scale forests in the Amu Darya lowlands of rural Uzbekistan. Forests, Trees and Livelihoods, 18(4), 373e386. Liaw, A., & Wiener, M. (2002). Classification and regression by randomForest. R News, 2(December), 18e22. http://doi.org/10.1177/154405910408300516. Lin, B. B. (2011). Resilience in agriculture through crop Diversification: Adaptive management for environmental change. BioScience, 61(3), 183e193. http://doi. org/10.1525/bio.2011.61.3.4. Louppe, G., Wehenkel, L., Sutera, A., & Geurts, P. (2013). Understanding variable importances in forests of randomized trees. Neural Information Processing Systems, 1e9. €w, F., Duveiller, G., Conrad, C., & Michel, U. (2015). Impact of categorical and Lo spatial scale on supervised crop classification using remote sensing. Photogrammetrie - Fernerkundung - Geoinformation, 2015(1), 7e20. http://doi.org/10. 1127/pfg/2015/0252. €w, F., Fliemann, E., & Chandraskekar. (2015). Interim report Fergana project. Lo €w, F., Michel, U., Dech, S., & Conrad, C. (2013). Impact of feature selection on the Lo accuracy and spatial uncertainty of per-field crop classification using support vector machines. ISPRS Journal of Photogrammetry and Remote Sensing, 85, 102e119. Magurran, A. E. (2004). Measuring biological diversity. Oxford: Blackwell Publishing. Martius, C., Rudenko, I., Lamers, J. P. A., & Vlek, P. L. G. (2012). In C. Martius, I. Rudenko, J. Lamers, & P. Vlek (Eds.), Cotton, water, salts and soums: Economic and ecological restructuring in Khorezm, Uzbekistan. Berlin Heidelberg New York: Springer. Matson, P. A., Parton, W. J., Power, A. G., & Swift, M. J. (1997). Agricultural intensification and ecosystem properties. Science, 277(5325), 1257e1258. http://doi. org/10.1126/science.277.5325.504. Naeem, S., Thompson, L. J., Lawler, S. P., Lawton, J. H., & Woodfin, R. M. (1994). Declining biodiversity can alter the performance of ecosystems. Nature, 368(6473), 734e737. http://doi.org/10.1038/368734a0. Oberkircher, L., Haubold, A., Martius, C., & Buttschardt, T. K. (2012). Water patterns in the landscape of Khorezm, Uzbekistan: A GIS approach to socio-physical research. In C. Martius, I. Rudenko, J. P. A. Lamers, & P. L. G. Vlek (Eds.), Cotton, water, salts and soums: Economic and ecological restructuring in Khorezm, Uzbekistan (pp. 285e307). Dordrecht: Springer Netherlands. http://dx.doi.org/ 10.1007/978-94-007-1963-7_18. Olofsson, P., Foody, G. M., Stehman, S. V., & Woodcock, C. E. (2013). Making better use of accuracy data in land change studies: Estimating accuracy and area and quantifying uncertainty using stratified estimation. Remote Sensing of Environment, 129, 122e131. http://doi.org/10.1016/j.rse.2012.10.031. Orlovsky, N., Glantz, M., & Orlovsky, L. (2001). Irrigation and land degradation in the aral Sea Basin. In S. W. Breckle, M. Veste, & W. Wucherer (Eds.), Sustainable land use in deserts (pp. 115e125). Berlin: Springer. Palmu, E., Ekroos, J., Hanson, H. I., Smith, H. G., & Hedlund, K. (2014). Landscape-

scale crop diversity interacts with local management to determine ground beetle diversity. Basic and Applied Ecology, 15(3), 241e249. http://doi.org/10. 1016/j.baae.2014.03.001. Pasher, J., Mitchell, S. W., King, D. J., Fahrig, L., Smith, A. C., & Lindsay, K. E. (2013). Optimizing landscape selection for estimating relative effects of landscape variables on ecological responses. Landscape Ecology, 28(3), 371e383. http://doi. org/10.1007/s10980-013-9852-6. Rahman, S., & Kazal, M. M. H. (2015). Determinants of crop diversity in the regions of Bangladesh (1990 e 2008) (Vol. 36, pp. 83e97). http://doi.org/10.1111/sjtg. 12086. Reddy, J. M., Jumaboev, K., Matyakubov, B., & Eshmuratov, D. (2013). Evaluation of furrow irrigation practices in Fergana Valley of Uzbekistan. Agricultural Water Management, 117, 133e144. http://doi.org/10.1016/j.agwat.2012.11.004. Reddy, J. M., Muhammedjanov, S., Jumaboev, K., & Eshmuratov, D. (2012). Analysis of cotton water productivity in Fergana Valley of central Asia. Agricultural Sciences, 03(06), 822e834. http://doi.org/10.4236/as.2012.36100. Richter, R. (2011). Atmospheric/topographic correction for satellite imagery. ATCOR-2/3 user guide 7.1. Richter, K., Hank, T. B., Atzberger, C., Locherer, M., & Mauser, W. (2011). Goodnessof-fit measures: What do they tell about vegetation variable retrieval performance from Earth observation data introduction - background. Proceedings of SPIE 2011 Remote Sensing for Agriculture, Ecosystems, and Hydrology, 8174. Rodriguez-Galiano, V. F., Chica-Olmo, M., Abarca-Hernandez, F., Atkinson, P. M., & Jeganathan, C. (2012). Random Forest classification of Mediterranean land cover using multi-seasonal imagery and multi-seasonal texture. Remote Sensing of Environment, 121, 93e107. http://doi.org/10.1016/j.rse.2011.12.003. €pfen, Wasser Rudenko, I., Grote, U., Lamers, J. P. A., & Martius, C. (2008). Wert scho sparen Effizienzsteigerung im usbekischen Baumwollsektor. Osteuropa, 58(4e5), 407e417. Russell, A. E. (2002). Relationships between crop-species diversity and soil characteristics in southwest Indian agroecosystems. Agriculture Ecosystems & Environment, 92(2e3), 235e249. http://doi.org/Pii S0167-8809(01)00295-X. SIC-ICWC. (2014). Regional information system on water and land resources in the Aral Sea basin (CAREWIB). Retrieved from: http://www.cawater-info.net (last call: 14 April 2017). Sichoongwe, K., Mapemba, L., Tembo, G., & Ng’ong’ola, D. (2014). The determinants and extent of crop diversification among smallholder farmers: A case study of southern province Zambia. Journal of Agricultural Science, 6(11), 150e159. http:// doi.org/10.5539/jas.v6n11p150. Siegfried, T., Bernauer, T., Guiennet, R., Sellars, S., Robertson, A. W., Mankin, J., et al. (2012). Will climate change exacerbate water stress in Central Asia? Climatic Change, 112(3e4), 881e899. http://doi.org/10.1007/s10584-011-0253-z. Simpson, E. H. (1949). Measurement of diversity. Nature. http://doi.org/10.1038/ 163688a0. Singh, N. P., Kumar, R., & Singh, R. P. (2006). Diversification of indian Agriculture: Composition, determinants and trade implications. Agricultural Economics Research Review, 19, 23e36. Smale, M., & King, A. (2005). Genetic resource policies, What is diversity worth to Farmers? (pp. 13e18) International Food Policy Research Institute (IFPRI). Available online: http://www.eldis.org/go/home&id¼23351&type¼Document #.WVF35VHiKUm. Spoor, M. (2007). Cotton in central Asia: “Curse” or “foundation for development”? In D. Kandiyoti (Ed.), Proceedings of a conference held at SOAS University of London 3-4 November 2005The cotton sector in central Asia economic policy and development challenges (pp. 54e74). London: The School of Oriental and African Studies, University of London. Stefanski, J., Kuemmerle, T., Chaskovskyy, O., Griffiths, P., Havryluk, V., Knorn, J., et al. (2014). Mapping land management regimes in western Ukraine using optical and SAR data. Remote Sensing, 6, 5279e5305. Strobl, C., Boulesteix, A., Kneib, T., Augustin, T., & Zeileis, A. (2008). Conditional variable importance for random forests. BMC Bioinformatics, 9(1), 307. http:// doi.org/10.1186/1471-2105-9-307. Strobl, C., Hothorn, T., & Zeileis, A. (2009). Party on! R Journal, 1(2), 14e17. Thrupp, L. A. (2000). Linking agricultural biodiversity and food security: The valuable role of agrobiodiversity for sustainable agriculture. International Affairs, 76(2), 265e281. http://doi.org/10.1111/1468-2346.00133. Tilman, D., Cassman, K. G., Matson, P. A., Naylor, R., & Polasky, S. (2002). Agricultural sustainability and intensive production practices. Nature, 418(6898), 671e677. http://doi.org/10.1038/nature01014. Tischbein, B., Manschadi, A. M., Conrad, C., Hornidge, A.-K., Bhaduri, A., Ul Hassan, M., et al. (2013). Adapting to water scarcity: Constraints and opportunities for improving irrigation management in Khorezm, Uzbekistan. Water Science & Technology: Water Supply, 13(2), 337e348. http://doi.org/10.2166/ws. 2013.028. Umarov, K., Kenjabaev, S., Stulina, G., & Dukhovny, V. (2010). Development of methods for irrigation monitoring in Central Asiad on example of Fergana Valley, Scientific report for the CAWa project. Tashkent: Uzbekistan. Uzbekistan Today. (17.1.2016). Uzbekistan reduces cotton production. Retrieved from: http://www.ut.uz/en/business/uzbekistan_reduces_cotton_production/? sphrase_id¼5769 (last call: 21 June 2017). Veldwisch, G. J. A. (2008). Cotton, Rice & Water: The Transformation of Agrarian Relations, Irrigation Technology and Water Distribution in Khorezm. PhD Dissertation. Germany: University of Bonn. Wardlow, B. D., Egbert, S. L., & Kastens, J. H. (2007). Analysis of time-series MODIS 250 m vegetation index data for crop classification in the US Central Great

C. Conrad et al. / Applied Geography 86 (2017) 102e117 Plains. Remote Sensing of Environment, 108, 290e310. Waske, B., & van der Linden, S. (2008). Classifying multilevel imagery from SAR and optical sensors by decision fusion. IEEE Transactions on Geoscience and Remote Sensing, 46, 1457e1466. Wehrheim, P., & Martius, C. (2008). Farmers, cotton, water and models: Introduction and overview. In P. Wehrheim, A. Schoeller-Schletter, & C. Martius (Eds.), Continuity and change- Land and water use reforms in rural Uzbekistan.

117

Socio economic and legal analyses for the region Khorezm. Studies on the Agricultural and Food Sector in Central and Eastern Europe (pp. 1e15). Leipzig: Leibniz Institute of Agricultural Development in Central and Eastern Europe (IAMO). Zhong, C., Wang, C., & Wu, C. (2015). MODIS-based fractional crop mapping in the U.S. Midwest with spatially constrained phenological mixture analysis. Remote Sensing, 512e529 (Cdl) http://doi.org/10.3390/rs70100512.