1.1.

Background

Agricultural production worldwide needs to sustainably increase for sustaining global food security for a growing human population and to stimulate economic development, especially in rural areas.¹ Although concerns of a global slowdown in growth rates of harvested yield are inconsistent with recent evidence of increased cropping intensity,² in fact, some regions are affected by a slowdown in agricultural growth and agricultural abandonment.^3–5 Land and water resources, especially in the drylands, are often sparse and exposed to various stress factors such as water scarcity, advancing land degradation, or negative effects on irrigation water availability by climate change or conflict,⁶ but also competitive water uses from other sectors. It is therefore important to monitor and if necessary adjust agricultural performance. This requires a comprehensive understanding of the pattern of agricultural performance across both space and time, to identify potentially vulnerable regions.

Measuring and quantifying agricultural performance is typically based on the relation of outputs (e.g., crop yield) to inputs (e.g., fertilizer or irrigation water, but also labor). Another measure for agricultural performance is the cropping intensity (the number of harvests per year). Spatially and temporally comprehensive and robust information must be collected frequently to trace back and better understand the determinants of the dynamics of agricultural performance,⁷ and to evaluate and possibly improve agricultural management. The challenge is retrieving reliable and spatiotemporally consistent information. Agricultural statistics can be outdated, of doubtful quality,⁸ or missing adequate sub-national historical inventory data.⁹ Doing in-situ sampling can become prohibitively expensive when the areas to be monitored are huge or inaccessible.

Satellite-based Earth observation (EO) is a well-known alternative for assessing large-scale agricultural systems. EO has a great potential for supporting agricultural monitoring due to its ability to provide consistent information across scales, from the level of single parcels to regional-level coverage. Previous studies illustrated that EO-based monitoring can provide detailed information on land use and productivity to support detecting critical zones that are characterized by low agricultural productivity¹⁰ or areas vulnerable to agricultural drought.¹¹ Such data provide valuable baseline information for decision makers to support implementing countermeasures to the adverse effects of land degradation or drought. Inversely, the detection of zones with a high agricultural productivity and high sustainability is useful to ensure their persistence.¹² The advantage of EO-based systems over conventional, terrestrial crop inventories is the timeliness and regularity, which is crucial for an operational land use monitoring system of a (near) real-time or retrospective concept.

Previous research has focused on land use monitoring and crop mapping using multitemporal satellite EO data, such as that from the Moderate Resolution Imaging Spectroradiometer (MODIS). Such studies present small-scale, regional^13–16 or large-scale, global^17–20 crop type mapping, where the majority of works clearly emphasizes the use of time series of normalized difference vegetation index (NDVI)^4,21 or the enhanced vegetation index.²² Another important agricultural application based on MODIS data is crop condition mapping addressing aspects of vegetation dynamics.²³ In the latter field, MODIS imagery is commonly employed for drought forecast or early warning applications^24,25 and it generally fulfills the requirements of agricultural production monitoring, namely having a quasigapless, consistent archive of regular observations with sufficiently high temporal resolution since the year 2000²⁶. Furthermore, the many contributions on agricultural applications employing MODIS data indicate that the relatively coarse spatial resolution provides an analytical scale of suitable spatial granularity for such applications.

Based on such advancements, operational monitoring systems have been developed that make use of meteorological data and remote sensing to support agricultural decision-making and research.²⁷ Examples are the Global Information and Early Warning System of the United Nations (UN) Food and Agriculture Organization (FAO)²⁸ or the Monitoring Agricultural Resources Program of the European Commission.²⁹ These monitoring systems focus on vegetation health, vegetation conditions, yield gap predictions, or assessments for drought risk estimation. Others are based solely on land degradation analysis caused by land abandonment, soil degradation, and desertification.³⁰ Previously developed monitoring systems for Africa focus on water productivity and take into account various sources of information, such as net primary productivity (NPP) or evapotranspiration (ET), e.g., the FAO water productivity open-access portal.³¹ Such monitoring systems usually provide long-term monitoring which not only enables the understanding of spatiotemporal patterns and land use change, but also constitutes a solid basis for predictive analysis.³² However, existing monitoring systems are mostly noncrop specific, hence, they do not inform about crop-specific land use intensity, or they are developed region by region and missing a comprehensive set of indicators that characterize agricultural performance.³³ In addition, they fall short in providing information about crop area and crop production. Moreover, when different basic (global) products are compared, e.g., cropland extent, crop types, or rainfall estimates, they often show large differences and show inaccuracies, which can partly be explained by restrictions due to the lack of in-situ data for calibrating crop growth models or crop classification, or barriers in operational adoption of EO-based methodologies. Relying on a single indicator to understand agricultural performance is likely insufficient to reveal a comprehensive understanding. For example, lower yields can be counterbalanced by higher intensity. Thus, it is recommended to combine various complementary indicators into an overall composite index of agricultural performance. Originally, the application of composite indices can be found in many studies related to agricultural, environmental, or socioeconomic vulnerability,^34,35 but has not been incorporated explicitly into an EO-based framework.

One region particularly affected by the adverse effects of an increasing food demand of a growing population and increasing agricultural performance is the irrigated agricultural system in the Aral Sea Basin (ASB) in Central Asia. It is one of the largest irrigated systems of the world, but only poorly equipped in terms of comprehensive monitoring systems that allow mapping and back tracing agricultural performance across the countries that belong to the ASB. Archive crop type or performance maps of the whole ASB do not exist yet, and previous studies on this issue were mostly developed region by region.

Against this background, we present a composite agricultural performance index (CAPI), based on satellite EO for assessing and mapping agricultural performance in the ASB. This index is based on a multitemporal sequence of maps that were derived from MODIS NDVI time series data using supervised, machine-learning-based time series classification. Since labeled ground truth data (i.e., crop type represented by a certain NDVI time series) for training and calibration at sufficient quality, quantity, and spatial granularity is often not available at a regional scale, we developed a framework that, first, generates and refines training data using a sequence of unsupervised and supervised classifiers in a hierarchical manner, and second, inputs that training data to a decision fusion classifier to estimate the underlying phenology of a given NDVI time series which at the same time can be used as a proxy for specific crop types (e.g., summer crops are depicted by the peak of season, i.e., maximum NDVI occurring during the summer season, whereas 80% of the summer crops are composed of cotton fields) and other cropland types (e.g., fallow and orchards). Herein, the specific technical goals are to

1.2.

Definition of Agricultural Performance

In this paper, we define agricultural performance as a combined metric of sustainability, efficiency, and productivity. This is an envisaged goal of the Central Asian countries to improve the efficiency of the cropping systems (i.e., the ratio between output and effort) and to increase the food production in order to ensure the food security for a growing population which is expected to reach $80\times {10}^6$ by 2030.³⁶ This aims at the UN Sustainable Development Goals to combat hunger and desertification and simultaneously achieving sustainable agricultural practices.³⁷ The combination of criteria such as sustainability, efficiency, and productivity into a composite agricultural index helps to better understand the patterns of agricultural development as well to identification of problematic (i.e., less sustainable, less efficient, and less productive) regions. The criteria are as follows.

High values of the proposed CAPI depict regions where all of the above criteria are met. Low values stand for the exposure to agricultural threats, such as land degradation or water scarcity, and the inversion of the CAPI can be used as a proxy to map agricultural vulnerability.

2.1.

Aral Sea Basin

The ASB is a transboundary river basin composed of the six countries Kazakhstan, Kyrgyzstan, Tajikistan, Turkmenistan, Uzbekistan, and part of Northern Afghanistan. The ASB is directly supplied by the Amudarya and the Syrdarya Rivers. Irrigated cropland in the ASB accounts for $\sim 8.5\mathrm{Mha}$ that stretches alongside these two major rivers.³⁸ Under Soviet leadership, the ASB experienced a pronounced increase of the irrigated cropland area from 5.4 Mha (1950s) to around 8 Mha (1990s), mostly with the aim of increasing the production of cotton.³⁹ However, the excessive withdrawal of freshwater and inefficient water use reduced the inflow of the rivers. As a result, the Aral Sea dried out almost completely, and the irrigated cropland became largely salinized⁴⁰ and partly abandoned.⁴¹ After the collapse of the Soviet Union and the following institutional and economic transformation, the agricultural production partly decreased sharply in some regions of the ASB. The socioeconomic and agricultural setting in each of these countries differs. Uzbekistan still has a state-regulated, planned economy, while there is an already well-established market economy in Kazakhstan, where the relative contribution of the agricultural sector to the gross national product has been decreasing since several years in favor of the secondary and tertiary economic sectors.⁴² This is reflected by the different water use policies and irrigation practices of the countries. Figure 1 shows the study area with the different irrigation zones within the ASB.

Fig. 1

The irrigation zones of the ASB. The red boxes depict the areas where in-situ validation data are available. Basemap: ArcGIS Online.

The climate in the irrigated regions of the ASB is dry-arid continental with 100 to 250 mm of annual precipitation mainly falling during the winter (December to February).⁴³ Annual precipitation in the mountains can exceed 1000 mm. Because of the aridity, agriculture in the study area is fully dependent on a dense irrigation and drainage network, which was extensively developed during the Soviet era. Tajikistan and Kyrgyzstan, both upstream countries, are rather water-rich, whereas Uzbekistan and Turkmenistan experience more often water scarcity due to their dependency on the water withdrawal from the main river systems. In the arid environment of the ASB, precipitation contributes only very marginal to the overall water availability of the croplands whereas irrigation is crucial for the agricultural productivity in the region.⁴⁴ To maintain sustainable water use management, it is important that the actual field ET is less than the potential ET. The ASB is characterized by a high variance in climatic conditions ranging from the Kazakh steppe, the irrigation zones in the periphery of the Pamir Mountains with higher irrigation water availability, to the Karakum and Kyzylkum deserts in Turkmenistan and Uzbekistan, respectively.

2.2.

Data

2.2.7.

In-situ validation data

Moreover, we used in-situ validation data from two different field campaigns conducted by the International Center for Agricultural Research in the Dry Areas in 2008 in Karakalpakstan, Uzbekistan, and 2014 in the Fergana Valley, Uzbekistan. The data consist of 516 surveyed fields in Karakalpakstan and 196 surveyed fields in the Fergana Valley based on GPS point measurements and include the respective crop types identified in the field at each location. The in-situ data extend approximately across 4% of the overall study area in the ASB (i.e., 350 ha) (Fig. 2).

Fig. 2

GPS-based in-situ validation data from two field campaigns: (a) Karakalpakstan, 2008 and (b) Fergana Valley, 2014. See Fig. 1 for the locations of these regions within the study area. Basemap: ArcGIS Online.

3.1.

Satellite Data Preprocessing

We calculated the NDVI from the red and near-infrared bands contained in the MOD09Q1 products. We excluded pixels flagged as no data and excluded snow/ice or clouds in the MOD09Q1 pixel reliability layer prior to filtering based on MODIS quality assurance information. Only pixels labeled “good data” or “marginal data” were retained. Then, the annual NDVI data vectors were smoothened through a Savitzky–Golay filter⁵³ to reduce data contamination by clouds, cloud shadows, climatic effects, and atmospheric noise. Several tests examined and determined the optimum size of the filter window to 5 and the degree of the polynomial to 4, resulting in a good fit of the original NDVI signal and retaining important peaks of the time series. All raster datasets (i.e., the cropland mask, NDVI, NPP, and ${\mathrm{ET}}_{\mathrm{a}}$ and ${\mathrm{ET}}_{\mathrm{p}}$) were resampled to the same spatial resolution as the cropland mask using nearest-neighbor interpolation. To keep the original values encoded in the MODIS grids, we kept the sinusoidal projection for the analysis. Based on the cropland mask (see Sec. 2.2.4), all nonagricultural pixels were masked out and not analyzed further.

3.2.

Multiannual Cropland Classification

Creating a set of multiannual cropland maps for a larger region, such as the ASB, requires the calibration of a consistent classifier algorithm that can be applied to data of different years and regions to ensure transferability and persistence. Due to the lack of a suitable set of in-situ training data, we employed a hierarchical approach employing unsupervised and supervised classifiers in a sequential manner to create labeled training datasets for the target cropland classes that describe the crop phenology. The assignation (labeling) to the target classes was done by manually selecting suitable clusters by visual interpretation of the respective NDVI time series response pattern, using the year 2016 as a baseline for all other years. Only clusters with explicit correspondence with the target class were considered and fed to a supervised classification algorithm. All other clusters were neglected. The classification algorithm classified the multiannual MODIS NDVI time series for all observation years for all the ASB in the defined cropland classes. The workflow is shown in Fig. 4 and described in the following.

Fig. 4

Cropland classification workflow.

3.2.1.

Training data generation

Based on the multitemporal MODIS data in 2016 consisting of 30 individual scenes, we extracted pixel-wise NDVI time series, each represented as a data point in a 30-dimensional feature space defined by the NDVI values from each point in time. We used $k$-means clustering, which is commonly applied to remote sensing time series data,⁵⁴ in order to detect clusters of pixels characterized by similar temporal signatures. To determine an optimum number of clusters, we performed $k$-means for a systematically varying number of clusters $k$, and we used silhouette analysis⁵⁵ to evaluate the quality of the assigned clusters for each $k$ by measuring the quality of the assigned clusters based on distances to the cluster centroid and to the nearest cluster for each data point. This yields an individual measure (i.e., silhouette score) for each data point, which is typically averaged over all data points to obtain a global separability measure. We visualized the clustering results by mapping a random subset of pixels (i.e., $N=5000$) from the 30-dimensional feature space into a two-dimensional (2-D) space using t-distributed stochastic neighbor embedding (t-SNE).⁵⁶ This allowed for visual inspection of the clustering results, as shown for a subset of the study area (i.e., Fergana Valley, Uzbekistan) in Fig. 5.

Fig. 5

T-SNE mapping of a random subset ($N=5000$) of the pixel-wise NDVI time series in Fergana Valley from the 30-dimensional feature space into a 2-D space: (a) color-coded by the assigned clusters from $k$-means using $k=14$ and (b) by corresponding silhouette scores.

To account for cropland variability across the different irrigation zones in the study area, we run separate $k$-means clustering for each irrigation zone, allowing for a different optimum number of clusters in each region. The resulting clusters were then assigned to seven target classes (see Table 1) by visually inspecting and interpreting the annual NDVI time series signature of each cluster and referring to literature research about major crop types in the ASB. The target classes are cropland classes based on phenology and growing season. They depict the major crop types by characteristic phenological signatures captured by the NDVI time series. Clusters that did not unambiguously represent one of these target classes were neglected. Fallow lands represents regions that are temporarily (or permanently) taken out of the production cycle, be it due to water shortage, farmer’s decision, land degradation, or land abandonment. On the latter, primary vegetation succession can be found. Due to this, fallow land encompasses several land cover types and has a temporal and spectral similarity with bare soil, shrubland, trees, or pastures and natural grassland. Figure 6 shows the NDVI signatures of the derived target classes.

Table 1

Target classes assigned to the clusters from unsupervised k-means clustering, and the number of selected training pixels in the ASB is given.

Class ID	Class name	Description	Major crops and vegetation types	No. of selected training pixels
1	Summer crops	Summer crops growing having the maximum NDVI values between June and September.	Cotton, summer maize, vegetables, and cucurbits	9642
2	Spring crops	Spring crops growing having the maximum NDVI values between March and June.	Winter wheat, winter barley, and oat	3505
3	Rice	Single season rice having the maximum NDVI values between June and October.	Rice	1972
4	Fallow	Fallow, not cultivated, but potentially arable land. Very often covered by herbaceous vegetation or shrubland.	Shrubland, grassland, steppe, scrubland, and weeds	9925
5	Double crops	Spring crops followed by summer crops, describing cropping intensity.	Spring crops followed by fodder maize, rice, sorghum, vegetables, or cucurbits	5670
6	Alfalfa	Characteristic “triple” or even more NDVI peaks.	Alfalfa, very seldom a combination of winter wheat, alfalfa, and vegetables	1103
7	Perennial crops	Perennial crops with a characteristic high NDVI peak throughout most of the season.	Mostly orchards, such as apple, apricot, peach, or vineyards.	4096

Fig. 6

Multitemporal NDVI signatures of the major cropland classes from the selected clusters, depicting typical phenological signatures, drawn from the Fergana Valley region: mean ($\mu$) and standard deviation ($\sigma$) values are visualized.

3.2.3.

Cropland classification validation

We validated the created cropland maps in a three-stage validation procedure:

• • Landsat cross-comparison: Cross-comparison of the MODIS-based cropland maps to a cropland classification created by the same classifier and same training labels, however, trained on Landsat 8 OLI NDVI time series of higher spatial resolution (i.e., 30 m).

• • Visual verification: Visual inspection and interpretation of the temporal signatures of the NDVI time series and Google Earth historic imagery based on a stratified random sampling scheme. We visually associated the selected samples with the defined classes and verified in available Google Earth imagery, whether the MODIS pixel is located within cropland or whether the land has been cultivated for this observation time.

• • External crop type validation: Accuracy assessment through GPS-based in-situ data associated with the target classes used herein.

Landsat cross-comparison: To assess effects of the relatively coarse spatial resolution of MODIS data (i.e., $\sim 230\mathrm{m}$) over time, we cross-compared them to high-resolution Landsat 8 data.^71–73 We randomly selected six Landsat 8 OLI scenes within the ASB (Fig. 7) from different years between 2013 and 2016 and applied the same proposed classification approach with the same target classes for each Landsat scene. A direct comparison between both datasets is difficult due to the large spatial resolution differences. The Landsat classification maps were aggregated and reprojected to MODIS spatial resolution and sinusoidal projection. We reassigned the majority class of the Landsat pixels located within a MODIS pixel by zonal statistics through an area majority rule. Based on these pairs of cropland classifications, we performed pixel-wise map comparison to derive agreement measures.

Fig. 7

MODIS and Landsat 8 tiles overlapping the ASB with the corresponding cropland mask (green) and the country and province boundaries.

Visual croptype verification: To assess spatial variations of accuracy across the study area, we estimated accuracy in the different subregions using a stratified random sampling approach⁷⁴ by visually inspecting the NDVI time series signatures of the samples and the corresponding classes. For each class, 250 samples were generated over the whole ASB for randomly selected years and assigned to the respective target class if the NDVI time series curves could be interpreted unambiguously. Such an approach can be also found in other related studies.^5,75,76

External croptype validation: We used GPS-based crop type in-situ data from two different field campaigns. To associate these discrete measurements with underlying agricultural parcels, we manually digitized the respective using Google Earth imagery. We then reclassified the in-situ labels to the phenologic classes used herein and discarded parcels that covered $<50\%$ of the overlapping MODIS cells. This procedure affected $\sim 5\%$ of the overall in-situ data.

3.3.

Agricultural Performance Indicator Derivation

Based on the multitemporal cropland information, several indices were calculated, each of them describing agricultural performance (Table 2).

4.1.

Training Data Generation

Figure 8 shows the distribution of the initial clusters that were associated with the respective cropland class. As the clusters are spatially not well distributed, additional training data were generated by training the voting classifier ensemble with the initial labeled clusters. The resulting classified pixels with more than 99.5% classification probability were added to the labeled training dataset to improve the spatial distribution of the training data and to account for spatial variability. The NDVI signatures of the newly selected training pixels were again visually inspected and used in the consequent cropland classification.

Fig. 8

Class labels resulting from the selected clusters associated with the target classes (a) and a preclassification step where pixels with a probability higher than 99.5% are added to the training dataset (b) to increase the sampling distribution. (Source: ArcGIS Online).

4.2.

Accuracy of the Cropland Maps

4.2.1.

Map comparison, MODIS versus Landsat

The accuracy assessment (Sec. 3.2.3) revealed a high overall agreement between the MODIS-based cropland maps from 2016 and the corresponding Landsat based maps (Table 7). Clearly, the distinction of summer crops and rice was sufficiently accurate. Lower precision and recall scores were achieved for the spring crops class that was confused with the double-crops class as both classes include winter wheat and other cereals of the same phenologic characteristics. Moreover, the pattern of the double crops (two maxima in the NDVI time series) is sometimes confused with the spring crops class when the lower second peak of the double-crops time series is similar to the lower second peak of the spring crops, which is often residual vegetation after the harvest. The lowest accuracy scores were achieved for the Landsat scene 158034 located in Turkmenistan due to the lower parcel sizes and a higher within-field heterogeneity. Soil salinity and natural vegetation mixed with agricultural land cause a more mixed NDVI response in the MODIS time series.⁹¹ The highest accuracy scores were achieved for the Landsat scene 157029 located in the Lower Syrdarya irrigation zone (Kazakhstan), where only 3 cropland classes were present (rice, alfalfa, and fallow). Overall, the classifications have a higher precision and a lower recall which means that the classifier identified most pixels semantically, but also missed a lot of actual true positives, resulting in more false negatives. An overall agreement of 74% between the Landsat and MODIS classification datasets can be regarded as sufficient. Figure 9 shows the classification examples and the agreement between both datasets cartographically.

Table 7

Overall and per-class agreement estimates obtained through cross-comparison to high-resolution cropland maps from Landsat 8 OLI where P=precision and R=recall.

Landsat tile	Year	Overall accuracy	Average F-score	Pavg	P1	P2	P3	P4	P5	P6	P7
				Ravg	R1	R2	R3	R4	R5	R6	R7
153032	2013	0.78	0.71	0.71	0.85	0.47	0.89	0.52	0.72	—	0.81
				0.73	0.68	0.61	0.94	0.53	0.66		0.94
160031	2014	0.72	0.61	0.62	0.88	0.32	0.85	0.56	0.48	0.51	0.73
				0.70	0.72	0.58	0.62	0.87	0.89	0.92	0.32
155032	2014	0.71	0.67	0.73	0.43	0.64	0.91	0.91	0.51	0.82	0.91
				0.73	0.92	0.93	0.93	0.34	0.66	0.98	0.35
158034	2015	0.58	0.52	0.59	0.55	0.81	—	0.82	0.45	—	0.34
				0.53	0.83	0.53		0.14	0.27		0.87
157029	2015	0.87	0.76	0.81	—	—	0.94	0.92	—	0.58	—
				0.79			0.98	0.46		0.94
154034	2016	0.74	0.71	0.73	0.82	0.69	0.75	0.98	0.82	0.41	0.65
				0.77	0.82	0.77	0.91	0.37	0.62	0.97	0.96

Fig. 9

Land use maps for the year 2016, Khorezm region (above) and 2016, Surxondarya region (below), respectively, derived from (a) Landsat and (b) MODIS.

A direct comparison is sometimes aggravated due to the different dimensionality of the temporal features of the Landsat dataset. For each year, much fewer Landsat observations (on average 11 observations within a growing season) were available due to cloud cover over the areas of interest. This reduced dimensionality (compared to the 30 available MODIS composite scenes per year) also may lead to a worse distinction of certain crop classes which require a higher data dimensionality for a better class separation, such as orchards and alfalfa.

4.3.

Classification Results

The output of the cropland classification is 17 cropland maps with 8 thematic classes for all years between 2000 and 2016. The classifications were used as input for the calculation of cropland and water use subindicators as already presented in Secs. 3.3.1 and 3.3.2. The results of the cropland classification were spatially aggregated to the first level of the GADM layer which is the country level (Fig. 10). The visual inspection of the multiannual crop share per country shows the water scarcity years 2001, 2008, and 2014. The crop share results for Northern Afghanistan, Tajikistan, and Uzbekistan show a significant increasing trend for double cropping systems. This suggests that land intensification has been targeted in order to increase food production which is driven by a growing population and at the same time decreasing cropland availability. The acreage of orchards and alfalfa show a positive trend in all countries for the observation period. It is clear that cropland dedicated for cotton growing has been reduced in favor of crop diversification as well as the focus on other crops with a more sustainable water footprint and a more efficient water use. This can be seen by the negative trend of the summer crops, of which cotton is the major crop. More precisely, the aggregated statistics corresponds to the agropolitical developments of Uzbekistan and Turkmenistan since 2011 to reduce cotton production in favor of cereals and orchards. Moreover, the occurrence of water-rich years has increased for the time period since 2008.

Fig. 10

Spatially aggregated cropland results for the entire observation period and for each country of the ASB. The cultivated land is shown as 100%. Values above 100% depict the gross irrigated area containing double cropping systems.

From the results for the Surxondarya and Vakhsh irrigation systems in Eastern Uzbekistan, Northern Afghanistan, and Tajikistan (Fig. 11), agricultural intensification and a decrease of unused fallow land have occurred during the observation period.

Fig. 11

Cropland maps for the year (a) 2000 and (b) 2016 for the transboundary irrigation systems of Uzbekistan, Afghanistan, and Tajikistan. This example reveals the extensification and the intensification of agriculture in Central Asia. Fallow lands have decreased massively in the past few years.

4.4.

Spatial Patterns of Cropland and Water Use Subindicators

The calculation of the subindicators yielded eight different maps as shown in Fig. 12 with each of them representing agricultural performance in a different manner. The ACI map reveals areas with a higher cropland frequency (or reversely, lower fallow land frequency) that are closer to the main irrigation systems of the Amudarya and Syrdarya rivers. The Fergana Valley mostly includes areas with a high ACI whereas this indicator draws a clear boundary between the provinces of Khorezm (UZB), Karakalpakstan (UZB), and Tashauz (TKM), showing high ACI scores in Khorezm only, whereas the areas peripheral to the deserts are affected with a higher fallow land frequency due to land abandonment, desertification, or soil salinization. The CII maps show a similar pattern of agricultural performance except the irrigation areas of the Lower Syrdarya, where no cropping intensity has been identified. High spatial crop diversity was found in similar areas, whereby cropping diversity in the Fergana Valley had less importance than in the provinces Tashkent, Jizzakh, or Samarkand. Usually, areas with high spatial crop diversity also achieve high temporal crop diversity scores, expressed by the TCDI as areas with a frequent and regular crop rotation and a higher crop variation over time. CLUI scores were high in most irrigation zones except in Turkmenistan, which can be explained through the relatively short duration of the growing season. The LCI shows hotspots of land reclamation in Karakalpakstan, Turkmenistan and other peripheral and arid regions prone to desertification and land abandonment.

Fig. 12

Subindicators results for the entire ASB, spatially aggregated on $25{\mathrm{km}}^2$ hexagons: (a) ACI, (b) CII, (c) spatial crop diversity index, (d) TCDI, (e) CLUI, (f) LCI, (g) WUEI, and (h) WAI.

5.1.

Correlation Analysis of the Subindicators

A correlation analysis between all subindicators (Fig. 13) shows the highest Pearson correlation coefficient scores between the CLUI and ACI indicators. The lowest scores were achieved between the CII and WAI indicators. In this sense, cropping intensity is independent from the actual and potential ET. Surprisingly, both water use indicators, WAI and WUEI, are not correlated. This can be explained with the relative high discrepancies between the actual and potential ET values of the WUEI indicator. Moreover, we generated scatterplots between each pair of subindicators and plotted them in combination with the cross-correlation matrix (Fig. 13), giving additional insights in the relationships between the subindicators.

Fig. 13

Cross-correlation matrix of the eight agricultural performance subindicators combined with pairwise scatterplots color-coded by the CAPI.

5.2.

Composite Agricultural Performance Index

The resulting map in Fig. 14 shows high CAPI values in proximity to the mountain areas in the Eastern ASB as well as in close proximity to the main irrigation channels. A very high correlation between both maps, especially in zones with a higher agricultural performance, is evident. In general, the usage of the additional MODIS products for NPP and ET for the calculation of the CAPI through all subindicators seems to suggest higher agricultural performance, even if crop diversity, cropping intensity, and crop rotations are relatively low, and fallow land occurrence relatively high. The computation of the CAPI solely through the cropland subindicators is more precise for the depiction of low agricultural productivity areas. A drawback of the CAPI is the high abstraction and generalization level. Areas that perform better within a specific subindicator might still be interpreted as vulnerable or less productive, e.g., areas with a high ACI (i.e., low fallow land occurrence or frequency) might still have low cropping intensity, low spatiotemporal crop diversity, low land utilization, or a low net irrigated area share. This could be monocropping areas without crop rotation, such as continuous summer cropland (in most cases cotton). The inverse CAPI can be also interpreted as an agricultural vulnerability index depicting critical areas with the following categorization:

Fig. 14

Mapping of CAPI, grouped in five qualitative categories (very low to very high). Cartographic color-coding reveals zones of high agricultural performance in green.

However, the CAPI does not provide information on the causality of the vulnerability or the drivers which have led to such circumstances. They can be numerous and often linked to socioeconomical factors. The benefit of a CAPI over single indicators is the capability to summarize and to reduce the dimensionality of the observed variables. Especially in the constantly increasing domain of EO data, generalization is essential to build robust and descriptive global or regional models. Moreover, land use is a very dynamic, spatially and temporally varying issue where a single indicator is not sufficiently descriptive enough. Composite indicators had their origin in socioeconomic sciences where a variety of indicators is merged into a more meaningful and comprehensive information. The CAPI can be further used by agrobusiness or agroinsurance companies to enhance the information on the performance, efficiency, vulnerability, and sustainability of the used cropland. As accurate crop yield predictions are still very difficult to obtain, the CAPI can be also used as a proxy indicator for crop performance, identifying hotspots of potentially higher or lower crop yields. If applied in a real-time or near-real-time manner, the CAPI could also be used as an early warning system for impending land degradation.