Introduction

Water mismanagement, pollution, rapid population growth, and water-related environmental and ecological problems are major issues in the Aral Sea basin, located in the semiarid and desert areas of Central Asia (Martius et al. 2004). Although these parts of the desert regions of Central Asia are associated with high evaporation and infiltration rates, the main reason for the shrinking of the Aral Sea (Ressl and Micklin 2004) and the resulting ‘Aral Sea Syndrome’ (WBGU 1998) is the human-induced diversion of river water for irrigated agriculture. Between 1960 and 1990, cotton cultivation in Central Asia expanded from 3 million ha to 8 million ha (Micklin 1991). Since independence from the Soviet Union in 1992, Uzbekistan has continued the cotton production practices developed under the Soviet planned economy and in addition introduced a state order for winter wheat from 1995 onwards (Ruzmetov et al. 2003). Presently, Uzbekistan and the other central Asian countries top the list of highest per-capita water use (WWF 2002).

Due to the aridity of the climate, agricultural production on more than 90% of the arable land in Uzbekistan is only feasible with irrigation (Chub 2000). The agriculture-based economy of the Khorezm region in northwest Uzbekistan highly depends on irrigation from the Amu Darya River. The development of extensive irrigation schemes but with improperly working drainage systems caused rising groundwater tables, and consequently drastically changed the local and regional hydrology (Ibrakhimov 2005). In addition, the irrigation water-use efficiency is very low (Dukhovny et al. 2004) and there is an urgent need to find options to “grow more crop per drop.”

One important means to approach this goal may be to increase the low water use efficiencies, which is feasible when management decisions are based on an analysis of accurate hydrological information at each scale. Unfortunately, until today water resource managers in Central Asia face both a paucity of reliable and consistent information (Dukhovny et al. 2004) and inadequate and inappropriate tools for analysis, especially at the irrigation system and watershed level (Chemin et al. 2004). The use of appropriate, reliable, and consistent information at the irrigation system level can improve water management significantly, e.g., by indicating necessary changes for managing water more beneficially, negative impacts in downstream areas can be minimized. An important challenge for water managers is, furthermore, obtaining timely, objective and accurate information on water usage at each spatial scale. Despite existing land and water regulations from field to basin scale, major problems in the Central Asia countries remain the inconsistent and incomplete information about the actual resource usage (Chemin et al. 2004) and the absence of monitoring or control facilities (Dukhovny et al. 2004). To bridge this gap, spatial and temporal information about the volume of water consumption for the land use classes in the basin are needed. Recent developments in the remote sensing sciences and their application to water resources management now allow the provision of accurate spatio-temporal information about land and water use patterns in large basin (Schmugge et al. 2002).

The current research was carried out under the framework of the ZEF/UNESCO project, which addresses among other concerns the assessment of valid and reliable information about availability, distribution and usage of water in Khorezm (Martius et al. 2004). To complement the findings on water use and efficiency on the field level (Forkutsa 2006), analyses of land and water use patterns for all of Khorezm have been initiated but still suffer from insufficient information (Schweitzer et al. 2004; Ruecker et al. 2005). The major objective of this study was to map and assess spatio-temporal water use patterns in the large scale irrigation system of Khorezm by utilizing MODIS remote sensing data and GIS to improve decision making. Additional objectives were (1) mapping and validating accurate land cover and land use based on time series of 8-day, 250 m MODIS MOD09 surface reflectance data and (2) assessing the validity of 1 km MODIS MOD11 land surface temperature (LST) data for mapping seasonal actual evapotranspiration by using SEBAL and METRIC.

Remote sensing for irrigation water management

Remote sensing techniques are highly suitable for classifying regional and global scale land cover, and for delineating irrigated areas (Droogers 2002) or land use patterns in agricultural regions (Thenkabail et al. 2005; Bastiaanssen et al. 2000). Several approaches exist to derive land cover and use classes from different types of remote sensing data (Richards and Xiuping 2005). In particular, high temporal resolution data in combination with multitemporal classification strategies have been found useful for describing the temporal dynamics of seasonal vegetation cover and land use in irrigated areas (Thenkabail et al. 2005).

Until 2000, the only sensor providing daily data with an acceptable spatial resolution of 1 km2 in an operational mode was the Advanced Very High Resolution Radiometer (AVHRR) on board TIROS-N and NOAA 8-15 (Jensen 2000). Essential improvements were obtained with the MODerate resolution Imaging Spectroradiometer (MODIS) on board the Terra (EOS-AM1) and Aqua (EOS-PM1) platforms launched in 1999 and 2002, respectively (Savtchenko et al. 2004). The high spectral resolution of 36 bands allowed improvements in atmospheric correction, which resulted in the reflectance product (Vermote et al. 1997) as a standardized input parameter of the MODIS land product suite (Justice et al. 1998). MODIS spatial resolutions of 250 m within the red and near infrared spectra and of 500 m in the visible and short wave infrared (five bands) spectra increased the level of detail for classification.

Unsupervised clustering techniques were applied to AVHRR time series of Normalized Difference Vegetation Index (NDVI; Loveland et al. 2000) or filtered data (Viovy 2000) to distinguish land cover classes on a continental scale. The results of a cluster analysis of 42 MODIS 500 m 8-day surface reflectance data sets and a complex set of features describing the location of a class within a feature space were integrated into a decision tree to classify different crop rotations in the Ganges and Indus River basins (Thenkabail et al. 2005). DeFries et al. (1998) applied supervised classification trees to metrics of 8 km AVHRR time series resulting in 57 nodes for 17 global land cover classes. One year of monthly 1 km AVHRR data made it possible to distinguish 12 classes by a cascading two-class hierarchy of classification trees (Hansen et al. 2000). MODIS 500 m reflectance data were combined with other GIS data in “classification and logistic regression” trees to derive land use and predict deforestation in Colombia (Etter et al. 2006). Remote sensing techniques also proved useful for estimating evaporation when solving the energy balance of thermodynamics fluxes at the earth surface. This approach has gained momentum since 1990 due to the relatively low cost of data collection, which amounts to $0.03/ha for irrigated lands (Sakthivadivel et al. 1999). Various methods for estimating actual evapotranspiration (ETact) have been investigated by combining satellite images and ground meteorological data for large areas (Vidal and Perrier 1989; Choudhury 1994; Granger 1997).

One method for estimating ETact is the Surface Energy Balance Algorithm for Land (SEBAL) developed by Bastiaanssen (1995, 2000) and Bastiaanssen et al. (1998). SEBAL is designed for flat terrain in semiarid regions and is based on the calculation of the latent heat of ET (λET) as a residual from the surface energy balance. In recent years, SEBAL has been widely used to model seasonal ETact by utilizing remote sensing data (Bastiaanssen 2002). METRIC extends SEBAL for mountainous relief and couples it with reference evapotranspiration for alfalfa and was used to predict seasonal ETact based on Landsat data in the USA (Allen et al. 2005). Tasumi et al. (2005) demonstrated the internal calibration of SEBAL and METRIC showing with sensitivity analyses the independence of atmospheric influences on the thermal radiation received at the sensor. Verstraeten et al. (2005) analyzed the suitability of different remote sensing approaches in view of operational estimations of ET from NOAA data over European forests.

SEBAL was applied in more than 15 case studies to all remote sensing sensors receiving thermal information (e.g., Landsat TM and ETM, NOAA, ASTER, MODIS, and Meteosat) and carried out worldwide in different climatic regions (Hafeez and Khan 2006). The accuracy of SEBAL measured on a field scale by lysimeter, scintillometer, Bowen ratio towers, or Eddy correlation fluxes varied between 1 and 33% for instantaneous ET and between 2 and 30% for 10-day ET (Bastiaanssen et al. 2005). At the catchment scale, accuracies were estimated by the water balance, which led to results of less than 11%.

Description of the study area

Research was conducted in the Khorezm region, located in the northwest of Uzbekistan and between 60.05 and 61.39 N and 41.13 and 42.02 E of the Greenwich meridian, approximately 225 km south of the present shores of the Aral Sea. Khorezm is part of the lower Amu Darya River floodplain (Fig. 1, upper right map). Besides the capital Urgench, the region encompasses ten administrative districts. Khorezm encompasses an area of about 5,600 km2, of which 260,000 ha (2,600 km2) can be irrigated.

Fig. 1
figure 1

The irrigation system within the administrative boundaries of Khorezm oblast, Uzbekistan. The small map in the upper right corner illustrates the location and dimension of Khorezm within Central Asia

Full size image

The climate in the Khorezm Region is extremely continental (Chub 2000). Potential evapotranspiration exceeds the annual precipitation of about 100 mm between 14 and 16 times (Glazirin et al. 1999). The mean annual air temperature is approximately 13°C and annually 205 frost-free days are recorded (Glazirin et al. 1999; Chub 2000). The main crops are cotton, winter wheat, rice and alfalfa, but livestock rearing is rapidly gaining importance (Mueller 2006).

The water of the Amu Darya River is distributed through six major inlet water constructions. It is supplied in an extensive, hierarchically constructed network of main, inter-farm and on-farm irrigation channels with a total length of 16,233 km and a drainage network of 7,679 km (Fig. 1, main map). Although pumps are used, water flow is overwhelmingly caused by gravity forces. Khorezm receives annually between 3.5 and 5 km3 water, of which about 95% is designated for agricultural purposes (Conrad 2006). Water is often led to the fields in open, nonlined channels (Ibrakhimov 2005), and losses from evaporation and groundwater discharge are estimated to be 40% or more (Martius et al. 2004).

Materials and methods

The study was based on MODIS remote sensing data comprising spectral reflectance (ρ), land surface temperature (LST) and emissivity (ɛ), surface albedo (α), Normalized Difference Vegetation Index (NDVI), and Leaf Area Index (LAI). Half-hourly measurements of prime meteorological data recorded at a meteorological station within the study area were used to estimate reference evapotranspiration for modelling. After data selection and preprocessing, the methods applied to derive agricultural land use classes and seasonal ETact are specified. Subsequently, a plausibility analysis for the validation of the modeling results is presented, followed by a description of the spatio-temporal analysis using Geographical Information Systems (GIS).

NASA MODIS products

The large MODIS product suite comprises land surface parameters with daily, 8-day, and 16-day temporal resolutions reprojected and gridded in 250, 500 m, and 1 km spatial resolutions (Justice et al. 1998). Overpass LST data necessary for the application of SEBAL (Bastiaanssen et al. 1998) are provided as swath products and needed to be separately preprocessed, which is described below.

Table 1 lists different Terra-MODIS products which were downloaded and preprocessed for the land cover/land use classification and for SEBAL. Detailed information about MOD09 surface reflectance products is provided in Vermote et al. (1997). Wan and Li (1997) describe the retrieval of MOD11 land surface temperature (LST) and emissivity from MODIS data. The algorithms of the vegetation index (MOD13) and the Leaf Area Index (MOD15) products are presented by Huete et al. (2002) and Myneni et al. (2002). The computation of broadband white sky albedo (stored in the MOD43B3 product) by integrating bi-hemispherical reflectance data modeled over MODIS channels 1–7 (0.3–5.0 μm) is explained in Schaaf et al. (2002).

Table 1 MODIS data products used to classify land cover/land use and to model seasonal actual evapotranspiration between the 1st of April and the 31st of October 2004 (DOY 92 to DOY 305)
Full size table

All NASA MODIS land products include so called Quality Assessment Science Data Sets (QA-SDS). Depending on the product, the QA-SDS considers the atmospheric conditions in terms of cloud cover and aerosol content, algorithm choices, processing failures, and error estimations. The software application ‘Time Series Generator’ (TiSeG, Colditz et al. 2006) was used to exclude invalid data by analyzing the MODIS QA-SDS. With TiSeG, improved time series were built from several gridded MODIS land data products by assessing data quantity against data quality (Conrad et al. 2005). The settings applied are also noted in Table 1. The albedo and the vegetation products NDVI and LAI were interpolated using temporal linear interpolation.

MOD11_L2 LST and emissivity data was preprocessed in three steps. First, the MODIS reprojection tool (MRTSwath) was used to transform the swath data into 1 km grids and for subsetting. Second, the QA-SDS were analysed and all scenes with valid data covering more than 75% of Khorezm were selected (see Fig. 2). This provided 110 overpasses during the 214 day vegetation period from April 1st to October 31st, 2004 for analysis. The third preprocessing step required sensor related considerations. MODIS is a whisk broom scanner recording one complete scan line after another. Due to the Earth’s curvature, the MODIS detector’s instantaneous field of view (IFOV) at the edge of the scan lines is 2.0 (4.8) times larger in the track (scan) direction than at nadir (Wolfe et al. 1998). The overlap of the scan lines increases with the scan angle and the same geographic location may be recorded twice. The MRTSwath (L2G approach, see Wolfe et al. 1998) eliminates this “bow-tie effect.” However, fitting pixels to a 1 km raster requires the integration of adjacent observations with an increasing error in off-nadir scans, because the IFOV also increases with the scan angle. To minimize this impact, pixels with a view angle higher than 30° off-nadir were removed. At 30° off-nadir view angles, the IFOV is approximately 1.1 km in the along-track and 1.25 km in the across-track direction (Wolfe et al. 1998). In addition, the 75% criteria was applied for each overpass. Finally, 62 MODIS LST observations were considered suitable for analysis of the 214 day vegetation period in 2004. Figure 2 shows the percentage of valid pixels for each day. The red line indicates the limit of 75%, below which the complete overpass was taken out of the further modeling process.

Fig. 2
figure 2

Daily percentage of valid pixel values for MOD11A1 data 2004 recorded over Khorezm

Full size image

Meteorological data

The application of SEBAL according to the METRIC variant requires several types of meteorological data. Air temperature, relative humidity, wind speed in 2 m height, and net radiation were recorded half-hourly at a meteorological station located in the center of the irrigation system (60°37′12″E, 41°39′10″N; Khamzina 2006). Reference evapotranspiration was obtained by applying the ASCE standardized reference evapotranspiration equation for alfalfa (ETrs; ASCE-EWRI 2005), as suggested by Allen et al (2005). Therefore, coefficients given by ASCE-EWRI were used to estimate ETrs on hourly data, whereas ground heat flux (G) was calculated according to FAO-56 (Allen et al. 1998).

Land cover/land use classification

For classification, temporal signatures of NDVI time series were derived from 46 8-day 250 m surface reflectance data (MOD09Q1, Fig. 3). Mono-modal distributions with one growing peak in the summer characterize cotton and rice (Fig. 3a, b), whereas the winter wheat (Fig. 3c) had its maximum in late spring. The crop rotations of wheat with rice, sunflower, alfalfa, and maize (Fig. 3d) typically showed a bi-modal NDVI development. Non-modal time series represented the land cover classes desert and permanent water surfaces (Fig. 3e, f). Water was spatially stratified into water from lakes and water from river and large channels, because the varying water levels of the rivers generated sudden shifts of the curve above the soil line and resulted in a signature with higher variances than shown by water from lakes. In total, the eleven classes included six nonagricultural surface types (‘desert’, ‘urban settlement’, ‘rural settlement’, ‘trees and gardens’, ‘permanent water surface’, and ‘rivers and channels’) and five agricultural land use types (‘cotton’, ‘rice’, ‘wheat-fallow’, ‘wheat-rice’, and ‘wheat-other’).

Fig. 3
figure 3

Temporal NDVI signatures from 8-day 250 m MOD09 surface reflectance data, mean values and standard deviation for cotton (a), rice (b), wheat (c), the crop rotation wheat–rice (d), and the land cover classes desert (e) and water (f)

Full size image

Up to 60 training and validation samples for each of the eleven classes were extracted from four ASTER scenes acquired on the 17th and 26th of August 2004. They were cross-checked with 120 field samples collected during the summer season of 2004 and with seven observation sites on large homogeneous cotton and rice areas.

For classification, recursive partitioning and regression trees (Breiman et al. 1984) were applied to metrics of MOD09 NDVI time series by employing the statistical open source software R (R Development Core Team 2005). The statistical gini index was used as impurity measure (split criterion); the minimum number within one observation in one node was set to 5 (stop criterion). The method used in this study resembles the approach presented by Hansen et al. (2000).

Applying the classification tree to the training samples led to a misclassification error rate of 3%. Validation samples were used to calculate the overall accuracy and the Kappa coeffi-cient according to Congalton (1991). During post-classification, errors within the settlement pixels were eliminated by using settlement polygons provided by the ZEF/UNESCO GIS Center in Urgench.

Modelling seasonal actual evapotranspiration

Basically, the SEBAL model solves the surface energy balance for each pixel, which results in the latent heat for evapotranspriation (λET). λET is derived by subtracting ground heat fluxes (G) and sensible heat fluxes (H) from the available energy; the net radiation (R n). MODIS LST, α, ɛ, and NDVI data were used to calculate R n and G according to the formulation of SEBAL (Bastiaanssen et al. 1998; Bastiaanssen 2000). The multitemporal approach and the application of SEBAL to large scale MODIS pixels required modifications for the modelling of H (Eq. 1), which is explained in the following.

$$ H = \rho _{{{\text{air}}}} c_{{\text{p}}} \frac{{{\text{dT}}}} {{r_{{{\text{ah}}}} }} $$
(1)

In SEBAL and METRIC (Allen et al. 2005; Tasumi et al. 2005), H usually results from dividing the gradient of vertical temperatures (dT) by the aerodynamic resistance of heat transport (r ah). ρ air is the air density and c p is the specific heat of air. Atmospheric conditions determine the aerodynamic resistance of heat transport (r ah). Unstable conditions initiate an upward force of air mass; no movement can be registered at neutral conditions, while stable conditions trigger a downward force of air mass. An iterative process integrating the Monin–Obhukov length to consider the stability of the atmosphere finally results in H (Bastiaanssen et al. 2005).

One basic assumption of SEBAL and METRIC is the linear relationship between dT and land surface temperature (Bastiaanssen et al. 1998; Allen et al. 2005). The linear equation is computed using two anchor points, where H can be calculated by using the surface energy balance. For a completely dry (hot) pixel the latent heat of evaporation is zero. Thus H can be calculated as the difference between R n and G, immediately resulting in dThot (Bastiaanssen et al. 1998). Both SEBAL and METRIC characterize hot pixels as nonvegetated dry areas and at best bare soil located close to the meteorological station (Allen et al. 2002). Completely cold pixels such as water surfaces are presumed to consume all available energy for evaporation processes, from which dTcold becomes zero (Bastiaanssen et al. 1998).

In METRIC, well-vegetated fields are selected as cold pixels. H at the cold anchor point (H cold) is expressed as R n − G − k × λETrs where ETrs is the standardized reference evapotranspiration for alfalfa (ASCE-EWRI 2005). The k-factor considers well-watered conditions at the cold pixel which outrange other densely vegetated fields due to higher surface wetness. Allen et al. (2005) assumed k to be 1.05 and included NDVI corrections of k for initial and late growing seasons, when well-watered vegetation pixels are absent.

The selection of homogeneous cold and hot pixels over the whole season by using low spatial resolution remote sensing data sets leads however to uncertainties in heterogeneous landscapes. One square kilometer pixels identified as well-watered vegetation are at least partly covered by roads and houses or smaller irrigation and drainage canals. Spatial variations within the agricultural plots are also possible. Similarly, pure nonvegetated pixels on the 1 km2 scale are only located in the desert, often far away from the meteorological station, whereas pixels representing nonvegetated land in the irrigated area again often include spots of vegetation, gardens, parts of vegetated fields and at least small or dry irrigation and drainage channels. In this study, a second k-factor for the hot pixel and a slightly finer specification of the k-factor for the cold pixel (Allen et al. 2005) are suggested.

For the hot pixel (Eq. 2) vegetation cover is considered in the case that the NDVI excels the average values of the sandy surface of the surrounding desert (0.15). The factor k hot is defined as a function of NDVI, with k hot = NDVI − 0.15 if the NDVI is greater than 0.15, and k hot = 0, otherwise.

$$ H_{{{\text{hot}}}} = R_{{\text{n}}} - G - k_{{{\text{hot}}}} {\left( {{\text{NDVI}}} \right)} * \lambda {\text{ET}}_{{{\text{ref}}}} $$
(2)

METRIC uses a soil water balance to account for possible effects of antecedent precipitation at the hot anchor point, which is not reflected by the NDVI. However, in the vegetation period 2004 total precipitation only reached 48 mm. Consequently, influences of soil moisture were assumed to be negligible. In addition, the introduced automated selection process of the anchor points screened all parts of the investigation area, where the probability to find effectively dry pixels is always high due to the patchiness of precipitation in arid regions. Corrections for soil moisture which are part of METRIC were therefore omitted in this approach due to local weather conditions in Khorezm.

For the computation of H cold at the climax of the vegetation period (NDVI ≥ 0.65) k cold was implemented as 1.05 for the same reasons given by Allen et al. (2005): the probability to find pixels within the 240,000 ha sized irrigated network surpassing other pixels with comparably dense and well watered vegetation due to surface wetness was presumed to be high. Adjustments were applied to k cold using the NDVI to account for low vegetation cover during initial and vegetative crop stages or after harvest (k cold = 1.05 − (0.65 − NDVI)/2). In addition rice, water, and other nonagricultural land cover were expected to be absent at the cold anchor point on the 250 m subpixel scale. Thus, H cold is a written as function of NDVI and the actual land cover (Eq. 3):

$$ H_{{{\text{cold}}}} = R_{{\text{n}}} - G - k_{{{\text{cold}}}} {\left( {{\text{NDVI,}}\,{\text{landcover}}} \right)} * \lambda {\text{ET}}_{{{\text{ref}}}} $$
(3)

To apply the model on daily data, automated selection of the hot and the cold pixels was necessary to reduce processing time. The presented automated selection process employed remotely sensed information on vegetation density and water conditions when assessing the pixels. Vegetation density can be biophysically related to the leaf area index (LAI, Huete et al. 2002). The land surface temperature (LST) describes the skin temperature of soil and vegetation cover, on which ET has a cooling effect (Verstraeten et al. 2005). For the automated calculation of ET over European forests using NOAA data, Verstraeten et al. (2005) suggested a combination of NDVI and LST as an indication of the vegetation cover and water conditions of a single pixel. LST and LAI were preferred for selecting the hot and cold pixels by Allen et al. (2002). In the presented approach, low LAI values and high land surface temperature characterize hot pixels; high LAI values describe dense vegetation, which in combination with relatively low surface temperatures indicates high ET values (Table 2).

Table 2 Suggested rules to automate the selection of cold and hot pixels
Full size table

Additionally, the land cover and land use classification is included. Both pixels should originate from the agricultural area in order to discard lakes, desert or settlement pixels from the analysis. Rice fields were unsuitable for the automation process, firstly because they are initially completely covered with water while METRIC requires vegetated areas, and secondly because the literature on the later growing stages gives inconsistent values for the actual water consumption of rice fields (see Mohan and Arumugam 1994).

Table 3 depicts the bi-monthly result of the automated selection process for each of the two model computations. Here, NDVI and LST reflect the introduced k-factors at the cold and hot anchor points. Due to the lack of well watered conditions and the absence of full vegetation cover within one entire MODIS pixel, k cold of less than 1 were found in the early vegetation period. The increase of k hot demonstrates the difficulty in finding 1 km2 with completely dry conditions in Khorezm during the cotton season, which made adjustments necessary.

Table 3 Environmental situation at the anchor points automatically selected from remote sensing data and resulting k-factors for each two computations of SEBAL per month
Full size table

Finally, λET transformed to daily ETact by using ETref as suggested by Tasumi et al. (2005) resulted in amounts of seasonal ETact. The relation ETact/ETref was interpolated linearly to close gaps caused by missing data. Accordingly, the data gaps in single overpasses resulting from omitting invalid MOD11_L2 data could be filled, except for minor errors owing to the scan angles, as mentioned before.

Validation of ETact

Due to absent lysimeters (Tasumi et al. 2005), scintillometers, Bowen ratio towers, or eddy covariance systems (Bastiaanssen et al. 2005) often used for validation, a stepwise plausibility analysis was carried out to validate the SEBAL modeling results. To account for possible errors caused when using MODIS data, in a first step, albedo and emissivity were compared with the biophysical values of cotton, rice and wheat taken from secondary sources. Secondly, class-A pan evaporation (E pan) was measured in close vicinity of the meteorological station in central Khorezm, which allowed comparison with ETact over water surfaces. The land use classification was used to obtain pixels from water surfaces within a 15 km circle around the location of the class-A pan (9 pixels). In a third step, ETact over vegetated areas was opposed to crop-specific reference evapotranspiration (ETcrop) calculated as follows:

$$ {\text{ET}}_{{{\text{crop}}}} = K_{{\text{c}}} * {\text{ET}}_{{{\text{OS}}}} $$

Crop coefficients (K c) and the cropping calendar of wheat and cotton for the Khorezm region were obtained from the Central Asian Scientific Research Institute of Irrigation (SANIIRI; Table 4). ASCE standardized reference evapotranspiration for small crops (ETos) similar to grass with an approximate height of 0.12 m was computed from daily meteorological data (ASCE-EWRI 2005). This formula for reference evaporation is equal to that presented in FAO-56 (Allen et al. 1998) and therefore valid for the correct derivation of ETcrop according to the FAO manual. The comparisons were made for each 10-day period.

Table 4 Growing stages and crop coefficients of winter wheat and cotton for the Khorezm region
Full size table

Practical application: spatio-temporal GIS analysis

Water use patterns in Khorezm were analysed in two steps. First, agricultural land use was compared with agricultural water consumption (ETact) to investigate seasonal variations in water distribution. Second, the inventory of agricultural land use and ETact for the Water User Associations (WUAs) was related to their absolute distance to the intake points of the irrigation system to assess upstream–downstream trends in water distribution. For the latter, pixels covering more than 75% agricultural land were included.

To analyze ETact on the MODIS 1 km scale, the 250 m pixels assigned to agricultural land use (cotton, rice, and wheat rotations) were counted within each 1 km2 pixel [=agricultural land use density (%)]. According to the cropping calendar (see Table 4) five configurations of agricultural land use density were extracted (Table 5). Until the first half of April, the only crop is winter wheat; from the middle of April, cotton is seeded, and rice growing usually starts from mid-May onwards. Wheat harvest occurs from the beginning of June, whereas rice is usually harvested by the end of September. Thus, cotton is the only crop remaining until the end of October.

Table 5 Relevant configuration changes in detectable land use classes during the 2004 vegetation period
Full size table

The GIS data used to calculate the upstream–downstream situation included the irrigation canal network of Khorezm, the Amu Darya River, and the administrative boundaries of the Water User Associations (WUAs). Intersections with the canal system were applied to identify the intake points of the WUA subsystem boundaries. The heads of those parts of the Khorezm irrigation system close to the Amu Darya River were digitized manually. Then, the reticular distance between each Khorezm system intake location and each WUA distribution point was calculated. The distances were averaged for WUAs having more than one subsystem intake point.

Results and discussion

Land cover and land use classification

For the validation of the land cover and land use classification, at best 30 homogeneously covered MODIS pixels were extracted visually for each class from high resolution ASTER data as validation samples. The overall accuracy was 89% and the Kappa coefficient, 0.88. Incorporating inhomogeneous pixels (dominated by more than one class) into the assessment decreased these accuracy indicators.

The classification results (Fig. 4) showed a clear view of distribution and patterns of actual land use in Khorezm. About 49% of the study area was classified as irrigated land (IL). Agricultural land expansion is mainly delineated by the desert and settlements. As expected, the dominating agricultural crops were cotton (124,000 ha, ∼54% of IL) and rice (73,500 ha, ∼32% of IL), particularly found in the vicinity of the Amu Darya River. Rice fields covered large parts of the northern rayon Gurlen (see GUR, Fig. 1), which historically is known for rice cultivation. The homogeneity of the agricultural landscape decreased with distance to the river. Concurrently, an increase in crop rotation with winter wheat was identified (54,899 ha, ∼24% of IL). Trees and gardens were scattered throughout the entire oblast. Even though settlements could be extracted from the GIS layer, the classes ‘rural’ and ‘urban settlements’ indicate a relation between green and impervious portions within a pixel.

Fig. 4
figure 4

Spatially distributed land cover and agricultural land use in Khorezm 2004

Full size image

Detailed GIS analysis of the classification results allows a comprehensive discussion, which is presented below, including spatio-temporal patterns of water consumption (ETact). The following estimations of ETact are based on these results of the land use classification.

Mapping seasonal actual evapotranspiration (ETact)

Comparisons with biophysical values for wheat, cotton, and rice taken from secondary sources indicated the validity of the input parameters albedo and emissivity (Table 6). Averaged MODIS emissivities overestimated the values found in literature, while in general 0.97 is suggested for modeling emissivity of nonmetal surfaces (Campbell and Norman 1998). However, simulations showed that overestimations in emissivity of 0.01 increased the available energy budget at the earth surface by less than 1%.

Table 6 Emissivity, albedo and surface roughness (z 0 m) of wheat, cotton, and rice used as input parameters for modeling ETact compared to literature values
Full size table

The r 2 value of 0.6 (n = 52 applications of SEBAL) indicated a moderate but significant correlation between ETact over water surfaces averaged for each overpass and daily measured E pan. The resulting trend (E pan = 1.04 ETact) showed smaller ETact than E pan because in contrast to the class-A pan, radiation deeply penetrates a natural water body, which makes less energy available for evaporation. Even though physical conditions of evaporation differ widely between water bodies and agricultural crops and SEBAL was designed for land rather than for water surfaces, the correlation between ETact and E pan indicates that the model output is plausible.

ETcrop was calculated for winter wheat and cotton to compare ETact over the irrigated land of Khorezm. ETcrop and ETact summarized for each 10-day periods and standard deviations of ETact are depicted in Fig. 5.

Fig. 5
figure 5

Comparison between total amounts of ETcrop and ETact, for wheat (a) and for cotton (b)

Full size image

In early April, 10-day ETact overestimated ETcrop for winter wheat between 3% and 11%. Underestimations between 18% and 32% were found during the ripening stage (Fig. 5a). The 105% deviation of ETact from ETcrop in the wheat harvest stage can be explained by initial leaching and initial growth stages of a second crop (on the sub-pixel level). When excluding the last 10-day period, the coefficient of variation (CV) of ETact ranged between 15 and 21%.

In the initial cotton phase, ETact exceeded ETcrop by more than 50% (Fig. 5b), most likely due to high soil evaporation after the leaching period, when high groundwater tables were recorded (Ibrakhimov 2005). In the beginning of July, ETact tends to be lower than expected. Forkutsa (2006) explained similar results from field level by inadequate irrigation within the vegetative growing stage of cotton. During the intensive irrigation phases (July until September), ETact virtually equaled ETcrop. Late-season leaching for field preparation explains the higher overestimations (68 and 78%) during harvest time, usually starting at the end of September. CV for 10-day ETact varied between 30 and 44% in April and between 12 and 26% throughout the later growing phases.

Figure 6 shows the spatial distribution of total ETact in the period from 1st of April until the 31st of October, including late leaching and the entire vegetation period. ETact was lowest in desert margins and urban settlements (0–300 mm) and increased in agricultural land to over 1,200 mm. As expected, there were high values for the lakes in southern Khorezm, also because of noise in the input data which occurred at the southern border to Turkmenistan. In central Khorezm (Khonka rayon, KHO, Fig. 1) and in the most northern rayon, Gurlen, (GUR, Fig. 1), minimum seasonal ETact was 700 mm. Well watered locations were also detected south of central Khorezm, where a large branch of the irrigation system is feeding a large fish farm with water (Fig. 6).

Fig. 6
figure 6

Seasonal actual evapotranspiration (mm) integrated for parts of the leaching season and covering the whole summer vegetation period in Khorezm 2004

Full size image

Crop specific statistics were derived by intersecting the aggregated land cover and land use map with seasonal ETact. Mean values and standard deviations of ETact were 768 and 248 mm on cotton fields, and 798 and 300 mm on rice fields, respectively. On wheat-rice crop rotations, where rice seedlings are commonly transplanted from the nursery to the rice paddies in mid-June, slightly lower average values (739 mm) were modeled.

The results for cotton compare well with the ETact of 706 mm measured in 2004 in an irrigation experiment with furrow-irrigated cotton fields near Tashkent (Ibragimov et al. 2007). Howell et al. (2004) reported an ETact of approximately 750 mm from cotton fields in the arid Texas High Plains.

On rice fields seasonal ETact varies with climatic conditions and vegetation period. In sub-humid India ETact of 596 mm (815 mm) were measured for the 90-day (125-day) period of dry (wet) season rice growth, whereas ETos achieved 470 mm (623 mm; Mohan and Arumugam 1994). Hafeez (2003) reported seasonal ETact of 569 mm (452 mm) for 118 days (130 days) during the dry (wet) season in the Upper Pampanga River Integrated Irrigation System of the Philippines in 2001. In rice growing areas of New South Wales in Australia ETact from 685 to 1,350 mm were measured during the 150 day rice growing season over 32 years (1962–1994), whereas ETos ranged between 960 and 1,412 mm (Humphreys et al. 1994). For rice field in semiarid environments of West Africa, Hendrickx et al. (1986) measured seasonal ETact of 706 mm (443 mm) for a 123 day (92 day) period during two different cropping seasons, whereas ETos was 665 mm (525 mm). The vegetation period for rice in arid Khorezm lasts 135 days.

The assumption of constant weather conditions for the entire irrigation system possibly caused errors especially for the southern parts near the desert. Due to missing reliable meteorological data, the quantification of deviations is yet to be accomplished. In the neighborhood to the desert, ETrs is expected to be higher than in the center of the irrigation system and underestimations of ETact can be a result. On the other hand, this transition zone was assumed to be small because the numerous lakes indicate increased humidity and decreased air temperature also affecting the outer boundaries of the irrigation system. The installation of a small network of meteorological stations could allow these spatial variations to be taken into account.

SEBAL is a worldwide validated model. By selecting MODIS data with scan angles close to nadir the total number of suitable overpasses decreased but potential errors resulting from high IFOV was minimized. In addition, the plausibility analysis confirmed the reliability of the study findings and the comparisons with field experiments also demonstrated the reliability of the presented SEBAL variant. However, for applications in irrigation management in Central Asia, significant optimizations are lacking.

Another relevant aspect for the implementation of this SEBAL variant is the dependence on the actual land use and land cover, which is essential for the selection of the anchor points in the presented approach. Here, reasonable considerations are required to overcome the situation of having a post seasonal land use classification derived from annual time series. Pre-season or early-season land use information would allow intra-seasonal derivations of ETact useful for operational irrigation management.

Spatio-temporal patterns of land and water usage

Multitemporal analysis of the land cover and land use classification result disclosed seasonal variations of water distribution in Khorezm. Figure 7 contrasts the temporal development of agricultural land use density with seasonal ETact. In spring, the homogeneous distribution of winter wheat caused an intermediate land use density throughout the study area, which increased with the seeding of cotton in the last decade of April. The highest density of non-fallow agricultural land occurred between mid-May and mid-June, and decreased slightly after wheat harvest. However, most wheat fields were used to produce a second crop. The map for October showed homogenous cotton distribution in Khorezm.

Fig. 7
figure 7

Actual land use density versus (left column) seasonal ETact (right column) summarized for the configurations according to the cropping calendar (Table 5). The a–e opposite these variables for the period from 1.4.2004–15.4.2004, 16.4.2004–15.5.2004, 16.5.2004–15.6.2004, 16.6.2004–30.9.2004, and 1.1.2004–31.10.2004, respectively

Full size image

The ETact map for the first half of April mainly highlights patterns of leaching. Water seemed to be distributed homogeneously over all of Khorezm in the initial cotton growing period, which substantiates the assertion that even in the outer rayons ETact achieve average values. With increasing dominance of rice fields, ETact maps indicated higher water allocations, in particular in the central parts of the irrigation system and in the rice dominated northern irrigation system. Also in the late season, when initial leaching was carried out, the water distribution pattern appeared constant.

Figure 8 summarizes agricultural land use density (a) and ETact on agricultural areas (b) for the entire cropping season resulting from the GIS analysis of the irrigation canal network (Fig. 9). The density of agricultural land within a WUA ranges between 35 and 65% and averages 49%, independent of the location of the WUA within the irrigated area. The low agricultural land use density close to the heads of the irrigation systems was mainly caused by a desert branch reaching deep into the irrigated land. A lower concentration of settlements close to the Amu Darya River explains the peak in the distance class “25–35 km”. Most WUAs within the “45–55 km” distance occur in the surroundings of Urgench (URG, Fig. 1), which are dominated by settlements. Seasonal ETact used for agriculture indicate a negative upstream–downstream trend of water availability. From class ‘<15 km’ to class ‘35–45 km’ ETact achieved average values between 891 and 975 mm, whereas in the outer reach (class ‘>95 km’) ETact averaged 712 mm.

Fig. 8
figure 8

Relating agricultural land use density and ETact to the reticular distance between WUAs and intake heads within the irrigation system

Full size image
Fig. 9
figure 9

Average distances (km) between the WUAs and the system intake points of Khorezm

Full size image

The classification substantiated the well-known (Mueller 2006), very intensive agricultural use even in regions bordering the desert. The land and water distribution in Khorezm is characterized by high land use density, also in the tail end regions of the irrigation system where low ETact was recorded during the vegetation period. In contrast, high water availability is reported for the downstream parts of the canal network of Khorezm (Mueller 2006; Conrad 2006).

The increase of heterogeneity in crop production in central and downstream Khorezm is evident; the reasons are speculative and may be related to informally driven water allocation and supply strategies, as recently postulated (Wegerich 2004). Intensive crop rotations with winter wheat might be used to optimize income, to fulfill the state order for wheat, to reduce risk, or as a strategy to cope with seasonal droughts. The heterogeneous environmental settings in Khorezm may also be caused by soil texture, and by the level and salinity of groundwater, any of which may provoke large variations in crop water demands. Time series of remotely sensed crop coefficients derived on the field level where the environmental settings are representative and well known, as recently demonstrated by Tasumi and Allen (2007), would make exhaustive and expensive hydrological measurements during the season unnecessary when assessing irrigation performance.

Conclusions

MODIS land products (Justice et al. 1998) provided insights into land and water distribution in Khorezm. Excellent atmospheric conditions in the central Aral Sea basin allowed the generation of nearly gapless time series of remote sensing parameters, especially throughout the vegetation period from April to October. Time series of the 250 m 8-day MODIS MOD09 surface reflectance product contained relevant features to classify and validate all important agricultural land use classes.

Furthermore, modelling ETact became possible for 62 of 214 days between the 1st of April and the 31st of October 2004. Here, extensive preprocessing and pixel selection of 1 km MODIS MOD11_L2 LST/emissivity swath data was necessary to fulfill the requirements of the SEBAL model where overpass records are requested for. As soon as collection 5 of daily MODIS level 3 LST products is available for 2004 (Wan 2007), this time-consuming preparation of level 2 data can be omitted. Interpolation between two overpasses made collection 4 LST data useless for SEBAL in middle and higher latitudes.

The presented results for the Khorezm region show the potential of remote sensing to overcome the lack of reliable, objective, exhaustive, and consistent information on water management issues of Central Asia at a regional scale. With a denser network of meteorological stations permitting better consideration of climatic variations in the irrigation system and in-situ validation facilities for optimizing the model parameterizations, ETact modeled on daily MODIS data and integrated to a monthly product could provide decision support for system-wide operational water management issues.