Reanalysis data underestimate significant changes in growing season weather in Kazakhstan

We present time series analyses of recently compiled climate station data which allowed us to assess contemporary trends in growing season weather across Kazakhstan as drivers of a significant decline in growing season normalized difference vegetation index (NDVI) recently observed by satellite remote sensing across much of Central Asia. We used a robust nonparametric time series analysis method, the seasonal Kendall trend test to analyze georeferenced time series of accumulated growing season precipitation (APPT) and accumulated growing degree-days (AGDD). Over the period 2000–2006 we found geographically extensive, statistically significant (p<0.05) decreasing trends in APPT and increasing trends in AGDD. The temperature trends were especially apparent during the warm season and coincided with precipitation decreases in northwest Kazakhstan, indicating that pervasive drought conditions and higher temperature excursions were the likely drivers of NDVI declines observed in Kazakhstan over the same period. We also compared the APPT and AGDD trends at individual stations with results from trend analysis of gridded monthly precipitation data from the Global Precipitation Climatology Centre (GPCC) Full Data Reanalysis v4 and gridded daily near surface air temperature from the National Centers for Climate Prediction Reanalysis v2 (NCEP R2). We found substantial deviation between the station and the reanalysis trends, suggesting that GPCC and NCEP data substantially underestimate the geographic extent of recent drought in Kazakhstan. Although gridded climate products offer many advantages in ease of use and complete coverage, our findings for Kazakhstan should serve as a caveat against uncritical use of GPCC and NCEP reanalysis data and demonstrate the importance of compiling and standardizing daily climate data from data-sparse regions like Central Asia.


Introduction
Kazakhstan's ecophysiography is largely ecotonal, ranging from forest steppe and steppe (Chibilyov 2002) into desert (Lioubimtseva 2002). Thus, it is especially vulnerable to desertification and climate change (Lioubimtseva and Henebry 2009). Over the period 1891-1990, Kazakhstan experienced a 1 • C increase in mean annual temperature, approximately twice the global rate of increase (Pilifisova et al 1997). By 2030, global climate models predict that temperatures in Kazakhstan will increase mostly during the spring months, with predicted springtime temperature increases ranging from a low of 2.3 • C (Geophysical Fluid Dynamics Laboratory model) to a high of 11 • C (Canadian Climate Center model; Pilifisova et al 1997). Kazakhstan was also subject to one of the more extreme land cover/land use change (LCLUC) events of the 20th century-Khrushchev's 'Virgin Lands' programwhere more than 13 million hectares of native steppe were converted to cereal cultivation (mostly spring wheat) over a dozen years, thereby transforming Kazakhstan into supplier of nearly one-third of the Soviet Union's wheat crop (McCauley 1976, Kaser 1997. Subsequent cessation of centralized planning following collapse of the Soviet Union in 1991 resulted in substantial de-intensification of the agricultural sector in Kazakhstan, including contraction of land area in wheat cultivation, reduction of livestock herds, and sharp declines in fertilizer and pesticide use (Baydildina et al 2000). This LCLUC event appears to have significantly altered land surface phenology at spatial resolutions sufficient to alter atmospheric boundary-layer processes. For example, Henebry (2004a, 2005) found faster spring green-up and an earlier peak normalized difference vegetation index (NDVI) in wheat-growing areas of Kazakhstan following the Soviet collapse (1995)(1996)(1997)(1998)(1999) relative to a pre-collapse period (1985)(1986)(1987)(1988). A more recent analysis (covering the period [2000][2001][2002][2003][2004][2005][2006][2007] shows that Kazakhstan is one of a number of global 'hotspots' exhibiting highly statistically significant ( p < 0.01) declines in growing season NDVI (de Beurs et al 2009).
In this study, we use recently compiled climatic data from individual weather stations in Kazakhstan (NCDC 2008) to analyze contemporary precipitation and near surface air temperature trends (here over the period [2000][2001][2002][2003][2004][2005][2006] that may be driving the recent NDVI declines observed in Kazakhstan (de Beurs et al 2009). In this data-sparse region, we also compare trend analysis results for precipitation and air temperature measured at stations with trend results for gridded monthly precipitation data from the Global Precipitation Climatology Centre (GPCC) Full Data Reanalysis v4 (Schneider et al 2008) and gridded daily near surface air temperature data from the National Centers for Climate Prediction Reanalysis v2 (NCEP R2; Kanamitsu et al 2002).
The significance of our contribution is twofold. First, we present a robust assessment of trends in growing season weather in Kazakhstan that are relevant to modeling vegetation dynamics: accumulated growing degree-days and accumulated precipitation. This assessment complements an earlier analysis published in Russian (Akhmadieva and Groisman 2008) of an important new dataset. Second, we compare the trends in the station data with well known global precipitation and temperature datasets and find them lacking.

Methods
For the years 2000-2006, a complete record of daily precipitation and near surface air temperature is available from 243 stations out of a total of 351 synoptic stations in the Kazakhstan subset of the Global Daily Climatology Network (NCDC 2008). At each station, we aggregate daily precipitation into a running sum of accumulated precipitation (APPT), restarting at 1 January of each year. Daily average temperature is calculated as an arithmetic average of eight 3 hourly synoptic observations. We generate three separate accumulated growing degree-days (AGDD) time series using different base temperatures: where t is the temporal index, AvTemp is daily average temperature, and BT is a base temperature of 0, 4, or 10 • C. A base temperature of 4 • C is typically used in modeling phenological phases of cool season crops likes wheat (Slafer and Savin 1991); whereas, a base temperature of 10 • C is commonly used to model the phenology of warm season crops like corn (Viña et al 2004). At each daily time step, AGDD is positively incremented if AvTemp t > BT, otherwise it remains unchanged.
For subsequent trend analysis of station data, we construct time series consisting of APPT, AGDD0, AGDD4, and AGDD10 values at the 1st and 15th day of each month from March through September. Each georeferenced time series is 98 values in length (7 years × 14 dates per year).
We also analyze precipitation data from the GPCC Full Data Reanalysis v4 dataset, a gridded monthly precipitation product at 0.5 • resolution (Schneider et al 2008). Individual APPT time series are extracted at each grid cell using APPT values from March through September (restarting each 1st January) from 2000-2007. Note that despite adding an additional year, GPCC-derived APPT time series are considerably shorter than their bi-monthly aggregated station equivalents, 56 values in each time series (8 years × 7 dates per year), given that precipitation is reported as monthly totals in GPCC data.
Lastly, we analyze near surface (2 m) air temperature data from the NCEP R2, a gridded daily near surface air temperature product at approximately 1.9 • spatial resolution (Kanamitsu et al 2002).
We calculate average daily temperature as the arithmetic average of maximum and minimum air temperature. Similar to trend analyses of surface temperature at individual stations, we construct three time series consisting of AGDD0, AGDD4, and AGDD10 values at the 1st and 15th day of each month from March through September (restarting each 1st January) from 2000-2007.
Distinguishing significant change from background variability requires either appropriate baselines for parametric analysis or robust nonparametric analysis. We adopt the latter course given the relatively short duration of time series since Kazakhstan achieved independence. Using a seasonal instead of an annual trend test increases the power of the analysis. Furthermore, changes in measurement protocols make baseline development problematic (see NCDC 2008). Given these considerations, trend analyses are conducted using the seasonal Kendall (SK) trend test, a nonparametric method well suited to identifying monotonic trends in time series containing a strong seasonal component. Our implementation of the SK test is also corrected for serial autocorrelation, particularly important when dealing with climate data and time series of accumulated quantities. For additional details on the SK test and its use in analysis of geospatial data, see Hirsch and Slack (1984), de Beurs and Henebry (2004b, 2005, and de Beurs et al (2009).
The sign of the SK test statistic indicates trend direction. The magnitude of the test statistic indicates the strength of the trend; however, it cannot be interpreted as a slope. To facilitate visual comparisons, we normalize SK test statistics on the real interval [−1, 1]. For each layer of georeferenced SK statistics, namely, trend results for station and GPCC APPT time series, station and NCEP R2 AGDD0, AGDD4, and AGDD10 time series, we divide negative SK statistics by the most negative value in the layer (i.e., the absolute minimum) then multiply by −1 in order to maintain the correct sign of normalized SK test statistics; conversely, positive SK statistics are simply divided by the absolute maximum (most positive) value in the layer. Thus, SK statistics for each layer are normalized relative to their maximum negative trend and maximum positive trend, not with respect to the absolute difference between these values (although maximum negative trends and maximum positive trends typically had very similar absolute values). Visually, normalized SK statistics identify where, geographically, trends are largest relative to maximum values for a given layer of time series analyses. However, between different layers (say, station APPT trends versus GPCC APPT trends), normalized SK test statistics do not allow a geographic comparison of trend magnitude on an absolute scale. Similar analyses of APPT trends in gridded GPCC data from 2000-2007 are consistent with negatively trending station results in far western and northeast Kazakhstan, but statistically significant ( p < 0.05) negative trends are not detected between these areas (figure 2), in contrast with station results (figure 1). In fact, GPCC data indicate three regions with statistically significant ( p < 0.05) positive APPT trends in northern Kazakhstan (figure 2). In the westernmost of these areas, there is simply no evidence of increasing APPT from station data, but rather the opposite, statistically significant ( p < 0.05) negative trends (

Discussion and conclusions
Seasonal Kendall trend analyses of recent climate trends complement a recently published analysis of Kazakhstan's climate from 1990-2006 relative to a 1960-1989 baseline (Akhmadieva and Groisman 2008). Using the same station data (NCDC 2008), they found a nationwide increase in near surface air temperature of 0.8 • C between the two periods with warming concentrated geographically in northern and eastern Kazakhstan, and seasonally more pronounced in winter (+1.2 • C) and spring (+0.9 • C) than in autumn (+0.6 • C) or summer (−0.2 • C) (Akhmadieva and Groisman 2008).
Our results concur in terms of showing recent temperature increases in northwestern Kazakhstan (figures 3-5), but also indicate that warming from 2000-2006 was concentrated toward higher temperature excursions (i.e., above a degreeday threshold of 10 • C) during the growing season. Localized cooling in eastern Kazakhstan from 2000-2006 concentrated at growing degree increments less than 4 • C also contrasts with reported warming in eastern Kazakhstan (Akhmadieva and Groisman 2008). In both cases, an examination of georeferenced AGDD trends given different base temperatures reveals localized information (both spatially and at different temperature extremes) that are obscured by statistical averaging.
Akhmadieva and Groisman (2008) report a weak increase (approximately 4%) in nationwide annual precipitation from 1990-2006 relative to the 30-year baseline that was not statistically significant (as were other seasonal and geographic patterns of change) given large variability of the precipitation field. In contrast, the SK trend analysis identified precipitation trends over much of Kazakhstan that are opposite in sign (i.e., negative), statistically significant, and spatially coherent (figure 1).
Contemporary negative precipitation trends across much of northern Kazakhstan (figures 1 and 2) and positive temperature trends in northwest Kazakhstan (figure 3) suggest that negative NDVI trends observed across northern Kazakhstan in recently released MODIS Collection 5 NBAR imagery (de Beurs et al 2009) cannot be attributed solely to legacy effects of agricultural de-intensification following the Soviet collapse Henebry 2004a, 2005). Significant negative NDVI trends in northern Kazakhstan reflect the effects of a regional drought that was sufficiently severe and extended to be detectable in a relatively short time series (8 growing seasons) of satellite observations (de Beurs et al 2009).
Importantly, recently compiled station data from Kazakhstan (NCDC 2008) conflict with gridded global reanalysis products-GPCC (Schneider et al 2008) and NCEP R2 (Kanamitsu et al 2002)-which suggests that these widely used global products do not accurately portray contemporary trends in precipitation and near surface air temperature across a substantial portion of Kazakhstan. In northern Kazakhstan, GPCC and NCEP R2 products substantially underestimate the geographic extent of recent drought and temperature increases. In southern Kazakhstan, GPCC based analyses of APPT trends generally agree in sign with station based trends (largely positive), but statistically significant ( p < 0.05) results from the two datasets rarely coincide spatially. In far eastern Kazakhstan, negative AGDD0 and AGDD4 trends found in the NCEP R2 data are inconsistent with positive AGDD trends evident at individual weather stations.
The result that SK statistics of GPCC time series are generally negative but not significant across northern and central Kazakhstan (like corresponding APPT time series at individual weather stations), may be due to the shorter duration of observations (56 values for monthly GPCC versus 98 values for semi-monthly station data; see section 2). Thus, the GPCC based time series may simply not be long enough to reveal signals that are statistically significant.
Globally, the total number of weather stations used to generate the GPCC Full Data Reanalysis v4 falls off sharply from more than 30 000 stations in 2000 to fewer than 10 000 stations in 2007 due to time lags in incorporating more recent data (Schneider et al 2008). Within Kazakhstan proper, only 52 weather stations were used to interpolate the 2000 GPCC product (figure 6), while data from 63 stations was incorporated in 2007 (albeit with fewer stations from surrounding countries, Uzbekistan in particular). This limitation emphasizes the importance of compiling and standardizing daily climate data from data-sparse regions like Central Asia.
Station data can be difficult to use over large areas due to missing data and measurement inconsistencies. Thus, both the remote sensing and climate modeling communities have an affinity for gridded climate products. However, the quality of the patterns of change and variability captured in gridded products must be evaluated by the end-user with respect to the particular questions under investigation. Here we show significant discrepancies between trends apparent in the gridded data and those found at individual weather stations. Although gridded products offer the allure of complete coverage, our findings in Kazakhstan should serve as a caveat against uncritical use of GPCC and NCEP reanalysis data.