Assessing satellite-based start-of-season trends in the US High Plains

To adequately assess the effects of global warming it is necessary to address trends and impacts at the local level. This study examines phenological changes in the start-of-season (SOS) derived from satellite observations from 1982–2008 in the US High Plains region. The surface climate-based SOS was also evaluated. The averaged profiles of SOS from 37° to 49°N latitude by satellite- and climate-based methods were in reasonable agreement, especially for areas where croplands were masked out and an additional frost date threshold was adopted. The statistically significant trends of satellite-based SOS show a later spring arrival ranging from 0.1 to 4.9 days decade−1 over nine Level III ecoregions. We found the croplands generally exhibited larger trends (later arrival) than the non-croplands. The area-averaged satellite-based SOS for non-croplands (i.e. mostly grasslands) showed no significant trends. We examined the trends of temperatures, precipitation, and standardized precipitation index (SPI), as well as the strength of correlation between the satellite-based SOS and these climatic drivers. Our results indicate that satellite-based SOS trends are spatially and primarily related to annual maximum normalized difference vegetation index (NDVI, mostly in summertime) and/or annual minimum NDVI (mostly in wintertime) and these trends showed the best correlation with six-month SPI over the period 1982–2008 in the US High Plains region.


Introduction
Numerous studies have documented the effects of recent climate change on the plant phenological indicator by using surface-based phenological data (Linderholm 2006 Vegetation phenology can be highly sensitive to climate variability and change (Schwartz et al 2006, Richardson et al 2013, and hence, may be a good indicator of changes in climatic drivers. For example, long-term data indicate increased normalized difference vegetation index (NDVI) trends ('greening') in cold arctic tundra (Goetz et al 2005, Verbyla 2008) and semi-arid areas across the globe (Fensholt et al 2012). On the other hand, some regions, including Chilean semi-arid zones, reported a decreased NDVI trend Environmental Research Letters Environ. Res. Lett. 9 (2014) 104016 (9pp) doi:10.1088/1748-9326/9/10/104016 Content from this work may be used under the terms of the Creative Commons Attribution 3.0 licence. Any further distribution of this work must maintain attribution to the author(s) and the title of the work, journal citation and DOI. ('browning', Baldi et al 2008). In North America such a 'browning' phenomenon has been associated with increasing drought stress (Beck and Goetz 2011). This research has also demonstrated a mismatch between the surface climate trends and satellite-based NDVI trends (greening or browning) in some regions. This mismatch could result from human activities, difficulties in extracting a clear satellite-based signal to match climate forcing ( In the northern hemisphere, satellite imagery indicates the SOS has become progressively earlier by 5.4 days and the end of season (EOS) has been delayed by 6.6 days from 1982 to 2008 (Jeong et al 2011), which is consistent with an overall warming climate over the northern hemisphere. Although significant on hemispheric scales, few studies have been conducted to investigate the cause of SOS mismatch between surface climate and satellite data on a sub-regional scale for the spatial heterogeneity of spring phenology. One study noted, for example, a delayed SOS in the eastern areas of the US High Plains region (in figure 14 The objective of this study is to assess the SOS trends detected by satellite from 1982-2008 in the US High Plains region. For this purpose, we used the satellite biweekly NDVI data to derive the SOS time series. The surface climate observations were then trimmed to the same time series interval for a trend and correlation analysis for climate-and satellite-based phenological indicators in the US High Plains.

Satellite and climate data
The study area extended from −104°to −94. The climate data were obtained from a total of 112 highquality surface climate stations (figure 1(a)) selected from the US Historical Climatological Network (USHCN) as described in Hubbard and Lin (2006), and Menne and Williams (2009). These data include monthly and daily maximum and minimum temperatures and precipitation. The standardized precipitation index (SPI) was calculated from the monthly USHCN precipitation data according to the procedure developed by Mckee et al (1993). The daily climate data quality was assured using the following criteria: (1) outliers in daily mean and minimum temperatures were identified as those that were more than 3.5 standard deviations away from the climatological mean temperature for the day (Frich for masking out the cropland for determining SOS trends without croplands. The croplands and grasslands cover 42% and 51% of the US High Plains area in this study, respectively (figure 1(b)).

Methods
The algorithm for determining satellite-based SOS dates used in our study is the Midpoint pixel method, which is one of the most reliable of 10 methods for comparing long-term surface phenological data (White et al 2009). It is found that this method achieved 65% acceptable SOS retrievals, correlations greater than 0.6, low offsets or biases, and regression slope near 1 (White et al 2009, de Jong et al 2011). The Midpoint pixel method is a local threshold method in which the SOS dates are determined as the time at which NDVI exceeds a locally tuned threshold for each pixel. The locally tuned threshold is determined from the annual maximum and minimum NDVIs (NDVI max , NDVI min ) in a sub-daily (halfday resolution) time series, obtained from a cubic smoothing spline interpolation (White et al 2009). The NDVI max and NDVI min are calculated using a 7-day moving average window (i.e. 14 half-day resolution data points). When both NDVI max and NDVI min are determined, the middle point of NDVI max and NDVI min earlier than NDVI max date is the SOS date (more details in White et al 2009).
There are a number of approaches for calculating the climate-based SOS (Walther and Linderholm 2006). We selected a '5C5D' method in the calculation (Frich et al 2002). The 5C5D method defines climate-based SOS as the date when mean air temperatures exceed 5°C for more than five consecutive days (Frich et al 2002). We then used an additional requirement for modifying the 5C5D method to incorporate the frost date criterion (5C5D FROST , adapted from Jones et al 2002). Under this criterion, the five-day period indicative of SOS had to occur after the last frost in spring, where the frost date is defined as the date at which the daily minimum temperature falls below 0°C.
The satellite-based SOS dates are calculated pixel by pixel and the climate-based SOS dates are calculated station by station. To compare satellite-based with climate-based SOS trends in a spatial domain, the climate-based SOS dates were interpolated from individual station data into 0.5°× 1.0°g rids in latitude and longitude. For the SOS trend analysis, the serial autocorrelation remains an issue with the use of linear regression models (de Jong et al 2011). The linear regression spuriously inflates the power of the significance test (Wilks 2006), making it challenging to delineate statistically significant changes (de Jong et al 2011). To address this concern and improve the robustness of the analysis, the nonparametric Mann-Kendall method (Wilks 2006) was selected for evaluating the statistical significance of all temporal SOS trends at the 95% confidence level. This method can also accommodate data that are not normally distributed and is not sensitive to outliers. In addition, a Pearson's correlation measure was used for evaluating correlations at the 95% confidence level.

SOS profiles
The average SOS profiles calculated from satellite-based and climate-based methods are shown in figure 2. The variations of climate-based SOS dates were smaller than that of satellitebased SOS for both non-croplands (figure 2(a)) and croplands ( figure 2(b)). The satellite-based SOS dates were clearly closer to the 5C5D FROST profiles across the study area especially for non-croplands in the US High Plains (figure 2). This is closer agreement than the comparison results from various satellite-based SOS methods reported by White et al (2009). This finding suggests that the SOS from the Midpoint pixel and 5C5D FROST methods are in reasonable agreement for the US High Plains on a sub-regional scale, especially when masking out croplands.
The standard deviations of satellite-based SOS for noncroplands (i.e. mostly grasslands) were smaller than that of croplands (figure 2). The standard deviations of satellitebased SOS dates were comparable with climate-based SOS method's standard deviations in grasslands. This result suggests that satellite-based SOS dates for croplands have larger longitudinal variations than grasslands. This may be attributable to different numbers of spatial samples ( figure 1(b)) but it more likely occurs because the native grasslands are relatively homogenous in terms of phenological responses to changes of climatic drivers and soil textural conditions.
Generally, the satellite-based SOS dates were slightly earlier than the SOS dates detected by the 5C5D FROST method in non-croplands (figure 2(a)). However, for croplands only, the satellite-based SOS dates were later than SOS dates by 5C5D FROST except for Kansas ( figure 2(b)). The most likely attributions to this inconsistent behavior are differences in changes of land use and cropping systems. For example, winter wheat and pasture in Kansas begin growth initiation and green-up around 4°C. In contrast, corn or soybean begin growth around 10°C, with much less vegetative material at planting than over wintering crops and pasture, creating an apparent delay in the SOS calculated from satellite in Nebraska. Since agricultural production in the Dakotas is primarily rainfed summer crops, there was close correspondence between the satellite and the 5C5D FROST SOS dates. Therefore, satellite-based SOS dates in the US High Plains croplands may be due in part to confounding factors such as alterations in cropping systems, production management choices, and land use changes at sub-regional scales.

Satellite-based SOS trends versus climate-based SOS trends
To examine the trends of phenological changes from 1982 to 2008, statistically significant trends (at 95% confidence levels) from both satellite-based SOS and the 5C5D FROST method are compared (figure 3(a)) (all displayed trends are statistically significant) and; 3(b) (only statistically significant at four stations, indicated by pink boxes)). Our results showed the spatial pattern of SOS trends were similar to those trend patterns observed in White et al 2009, in which a linear regression model was used. Clearly, the satellite-based SOS has trended toward later in most of the ECO 46 and ECO 47 regions, with the ecoregion averages of 4.9 and 2.5 days decade −1 , respectively. It is clear that these satellite-based SOS trends toward later in the season were mostly located in the croplands (figures 3(a) and 1(b)).
Less significant trends toward later SOS on average were observed in the Great Plains ECO 43 , Nebraska Sandhills EC 44 , and ecoregions in Kansas ( figure 3(a)); all of these ecoregions are mostly covered by grasslands. It should be noted that ecoregion averages may suffer from statistical scaling issue (figure 3(a)) due to a considerable latitudinal gradient. Our results from individual pixel trends support   findings of average SOS profiles shown in figure 2. Most of the croplands showed a significantly delayed SOS trend (later arrival) by satellite in terms of trend magnitudes. However, the grasslands for the most part showed no significant SOS trends ( figure 3(a)).
For the climate-based SOS trends using 5C5D FROST method, the SOS time series showed a statistically significant delay (later arrival) at only four stations across all ecoregions ( figure 3(b)). Some stations showed earlier spring arrivals, especially in the eastern areas, but they were not statistically significant in terms of the non-parametric statistics. It should be noted that our 27-year study in the US High Plains region is shorter in length than previous research by Frich et al (2002) and Schwartz et al (2006) for climate-based SOS. Their research demonstrated physical environmental responses towards overall warming trends with an earlier SOS over a larger geographic scale, such as global or hemispherical.
The US High Plains region is unique in terms of satellitebased SOS, with a delayed SOS ( figure 3(a)). Figure 4 showed four area-averaged time series of satellite-based SOS and climate-based SOS by 5C5D FROST method for noncroplands (after masking croplands) ( figure 4(a)) and croplands ( figure 4(b)) from 1982 to 2008. For croplands, only satellite-based SOS time series exhibited a statistically significant delayed trend (later arrival, p value of 0.008) but climate-based SOS time series was not statistically significant. Part of this later arrival of SOS in croplands ( figure 4(b)) could be related mainly to changes in cropping systems and land management practices (Mahmood andHubbard 2002, Mahmood et al 2006). In addition, this 'delayed spring' occurred in croplands was not statistically correlated with the climate-based SOS (r = 0.34 and p = 0.08) ( figure 4(b)). In contrast to croplands' SOS trends, two area-averaged SOS time series for grasslands showed no statistically significant trends and the correlation (r = 0.37 and p = 0.05) between satellite-based SOS and climate-based SOS became statistically significant ( figure 4(a)).

Satellite-based SOS trend related to climatic drivers
Changes in monthly average temperatures, precipitation and drought index from 1982 to 2008 were displayed for each station to examine these climatic drivers in relation to SOS trends and variations (figures 5 and 6). Few stations showed statistically significant warming or cooling trends from January to May (figures 5(a)-(e)). A correlation between the satellite-based SOS and monthly average temperatures showed no significant relationship. Similarly, there were no significant correlations evident between monthly precipitation and satellite-based SOS (figures 5(f)-(j) for trends, correlations not shown). In addition, six-month averages of temperature and precipitation were not significantly correlated with satellite-based SOS (not shown).
To further explore any possible impacts of climatic drivers on the satellite-based SOS, the multi-scale (1-8 month) SPIs were calculated and assessed in relationship to the satellite-based SOS. Unlike the strong correlations between NDVI and three-month SPI in the High Plains region found by Ji and Peters (2003), our results indicate that the six-month SPI (SPI6) had the highest correlation to the satellite-based  figure 6(a)). Again, these trends were not significant. Negative correlations in figure 6(b) suggest that when SPI6 decreased (became drier), the satellite-based SOS dates were delayed (later spring arrivals).
To further assess the satellite-based SOS trends ( figure 3(a)), trends of both annual NDVI max and NDVI min time series obtained from the Midpoint pixel method were examined (figures 7(a) and (b)). The spatial pattern of satellite-based SOS trends (figure 3(a), earlier arrival or later arrival) was similar to the annual NDVI max trends ( figure 7(b), greening or browning). When the NDVI max had a positive trend (greening at timing of NDVI max in the season, figure 7(b)), for example in the croplands (or in ECO 46 , Nebraska areas of ECO 47 , and ECO 27 ), the satellite-based SOS was trending towards later ( figure 3(a)). On the other hand, when the NDVI min had a positive trend, (greening at timing of NDVI min in the season, figure 7(a)), for example in the grasslands (or in ECO 43 and Sandhills ECO 44 ), the satellite-based SOS was slightly and sparsely trending toward later or trending earlier ( figure 3(a)). This result indicates that the satellite-based SOS trends (either earlier or later) are much better correlated with summertime and wintertime greenness (NDVI values) (figures 7(c) and (d)). The statistically significant correlation coefficients between NDVI max or NDVI min and satellite-based SOS were higher, up to 0.9, indicating that the summertime NDVI or wintertime NDVI values were primarily related to satellite-based SOS dates detected from 1982-2008 in the US High Plains region.

Summary and conclusion
The SOS dates calculated using satellite-based Midpoint pixel and the 5C5D FROST methods are in reasonable agreement for the US High Plains region on a statewide scale, especially for areas after masking out the croplands. The SOS difference (earlier or later) between satellite-based and 5C5D FROST on a state-wide scale depends at least in part on confounding factors such as cropping systems and land management. Winter grains green-up as soon as temperatures approach 5°C. Soybean and corn, however, do not green-up until after planting and emergence (around 10°C). The vegetative canopy of row crops also take more time to develop and cover the ground because of wider row spacing and less dense plant populations compared to grasses. Use of earlier-maturing varieties and changes of planting dates could further impact the satellite-based SOS. Thus, the satellite-based SOS signals are impacted by land management decisions such as what crops are grown and when they are planted.
The short time period of satellite observations relative to the climate drivers may limit their use in developing realistic and accurate correlations to study long-term warming trends. While more than 100 year climatological records clearly demonstrate long-term warming trends in the US High Plains region, the shorter length of time that satellite observations have provided region-wide NDVI data may lead to some inconsistencies due to spatial heterogeneity for spring phenology (Schwartz et al 2006, Beck andGoetz 2011). This is despite the fact that we used the longest available consistent satellite observation in this study for interpreting anthropogenic warming responses to spring phenology at the Level III ecoregion. Unlike the relationship between satellite NDVI and climatic variables found in previous studies (Ji andPeters 2003, Richardson et al 2013), the satellite-based SOS in our study area did not have a strong relationship with the same period of temperatures and precipitation. The six-month SPI presented the best but still weak correlation with satellitebased SOS among multi-scale SPIs from 1-8 months. The six-month SPI had the highest correlation with the satellitebased SOS for spring owing to impacts of water deficit on the vegetation and the associated lag in the spring phenology. Finally, changes of annual NDVI max (mostly in summertime) and NDVI min (mostly in wintertime) were correlated to the statistically significant satellite-based SOS trends in the US High Plains regions over 1982-2008. The increase of annual NDVI max was positively, spatially, and statistically significantly correlated to satellite-based SOS dates towards later