Land surface phenologies and seasonalities using cool earthlight in mid-latitude croplands

The advanced microwave scanning radiometer (AMSR-E) was launched onboard the NASA-EOS Aqua satellite in May 2002. Data from AMSR-E were acquired at both daytime (∼1330) and nighttime (∼0130) overpasses from mid-June 2002 until its failure in early October 2011. The Numerical Terradynamic Simulation Group (NTSG) at the University of Montana has produced an enhanced land surface parameter suite from AMSR-E data for 2003 through 2010. The data product includes twice-daily air temperatures (ta; ∼2 m height), fractional coverage of open water over land (fw), vegetation canopy transmittance (tc) at three microwave frequencies, surface soil moisture (mv; ≤2 cm soil depth), and integrated water vapor content of the intervening atmosphere (v) for the total column (Jones and Kimball 2011).

Soil moisture measurement was the primary land surface objective of AMSR-E. Accurate soil moisture retrieval is constrained by vegetation opacity as it reduces the observed microwave sensitivity to soil moisture. The transparency (or transmissivity) of the canopy is inversely related to canopy thickness or vegetation optical depth (VOD) (Owe et al 2001). The VOD parameter is a frequency dependent measure of canopy attenuation of microwave emissions due to vegetation biomass structure and water content (Jones et al 2012). Lower VOD (higher transmissivity) indicates lower attenuation of soil-emitted microwave radiation by overlying vegetation canopy and vice versa. VOD equal to 0 corresponds to a transmissivity of 1 indicating bare soil, and for dense vegetation the transmissivity gets close to 0 (Owe et al 2001, Liu et al 2011). A long-term (1988–2008) global vegetation biomass change study on major world biomes found correspondence between VOD and production of major crops (Liu et al 2013).

We evaluated eight years (2003–2010) of the air temperature (ta) and canopy transmittance data in three microwave frequencies: 6.925 GHz (tc06), 10.65 GHz (tc10), and 18.7 GHz (tc18). Over North America, there were many gaps in the VOD retrievals at 6.925 GHz, likely due to radio frequency interference (RFI; Li et al 2004, Njoku et al 2005). We omitted these data from further analysis and restricted our focus to the two higher frequency VOD retrievals at 10.65 and 18.7 GHz.

To select AMSR-E pixels for analysis, we used the international geosphere biosphere programme (IGBP) global land cover classification scheme in the MODIS land cover product at spatial resolution of 0.05° (MCD12C1). At this coarser resolution each land cover class is available as a percentage cover. In Russia, we selected three AMSR-E pixels near Saratov and one near Volgograd, along the Volga River. For the USA, four pixels were selected in the major spring wheat producing state of North Dakota. In the USA, we also use finer resolution USDA NASS county-level crop maps (www.nass.usda.gov/Charts_and_Maps/Crops_County/index.asp) to identify spring wheat areas. Five AMSR-E pixels from Canada were selected in the Prairie Provinces of Manitoba, Saskatchewan, and Alberta (figure 1).

Zoom In Zoom Out Reset image size

The AMSR-E parameters were analyzed from 1st April to 31st October to avoid the frozen season. We applied a 10-day retrospective moving average filter to each parameter time series to minimize data gaps due to orbit and swath width. Meteorological station temperature and rainfall data from NOAA National Climatic Data Center was used for comparative assessment. Additionally, NDVI derived from MODIS MYD09A1 was obtained from the Oak Ridge National Laboratory Distributed Active Archive Center (http://daac.ornl.gov/cgi-bin/MODIS/GR_col5_1/mod_viz.html) for one AMSR-E pixel in one of the Russian cropland sites for comparison with the microwave data.

The thermal regime of growing season can be characterized in terms of accumulated growing degree-days (AGDD; de Beurs and Henebry 2004, 2010, de Beurs et al 2009). Growing degree-days were calculated from AMSR-E air temperature data (ta) using a base temperature of 273.15 K (0 ° C) as follows:

$$\begin{eqnarray} &&\mathrm{GDD}=\max \left [\frac{{\mathrm{ta}}_{\mathrm{day}}+{\mathrm{ta}}_{\mathrm{night}}}{2}-2 7 3.1 5,\hspace{0.167em} 0\right ] \end{eqnarray} \tag{ 1 }$$

where ta_day and ta_night are the air temperatures retrieved during the ascending pass (day) and descending pass (night). The AGDD were derived from simple summation of the GDD:

$$\begin{eqnarray} &&{\mathrm{AGDD}}_{t}={\mathrm{AGDD}}_{t-1}+{\mathrm{GDD}}_{t} \end{eqnarray} \tag{ 2 }$$

where GDD_t is the daily increment of growing degree-days at day t, and AGDD_t−1 is the growing degree-days accumulated from the beginning of the study period (here 1 April) until day t.

GDD as a function of AGDD was fitted with a quadratic model:

$$\begin{eqnarray} &&{\mathrm{GDD}}_{t}=\alpha +\beta {\mathrm{AGDD}}_{t}+\gamma {\mathrm{AGDD}}_{t}^{2}. \end{eqnarray} \tag{ 3 }$$

These three parameters have straightforward interpretations: (1) the intercept α indicates the background GDD value at the beginning of the observation period; (2) the linear parameter β affects the slope; and (3) the quadratic parameter γ controls the curvature. When the fitted model (equation (3)) is convex in shape, i.e., the sign of the β is positive and the sign of the γ is negative, the curve first rises and then falls as thermal time advances.

LSP studies have shown that vegetation index time series, such as the NDVI of herbaceous vegetation in temperate and boreal ecosystems can be readily modeled by a quadratic function of AGDD (de Beurs and Henebry 2004, 2010, Henebry and de Beurs 2013). Here we test if VOD can be modeled in a similar fashion:

$$\begin{eqnarray} &&{\mathrm{VOD}}_{t}=\alpha +\beta {\mathrm{AGDD}}_{t}+\gamma {\mathrm{AGDD}}_{t}^{2}. \end{eqnarray} \tag{ 4 }$$

The parameter coefficients of the fitted convex quadratic land surface phenology (CxQ LSP) model yield two phenometrics: (1) the peak height (PH), which is the maximum value in the fitted model or the vertex of the parabola (equation (5)); and (2) the thermal time to peak (TTP), which is the amount of accumulated growing degree-days required to reach the peak height (equation (6)):

$$\begin{eqnarray} \mathrm{PH}&=&\alpha -({\beta }^{2}/4 \gamma ) \end{eqnarray} \tag{ 5 }$$

$$\begin{eqnarray} \mathrm{TTP}&=&-\beta /2 \gamma . \end{eqnarray} \tag{ 6 }$$

Note that these phenometrics are derived from the parametric model fitted to the data rather than from the data directly.

The VOD time series were modeled in two phases. Convex quadratic models can parsimoniously link GDD as a function of AGDD (figure 2). Thus, in the first phase the TTP_GDD was used as a starting point to model VOD, but the duration of the growing period differed between study areas. In the North American sites, peak VODs nearly co-occurred with peak GDDs. Accordingly, we considered the core growing season (CGS) as the period from 0.5∗TTP_GDD through 1.5∗TTP_GDD, and we fitted the CxQ LSP model using the VOD and AGDD time series from this period. In the Russian sites crops attain their peak VOD very much earlier than peak GDD. Accordingly, we considered core growing season in these croplands to extend from 1st April to date of TTP_GDD, and we fitted the CxQ LSP model using the VOD and AGDD time series from this period. The PH_VOD and TTP_VOD for all sites were derived using equations (5) and (6), respectively (table 1). Since we want to model VOD based on its own behavior, the TTP_VOD derived from the first phase was used to refine the fit of the VOD models. In the second phase, we considered the CGS as the period from 0.5∗TTP_VOD through 1.25∗TTP_VOD, and we fitted the CxQ LSP model using the VOD and AGDD time series from this period. The PH_VOD and TTP_VOD metrics reported in table 2 were derived from these second phase models.

Zoom In Zoom Out Reset image size

Table 2.  Study site name, coefficient of determination, peak height of VOD, thermal time to peak VOD, and day of year at VOD peak for AMSR-E frequencies 10.65 GHz and 18.7 GHz, respectively, and the difference in growing degree-days between TTP_VOD at the two frequencies.

Country	Site	${r}_{\mathrm{VOD}1 0}^{2}$	PHVOD10	TTPVOD10 (° C)	DOY@PHVOD10	${r}_{\mathrm{VOD}1 8}^{2}$	PHVOD18	TTPVOD18 (° C)	DOY@PHVOD18	ΔTTPVOD18−VOD10 (° C)
USA	Turtle Lake	0.99	0.52	1679	205	0.99	0.53	1721	207	42
USA	Grafton	0.99	0.57	2219	228	0.99	0.62	2153	231	66
USA	Maxbass	0.99	0.56	1816	209	0.99	0.57	1843	209	27
USA	Munich	0.99	0.55	1741	217	0.99	0.56	1793	220	51
Canada	Lethbridge	0.98	0.50	1820	223	0.99	0.53	1808	207	−12
Canada	Regina	0.99	0.51	1853	210	0.99	0.54	1843	210	−10
Canada	Winnipeg	0.98	0.53	1735	215	0.99	0.62	2051	231	316
Canada	Prince Albert	0.97	0.60	1686	221	0.98	0.60	1710	223	24
Canada	Edmonton	0.98	0.52	1619	219	0.99	0.55	1629	220	10
Russia	Volgograd	0.97	0.41	1113	168	0.92	0.44	1004	163	−109
Russia	Saratov 1	0.98	0.44	1096	163	0.99	0.47	1067	161	−29
Russia	Saratov 2	0.97	0.42	1032	168	0.94	0.44	991	166	−40
Russia	Saratov 3	0.98	0.43	1142	174	0.98	0.45	1150	175	8

GDD from AMSR-E and that from weather station data at two study sites compared favorably (figure 2). However, there is some divergence towards the end of the season that likely arises from data gaps due to a lack of AMSR-E retrievals over frozen surfaces.

Behavior of the GDD as a function of AGDD displayed strong seasonality with a convex quadratic shape at each study site (figures 3(a) and 4). Given that mid-latitudes are temperature limited, this quasi-parabolic relationship during the frost-free period is expected, and the coefficients of determination are uniformly high (r² > 0.90; table 1, figure 4). In the USA and Canada croplands, the modeled peak GDD occurred on 27 July (DOY 208) on average; whereas, it occurred four days earlier, on average, in the Volga River basin of Russia (DOY 204) (table 1). While there is an expected inverse relationship between the thermal time at peak GDD (TTP_GDD) and latitude, the pattern is relatively weak suggesting that other factors influence the seasonal progression of thermal time (figure 3(b)). The GDDs were higher in Russia than in the North American sites (figures 3 and 4, table 1). The northernmost study sites in Canada, Edmonton and Prince Albert, experienced much lower GDDs. Site Saratov 1 had higher GDD compared to the other three Russia croplands (figure 3(a), table 1); it is located relatively far from the Volga River, while the other three Russian sites are adjacent to the river. Interannual variation in GDD as a function of AGDD appeared higher during the transitional seasons of spring and fall (figure 4).

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

The behaviors of VOD as a function of AGDD exhibit distinct LSPs. During the growing season, VODs ascend to a unimodal peak value and then decline gradually. Peak VOD in the USA and Canada croplands occurred in early August for both frequencies on average. In contrast at the Russian sites the VOD peaks significantly earlier in mid-June (figure 5, table 2). VOD time series showed steep slopes towards their annual peak and then descended gently. The shapes of these seasonal trajectories may reflect microwaves' sensitivity to the timing of vegetation biomass growth and associated changes in canopy water content and the later season drydown and harvest. The pace of fractional green vegetation cover development is quicker than during its disappearance. TTP_VOD differences between the AMSR-E two frequencies, 10.65 and 18.7 GHz, was notably higher for Winnipeg (316 AGDD ° C) (table 2).

Zoom In Zoom Out Reset image size

The fitted convex quadratic LSP models for every site for both frequencies showed high coefficients of determination (r² > 0.90, figure 6, table 2). The Russian sites displayed lower peak VODs that occurred earlier; whereas, the North American sites displayed higher VOD peaks that occurred at higher AGDD (figures 5 and 6, table 2). At each study site the VOD peak at the higher frequency (18.7 GHz) was higher than the VOD peak at the lower frequency (10.65 GHz), with the exception of the Prince Albert site, where the peaks were equal in magnitude (table 2). This pattern accords with the fact that the shorter wavelengths are more attenuated by the vegetation canopy and, thus, register higher VOD.

Zoom In Zoom Out Reset image size

Peak VOD in Russia occurs much earlier (mid-June) compared to that of the North America (early August) (figure 5, table 2). The North American sites were planted to spring wheat that matures in the first half of August; whereas, the Russian sites were planted to winter grains and some grasses that grow and mature soon after the end of frozen season. Major crops that covered the Volgograd site include winter wheat, while the three Saratov sites were covered by winter wheat and some spring wheat. According to USDA Foreign Agricultural Service (FAS), each of these Russian sites was within the five major winter wheat producing oblasts that produce more than 60% of the Russian winter wheat crop (USDA FAS 2009). The USDA FAS also indicates that these Russian sites had minimal cover in rye (USDA FAS 2005). The North America croplands had longer growing season and much higher VOD peaks while the Russia croplands exhibited shorter growing season and lower peak VODs. While the difference in peak VOD timing results from phenological differences in winter versus spring grains, the magnitude of the VOD peak suggests differences in cultivation practices and grain varieties as well.

Some cropland sites show distinctive VOD trajectories. For example, the Edmonton site exhibited higher VOD much earlier in the season compared to other Canadian sites. Land cover information, gleaned from MODIS MCD12C1 IGBP 0.05° LC Type 1 percentage, indicates that ∼16% of the Edmonton study site falls within suburbs of Edmonton. The earlier season VOD may then result from woods, lawns, parks, and cemeteries in the suburban areas. Other sites having an urbanized component include Winnipeg and Volgograd (∼12% each). A study conducted in eastern North America using MODIS NDVI found the effects of urbanized areas on land surface phenology to decay exponentially from the urban boundary to 10 km into rural land covers (Zhang et al 2004). Vegetation in urbanized areas can experience longer growing season and lower canopy density compared to those in rural areas, i.e., they have earlier green-up and later dormancy (White et al 2002, Zhang et al 2004).

The effects of the 2010 Russian heat wave (Trenberth and Fasullo 2012) were evident in the AMSR-E time series. At the three Saratov sites, the GDDs were well above the expectation based on the 2003–2009 period, especially at Saratov 1 (figure 7(a)). The VODs were below expectation for much of the growing season as croplands were negatively impacted by the heat wave (Grumm 2011). Curiously, the VOD trajectory at Saratov 1 was closer to expectation than the other two sites that experienced more modestly elevated temperatures. At two Canadian sites at comparable latitude, the thermal regime was substantially lower than expectation (figure 7(c)). The VODs were significantly depressed at one location (Regina) but not the other (figure 7(d)), despite comparable thermal regimes (figure 7(c)), suggesting that factors other than the temperature affected the Regina LSP. The third Canadian site at the same latitude, Winnipeg, was not included in the comparison due to the influence of (sub)urban land uses within the AMSR-E pixel.

Zoom In Zoom Out Reset image size

Time series analysis of NDVI from Aqua MODIS (calculated from 8-day reflectance) and VOD from AMSR-E (8-day retrospective moving average) for the Saratov 1 site in Russia showed similar LSPs (figure 8). Both indicators of the vegetated land surface displayed correspondence with temperature data from a nearby meteorological station. The peak in the VOD time series lagged those in the NDVI. This behavior was expected since the VOD is sensitive to canopy water content; whereas, the NDVI is sensitive to photosynthetically active radiation (PAR) absorption (Viña et al 2004). Noise in the NDVI time series likely resulted from atmospheric effects, such as sub-pixel cloud contamination. While the NDVI displayed a higher dynamic range than the VOD, the late season secondary peak in winter grains is captured more consistently by the VOD than the NDVI (figure 8). Note how the dynamics in 2010 differ from the earlier years, particularly the long decline in VOD.

Zoom In Zoom Out Reset image size

An important limitation of the microwave data is its relatively coarse spatial resolution necessitated by the earth's surface low microwave energy emissions, limiting its ability to detect fine spatial scale changes (Liu et al 2011, 2013, Jones et al 2012). At the same time, the higher temporal resolution of the microwave data offers different insights in the land surface dynamics. Particularly useful is the retrieval of air temperature (rather than skin temperature) at much fine spatial resolution than ground-based meteorological station networks over most of the planetary land surface (although these temperature retrievals are limited in temporal resolution). Although the VOD retrievals span a very large area—nominally 625 km²—the sensitivity to the water content in the vegetated land surface offers a different perspective that may be more temporally responsive to root-zone soil moisture changes than the absorption of PAR which the NDVI indicates. Additional research is needed to investigate how best to use these complementary perspectives for monitoring and modeling the dynamics of the vegetated land surface.