Global vegetation phenology from Moderate Resolution Imaging Spectroradiometer (MODIS): Evaluation of global patterns and comparison with in situ measurements

Authors


Abstract

[1] In the last two decades the availability of global remote sensing data sets has provided a new means of studying global patterns and dynamics in vegetation. The vast majority of previous work in this domain has used data from the Advanced Very High Resolution Radiometer, which until recently was the primary source of global land remote sensing data. In recent years, however, a number of new remote sensing data sources have become available that have significantly improved the capability of remote sensing to monitor global ecosystem dynamics. In this paper, we describe recent results using data from NASA's Moderate Resolution Imaging Spectroradiometer to study global vegetation phenology. Using a novel new method based on fitting piecewise logistic models to time series data from MODIS, key transition dates in the annual cycle(s) of vegetation growth can be estimated in an ecologically realistic fashion. Using this method we have produced global maps of seven phenological metrics at 1-km spatial resolution for all ecosystems exhibiting identifiable annual phenologies. These metrics include the date of year for (1) the onset of greenness increase (greenup), (2) the onset of greenness maximum (maturity), (3) the onset of greenness decrease (senescence), and (4) the onset of greenness minimum (dormancy). The three remaining metrics are the growing season minimum, maximum, and summation of the enhanced vegetation index derived from MODIS. Comparison of vegetation phenology retrieved from MODIS with in situ measurements shows that these metrics provide realistic estimates of the four transition dates identified above. More generally, the spatial distribution of phenological metrics estimated from MODIS data is qualitatively realistic, and exhibits strong correspondence with temperature patterns in mid- and high-latitude climates, with rainfall seasonality in seasonally dry climates, and with cropping patterns in agricultural areas.

1. Introduction

[2] Vegetation dynamics at seasonal and longer timescales reflect large-scale interactions between the terrestrial biosphere and the climate system. As a result, phenology has recently emerged as an important topic with relevance to a wide array of climate and ecological research including regional and global carbon modeling, ecological assessment, and agricultural monitoring, to name only a few [Asner et al., 2000; Parmesan and Yohe, 2003; Lucht et al., 2002]. Site and species-specific phenological measurements such as the timing of budburst and flowering are now widely used to detect and study the effect of climate variability and trends on vegetation [Kramer et al., 2000; Beaubien and Freeland, 2000; Primack et al., 2004], and phenology has been shown to be an effective indicator of urban heat islands [Roetzer et al., 2000; Zhang et al., 2004a], the El-Niño–Southern Oscillation [Keeling et al., 1996; Asner and Townsend, 2000; Vicente-Serrano et al., 2006], forest and crop management [Keatinge et al., 1998; Hartkamp et al., 2002], and the impact of allergies on human health [Van Vliet et al., 2002]. Moreover, accurate prescription (or prognosis) of vegetation phenology is important for calculating key quantities in Earth system models related to surface energy and water fluxes [Arora, 2002] and the global carbon cycle [Running and Nemani, 1991; Wilson and Baldocchi, 2000]. Accurate knowledge and understanding of phenology is therefore important to atmospheric general circulation models, which are now commonly coupled to dynamic global vegetation models [Sellers et al., 1996; Foley et al., 2000].

[3] In recent decades, remote sensing has become a widely used mechanism for monitoring the activity of vegetation at large spatial scales. Vegetation indices, which provide an indication of the canopy “greenness,” are commonly used for this purpose [Curran, 1983; Sellers, 1985], and in particular, time series of normalized difference vegetation index (NDVI) data derived from the Advanced Very High Resolution Radiometer (AVHRR) have been used extensively for monitoring vegetation phenology [Lloyd, 1990; Reed et al., 1994; White et al., 1997]. Indeed, AVHRR data provide the only source of global data that can be used to analyze seasonal- to decadal-scale dynamics in global vegetation over the last two decades [e.g., Zhou et al., 2001].

[4] More recently, the VEGETATION instrument onboard the SPOT 4 spacecraft, and the Moderate Resolution Imaging Spectroradiometer (MODIS) onboard NASA's Terra and Aqua spacecraft have provided a new era of global remote sensing data products. In particular, MODIS provides daily reflectance data at spatial resolutions of 250 m, 500 m, and 1 km globally with substantially improved geometric and radiometric properties relative to AVHRR data. Further, MODIS data have been atmospherically corrected and screened for clouds in a fashion that is not possible using AVHHR data. MODIS data therefore provide an improved basis for monitoring global ecosystem dynamics relative to previously available data. Having said this, global monitoring of phenology from remote sensing remains a significant challenge.

[5] With these issues in mind, the objective of this paper is to describe results from recent efforts to develop robust and repeatable methods to estimate and map global vegetation phenology using data from MODIS, which have been used to produce the MODIS land cover dynamics product (MOD12Q2). In the next section we provide a short background describing previous research efforts related to this work. We then briefly describe the algorithm being used to detect vegetation phenology from MODIS data, followed by results comparing MODIS-derived estimates of phenology against in situ measurements collected in the field. In the penultimate section of this paper we describe patterns in vegetation phenology observed at continental scales, and link these patterns with regional climate and environmental forcing. We conclude with a discussion of key technical and scientific issues that we are currently examining.

2. Background

[6] A variety of methods have been developed over the past decade to detect the timing of vegetation phenology from satellite data. To estimate the start and the end of the growing season the simplest approaches use prescribed thresholds in time series of a remotely sensed vegetation index (VI) [Lloyd, 1990; White et al., 1997; Jönsson and Eklundh, 2002; Suzuki et al., 2003]. These methods can work well at local scales and for specific vegetation types. However, they are difficult to implement at global scales because they do not allow for differences in vegetation characteristics across ecosystems, nor do they allow for variation associated with soil and snow backgrounds. An alternative approach, which is used in this work, is to identify inflection points in time series of VI measurements [Badhwar, 1984; Moulin et al., 1997; Zhang et al., 2003].

[7] In addition to the timing of vegetation phenology, several other phenological metrics have been estimated from remote sensing. These metrics include the minimum, maximum, and amplitude of “greenness” (quantified using a VI), and the growing season integrated greenness. The minimum and maximum greenness are important metrics in land cover classification [DeFries et al., 1995, 1998; Hansen et al., 2000], and the growing season integrated greenness has been widely used for estimating crop yield [e.g., Knudby, 2004; Labus et al., 2002; Rasmussen, 1992], ecosystem net primary production [Posse and Cingolani, 2004; Goward et al., 1985; Tucker et al., 1986; Paruelo et al., 1997], annual evapotranspiration [Sun et al., 2004], and land cover type [Rasmussen, 1992]. Further, interannual variation in integrated greenness has also been used to identify anthropogenic activity including agricultural and urban expansion, maintenance of protected areas, and ecosystem degradation [e.g., Budde et al., 2004].

[8] While a significant amount of research has focused on the use of satellite data to estimate vegetation phenology, relatively few data sets characterizing this important property of ecosystems are available at global scales, and those that are available are typically produced via retrospective analysis of AVHRR or VEGETATION data at very coarse resolutions. Here we describe results from the MOD12Q2 data product, which is part of the MODIS land product suite. This product is designed to be operational (i.e., produced via repeatable methods on a recurring basis), and provides information related to global vegetation phenology at a spatial resolution of 1 km at biannual time steps.

3. Methodology

3.1. Remote Sensing Data Sources

[9] The method we use to characterize phenology is presented by Zhang et al. [2003]. In this approach, vegetation growth cycles are characterized by four transition dates, which correspond to key phenological phases: (1) greenup: the date of onset of greenness increase; (2) maturity: the date at which canopy greenness approaches its seasonal maximum; (3) senescence: the date at which canopy greenness begins to decrease; and (4) dormancy, the date at which canopy greenness reaches a minimum. Note that because we are using remotely sensed surrogates (i.e., we are not measuring canopy variables directly) our terminology specifically refers to the phenology of canopy greenness.

[10] The primary data source used to estimate the aforementioned transition dates is the enhanced vegetation index (EVI) [Huete et al., 2002], which we compute using data from the MODIS BRDF/albedo product [Schaaf et al., 2002]. Specifically, we use nadir BRDF-adjusted reflectance (NBAR) data from MODIS (MOD43B4, version 4). NBAR data provide surface reflectances in which view angle effects have been removed, and both cloud and aerosol contamination is minimized. In addition, we also use the MODIS land surface temperature (LST) product (MOD11A2, version 4), which provides land surface skin temperature at a 1-km spatial resolution and 8-day time intervals [Wan et al., 2002]. Specifically, we produce annual time series of LST for each pixel by replacing missing values (due to clouds) using linear interpolation of the nearest available values (in time).

[11] Because the EVI data include gaps due to clouds and are inherently noisy, the time series at each pixel are first preprocessed. To do this, single dates of missing data are replaced using a moving-window average based on the two nearest neighbors with valid data (i.e., linear interpolation in time). To reduce high-frequency temporal variability we have implemented a set of filters, focusing particularly on the effect of snow. Specifically, the presence of snow (either on the canopy or beneath it) acts to decrease the value of vegetation indices such as the EVI [e.g., Dye and Tucker, 2003; Zhang et al., 2004b; Delbart et al., 2006]. To minimize this effect, two main strategies have been used. First, the MODIS BRDF/albedo product provides a snow and ice flag in its quality assurance field, indicating whether the data are acquired over a snow-covered or snow-free surface. Using this flag, all values in a time series containing snow are replaced with their most recent snow-free value. Second, because the snow flag is not perfect, we screen for pixels that have LST values less than 5°C. Where such cases occur, we replace them using the nearest valid EVI value in the time series. As a final step, a three-point median-value moving-window technique is applied to the data.

[12] For the results presented in this paper, we used global MODIS data acquired between July 2000 and March 2004. In addition to the MODIS data sets described above, we also used data from (1) the MODIS land cover product [Friedl et al., 2002]; (2) the MODIS net primary productivity product [Running et al., 2000; Zhao et al., 2005]; and (3) daily precipitation data acquired from the tropical rainfall measuring mission (TRMM) product (3B42RT).

3.2. Detection of Phenological Metrics From MODIS Data

[13] Vegetation growth and senescence are identified by periods of sustained EVI increase and decrease, respectively. Specifically, transitions from periods of increasing EVI to periods of decreasing EVI are identified by changes from positive to negative slope within time windows consisting of five 16-day EVI values, and vice versa. Because slight decreases or increases in EVI can be caused by local or transient processes unrelated to vegetation-growth cycles, we use two heuristics to exclude such variation: (1) the change in EVI within any identified period of EVI growth or decrease should be larger than 35% of the annual range in EVI for that pixel; (2) the ratio of the local maximum EVI to the annual maximum EVI should be at least 0.7. This approach is able to screen out short-term variation unrelated to growth and senescence cycles in EVI data, while at the same time identifying multiple growth cycles within any 12-month period. This is important in some tropical and subtropical regions where crop cycles and precipitation regimes can result in more than one annual vegetation growth cycle.

[14] To identify specific transition dates, sigmoidal functions are used to model the temporal trajectory of NBAR EVI data. Specifically, individual growth cycles, which include greenup, maturity, senescence and dormancy, are represented using pairs of logistic models [Zhang et al., 2003]:

equation image

where t represents time (day of year - DOY), a and b are parameters associated with the timing and rate of change in EVI, c + d is the maximum EVI value for a given period of vegetation growth, and d represents the initial background value. Using this representation, phenological transitions are determined using the curvature-change rate (CCR) [Zhang et al., 2003]. Specifically, transition dates correspond to the time at which the CCR exhibits local minima or maxima. During greenup, the two maximum CCR values correspond to the onset of greenness increase and the onset of greenness maximum. Similarly, during periods of EVI decrease the two minimum CCR values identify the onset of greenness decrease (senescence) and the onset of greenness minimum (dormancy). NBAR EVI values corresponding to these transition dates are also recorded. In addition, we also compute the growing season integrated NBAR EVI (INEVI), which is computed as the sum of the modeled daily EVI values from the onset of greenness to the onset of dormancy. As a final step, a three-by-three spatial moving window technique is used to reduce high spatial frequency noise in estimated phenological transition dates, accounting for multiple growth cycles.

3.3. In Situ Measurements

[15] The most obvious and important way to assess the quality of phenology retrievals from MODIS is to validate them against field observations. Collection of field data in support of this, however, is a challenging task because the MODIS NBAR data are produced with a spatial resolution of 1 km. Thus each pixel includes a mosaic of land cover and vegetation types. Further, even areas that are homogeneous in regards to land cover can exhibit significant spatial variability in phenology. Thus point measurements may not be representative of the 1-km pixel in which they are located. Recognizing these limitations, here we compare three field data sets against estimates of phenological transition dates estimated from MODIS data.

[16] Observations collected at the Harvard Forest long-term ecological reserve (LTER) site include samples collected at two to five individuals of 33 native woody species located along a 2 km loop transect near the Harvard Forest headquarters (42°32′N, 72°11′W). Field measurements include budburst and leaf development from April to June, and leaf coloration and leaf drop from September to November at 3–7 day intervals. These data were collected by staff at the LTER site between 2001 and 2003 and were obtained from the Harvard Forest LTER Web site (http://www.lternet.edu/hfr/). Phenological data for each species were averaged and treated as representative of this region (Figure 1). The data were collected over a large area encompassing several pixels. To reduce potential geolocation errors, MODIS results are presented using the mean and standard deviation for the five pixels (including the four nearest neighbors) to the locations on the ground where data were collected.

Figure 1.

Field observations of leaf phenology in 2001, 2002, and 2003 at Harvard Forest.

[17] Vegetation phenology measurements at the Hubbard Brook Experimental Forest (43°56′N, 71°43′W) were collected by the USDA Forest Service (A. Baily, unpublished data, http://www.hubbardbrook.org/research/data/veg/phn/phndoc.htm). The most common tree species at this site are American beech (Fagus grandifolia), sugar maple, (Acer saccharinum) and yellow birch (Betula lutea). Vegetation phenology is measured every week at Hubbard Brook at specific sites located in five small watersheds using the criteria in Table 1. Data from 2001 are presented in Figure 2. Since these field sites are located within several MODIS pixels, the MODIS data for these pixels are averaged and used to assess the results.

Figure 2.

Field observations of a phenological index measured at Hubbard Brook in 2001. Different curves represent different locations where observations are made (identified as 1B, 6T, 4T, 5B, 5T, 7B, and 7T).

Table 1. Criteria Used to Define Indices Used for the Field Measurements of Phenological Development in Hubbard Brook Experimental Foresta
IndexSpring CriteriaFall Criteria
0.5 most leaves fallen
1bud swelling noticeableno more green leaves in canopy, half of leaves fallen.
2small leaves or flowers visible, initial stages of leaf expansion (about 1 cm long)most leaves becoming yellow or red, and a few fallen leaves
3leaves with 1/2 of final length, or 5 cm long, or canopy obscuring 1/2 of skymany leaves having noticeable reddening or yellowing
4leaves fully expanded with little sky visible through crownsonly scattered leaves or branches changing color

[18] The final data set that we used as a basis for assessing the MODIS phenology retrievals is provided by Plantwatch (http://www.devonian.ualberta.ca/pwatch/index.htm [Beaubien, 1997]). This data set provides spring flowering time (first bloom date and full bloom date) for 182 sites distributed across Canada. At each site, the flowering time for key indicator plant species is recorded by students and the general public, typically for one plant at most sites. However, up to 6 plants are observed at some sites. For the comparison presented here, flowering times in 2001 were acquired from the Plantwatch Website and compared with MODIS-derived estimates for the timing of greenup. Flowering time is clearly different from the “greenness” phenology that the MOD12Q2 product measures. However, the phenology of leaves and flowers is highly correlated, and flowering dates therefore provide a useful surrogate for leaf phenology at the start of the growing season. Note, however, that there is a significant mismatch in the measurement scales between the data collected on the ground and the 1-km2 MODIS footprint.

4. Results

4.1. Limitations Imposed by Cloud Cover

[19] The accuracy of retrieved phenological metrics depends on a variety of factors including the quality and temporal resolution of MODIS data. In addition, missing data during key parts of the growing season also strongly influence our results (Figure 3). The magnitude of this issue varies geographically and depends on the climatology of cloud cover. For example, in 2001 27%, 11%, 5%, 4%, and 16% of the Earth's surface was sufficiently obscured by clouds to prevent acquisition of NBAR data for 1, 2, 3, 4, and greater than 4 consecutive 16-day periods, respectively. Not surprisingly, geographic regions that are most problematic in this regard include the tropics (Amazonia, West Africa, India, East Asia), and high latitudes because of low (or zero) illumination conditions. Obviously, it does not make sense to retrieve phenologic transition dates under these conditions. Therefore our algorithm is not applied in cases where there are more than two consecutive missing 16-day NBAR EVI values during snow free periods.

Figure 3.

Maximum number of consective16-day periods with missing NBAR data in 2001. Winter periods with snow cover were excluded.

4.2. Comparison With In Situ Measurements

4.2.1. Harvard Forest

[20] Summary results comparing phenological transition dates estimated using the methods described above versus those observed at Harvard Forest are presented in Table 2. The MODIS retrievals differ from in situ data by about 3–4 days for each of the three years, and correspond to the timing when budburst has occurred for 20% to 50% of the canopy. Results also show that the estimated onset of greenness maximum corresponds to the timing at which 84–90% of individual leaves reached their final size. It is also interesting to note that leaf growth was only about 77 percent complete at the estimated date for the onset of greenness maximum in 2001. This result likely reflects the fact that MODIS data were not acquired between DOY 166 (15 June) and 184 (11 July 2001) owing to MODIS instrument problems.

Table 2. Comparison Between Phenological Transition Dates Retrieved From MODIS Data and In Situ Data Collected at Harvard Foresta
 GinGmaGdeGmi
MODISABDBBDMODISFL1FL2MODISCLMODISAGDGLTL
  • a

    Values in the columns are date (DOY). For the MODIS-based estimates we provide the mean and standard deviation for the five nearest neighbor 1-km2 cells. Gin, date of greenness increase; Gma, date of greenness maturity; Gde, date of greenness decrease; Gmi, date of greenness minimum; ABD, average date of first budburst observed in field; BBD, percent budburst corresponding to the timing of MODIS retrieval; FL1, percent of leaves in canopy reaching their final size at the time of mean MODIS Gma; FL2, percent of leaves in canopy reaching their final size at the time of maximum MODIS Gma; CL, average percentage coloring of leaves at the time of mean MODIS Gde; AGD, average date of last green leaf coloring for each individual plant; GL, the percent of green leaves in the canopy at the time of mean MODIS Gmi; TL, the percent of leaves on trees at the time of mean MODIS Gmi.

2001120 ± 5123 ± 420157 ± 77781236 ± 41317 ± 2297 ± 120.253
2002122 ± 4118 ± 1050166 ± 38590234 ± 91315 ± 1304 ± 10110
2003130 ± 2127 ± 950161 ± 68490230 ± 130.2304 ± 3295 ± 9215

[21] Conceptually, the onset of greenness decrease estimated from MODIS corresponds to the onset of leaf senescence. At this stage, decreases in leaf chlorophyll and water are manifested in lower MODIS EVI values. Results shown in Table 2 suggest that MODIS data are able to identify this transition very well, and indeed, may be sensitive to changes in leaf water content that are precursors to changes in leaf color. The onset of greenness minimum is estimated to occur about 10 days later than the average date at which green leaves were still observed in the canopy, but agrees well with the timing for which all leaves have changed their colors (green leaves constitute less than 2% of the canopy).

4.2.2. Hubbard Brook Forest

[22] Data collected at Hubbard Brook show that MODIS-derived phenology in 2001 agrees quite well with field measured indices of phenology (Figure 2 and Table 3). For example, the onset of greenness was estimated to occur on DOY 128, which corresponds to an average index of 1.7 measured in the field. Inspection of Table 1, which presents information related to the ecological meaning of different values for the field index, suggests that this result is reasonable. The field measurements reach their maximum value (4 = full canopy cover) about 10 days earlier on average than the estimated onset of greenness maximum. The first field observation of leaf coloration is on DOY 267 (which is much later than the MODIS onset of greenness decrease), and the end of the growing season is measured in the field about 13 days before the onset of greenness minimum estimated from MODIS.

Table 3. Phenological Transition Date (DOY) Retrieved From MODIS Data for Several Small Watersheds of Hubbard Brook Experimental Forest in 2001
 Locations
1B6T4T5B5T7B7T
Gin127128123136128124129
Gma156156155161156158169
Gde250261263230261251247
Gmi305304304317304305309

4.2.3. Spring Bloom Data

[23] Results comparing MODIS retrievals with the Spring Bloom data are presented in Figure 4 and clearly show that the MODIS phenology retrievals and ground samples are strongly correlated and distributed along a 1:1 line. The onset of greenness increase is significantly correlated with both the date of first bloom (R2 = 0.60), and the date of full bloom (R2 = 0.64). However, the root mean square error (RMSE) for each reveals that the onset of greenness increase is closer to the date of first bloom (RMSE = 10.8 days) than to the date of full bloom (RMSE = 14.0).

Figure 4.

Correlation between flowering time measured in the field and the onset of greenness increase retrieved from MODIS data (data from Plantwatch).

[24] Taken together, the results presented above for the Harvard Forest, Hubbard Brook, and Spring Bloom data are quite encouraging. On the basis of these results we conclude that the four transition dates retrieved from MODIS agree well with in situ observations of vegetation phenology. In particular, estimates for the timing of greenup, maturity, and dormancy all agree very well with in situ measurements. The quality of the retrievals at the beginning of the growing season (onset of greenness increase and maximum) appears to be higher than those at the end of the growing season (onset of greenness decrease and minimum), which reflects the fact that the transition during greenup is more pronounced, and therefore easier to measure, both in the field and using remote sensing. Given the limitations imposed by the data (temporal sampling, restricted numbers of samples and geographic extent, not all species sampled, etc.) the results presented above suggest that the algorithm is performing well.

4.3. Observed Patterns in Global Vegetation Phenology

4.3.1. Overview of Global Results

[25] Figure 5 presents several representative EVI time series for different land cover types, including a location characterized by double-cropping in China. The estimated phenological transition dates for each are qualitatively realistic, and comparison of the fitted sigmoidal curves with the NBAR EVI data reveals RMSE's smaller than 0.04 and that the models explain more than 95 percent of the variance in EVI for evergreen needleleaf forests, deciduous broadleaf forests, mixed forests, grasslands, and croplands during the growing season.

Figure 5.

Logistic curves (solid lines) fitted to individual NBAR EVI values (dots) for several IGBP land cover types in the northern United States. The asterisks identify EVI values during snow periods and diamonds represent missing values. Abbreviations: Gin, onset of greenness increase; Gma, onset of greenness maximum; Gde, onset of greenness decrease; Gmi, onset of greenness minimum. Figure 5f shows an example of a cropland in China in which two crops are grown in a single twelve month period.

[26] Figure 6 presents images showing the timing of phenological transitions at global scales from MODIS. The estimated transition dates reflect both broad-scale patterns in controlling mechanisms related to climate, and more local factors related to land cover and human activities that contribute to spatial variation in phenology at this scale. Because of limited vegetation growth, phenological transitions are not identified in many arid regions (i.e., subtropical deserts), and in permanently snow-covered regions at high latitudes. Further, cloud cover severely compromises the quality of NBAR EVI time series in extensive areas of the tropics and subtropics, and as a consequence, estimates of phenological transition dates are missing in many of these areas. Indeed, in many areas with evergreen vegetation, the annual variation in EVI is too subtle to retrieve phenology.

Figure 6.

Global maps of phenological transition dates. For pixels where multiple growth cycles are present, only one cycle is displayed. The areas in grey do not have sufficient data or do not have a strong detectable annual cycle in vegetation growth.

4.3.2. Global Patterns in Phenology Associated With Climate

[27] Geographic patterns in the timing of phenological transitions reflect spatial patterns in climate, biome types, and land use. Vegetation phenological patterns in the northern mid and high latitudes are mostly associated with seasonal variation in temperature (Figure 6). As a consequence, the observed patterns depend strongly on latitude. For example, the onset of greenness increase occurs in March in the southern United States (south of roughly 40°N), in April and early May in the northern United States and southern Canada, and in June in northern Canada. Conversely, the onset of greenness minimum spreads southward beginning in late September around 65°N, and ends in late November in the southern midlatitudes.

[28] Inspection of zonal patterns in the onset of greenness increase and the onset of greenness minimum indicates that this transition varies at a rate of about 2 days per degree of latitude in North America, Europe, and Asia [Zhang et al., 2004b]. Further, the average growing-season length (estimated as the length of time between the onset of greenness increase and the onset of greenness minimum) is strongly correlated to mean annual land surface temperature, with an average overall dependence of about 5 days/°C for natural vegetation classes (Figure 7). Note, however, that this relationship is complex and spatially variable because of the spatial complexity in climate and land cover. In particular, the observed phenology of forests depends more strongly on latitude and temperature than do shrublands, savannas, and grasslands [Zhang et al., 2004b].

Figure 7.

Zonal averages (0.1 degrees) for phenological transition dates and MODIS land surface temperature in (a) North America and (b) Europe and Asia.

[29] In arid and semiarid regions, vegetation phenology depends strongly on rainfall. In Mediterranean climates and the southwestern United States, vegetation growth occurs mainly in the winter and spring, and during the summer monsoon season. In Africa (outside of the humid tropical regime) and in Australia and southern South America grasses, shrubs and savannas are the dominant vegetation types. In all of these regimes, the onset of greenness increase depends on the timing of the rainy season. To illustrate, Figure 8 shows results comparing the timing of greenness increase from MODIS against the onset of the rainy season estimated from TRMM for arid and semiarid regions at global scales, and clearly illustrates the strong dependence of phenology on precipitation regimes in these climate domains.

Figure 8.

Scatterplot showing the onset of greenness increase from MODIS on the onset of the rainy season from TRMM in seasonally dry climates at global scale.

[30] Closer examination of the results for Africa suggests some more specific conclusions [Zhang et al., 2005]. In particular, the timing of greenness onset shifts from early March around 6.5°N to late July at the southern edge of the Sahara Desert (Figure 6). The onset of greenness minimum occurs in early November just south of the Sahara Desert, but occurs in late January further to the south. Both of these patterns reflect the timing of the rainy season in this region, which is controlled by the Intertropical Convergence Zone (ITCZ). In contrast, the pattern of phenology in southern Africa is much more complex (southward of roughly 5°S). In the eastern part of this region, vegetation growth generally starts between September and November, whereas the onset of greenness increase tends to occur in February and March in southwestern Africa (west of the Kalahari Desert). In the Greater Horn of Africa, two cycles of vegetation growth are evident, which reflect the bimodal precipitation regime in this region.

[31] In South America and Australia the patterns are similarly complex, with no well defined geographic patterns in phenology at continental scales. However, some regional patterns are clearly evident. For example, the onset of greenness increase in central north Australia (13–21.5°S and 128–140°E) varies from October to late January as a function of latitude, which follows the north-south gradient in rainfall seasonality arising from the Australian summer monsoon and extra-monsoonal rainfall events [e.g., Hendon and Lebmann, 1990]. Note, however, that because of the arid climate of this region, the magnitude of seasonal variation in EVI is relatively small.

4.3.3. Biogeographic Patterns in Annual Metrics of Phenology

[32] Finally, we consider geographic patterns in the so-called annual metrics of phenology, which include (1) the EVI at the onset of greenness increase (GIEVI), (2) the EVI at the onset of greenness maximum (GMEVI), and (3) the growing season summation in EVI (INEVI). Figure 9 presents global maps for each of these metrics. In each case, geographic patterns are clearly evident, often with clear interpretations. For example, GIEVI and GMEVI clearly vary as a function of biome or land cover type, where values for each are high in forested regions, and low in more arid regions such as the western United States.

Figure 9.

Global maps of annual metrics for vegetation phenology. Note that INEVI was not calculated if the given time series of data did not cover a complete growing season. The areas in grey do not have sufficient data or do not have a strong detectable annual cycle in vegetation growth.

[33] To illustrate this more clearly, boxplots for GIEVI and GMEVI are shown in Figure 10, where the data have been stratified on the basis of land cover type for North America. Average values in GMEVI vary between 0.28 and 0.65, with the highest values exhibited by deciduous broadleaf forests, followed by mixed forests, savannas, evergreen needleleaf forests, shrublands, and grasslands. Similarly, GIEVI is highest in deciduous broadleaf forests (0.31 ± 0.04) and lowest in grasslands (0.16 ± 0.04). Values for the INEVI reflect these patterns, but also contain information related to the growing season average EVI and the length of the growing season. INEVI is very high in tropical and subtropical forests, and is low in arid, semiarid regions, cold climates. Of course, these results are predictable, but the patterns are striking nonetheless.

Figure 10.

Box plots for NBAR EVI values at (a) the onset of greenness increase and (b) greenness maximum, stratified by land cover type. Data include 1000 points for each land cover type randomly selected in North America. Abbreviations: v1, evergreen needle forests; v2, deciduous forests; v3, mixed forests; v4, shrublands; v5, savannas; v6, grasslands.

[34] In addition to reflecting biome-specific differences, it has been widely suggested that the INEVI (or its counterpart based on the NDVI) reflects large-scale patterns in ecosystem primary productivity. Figure 11, for example, presents the relationship between INEVI and annual net primary production (NPP) from the MODIS NPP algorithm (MOD17), again stratified by land cover type for North America. These plots clearly suggest that a significant relationship exists between these two variables, although the nature and strength of the correlation depends on land cover type. For example, in deciduous forests and croplands, INEVI appears to be exponentially related with NPP, while the relationship is linear in other land cover types. These results are similar to results from previous studies that have compared AVHRR data against estimates of both net primary productivity (NPP) and aboveground NPP (ANPP) [e.g., Posse and Cingolani, 2004; Tucker et al., 1986].

Figure 11.

Correlation between INEVI and MODIS NPP for different land cover types in North America.

5. Discussion and Conclusions

[35] This paper provides an overview of methods and results from recent efforts to estimate global transition dates in vegetation phenology using data from MODIS. The method we use is one of several that have been developed for this purpose, and provides a flexible, repeatable, and realistic means to monitor seasonal and interannual dynamics in vegetation from remote sensing data at global scales. However, more work is required to fully understand what estimates of land surface phenology measured from MODIS (and other data sources) are providing.

[36] Validation is always a significant concern in any remote-sensing-based analysis, especially one performed at global scale. The validation results presented in this paper, which used three independent sources of field measurements, suggest that the methodology we have developed provides realistic estimates of transition dates. Comparisons of MODIS-derived estimates of phenological transition dates to data collected at Harvard Forest and Hubbard Brook suggest that onset of greenness increase measured from MODIS corresponds quite closely to the start of budburst (or onset of vegetation greenup) in both of these locations. Estimates of the onset of greenness maximum differ by less than ten days from in situ observations of maximum green vegetation cover and the occurrence of maximum leaf size in the canopy (maximum leaf area index). Since vegetation indices are sensitive to leaf chlorophyll and moisture content [Hardy and Burgan, 1999; Ceccato et al., 2002], the onset of greenness decrease may reflect decreases in leaf chlorophyll content and leaf water stress just before the start of visible leaf coloring, which may influence our results. Our comparisons also suggest that the estimated onset of greenness minimum agrees closely with the timing when the green leaf area of the canopy in deciduous trees effectively reaches zero.

[37] While the assessment results presented in this paper are quite encouraging, it is important to note that substantially more field data are required to fully asses the nature and quality of the MOD12Q2 product, and looking forward, field data collection in support of this effort will become increasingly important. Specifically, because vegetation and land cover are rarely uniform at the scale of 1 km, the exact meaning of remotely sensed phenological metrics at this spatial resolution is somewhat unclear. Further, most current field measurements are conducted to identify the timing of budburst or flowering for one, or at most a few plants, at each site. The temporal and spatial scales of these measurements make them difficult to directly compare with MODIS retrievals. To address this, we are currently developing a set of field protocols and sampling designs that will provide a better basis for comparing in situ measurements against remotely sensed observations. Other large-scale efforts such as the National Phenology Network (http://www.uwm.edu/Dept/Geography/npn/) and the European Phenology network (http://www.dow.wau.nl/msa/epn/index.asp) should also help in this regard.

[38] At global scales the results presented in this paper suggest that MODIS-derived estimates of phenological transition dates are geographically and ecologically realistic. In particular, northern hemisphere patterns in the estimated start and end of the growing season correlate strongly with temperature patterns. Similarly, spatial variability in the phenology of seasonally dry climate regions is strongly correlated with precipitation regimes. Clearly, neither of these results constitutes a new discovery per se. However, the agreement and realism of these results provides confidence regarding the quality of results from MODIS. We therefore conclude that the global data sets described in this paper provide substantial information for understanding and modeling the effects of climate variability on ecosystem processes including energy, water, and carbon fluxes at regional to global scales.

[39] The presentation of large-scale patterns in phenology focused on climate, and to a lesser degree, biome type. However, land use is also an important determinant. In particular, agricultural land use is one of the most geographically extensive land cover types on Earth. Because agricultural land areas are managed by humans, their phenological behavior is frequently distinct from that of surrounding natural vegetation. This pattern is evident in estimates of phenological transition dates from MODIS, especially in regions of large-scale intensive agriculture. For example, Figure 12 illustrates the magnitude of differences in phenology between agricultural areas and naturally vegetated areas in the central U.S. The onset of greenness increase and the onset of greenness maximum occur much later in the Mississippi River valley and in the midwestern agricultural heartland relative to surrounding natural vegetation. Similarly, the onset of greenness minimum occurs much earlier in croplands. Both patterns reflect sowing and harvest practices, and as a result, the exact timing of crop phenology in these areas can be quite variable and depends strongly on crop type and human management. Moving forward, the role of land management by humans in controlling large-scale phenology is likely to become an increasingly important topic of research.

Figure 12.

Spatial correspondence between phenology and agricultural land use in central North America. (a) Onset of greenness increase. (b) Onset of greenness maximum. (c) Onset of greenness decrease. (d) Onset of greenness minimum. (e) Agricultural areas; yellow indicates croplands and maroon indicates cropland and natural vegetation mosaic.

[40] Finally, it is important to note that the algorithms and data sets we describe in this paper are in relatively early stages of development and a number of challenges still need to be addressed. In particular, persistent cloud cover precludes retrieval of MODIS reflectance data in many parts of the world, especially in the tropics. Similarly, while we have attempted to account for the effect of snow on the EVI, it is likely that residual contamination is present in the data. Perhaps most importantly, because the current phenology is based on 1-km data collected at 16-day intervals, the MOD12Q2 product currently provides a generalized characterization of global land surface phenology. In the near future we will begin using eight-day, 500-m data to produce maps of global phenology, which should significantly help in this regard.

Acknowledgments

[41] This work was funded under NASA contract number NNG04HZ71C. The authors thank John C. F. Hodges for help in preparing data sets. We also wish to thank John O'Keefe at the Harvard Forest LTER and Amey Bailey of the USDA Forest Service Hubbard Brook Experimental Forest for providing phenology data, and Maosheng Zhao who provided MODIS NPP data.

Ancillary

Advertisement