Long-Term Detection of Global Vegetation Phenology from Satellite Instruments

Vegetation phenology is the expression of the seasonal cycles of plant processes and their connections to climate change (temperature and precipitation). The timing of phenological events can be used to document and evaluate the effects of climate change on both individual plant species and vegetation communities. Thus, vegetation phenology (including shifts in the timing of bud burst, leaf development, senescence, and growing season length) is considered as one of the simplest and most effective indicators of climate change (IPCC, 2007). Long-term observing and recording of changes in plant phenology support efforts to understand trends in regional and global climate changes, to reconstruct past climate variations, to explore the magnitude of climate change impacts on vegetation growth, and to predict biological responses to future climate scenarios.


Introduction
Vegetation phenology is the expression of the seasonal cycles of plant processes and their connections to climate change (temperature and precipitation).The timing of phenological events can be used to document and evaluate the effects of climate change on both individual plant species and vegetation communities.Thus, vegetation phenology (including shifts in the timing of bud burst, leaf development, senescence, and growing season length) is considered as one of the simplest and most effective indicators of climate change (IPCC, 2007).Long-term observing and recording of changes in plant phenology support efforts to understand trends in regional and global climate changes, to reconstruct past climate variations, to explore the magnitude of climate change impacts on vegetation growth, and to predict biological responses to future climate scenarios.vegetation indices at spatial resolutions of 250 m, 500 m, and 1 km globally, with substantially improved geometric and radiometric properties (Huete et al., 2002).
Various phenology products have been developed from satellite data at regional and global scales.These products include: (1) the MODIS Land Cover Dynamics Product (MCD12Q2) derived from MODIS NBAR (nadir bidirectional reflectance distribution function adjusted reflectance) EVI (enhanced vegetation index) (500m-1000m), which is the only global product that is produced on an operational basis from 2001 to present (Zhang et al, 2006;Ganguly et al, 2010); (2) the MODIS-based product generated at NASA-GSFC (Goddard Space Flight Center) in support of the North American Carbon Program, which was produced using MODIS data at a spatial resolution of 250m-500m (Morisette et al., 2009;Tan et al., 2011); (3) the MODIS phenology product being generated for the contiguous United States (CONUS) by the US Forest Service (Hargrove et al., 2009); (4) the USGS long-term 1km AVHRR phenology product for CONUS (1989-present;Reed et al., 1994); (5) the NOAA 4-km GVIx phenology over North America from 1982-2006(Zhang et al., 2007); (6) the global 4.6 km product for 2005 from the Medium Resolution Imaging Spectrometer (MERIS) Terrestrial Chlorophyll Index (MTCI) (Dash et al., 2010); and (6) the global product based on FPAR (Fraction of Photosynthetically Active Radiation) developed by the European Space Agency (Verstraete et al., 2008).
Satellite-derived phenology demonstrates recent climate change at a large spatial coverage.Using AVHRR NDVI between 1981and 1991, Myneni et al. (1997) have estimated an advance of 8 ± 3 days in the onset of spring and an increase of 12 ±4 days in GSL in northern latitudes (45-70°N).An extended comparison of average AVHRR-NDVI values from July 1981 to December 1999 has shown that the duration of growing seasons increased by as much as 18 days in Europe and Asia, and by 12 days in northern North America (Zhou et al., 2001).Furthermore, analysis of phenology derived from AVHRR NDVI between 1981 and 2006 across North America indicates that vegetation greenup onset advanced by 0.32 days/year in cold and temperate climate regions because of spring warming temperatures, while it changed progressively from an early trend (north region) to a later trend (south region) in subtropical regions because the shortened winter chilling days were insufficient to fulfill vegetation chilling requirements (Zhang et al., 2007).However, little significant phenological trend has been found using the phenology detection capabilities of AVHRR NDVI during 1982-2006 over North America (White et al., 2009).
Monitoring of vegetation phenology from remote sensing remains a significant challenge, although this technique has been demonstrated to be a robust tool.This is because satellite observations are frequently interfered with various abiotic factors, and a satellite footprint covers a large vegetation community at landscape scales.This chapter briefly introduces current methods in phenology detection from satellite data, and further presents long-term variation in satellite-derived vegetation phenology at the scale of global coverage.

Vegetation index for phenology detection
Vegetation index (VI) derived from satellite data has been widely applied to monitor vegetation properties.The most commonly used vegetation index is the Normalized Difference Vegetation Index (NDVI).It was first formulated by Rouse et al. (1973) using the following formula: where ρ NIR and ρ red stand for the spectral reflectance measurements acquired in the nearinfrared and red regions.
The NDVI derived from satellite data has been proved to be a robust tool for retrieving local and global vegetation properties, including vegetation type, net primary product, leaf area index, foliage cover, phenology, photosynthetically active radiation absorbed by a canopy (FPAR), evapotranspiration (ET), and biomass (e.g.Tucker et al., 1986;Unganai and Kogan, 1998;Loveland et al., 1999;Myneni et al., 2002, Friedl et al., 2002).More importantly, a long time series of AVHRR NDVI data has been widely applied for exploring global climate change reflected by variation of inter-annual vegetation phenology (Read et al., 1994;Myneni et al., 1997;Zhou et al., 2001;Nemani et al., 2003;Zhang et al., 2007).Although NDVI provides researchers with a way to monitor vegetation characteristics, the use of NDVI across a variety of vegetation types may be limited by sensitivity to background reflectance (soil background brightness and moisture condition) (Huete et al., 1985;Bausch, 1993), the attenuation caused by highly variable aerosols (Kaufman and Tanré, 1992;Miura et al., 1998;Ben-Ze'ev et al., 2006), and the saturation at densely vegetated areas (Huete et al., 2002;Gitelson, 2004).
The enhanced vegetation index (EVI) has been developed to improve the quantification of vegetation activity (Huete et al., 2002).EVI reduces sensitivity to soil and atmospheric effects, and remains sensitive to variation in canopy density where NDVI becomes saturated (Huete et al., 2002).It is calculated from reflectance in blue, red and near-infrared bands, using the formula: 12 where  blue ,  red and  NIR are values in the blue, red, and near-infrared bands, respectively, L (=1) is the canopy background adjustment, C 1 (=6) and C 2 (=7.5) are aerosol resistance coefficients, and G (=2.5) is a gain factor.
As described in the above equation, EVI requires information on reflectance in blue wavelengths, which is not available on some satellite instruments, including SPOTVGT, SeaWiFS, ENVISAT-MERIS, GLI, and AVHRR.To overcome this limitation, a two band EVI (EVI2) has been proposed (Huete et al., 2006;Jiang et al., 2008), which is described as: where C 3 is a coefficient (2.4).
The two-band adaptation of EVI2 is fully compatible with EVI (Huete et al., 2006;Jiang et al., 2007).The EVI2 remains functionally equivalent to the EVI, although slightly more prone to aerosol noise, which is becoming less significant with continuing advancements in atmosphere correction.Similar to EVI, EVI2 is less sensitive to background reflectance, including bright soils and non-photosynthetically active vegetation (i.e.litter and woody tissues) (Rocha et al., 2008).Thus, it could be used to monitor vegetation phenology and activity across a variety of ecosystems (Rocha and Shaver, 2009).

Algorithm of phenology detection
Phenology detections from time series of satellite data are commonly composed of two steps: modeling of the temporal VI trajectory and identification of the timing of phenological phases.Modeling (or smoothing) of the temporal VI trajectory is to reduce non-vegetative information (noise) in the satellite observations.The noise in an annual time series is mainly caused by environmental impacts: cloud cover, atmospheric effects, and snow cover.To minimize cloud and atmospheric contamination, the maximum value composite (MVC) (Holben, 1986) and best index slope extraction (BISE) (Viovy et al., 1992) are commonly applied to create weekly, biweekly, or monthly composites.To further reduce noise, time series of VI data are often smoothed using a variety of different methods including Fourier harmonic analysis (Moody and Johnson, 2001), asymmetric Gaussian function-fitting (Jonsson and Eklundh, 2002), piece-wise logistic functions (Zhang et al., 2003), Savitzky-Golay filters (Chen et al., 2004), degree-day based quadratic models (de Beurs and Henebry, 2004), and polynomial curve fitting (Bradley et al., 2007).In mid-and high latitudes, vegetation signals are also contaminated by snow cover during winter.To reduce snow contamination, which generally results in a dramatically steep drop in NDVI and irregular variation in EVI (Zhang et al., 2006), snow cover observations are explicitly removed or replaced.This is done using nearest non-snow observations in a temporal VI trajectory after winter periods are determined using ancillary data of land surface temperature and snow detection (Zhang et al., 2004a;Tan et al., 2011) or high values of NDWI (Delbart et al., 2005).
For long-term VI data record, the noises also result from instrumental uncertainties related to sensor decay and inconsistency among multi-sensors.A variety of studies have simulated VI values across different sensors to investigate the uncertainty caused by various impact factors and to establish VI translation equations.Generally, the VI values from various instruments are continued using a set of linear or quadric equations (Steven et al., 2003;Fensholt and Sandholt, 2005;Miura et al., 2006).
The modeled annual time series of VI data is not necessary for the accurate reflection of seasonal vegetative signals because of the complex abiotic influences.The degree of vegetation representation is strongly dependent on the model approaches used.The uncertainty in the temporal VI trajectory is generally the main source of errors in the detection of vegetation phenologic metrics, which is currently lack of detailed investigations.
A number of methods have been developed to identify the timing of phenological phases (or metrics) from the modeled/smoothed temporal VI trajectory at regional and global scales.
Various approaches in detecting phenological timing, particularly the greenup onset, are compared using the same dataset (de Beurs and Henebry, 2010; White et al., 2009).Evidently, most of the methods work well at local and regional scales, or for specific vegetation types.However, they are difficult to implement globally since empirical constants are involved and generally do not account for ecosystem specific characteristics of vegetation growth.

Global vegetation phenological metrics
Phenology observed from satellite data is usually defined as land surface phenology (de Beurs and Henebry, 2004;Friedl et al., 2006) because an annual cycle of satellite data reflects seasonal variation composed of vegetation, atmosphere, snow cover, water conditions, and other land disturbance.However, vegetation seasonal dynamics are generally the parameters of interest to retrieve, whereas the abiotic signals in the temporal satellite data are considered to be noise.As a result, long-term global satellite-based phenological metrics in this chapter are defined according to vegetation seasonal cycles.Briefly, a seasonal cycle of vegetation growth consists of a greenup phase, a maturity phase, a senescent phase, and a dormant phase (Figure 1, Zhang et al., 2003).These four phases are characterized using four phenological transition dates in the time series of VI data: (1) greenup onset (leaf-out): the date of onset of VI increase; (2) maturity onset: the date of onset of VI maximum; (3) senescence onset: the date of onset of VI decrease; and (4) dormancy onset: the date of onset of VI minimum.Furthermore, the time series of VI data provides the integrated VI for the growing season (the sum of daily VI values varying from greenup onset to dormancy onset), maximum and minimum VI values during a growing season, and the length of the vegetation growing season.
During a senescent phase, foliage senescent development consists of several coloration statuses (Zhang and Goldberg, 2011).Fall foliage coloration is a phenomenon occurring in many deciduous trees and shrubs worldwide.Fall foliage status is a function of the colored leaves on the plant canopy.With the spread of colored foliage, the percentage of fallen leaves increases.Their difference represents relative variation in colored leaves on plant canopy, which can be quantified using a temporally-normalized brownness index.The occurrence of the maximum relative variation derived from the brownness index is considered to be a critical point in foliage coloration status, this being the onset timing of peak foliage coloration.Prior to this point, foliage status is generally defined using the categories of little/no change, low coloration, moderate coloration, and near-peak coloration.Following the critical point, it is divided into peak coloration phase and postpeak coloration phase (Figure 2).More than one set of vegetation phenological metrics could occur within a one-year period because of the complexity of phenological cycles across the globe.Vegetation growth can undergo one or more cycles, and may include an incomplete cycle (truncated at the beginning or end) during a year (Figure 3).The simplest case is illustrated in Figure 3a, where a single and complete growth cycle centers near the mid-point of a 12-month period.Two partial cycles are recorded in Figure 3b, 3c, and 3d. Figure 3e illustrates the situation where two complete growth cycles are finished, which leads to two complete sets of phenological metrics.Figure 3f-3h shows examples of two incomplete cycles and one complete cycle.To capture vegetation phenological timing properly from the complex cycles within a given one-year period, the satellite data should be extended by periods of a halfyear prior to and following the period of interest, separately.

Detection of global vegetation phenology
To determine the global phenological metrics described above, the following approaches are conducted.Temporal VI data are first preprocessed to remove or reduce the impacts of clouds, atmosphere, snow cover, etc.Specifically, the data gaps caused by clouds-creating isolated missing values-are filled by linear interpolation using neighbor good quality data.
The time series of VI data at each pixel is then smoothed using a Savitzky-Golay and running local median filter.The background VI value at each pixel, which represents the minimum VI of soil and vegetation in an annual time series (Zhang et al., 2007), is identified and it is used to replace VI values in the time series flagged as snow contaminations.
Vegetation growth cycle is identified using a moving slope along the VI time series.The periods with sustained VI increase and decrease at each pixel are determined using a fivepoint moving slope technique, where transitions from periods of increasing VI to periods of www.intechopen.comdecreasing VI are identified by changes from positive to negative slope, and vice versa.
Because slight decreases or increases in VI can be caused by local or transient processes unrelated to vegetation-growth cycles, two heuristics are applied to exclude such variation: (1) the change in VI within any identified period of VI increase or decrease must be larger than 35% of the annual range in VI for that pixel; and (2) the ratio of the local maximum VI to the annual maximum VI should be at least 0.7.This approach screens out short-term variation unrelated to growth and senescence cycles in VI data, while at the same time identifying multiple growth cycles within any 12-month period.
VI time series in the growing phases (VI consistent increase) and senescent phases (VI consistent decrease) is modeled using a sigmoidal vegetation growth function (Zhang et al., 2003).The specific sigmoid function used to model temporal VI dynamics is the logistic function of vegetation growth: () where t is time in days, y(t) is the VI value at time t, a and b are free parameters that are fitted using a non-linear least sqaures approach, c is the amplitude of VI variation and d is the initial background VI value.The advantages of the sigmoidal model are that: (1) it provides a simple, bounded, continuous function for modeling growth and decay processes; and that (2) each parameter can be assigned a biophysical meaning related to vegetation growth or senescence.
This sigmoidal model has been demonstrated to be effective in depicting seasonality of vegetation growth as a function of time (or cumulative temperature) in various ecosystems and data measurements.It was originally developed for monitoring crop growth based on field measurements (e.g., Richards, 1959;Ratkowsky, 1983) and adopted to simulate temporal satellite vegetation index (Zhang et al., 2003).It has then been applied to investigate seasonal vegetation growth using webcam data (Richardson et al., 2006;Kovalskyy et al., 2012), Landsat TM data (e.g., Fisher et al., 2006;Kovalskyy et al., 2011), AVHRR data (e.g., Zhang et al., 2007), and MODIS data (e.g., Zhang et al., 2003Zhang et al., , 2006;;Ahl et al., 2006;Liang et al., 2011).Moreover, studies have shown that the sigmoidal model performance is superior to both Fourier functions and asymmetric Gaussian functions for dictping remotely sensed phenology (Beck et al., 2006).Thus, the physically-based sigmoidal model is applicable for the detection of global vegetation phenology.
Phenological transition dates within each growth or senescence phase are identified using the rate of change in the curvature of the modeled sigmoidal curves (Zhang et al., 2003; Figure 1).Specifically, transition dates correspond to the day-of-year (DOY) on which the rate of change in curvature in the VI data exhibits local minima or maxima.These dates indicate when the annual cycle makes a transition from one approximately linear stage to another.Formally, at any time t, the curvature (K) for the sigmoidal function given above is: where  is the angle (in radians) of the unit tangent vector at time t along a differential curve, and s is the unit length of the curve.Setting z = e a+bt , the rate of change of curvature (K') is: During the growth period, when vegetation transitions from a dormant state to a growth phase, three extreme points in a VI curve can be identified using the equation 6 (Zhang et al., 2003).The two maximum values correspond to the onset of greenup (onset of VI increase) and the onset of maturity (onset of VI maximum), respectively (Figure 1).Similarly, the extreme points during the senescent phase represent the transition dates of the senescent onset (onset of VI decrease) and the dormancy onset (onset of VI minimum).
To determine foliage coloration status, a temporally-normalized brownness index is derived from the relative percentage dynamics of the fraction of colored foliage (Zhang and Goldberg, 2011).This brownness index is described as: where F cbmin =F cmin +F b ; F cbmax =F cmax +F b ; F cb(t) =F c(t) +F b ; TNBI b(t) is defined as the temporallynormalized brownness at time t; F b is the exposed surface background; F c(t) and F cb(t) are the fraction of colored foliage on plant canopy and total brown material at time t, separately; F cmin and F cmax are the maximum and minimum fractions of colored leaf cover; and F cbmin and F cbmax are the minimum and maximum fractions of brown material during the senescent phase, separately.
The temporally-normalized brownness index is directly linked to the temporal trajectory of vegetation index (Zhang and Goldberg, 2011).Specifically, the colored foliage is determined after the modeled temporal VI trajectory during the senescent phase is further combined with a linear mixture model of surface components consisting of green (or photosynthetic) vegetation, colored (or non-photosynthetic) vegetation, and exposed surface background (bare soil and rock).As a result, the temporally-normalized brownness index is deduced as: The temporally-normalized brownness index represents relative changes in colored foliage, and varies with time in each pixel individually.It is independent of the surface background, vegetation abundance, and species composition.Thus, it is robust to divide the foliage coloration status, as displayed in Figure 2, into separate categories of little coloration, low coloration, moderate coloration, near-peak coloration, peak coloration, and post-peak coloration.

www.intechopen.com
Long-Term Detection of Global Vegetation Phenology from Satellite Instruments 307

Vegetation index for global phenology detection
A long-term dataset of global EVI2 has been generated from the daily land surface reflectance from the AVHRR Long-Term Data Record (LTDR) and the MODIS Climate Modeling Grid CMG) records.AVHRR LTDR provides daily surface spectral reflectance at a spatial resolution of 0.05 degrees from various AVHRR sensors from 1981-1999 (Vermote and Saleous, 2006), which is available at the NASA funded REASoN project web site (http://ltdr.nascom.nasa.gov/).The MODIS CMG dataset provides Terra and Aqua MODIS daily CMG surface reflectance (Collection 5.0) at a spatial resolution of 0.05 degrees, covering the period from 2000 to 2010, which is available at the USGS for EROS DAAC (http://edcdaac.usgs.gov/main.asp).From these daily surface spectral reflectance, the longterm daily EVI2 has been calculated and available for the last 30 years (http://vip.arizona.edu/viplab_data_explorer).

Results in global vegetation greenup onset
Global vegetation phenological metrics during the last three decades were detected from global daily EVI2 using a series of piece-wise logistic models.Here, only the greenup onset is presented and discussed because it is the most important parameter in a vegetation seasonal cycle.

Spatial pattern in the timing of greenup onset
Figure 4 sets out the average onset of vegetation greenup in the 1980s, 1990s, and 2000s.If there were multiple seasonal cycles in a given calendar year, the first occurrence of greenup onset was selected.As expected, the spatial pattern in the three periods is very similar.However, the spatial variation in phenological transition dates reflects both broad-scale patterns in controlling mechanisms related to climate, and more local factors related to land cover and human activities.
Several spatially distinctive properties of greenup onset are evident.Changes in phenology with latitude are apparent in most of the northern hemisphere, from 30°N northwards (Figure 4).Greenup onset occurs in early March in the southern USA (south of 40°N), April in the northern USA, and at the end of June in northern Canada.Zonal patterns in the timing of greenup onset indicate that the transition date of greenup varies at a rate of about 2-3 days per degree of latitude in North America, Europe, and Asia (Figure 5a, 5b).This latitude dependence is assumed to be a function of temperature variation (Myneni et al., 1997;Zhang et al., 2004a).
The dependence on latitude is spatially variable because of the spatial complexity in elevation and human activities.For example, the timing shift in greenup onset is about one and half months from bottom to top of the Carpathian Mountains and Dinaric Alps in Europe (Figure 5b).This reflects that the timing of greenup onset is also a function of elevation in mountains, because temperature decreases with increasing elevation.Moreover, agricultural land use is one of the most geographically extensive land cover types on the Earth.Their phenological behavior is frequently distinct from that of surrounding natural vegetation because of controls applied by human management.It is highly evident in central North America, where the onset of greenup occurs much later in the Mississippi River valley and the mid-western agricultural heartland, relative to the surrounding natural vegetation (Figure 4).This pattern depends strongly on crop type and human management.Moreover, urban lands advance greenup onset relative to rural areas surrounding the urban regions because of the urban heat island effects (Zhang et al., 2004b), although this is not clearly visualized on the 0.05 degree maps.
In dry climate (arid and semi-arid regions), the spatial pattern in vegetation greenup onset is very complex because it is generally controlled by water availability.In Mediterranean climates and the southwestern United States, the start of vegetation growth occurs mainly in winter and early spring and, in some cases, during the summer monsoon season.Outside of the humid tropical regime in sub-Saharan Africa, Australia, and southern South America, the dominant vegetation types are grasses, shrubs and savannas.The onset of vegetation greenup in these vegetation types generally depends on timing of the rainy season.
Inspection of the greenup onset in dry climates reveals several regular patterns in local regions.The most notable pattern is present in northern Africa (the Sahelian and sub-Sahelian region).The timing of greenup onset shifts smoothly from early March, at around 6.5ºN, to mid-October in the boundary between the Sahel and the Sahara desert (17.9ºN, Figure 5b).The shift rate is about 20 days per degree of latitude, which is about 10 times slower than that in temperate North America and Eurasia.This pattern reflects the start of the rainy season, which triggers the onset of vegetation growth in this region (Zhang et al., 2005), which is in turn controlled by the migration of the Intertropical Convergence Zone (ITCZ).In contrast, the phenological pattern found in southern Africa is much more complex (roughly 1ºS southward), although greenup onset shows a regular delayed shift from 1ºS to 22ºS and an advanced shift of 22ºS southwards (Figures 4 and  5b).In the eastern part of this region, vegetation growth generally starts between September and November, whereas it tends to occur in February and March in southwestern Africa (west of the Kalahari Desert).In the Great Horn of Africa, two cycles of vegetation growth are evident, which reflects the bimodal precipitation regime in this region.These irregular patterns coincide strongly with patterns evident in the arrival of the rainy season (Zhang et al., 2005).
In South America, four different phenological regions follow the variation in the onset of vegetation greenup.Greenup onset occurs in the boreal winter, with no obvious gradient in the northern Andes mountainous region.In southern South America, green leaves emerge in the boreal summer and gradually push northward at a rate of about three days per latitude (Figure 5c).However, a remarkable phenological trend exists along the Brazilian Highlands (in the direction from 60ºW and 39ºS to 35ºW and 5ºS), where the greenup onset shifts from July to next February at a rate of about 0.12 days/km.In contrast, the timing of greenup onset is very irregular in the Amazon rainforest, where the values are of poor quality because of high frequencies of cloud cover and weak seasonality in vegetation index.
Overall, the complex phenological pattern is likely to be associated with precipitation and latitude-elevation-dependent temperature.
Phenological variation in Australia divides into three distinct regions.Greenup onset occurs in the late boreal autumn and winter in northern areas, in the boreal summer in southern areas, and in the boreal spring, or with no clear phenology, in central Australia.Although Specifically, the onset of greenup occurs mainly in January over northern Australia, while phenological phases occur about six months later in southern Australia.For example, the timing of greenup onset in central north Australia (13-21.5°S and 128-140°E) shifts at a rate of 0.1 days/km from October to late January.This trend is controlled by the Australian summer monsoon and extra-monsoonal rainfall events (e.g., Hendon and Lebmann, 1990), and also reflects the changes in species composition and a decrease in both tree biomass and diversity (Cook and Heerdegen, 2001).Note that the detected phenology metrics are of poor quality for evergreen vegetation (tropical rainforests and boreal forests) in many areas.This is because the annual variation in EVI2 is too subtle to retrieve phenology effectively.Moreover, there is no vegetation growth in tropical desert areas and polar regions of permanent snow cover.

Inter-annual variation in greenup onset
Inter-annual variation in greenup onset is limited in temperate and cold climate regimes (Figure 6).In the northern hemisphere, the standard deviation is generally less than 10 days, although there are several locations with a standard deviation of about 11-15 days in evergreen needle leaf forest where EVI2 seasonality is weak.This suggests that spring occurrences of greenup onset are regularly triggered by an increase in spring temperature, which leads to a comparable annual EVI2 trajectory (Figure 7a).
In contrast, the inter-annual timing of greenup onset varies considerably in arid and semiarid climate regimes (Figure 6).The standard deviation is generally larger than 15 days within each decade.This is probably associated with the fact that vegetation greenup onset strongly tracks rainy season occurrence, which can change greatly between years (Zhang et al., 2005).For example, a temporal EVI2 trajectory in shrubland in the southwestern United States clearly indicates the variability of inter-annual vegetation growing cycles (Figure 7b), with the timing of greenup onset varying from DOY 85 to 213 during the period from 2001 to 2009.

Shift in greenup onset
Figure 8 sets out the shift of greenup onset during the past three decades.From the 1980s to the 1990s, the onset of vegetation greenup became advanced in most of the northern hemisphere, South America, and the Sahelian and sub-Sahelian regions.However, delayed shifts appeared in relatively small regions in each continent, except for the southern semiarid region in Africa.
From the 1990s to the 2000s, shifts in greenup onset were basically opposite to those in the previous period in large parts of South America, Africa, and North America.In contrast, persistent trends during the three decades occurred in relatively small regions.In particular, an advanced trend was evident in most of Eurasia.
It is worth noting that the vegetation greenup occurred in a much larger area across the Sahel in the 1990s than in the 1980s.This trend agrees with the result derived by Tucker and Nicholson (1999) and Olsson et al. (2005).However, the region with greenup occurrence was reduced during 2000s, which is probably associated with retreat of the ITCZ migration.

Discussion and conclusions
This chapter provides an overview of methods and results in the detection of vegetation phenology using satellite data.Various vegetation indices derived from satellite data reflect seasonal dynamics in vegetation growth with reasonable accuracy, and a variety of methods have been developed for detecting vegetation phenological metrics.In the detection of longterm global vegetation phenology, EVI2 from AVHRR and MODIS data has advantages over NDVI and EVI (Rocha and Shaver, 2009) and a series of pieces-wise sigmoidal models of vegetation growth provide a flexible, repeatable, and realistic means to monitor seasonal and inter-annual dynamics in vegetation using remote sensing data across the globe.
At global scales, vegetation greenup onset during the past three decades suggests that AVHRR and MODIS-derived estimates are geographically and ecologically realistic.In particular, patterns in the timing of greenup onset are strongly dependent on latitude (temperature patterns) in temperate and cold climate regimes across the northern hemisphere, although these patterns are also impacted by elevation and human activities locally.Their inter-annual variance is relatively small, with a value generally less than 10 days within each decade.In contrast, greenup onset in arid and semiarid climate regions is very complex.The regular spatial gradient only occurs in local regions, such as the Sahelian and sub-Sahelian region.The inter-annual variance of phenological timing could be larger than one month.This is probably in response to precipitation regimes and rainfall seasonality migrations (Zhang et al., 2005).
The long-term shifts of vegetation phenology in most parts of the globe are generally episodic rather than persistent in response to climate changes.Early trends of greenup onset from the 1980s to the 1990s appear across most of the northern hemisphere, which agrees with previous findings (Zhou et al., 2001;Myneni et al., 1997;Zhang et al., 2007).The consistent advanced trends from the 1980s-1990s-2000s only occur in large parts of Eurasia and small parts of North America.In most regions of South America, the timing of greenup onset shifted from an early trend to a late trend while an opposite shift occurred in Africa.The detailed mechanisms driving these complex trends will be further investigated.
Finally, it is critical to provide the quality and accuracy of satellite global vegetation phenology detections.Without this, trends derived to predict the response to climate change are less reliable.The quality of phenological detection is strongly dependent on the temporal VI trajectory, which is generally affected by the frequency of cloud cover and snow appearance, and by the model efficiency in removing abiotic noise.To validate accuracy, sufficient field measurements comparable to a satellite footprint are required.This requires field data to reconcile with satellite-based phenological observations, which is currently extremely challenging.The validation effort will become more practical, with the inclusion of observations from webcam (Richardson et al., 2009) and the landscape measurements upscaled from field observations (Liang et al., 2011).Currently, the effort to assess the quality and accuracy of global vegetation phenology is underway.

Fig. 1 .
Fig. 1.Key phenological metrics in an annual trajectory of satellite vegetation index.

Fig. 5 .
Fig. 5. Profiles of the shift of greenup onset: a) along a meridian of 100°W in North America, b) along a meridian of 20°E in Europe and Africa, and c) along a meridian of 65°W in South America.The geographic locations are displayed on Figure 4a.

Fig. 6 .
Fig. 6.Inter-annual variation (standard variation) in the timing of greenup onset in a) 1980s, b) 1990s, and c) 2000s, separately.The color legend is the number of days.

Fig. 7 .
Fig. 7. Time series of daily EVI2 from 2001-2009 in two sample pixels.Solid line is the modeled vegetative EVI2 while the asterisks are the raw EVI2.a) Deciduous forests in northeastern North America and b) shrubland in the semiarid region of southwestern North America.

Fig. 8 .
Fig. 8. Shift in greenup onset.a) The difference between the 1990s and 1980s; b) the difference between the 2000s and 1990s.The green color indicates the number of advanced days, while the red color shows delayed days.