Modeling seasonal vegetation variation and its validation against Moderate Resolution Imaging Spectroradiometer (MODIS) observations over North America



This article is corrected by:

  1. Errata: Correction to “Modeling seasonal vegetation variation and its validation against Moderate Resolution Imaging Spectroradiometer (MODIS) observations over North America” Volume 110, Issue D7, Article first published online: 7 April 2005


[1] Seasonal variability of vegetation, determined by plant phenology, impacts the seasonality of surface and atmospheric water cycles as well as the seasonality of surface energy budget. At the same time, leaf seasonal variations respond to both cumulative and concurrent hydrometeorological conditions. In order to account for this vegetation feedback at the seasonal timescale, a predictive phenology scheme for various plant functional types is developed on the basis of previous studies, and a methodology for crop simulations is proposed and implanted to supplement this phenology scheme. The phenology scheme is then incorporated into the Community Land Model (CLM). The geographic focus of this study is on the United States where the need for seasonal prediction is urgent and vegetation seasonal characteristics have been shown to significantly influence summer precipitation and temperature. Comparison of the model simulation with Moderate Resolution Imaging Spectroradiometer (MODIS)-derived leaf area index data indicates that our model reproduces the observed vegetation seasonality reasonably well. Subsequent experiments demonstrate the interannual variability of vegetation phenology and its impact on surface water and energy budgets using the 1988 drought and 1993 flood in the U.S. Midwest as examples.

1. Introduction

[2] Soil moisture and vegetation are two important and closely related components of land surface conditions. Soil moisture supplies water for vegetation transpiration thus supporting vegetation growth; transpiration depletes the soil water storage, which then stresses the vegetation growth. At the same time, the soil-vegetation state is strongly coupled with the overlying atmosphere through mass, momentum and energy flux exchanges. Precipitation and potential evaporation regulate the state of soil moisture and vegetation; the resulting condition of vegetation and soil moisture controls the actual water and energy supplies from land surface to the atmospheric boundary layer, which then impacts the atmospheric moist convection. These processes form a complex soil moisture-vegetation-precipitation feedback mechanism, which tends to promote the land surface memory of climate anomalies, thus causing a certain degree of climate persistence. Such land memory and the resulting climate persistence are important in seasonal climate predictions. However, the soil moisture-vegetation-precipitation feedback at the seasonal timescale is not fully represented in many climate models because of the lack of model capability to realistically simulate the vegetation phenology, i.e., the seasonal variation of vegetation in response to climate and environmental conditions.

[3] The remotely sensed, normalized difference vegetation index (NDVI)-based leaf area index (LAI) is commonly used in climate models to describe the seasonal and/or interannual variation of vegetation. Many previous studies suggested that realistic description of temporal and spatial vegetation variation could improve the model simulation of surface and atmospheric conditions at the seasonal and interannual timescales [Bounoua et al., 2000; Buermann et al., 2001; Lu and Shuttleworth, 2002; Lawrence and Slingo, 2004]. However, when climate models are used for prediction purposes, observational data are obviously not available. A predictive phenology scheme capable of predicting the vegetation seasonality is therefore essential for seasonal climate predictions. It is also important for assessing the significance of land-atmosphere feedback at the seasonal timescale that gives rise to the existence of land memory for precipitation anomalies. The latter provides the foundation of a new predictor for seasonal precipitation prediction.

[4] There are two different approaches for predicting the seasonal variation of vegetation. One commonly used approach in ecological studies is rule based. It distinguishes between temperature-controlled phenology and water-controlled phenology, predicting leaf on and off once a certain predictor (often a cumulative function of one or several climate variables) reaches an empirically defined threshold [e.g., Murray et al., 1989; Foley et al., 1996; Chuine, 2000; Levis and Bonan, 2004]. The other commonly used approach is based on plant carbon budget [e.g., Dickinson et al., 1998; Tsvetsinskaya et al., 2001; Lu et al., 2001]. It predicts phenology on the basis of products resulting from plant physiological processes (e.g., carbon assimilation during photosynthesis, carbon loss during respiration, carbon allocation to leaf) that respond to various stresses caused by factors such as temperature, moisture, and sometimes nutrient. In this study, we take the rule-based approach.

[5] Most rule-based leaf phenology models were developed at the species level [e.g., Murray et al., 1989]. The availability of remotely sensed vegetation data (e.g., Advanced Very High-Resolution Radiometer (AVHRR) NDVI and Moderate Resolution Imaging Spectroradiometer (MODIS) NDVI) has provided the opportunity to develop regional and global phenology models through extrapolating the existing species-specific phenological models to regional and global ones [Chuine, 2000]. Such large-scale, biome-specific phenology based on satellite observations describes the mean phenological changes at the grid level such as 1° × 1° where diverse vegetation types are aggregated [White et al., 2002], which is consistent with the scale of GCMs. Together with the hydrometeorological data, the remotely sensed NDVI has been used to enhance the prediction of the timing of leaf green up and senescence at the regional [White et al., 1997, 2002] and global scales [Botta et al., 2000]. However, most phenological studies using satellite observations at the biome level only detect the timing of leaf onset/offset, rather than presenting trajectories of LAI.

[6] The recently developed dynamic global vegetation models (DGVMs) (documented by Cramer et al. [2001]), which simulate the biogeographic distribution of vegetation type, include predictive vegetation phenology schemes. These schemes have been parameterized on the basis of either local-, regional-, or global-scale phenological studies in order to simulate the trajectories of LAIs, but the model simulated LAIs have rarely been validated against observations [Wythers et al., 2003]. For over 20 years, National Oceanic and Atmospheric Administration (NOAA)-AVHRR has provided NDVI in fine spatial resolution (1 km), describing the spatially and temporally varying characteristics of the Earth's land surface. More recently, MODIS provides improved NDVI relative to the AVHRR [Justice et al., 1998; Tian et al., 2004]. This improved land surface observing capability as afforded by MODIS provides a good opportunity for validation of existing phenology models and for further model development. In this study we incorporate into the Community Land Model (CLM) version 2.1 a phenology scheme used by several existing phenology models, validate it against the MODIS-LAI data, and further enhance this phenology scheme using paramerizations and dates derived from satellite observations. In doing so, we focus on the region of United Sates for two reasons: First, within the United Sates, severe drought and flood events in the past two decades have caused huge loss and damages. The 1988 drought alone caused more than 40 billion dollar in monetary loss [Ross and Lot, 2003]. This makes accurate seasonal precipitation prediction an urgent need. Second, in this region, vegetation seasonal characteristics have been shown to significantly influence summer precipitation and temperature [Dirmeyer, 1994; Xue et al., 1996; Pielke et al., 1999]. This indicates that introducing the representation of seasonal vegetation-precipitation feedback is important for improving seasonal precipitation prediction.

2. Model and Methodology

2.1. Land Surface Model

[7] In this study, we use the Community Land Model (CLM) version 2.1 that is a recent state-of-the-art land surface model [Bonan et al., 2002; Zeng et al., 2002; Dai et al., 2003]. The model is available at CLM simulates energy, moisture and momentum fluxes between vegetation, soil and the atmosphere. It has 10 unevenly spaced vertical soil layers, up to 5 snow layers and 1 vegetation layer. Land surface within each grid cell is represented by the fractional coverage of four types of patches (glacier, lake, wetland, vegetated), and the vegetated portion of the grid cell is by the fractional coverage of up to four different plant functional types (PFTs). The total of 15 PFTs are described with respect to their ecological characteristics. The seven primary PFTs are needleleaf evergreen and deciduous trees, broadleaf evergreen and deciduous trees, shrubs, grasses and crops. These are further categorized into 15 physiological variants on the basis of climate rules (artic, boreal, temperate and tropical) and supplemented by discrimination of C3 and C4 grasses and crops. Leaf phenology in this version of CLM is prescribed, and the seasonal course of leaf area index (LAI) for each PFT is derived through interpolating the monthly PFT-specific LAI from the AVHRR data as described by Bonan et al. [2002].

2.2. Predictive Phenology Scheme

[8] Phenological behaviors are often divided into temperature-controlled ones and water-controlled ones. Generally, phenology of the boreal and temperate biomes is controlled by temperature, and phenology of tropical biomes depends largely on water availability and genetic control. Variables used to quantify this relationship in phenology models, for temperature-controlled biomes, include temperature threshold, chilling requirement, growing degree day requirement and photoperiod requirement; and for water-controlled biomes, include soil moisture, precipitation and photosynthesis. In this study, we adopt the approach used in the Integrated Biosphere Simulator (IBIS) [Foley et al., 1996; Kucharik et al., 2000], one of the first generation of DGVMs. The same approach is also used in the recently developed DGVM coupled with CLM [Bonan et al., 2003; Levis et al., 2004; Levis and Bonan, 2004]. IBIS (version 1) takes a simple, traditional rule-based approach. It describes the winter deciduousness using temperature threshold and growing degree day requirement, and describes the drought deciduousness using soil moisture. To improve the model performance for certain biomes, we modified this phenology scheme on the basis of the study by White et al. [1997]. In addition, we propose a crop phenology scheme that allows crops to respond to environmental stresses.

[9] The phenology model updates the daily change of LAI through scaling the annual maximum PFT-specific LAI (LAImax) as a response to seasonal climate and environmental conditions. Here we derive the LAImax from the monthly PFT-specific LAI at 0.5° × 0.5° resolution created by Tian et al. [2004]. They created this PFT-specific data set on the basis of the newly available satellite data from MODIS, using procedures similar to those described by Bonan et al. [2002]. Figure 1 presents the global distribution of the percentage of the grid cell occupied by deciduous PFTs and their LAImax on a 0.5° × 0.5° grid.

Figure 1.

Percentage of the grid cell occupied by and the annual maximum LAI of (a) needleleaf deciduous trees, (b) broadleaf deciduous trees, (c) shrubs, (d) grasses, and (e) crops on 0.5° × 0.5° grids.

[10] The PFT-specific LAI is updated daily according to the predicted phenology factor (D), ranging from 0 to 1, as follows:

equation image

The phenology factor (D) depends on the temperature threshold and growing degree days for winter deciduous plants and on plant water stress for drought deciduous plants (see Table 1). For PFTs that respond to both coldness and drought (i.e., grasses and crops), the canopy deciduousness depends on which factor is limiting at the specific time considered. When both temperature and moisture level are low enough to cause leaf shedding, we determine the degree of leaf shedding on the basis of the multiplicative effect of the temperature and moisture stresses instead of on the basis of whichever is more limiting. Our approach is justified by observations indicating that water-stressed plants are more vulnerable to freezing damages [e.g., Langan et al., 1997]. For evergreen trees, their LAI seasonality is prescribed on the basis of observations.

Table 1. Phenology of Plant Functional Types
Plant Functional TypePhenologySupplemental Rules
Needleleaf evergreen tree, temperate 
Needleleaf evergreen tree, boreal 
Needleleaf deciduous treewinter deciduous (W) 
Broadleaf evergreen tree, tropical 
Broadleaf evergreen tree, temperate 
Broadleaf deciduous tree, tropicaldrought deciduous (D) 
Broadleaf deciduous tree, temperateW 
Broadleaf deciduous tree, borealW 
Broadleaf evergreen shrub, temperate 
Broadleaf deciduous shrub, temperateW 
Broadleaf deciduous shrub, borealW 
Grass, C3 articW or D 
Grass, C3W or D 
Grass, C4W or D 
Crop, C3W or Dplantation and harvest dates from the MODIS-derived NDVI
Crop, C4W or Dplantation and harvest dates from the MODIS-derived NDVI

2.2.1. Winter Phenology

[11] The onset of greenness for winter deciduous trees has been successfully modeled using cumulative thermal summation [e.g., Murray et al., 1989]. Simple models consider only the temperature accumulated from a fixed date (usually 1 January), while more sophisticated models consider the additional chilling requirements in the break of buds dormancy [Chuine, 2000]. The winter deciduous phenology scheme in IBIS predicts leaf green up, development and senescence, using 10-day average air temperature (T10) and accumulated growing degree day (AGDD) from 1 January. The base temperatures for AGDD are 0°C for trees and −5°C for grasses. For trees (grasses), the leaf onset takes place during a period of 50° days after AGDD exceeds 100 (150). Equations (2) and (3) describe the winter phenology for trees and grasses:

equation image
equation image

where Tc is the coldest monthly temperature based on climatology and dday is a constant representing the rate of leaf growing or dropping. Here, dday is set to be 1/15, which means that the complete leaf offset from D = 1 to D = 0 or leaf onset from D = 0 to D = 1 takes 15 days. The rate of leaf onset has been modified from the original methodology used in IBIS. Instead of using a variable rate based on the AGDD that allows 50 growing degree days for leaf development, we specify a constant leaf onset rate of 1/15 (d−1). With the original IBIS methodology, when the temperature threshold is reached later than the AGDD threshold, the difference between AGDD and its threshold at the time of leaf onset can be larger than 50 in some cases, causing leaf to develop from nothing to full display within one single day in the model. The constant rate was introduced to avoid such unrealistic situation, even though it does not capture the interannual variability and geographical differences of leaf development rate as the variable rate does. ΔT is a tunable parameter, set to be 5 in IBIS.

[12] Moreover, we consider an additional factor to supplement the winter deciduous phenology scheme for the broadleaf deciduous tree (BDT) over the United States. White et al. [1997] found that the date of leaf offset of deciduous broadleaf forest could be best predicted with a photoperiod and soil temperature. To synthesize the findings of White et al. and the IBIS phenology scheme, we modify equation (2) to the following over the United States:

equation image
equation image

where Ld is the length of day time (photoperiod) in minute and Tsoil is the daily soil temperature. As shown in section 3, adding the photoperiod factor improves our model performance in predicting the leaf senescence for broadleaf deciduous trees.

2.2.2. Drought Phenology

[13] Moisture has been shown to be the primary control for both the leaf onset and offset of drought deciduous species [e.g., Borchert, 1994; Seghieri et al., 1995; Jolly and Running, 2004]. However, very little work has been carried out for the biome-level phenology controlled by water availability [Botta et al., 2000]. Since soil water availability is unquestionably regulating plant growth in dry regions, here we use the whole plant water stress factor (W), which depends on soil water potential, to model the drought deciduousness. It ranges from zero at the permanent wilting point to one at saturation. The drought deciduous phenology scheme predicts leaf shedding on the basis of the 10-day running mean of plant water stress (W10) factor as

equation image
equation image

where froot,j is the fraction of the root biomass within soil layer j, ψ is the soil potential, and Wth is the threshold for whole plant water stress. Here, the whole plant water stress factor (W) is originally designed to account for soil water limitations on photosynthesis and transpiration. Since such processes are not yet clearly understood, a simple heuristic approach, using the soil water stress factor, has been adopted in IBIS [Foley et al., 1996]. CLM takes the same approach, but uses a slightly different mathematical function to estimate W (equation (5) versus equation of Foley et al. [1999, p. 613]). The water stress threshold Wth is a tunable parameter, and a value of 0.333 was used in IBIS. We carried out a group of sensitivity experiments with Wth varying from 0.3 to 0.4, and found little difference in the results among different experiments. This indicates a low model sensitivity to the choice of this parameter within the range experimented. In this paper, we choose to present results from simulations using Wth = 0.4.

2.2.3. Crop Phenology

[14] A reasonable treatment of crop is required to study the land surface processes over regions such as the U.S. Midwest where a considerable fraction of land cover is crop. Numerous models (e.g., CERES-Maize by Jones and Kiniry [1986]; GLYCIM by Acock and Trent [1991]) have successfully simulated the crop phenology, but they often require knowledge of factors such as genotype, planting time, fertilization, irrigation and harvest time that significantly vary in space and are hard to quantify at large scales. Since crops are heavily anthropogenically-controlled, in our study we prescribe the plantation and harvest times on the basis of the NDVI data but let the leaf development respond to hydrometeorological conditions between plantation and harvest in the same way as grasses would. For example, in a year of extreme drought, LAI in an affected area may be zero even though it is the peak of growing season and the plantation has long passed. This approach can be applied globally. In our study we only focus on the U.S. region. While irrigation is routine in some areas of the United Sates, in our study no irrigation is considered.

[15] White et al. [1997] introduced a methodology to extract the dates of leaf onset/offset from the AVHRR NDVI data over the United States, assessing the state of ecosystem with a transformation of the daily NDVI:

equation image

where NDVIratio ranges from 0 to 1, NDVImax is the annual maximum daily NDVI, and NDVImin is the annual minimum daily NDVI. They suggested that a threshold of 0.5 for NDVIratio was the most appropriate to capture the onset and offset dates for deciduous broadleaf forest and grassland sites over the United States. Here we assume that this can be applied to crops as well. Therefore we apply this methodology to the cropland over the United Sates, and treat the so-derived leaf onset/offset dates as the plantation/harvest dates.

[16] We used MODIS level 4 16-day composite NDVI and land cover type (IGBP classification) at 1-km resolution (available at to extract the spatial/temporal distribution of crop NDVI. It has been shown that the AVHRR-derived and MODIS-derived NDVI are linearly correlated, with the coefficient of determination exceeding 90% [Gallo et al., 2004]. The MODIS data are much improved over AVHRR in georeferencing, atmospheric corrections and sensor calibration [Justice et al., 1998]. Although White et al.'s [1997] method for extracting the date of onset/offset was originally developed from the AVHRR-derived NDVI, here we apply it to the MODIS data. We use the MODIS NDVI data of the years 2001, 2002, and 2003 to derive the climatological plantation and harvest time. Moreover, only ideal cloud-free pixels, identified by the Quality Assurance (QA) flag of MODIS NDVI, are used since a cloud contamination of 10-day composites in a Northern Hemisphere study is 50% during spring and fall when the phenological events mostly occur [White et al., 2002]. The NDVI at the 1.0° resolution is extracted by averaging the NDVI over 1-km pixels in which the land cover type is classified as cropland. The 16-day composite at the 1.0° resolution is linearly interpolated to generate the daily NDVI and finally NDVIratio. The plantation and harvest dates are then extracted using the methodology suggested by White et al. [1997] (Figure 2).

Figure 2.

MODIS NDVI-derived dates of (a) plantation and (b) harvest of crop on 1° × 1° grids.

2.3. Atmospheric Forcing Data

[17] To drive the CLM, hourly atmospheric forcing data were derived from the daily products of the National Center for Environmental Prediction (NCEP) reanalysis data [Kalnay et al., 1996]. The daily global atmospheric forcing data, including precipitation, air temperature, air temperature range, relative humidity, wind speed, and incoming shortwave radiation, at the 2.5° resolutions were used. The 2.5° resolution data for precipitation, incoming shortwave radiation, and incoming longwave radiation were derived using interpolation from the T62 spatial resolution data. A diurnal cycle at hourly resolution is then generated from the daily data according to the empirical equations used in IBIS [Foley et al., 1996]. The length of a precipitation event is randomly determined between 4 and 24 hours, assuming it follows uniform distribution; the starting time of a precipitation event is assigned between the beginning of the day and the possible latest starting time, ensuring that a precipitation event ends before the next day, by the uniform-random number generation. Hourly temperature is derived on the basis of the daily maximum/minimum temperature and an assumed shape of diurnal cycle, which is a sinuous curve. The maximum temperature is set to occur at 2 pm. The diurnal cycle of humidity also follows the same sinuous curve as that of temperature. Daily solar radiation is interpolated into hourly solar radiation according to the cosine of the solar zenith angle.

3. Results

3.1. Model Validation

[18] The CLM is run with the predictive phenology scheme to validate our model against the MODIS-derived LAI. Global simulations are carried out at the resolution of 2.5° by 2.5° for 9 years. The first 4 years are for model spin-up, and atmospheric forcing from 1999 to 2003 is used to drive the model in the last 5 years. The comparison with observations is carried out for the monthly averaged values based on the time period 2000, 2001, and January through June 2003, for which the PFT-specific MODIS-based LAI is available from Tian et al. [2004]. Figure 3 presents the observed and predicted grid-averaged LAIs over the globe, averaged in January, April, July, and October (JAJO LAIs). The modified CLM with the phenology model reproduces the seasonality of LAI and its global pattern with reasonable accuracy. However, there are some discrepancies between the predicted and observed LAI at local-regional scales. For instance, the simulated LAIs show a late leaf senescence bias in the northeastern Eurasia (90°E–140°E, 50°N–60°N), and early leaf-on and late leaf-off biases in the eastern United States. To look at further details, we focus on the U.S. region in the following. Previous studies demonstrated a high sensitivity of seasonal precipitation prediction to land surface conditions over the United States. It is therefore important for climate models to have a well-validated predictive phenology scheme for this region, as explained in section 1.

Figure 3.

Monthly averaged observed and simulated LAI in (a) January, (b) April, (c) July, and (d) October.

[19] Over the eastern United States, where the model prediction features an early leaf onset bias and late leaf offset bias as shown in Figure 3, broadleaf trees dominates land cover (Figure 1). Photoperiod has been known to be a control of deciduous broadleaf forests at local scales [e.g., Nizinski and Saugier, 1988]. White et al. [1997] suggested that photoperiodic control and temperature control together provided an effective predictor for the time of leaf off, and the photoperiodic threshold showed no significant variation over the United States as introduced in equation (4). The photoperiodic control is expected to vary with latitude, and it has not been adequately studied at the regional level except for the United States. We therefore apply the photoperiod function by White et al. [1997] only for the broadleaf deciduous trees of the United States. First, we carry out a control simulation similar to the global runs but for the U.S. region only, at the resolution 1.0° by 1.0° (equation (2); CONTROL). We then carry out a simulation with the photoperiodic control for the broadleaf deciduous trees (equation (4); EXP_P) (see Table 2). Although White et al. [1997] suggested that photoperiod controls only the timing of leaf off, we applied that to both the leaf on and leaf off, and found that photoperiodic control does not reduce the early leaf-on bias. Figure 4 shows the observed grid-averaged LAI and the model bias in October, November, and December over the United States, where Figure 4 (left) presents the observed LAI, Figure 4 (middle) presents the difference between simulated LAI without the photoperiod impact and observations (i.e., only with the temperature impact) (CONTROL), and Figure 4 (right) presents the difference between the simulated LAI with the photoperiod impact added and observation (EXP_P). The biggest LAI difference is seen in November of CONTROL due to the late leaf-off bias, with the maximum difference exceeding 3.5. This bias is reduced in the EXP_P simulation, with the maximum being less than 2.0. Speaking of North America as a whole, introducing the photoperiod impact significantly improved the model simulation.

Figure 4.

Monthly average of observed (left) LAI and (middle and right) the LAI difference between model simulation and observations in different months during leaf off for (a) October, (b) November, and (c) December. The only difference between CONTROL (Figure 4, middle) and EXP_P (Figure 4, right) simulations is that the latter includes the impact of photoperiod.

Table 2. Lists of Simulations for the Model Validation Over the United States
ExperimentWinter Phenology RulesParameterization
CONTROLtemperature and AGDD thresholdsequation (2) with ΔT = 5
EXP_PCONTROL plus photoperiod and temperature thresholdsequation (4) with ΔT = 5
EXP_PTEXP_P plus adjusted ΔTequation (4) with ΔT = 12

[20] The modification based on the study by White et al. [1997] improved the model prediction to some degree. However, the early leaf-on and late leaf-off biases still exist. We therefore examined the relationship between the 10-day average air temperature (T10) at the leaf onset/offset and the coldest monthly temperature (Tc) over the United States. As suggested in equation (2), the 10-day average air temperature at the time of leaf onset/offset is linearly related to Tc:

equation image

where ΔT is the intercept, set to 5°C in IBIS. To evaluate this relationship, we first extract the dates of the leaf onset/offset for deciduous trees using the same methodology as we used for the crop plantation/harvest time. We then determine the value of T10 by calculating 10-day average air temperature at the time of leaf onset/offset from the daily air temperature of NCEP reanalysis data. Moreover, the pixels having AGDD less than 100 on the leaf-on time are excluded, and it is therefore possible to find the pixels where the leaf onset is regulated only by the temperature threshold. The scatterplot of T10 and Tc for the year of 2003 is presented in Figure 5. A linear regression analysis indicates that ΔT of 12°C provides the best fit in average of 2001, 2002 and 2003. Therefore we reran the simulations for the United Sates with ΔT equal to 12 in equation (4) (EXP_PT in Table 2). Figure 6 presents the LAI difference between EXP_PT and observation during leaf off. On the basis of the comparison between Figures 6 and 4, it is clear that the photoperiod impact and the recalibrated value of ΔT both lead to considerable improvement of the phenology model. Figure 7 shows the results during leaf on. Using the adjusted ΔT value reduces the early leaf-on bias substantially.

Figure 5.

Scatterplot of Tc and T10 at the time of (a) leaf on and (b) leaf off for deciduous trees over the United States in 2003. Each datum represents one grid box of 2.5° × 2.5°.

Figure 6.

Monthly averaged LAI difference between simulated and observed LAI after recalibrating the TcT10 relationship in (a) October, (b) November, and (c) December. Comparison with Figure 4 (right) reflects the model improvement due to the recalibration.

Figure 7.

Monthly average of observed (left) LAI and (middle and right) the LAI difference between model simulation and observations in different months during leaf on for (a) March and (b) April. The EXP_P (Figure 7, middle) and EXP_PT (Figure 7, right) simulations differ only in that the latter uses the recalibrated TcT10 relationship.

[21] Furthermore, crop phenology scheme over the United States is implemented as suggested in section 2. Figure 8 presents the monthly averaged predicted phenology factor for drought (D in equation (6)) in August, September and October, which features a strong gradient in the SW-NE direction over the upper Midwest. While crops in the southwestern part of the upper Midwest experience water stress, crops in the northeastern part do not. Such spatial pattern is also seen in the NDVI-derived crop harvest time (Figure 2). The harvest time lags behind drought. Drought in early fall leads to early harvest time in the southwestern part relatively to the northeastern part. The consistency in spatial pattern between Figures 2 and 8 validates our methodology to extract the crop onset/offset dates.

Figure 8.

Monthly averaged drought phenology factor (D in equation (6)) in (a) August, (b) September, and (c) October on 1° × 1° grids.

3.2. Interannual Variability of Vegetation Seasonality

[22] To investigate the impact of phenology prediction on the model simulation of land surface conditions, here we carried out two 11-year simulations over the United States at the resolution of 1.0° by 1.0°. This includes a 4-year spin-up. For the last 7 years, the atmospheric forcing from 1987 to 1993 is used. In the control run, the phenology is prescribed with the climatology of MODIS-derived LAI; in the experiment run, it is simulated with the predictive phenology scheme. These simulations are designed to investigate the impact of the response of vegetation or leaf phenology to different hydrometeorological conditions. Therefore, in the result analysis we select two contrasting years, 1988 and 1993. The U.S. Midwest experienced a severe summer drought in 1988 and flood in 1993. Crop dominates a large fraction of the land cover in this region. Figure 9 shows the averages of crop LAI in July. Simulated crop LAI is less green in 1988 than the climatology, but is greener in 1993. Qualitatively, the predictive scheme captures the interannual variability of the vegetation phenological response to environmental conditions.

Figure 9.

Monthly averaged crop LAI in July: (a) climatology based on MODIS data, (b) simulated in 1988, and (c) simulated in 1993 on 1° × 1° grids.

[23] Figure 10 shows as an example results averaged over a cropland area of 7° × 7° (39°N–46°N and 89°W–90°W) in the U.S. Midwest. In the control run, LAIs for the two different years (1988 and 1993) (Figure 10b) are prescribed to be the same as the LAI climatology. In the experiment run in which LAI is predicted by the phenology scheme, however, the land surface during summer is much less green in the drought year (1988) than in the flood year (1993). Transpiration in the drought year is much lower than in the wet year in both the control and experiment simulations (Figure 10c). For the drought year (1988), LAI in the experiment is ∼40% less in July than in the control simulation, but transpiration is only ∼10% less. In the control run, soil water stress limits transpiration from a vegetation canopy prescribed to be at full leaf display, causing a transpiration drop in summer; in the experiment run, the drought-induced leaf shedding also contributes to the transpiration drop in summer, as it does in reality among drought deciduous species like crops. The inconsistency caused by prescribing vegetation phenology in land surface model is addressed in the predictive phenology model.

Figure 10.

Monthly (a) rainfall, (b) LAI, (c) transpiration, (d) interception loss, (e) soil evaporation, (f) total evapotranspiration (latent heat), and (g) sensible heat for 1988 and 1993. Circled line is for the simulated results with the prescribed LAI (Control) and dash-dotted line for the predictive phenology (Experiment).

[24] The difference between the predicted LAI and prescribed LAI results in differences in surface water and energy budgets, and therefore differences in land-atmosphere water and energy exchanges. However, there is no significant difference in the overall evapotranspiration (ET, Figure 10f) amount between the control and experiment runs in this region primarily because the ET is limited by water. Despite the similarity in the ET, its partitioning among different components is different between the control and experiment runs. As presented in Figure 10, the severe drought in the late spring and summer caused a certain degree of plant dieback in July 1988 (Figure 10b). Consequently the interception loss (Figure 10d) and transpiration (Figure 10c) decrease, and soil evaporation (Figure 10e) increases in comparison with the control run. Since each component of the ET (soil evaporation, transpiration, and interception loss) has a different response timescale, the different partitioning among them can alter the degree of climate persistence [Scott et al., 1997]. The timescale increases from minutes and hours for interception loss to days for transpiration and soil evaporation. Therefore the overall residence time of water at the land surface differs between the experiment and the control run, which may further influence the precipitation persistence through strong land-atmosphere interactions. Moreover, the partitioning of available energy into the latent heat (Figure 10f) and sensible heat (Figure 10g) fluxes are also changed. The latent heat does not differ between the control and experiment because water is the limiting factor for the overall evapotranspiration, but the sensible heat flux differs to some degrees. Compared with the control simulation, the reduced summer LAI in the experiment for 1988 results in an increased albedo (i.e., reflectivity) and thus reduces the net radiation at the land surface. Since energy partitioning at the land surface favors latent heat over sensible heat when water is limiting, the reduced energy availability leads to less sensible heat flux. However, for the flood year 1993, the LAI increase in the experiment (compared with the control) does not significantly alter the surface energy flux partitioning. Such insensitivity is understandable: once LAI exceeds a certain threshold, albedo (therefore net radiation) is no longer sensitive to further LAI increases.

[25] Here we use off-line land models to study phenology modeling and its impact. Comparison with MODIS data indicates that the phenology model reproduce the LAI seasonality reasonably well. Compared with land models that prescribe LAI seasonality, including the predictive phenology scheme causes changes in the surface water and energy budgets. When such land models are coupled with atmosphere models, the phenology-induced changes in surface water and energy fluxes will impact the atmosphere, which then feeds back to further impact the leaf phenology. The full scope of the impact of including the predictive phenology on climate simulations will be evaluated using coupled land-atmosphere models in future studies.

4. Conclusions and Discussion

[26] In this study we first developed the predictive phenology model on the basis of previous studies of Foley et al. [1996] and White et al. [1997], and supplement this scheme with our proposed methodology to simulate crops. We then incorporate the phenology scheme into CLM and compare the model simulation with MODIS-derived LAI data. This comparison shows that our model reproduces the observation reasonably well. Moreover, using the predictive model instead of prescribing the climatological LAI results in greater physical consistency among different model components, between vegetation and atmosphere/land conditions in particular. Its impact on water and energy cycles has been shown using the 1988 drought and 1993 flood in the U.S. Midwest as examples.

[27] However, some biases still exist. In Figure 3, for example, the simulated LAIs show a late leaf senescence bias compared to observations in the northeastern Eurasia and early leaf-on and late leaf-off biases in eastern United States. Both regions are dominated by the broadleaf winter deciduous trees, whose phenology is controlled by temperature. In modeling the temperature-controlled phenology, we only distinguish the difference between tree PFTs and herbaceous PFTs. However, the winter phenology can be better parameterized with further discrimination among the different climate regimes, such as tropical, temperate and boreal climate [Botta et al., 2000]. For example, AGDD requirement and ΔT can be parameterized differently for different climate regions (see equation (2)).

[28] Our phenology model captures the seasonality of leaf display, yet the leaf onset/offset occurs more rapidly than what the satellite observation suggests. For example, the complete leaf fall of winter deciduous PFTs assumed to take 15 days in the model. While our parameterization has been originated from the site-specific phenology model, the satellite observation detects the mean phenological behavior in a grid cell. Further, our model assumes the annual maximum LAI to be constant, and only predicts leaf onset, offset and shedding. The interannual variation of maximum LAI could be simulated on the basis of the annual carbon budget as is in dynamic global vegetation models (e.g., IBIS and LSM-DGVM [Bonan et al., 2003]). Since the CLM version 2.1 does not fully simulate the carbon exchanges between soil, vegetation and atmosphere, here the interannual variation of maximum LAI are not explicitly parameterized. Moreover, our study concentrates on the vegetation variation at the seasonal timescale rather than the decadal or longer timescale as does the DGVMs. It is therefore reasonable to only include the daily vegetation phenology.

[29] This work provides a tool to study the two-way interactions between vegetation and the rest of the climate system. In our simulations, atmospheric forcings from the NCEP/NCAR are used to drive the modified CLM, and thus the simulated phenological variation of vegetation state only impacts the water and energy balances of the land surface. It does not account for the feedback from the atmosphere. In our future research, we plan to investigate the full scope of the soil-vegetation-atmosphere interactions using coupled land-atmosphere models.


[30] The authors would like to thank Yuhong Tian at Georgia Institute of Technology for providing the PFT-specific MODIS vegetation data. The authors also thank Samuel Levis at NCAR and two anonymous reviewers for their helpful comments on an earlier version of this paper. This work is funded by NOAA GAPP program (NA03OAR4310080).