Sensitivity of the NCEP/Noah land surface model to the MODIS green vegetation fraction data set



[1] Land surface processes are strongly controlled by vegetation cover. Current land surface models represent vegetation as a combination of leaf area index (LAI) and green vegetation fraction (GVF) parameters. The purpose of the study is to examine the impact of a spatially and temporally detailed Moderate Resolution Imaging Spectroradiometer (MODIS)-based GVF on surface processes in the NCEP Noah land surface model. The largest differences between the GVF data set currently used by the Noah model and the new MODIS GVF data set occur in winter and for tree-dominated vegetation classes. The greatest impact of the new GVF data on the surface energy and water balance is seen during the summer, when the transpiration is increased by more than 10 W/m2 on average for most vegetation types and the July averaged daily transpiration rate is increased by up to 50 W/m2 for evergreen needleleaf sites.

1. Introduction

[2] The horizontal distribution and vertical thickness of vegetation canopies and their seasonal variation play a central role in the atmosphere-land exchanges of energy, water, and trace gases. Many land surface models represent the vegetation layer by a spatially variable but temporally fixed fractional vegetation cover and a spatially and temporally varying leaf area index (LAI). Yet some land models let both green vegetation fraction cover (GVF) and LAI vary simultaneously in time. Still others, like Noah, use a third approach of applying a spatially and temporally varying GVF but a temporally-constant universal value of LAI, following the methodology and reasons presented in Gutman and Ignatov [1998]. In the National Land Data Assimilation System (NLDAS) presented by Mitchell et al. [2004], of the three land models of NLDAS that carried explicit vegetation treatment, one let both GVF and LAI vary temporally (Mosaic), one let only LAI vary temporally (VIC), and one let only GVF vary temporally (Noah). Hence, all three approaches were represented.

[3] The global GVF data set currently used in the Noah model was derived by Gutman and Ignatov [1998] using the normalized difference vegetation index (NDVI) data for five years at 0.144° resolution from the NOAA Advanced Very High Resolution Radiometer (AVHRR). For the current study, the NLDAS domain (i.e., continental United States, or CONUS), consisting of 464 zonal and 224 meridional 1/8° (0.125°) points, is used. The AVHRR data have been spatially interpolated to this grid. For the CONUS, annual GVF cycles for representative locations with the seven dominant vegetation types are given in Figure 1. One deficiency of the AVHRR-based data set is that its seasonality is too strong over evergreen needleleaf forests, for which at some evergreen needleleaf locations the GVF becomes as small as 0.0–0.1 in winter, while reaching as high as 0.7–0.8 in summer.

Figure 1.

Annual GVF Cycle for both old (dashed) and new (solid) data sets averaged over all pixels of specified vegetation type within a 2-degree box centered at the given latitude and longitude. Vegetation Type 2: Dryland Cropland and Pasture; Type 5: Crop/Grassland Mosaic; Type 7: Grassland; Type 8: Shrubland; Type 11: Deciduous Broadleaf; Type 14: Evergreen Needleleaf; Type 15: Mixed Forest.

[4] In this preliminary study, a global 1-minute (∼2 km) GVF data set is derived using the 16-day NDVI data obtained during the year 2002 from the Moderate Resolution Imaging Spectroradiometer (MODIS) and compared with the current GVF data as used in Noah. The sensitivity of an offline Noah land model to this data set over the CONUS is also evaluated.

2. Derivation of the MODIS-Based GVF

[5] GVF can be derived from

equation image

where N denotes NDVI at each pixel for a given period from AVHRR or MODIS, Ns is the NDVI value over a sparsely vegetated or barren vegetation class, and Nv is the NDVI value corresponding to full vegetation cover (i.e., GVF = 1). Gutman and Ignatov [1998] used this equation and the assumption of a spatially and temporally constant Nv and Ns in their derivation of GVF, as mentioned above.

[6] In contrast, the current study adopts the method of Zeng et al. [2000] in which Ns is a global constant within each year but varies between years, and Nv is a function of vegetation type and varies from year to year. To compute Ns and the spatially variable Nv, the annual maximum MODIS NDVI at each pixel, Nmax, is found. Then, the histogram of Nmax for each vegetation type around the globe is computed. Following Zeng et al. [2000], Ns for the barren land cover type is taken as the 5th percentile of Nmax, while Nv is taken as the 75th or the 90th percentiles for most non-sparse vegetation types. For some vegetation types that never have a dense vegetation cover (e.g., barren), Nv is taken from similar vegetation types with known measurements. Overall, the results are insensitive to the exact percentile chosen [Zeng et al., 2000]. The method is also shown in Zeng et al. [2003] to remain robust against the impact of spurious interannual NDVI on the computation of vegetation fraction data sets spanning multiple years.

[7] The original MODIS NDVI data [Huete et al., 1994] are discontinuous in both space and time because of cloud contamination. Moody et al. [2005] have developed a spatially and temporally complete data set of MODIS NDVI at 1-minute (∼2 km) resolution. They employed a technique for filling missing pixels with the phenological curves taken from non-missing pixels nearby that have an identical or similar ecosystem class as the missing pixel. For this study, MODIS-based NDVI from the year 2002 and USGS-based ecosystem land use data from the year 2000 were used, both available for download (

[8] To compare the above new GVF data with the original AVHRR data over the CONUS, the new data were spatially aggregated to 0.125° and temporally interpolated to 12 monthly values for each grid cell (see auxiliary material). Figure 1 shows the annual cycle of GVF from both old and new data sets for the seven dominant vegetation types in the CONUS. To obtain the curves, the 2-degree box within the CONUS that had the highest count (at least 92% of pixels in the box) of the chosen vegetation type is found. Then the annual cycle of GVF averaged over all pixels of that type within the 2-degree box is calculated.

[9] In general, the new MODIS-based GVF is higher than the old AVHRR-based data. This is seen in whole CONUS plots of vegetation class averages (not shown) as well as the geographically focused plots in Figure 1, which show more distinct GVF cycles. The difference in data sets is most pronounced during the winter months and for tree-dominated vegetation classes such as evergreen needleleaf and mixed forest. As mentioned earlier, the wintertime GVF for evergreen needleleaf trees is too low in the old data set. It is likely that the averaged summertime GVF (less than 0.8) for other tree-containing vegetation types is too small as well. The new data set provides a higher summertime GVF and weaker seasonal cycle of GVF for the evergreen needleleaf vegetation class. In addition, the averaged GVF for grassland in the old data set is too small because of the use of a global constant Nv [see Zeng et al., 2003]. For instance, Zeng et al. [2000] showed, using Landsat images and ground-based surveys, that Nv for grass is much smaller than that for trees. The new data set gives a higher and more realistic GVF for grassland as well. On the other hand, the wintertime MODIS-based GVF is too high for deciduous and mixed forests, grasslands, and croplands in Figure 1, primarily caused by the relatively high MODIS NDVI values from Moody et al. [2005], suggesting the need to further improve the wintertime MODIS NDVI data for these types.

3. Noah Sensitivity to the New GVF Data Set Over the CONUS

[10] To perform the preliminary sensitivity tests, the Noah land model was run using the old and new GVF data sets, hereafter called OLD and NEW, respectively, in the uncoupled mode of the NLDAS testbed described by Mitchell et al. [2004]. All other aspects of the model and boundary data remained the same between runs. Atmospheric forcing data, which were prepared as part of the NLDAS, are from October 1997 through September 1999 over the CONUS. All results shown for comparisons are from the last year of the simulation.

[11] In general, the land surface energy balance can be written as

equation image

where all upward fluxes are defined as positive. The downward solar flux, SWd, and downward longwave flux, LWd, remain the same for both runs. The reflected solar flux, SWu, also remains the same because the snow-free and maximum-snow albedos are prescribed independently of the GVF in Noah. The upward longwave flux, LWu, is directly related to calculated skin temperature (Ts), where a 1 W/m2 difference in LWu is equivalent to 0.25, 0.20, and 0.16 K in Ts for Ts = 260K, 280K, and 300K, respectively. The latent heat flux, LH, consists of three components;

equation image

where TR is the transpiration, Ec is the evaporation of precipitation intercepted by the canopy, and Eg is the bare soil evaporation. In general, since the precipitation remains the same for both OLD and NEW runs, both TR and Ec should increase proportionately with increased GVF. In contrast, Eg should decrease simply because the bare soil fraction is reduced by the higher GVF.

[12] Whether LH increases or decreases with higher GVF depends on which term in (3) is dominant. The variation of LH due to the use of the old and new GVF data sets is primarily balanced by sensible heat flux (SH) and ground heat flux (G) in (2). Ground heat flux (G) is also directly affected by GVF because of a parameterization in Noah whereby vegetation causes a shading effect and a reduction in the soil thermal conductivity [Ek et al., 2003]. The last term in (2), R, represents remaining terms of the energy balance (e.g., energy associated with snowmelt; see the Noah user's guide at Equation (2) is also used to compute the skin temperature (Ts). For a small change (e.g., 0.2 K) in Ts, the change in LWu is around 1 W/m2, depending only upon the actual Ts value (as mentioned earlier), while the change in SH and LH could be much larger or smaller, depending on environmental conditions (e.g., wind speed and atmospheric stability).

[13] Figure 2 shows the seasonal cycle of some components in (2) and (3) for all evergreen needleleaf grid cells averaged over the CONUS. Due to the new GVF being higher (Figure 1), TR in Figure 2b and Ec in Figure 2f are higher in the NEW run while Eg in Figure 2d is lower throughout the year, consistent with the above discussion.

Figure 2.

Annual Cycle of monthly-averaged variables for all evergreen needleleaf grid cells averaged over the NLDAS domain.

[14] Furthermore, TR and Ec are dominant over Eg, leading to a higher LH throughout the year from the NEW run (Figure 2c). This increase of LH is compensated for by the decrease in SH (Figure 2a) and the decrease in magnitude of G (Figure 2e). The monthly difference in LWu between NEW and OLD runs is small, varying from −0.35 W/m2 in January to −1.5 W/m2 in August (figure not shown).

[15] Figure 3 further illustrates the response of surface energy balance variables to change in GVF. All pixels of dryland cropland in the same 2-degree box as in Figure 1 are shown for the month of July, when the available surface energy peaks and the maximum differences for most components of the energy balance occur. The magnitude of the difference in each component increases nearly linearly with the increase of GVF in Figure 3. For these vegetation types as well as most others (not shown) TR exhibits the greatest sensitivity to changing GVF followed by Eg of opposite sign because of the direct effect of GVF in the computation of these components. The increase of LH is less than the TR difference because of the decrease of Eg in Figure 3, which is also similar to the results in Figure 2. This increase is largely compensated for by the decrease of SH. Overall, when averaged by vegetation types over CONUS, the changes in LWu, G, and Ec are small with the G and LWu differences within 2 W/m2 and the Ec differences within 4 W/m2 for the month of July (not shown). The flux differences are the smallest for shrubland because of its relatively small GVF and the lack of soil water for evapotranspiration in the southwestern U.S., where shrubland is concentrated.

Figure 3.

July (new–old) difference in monthly-averaged variables for all pixels of dryland cropland in the same 2-degree box as in Figure 1.

[16] Most of the components in (2) and (3) have a substantial diurnal cycle. Therefore, the effect of GVF on the surface energy balance is expected to have a strong diurnal cycle as well. As an example, Figure 4 shows that, for evergreen needleleaf trees, July-averaged differences in TR, Eg, and LH (as well as Ec, not shown) are all much larger in magnitude in the daytime. During the daytime the increase of LH in the NEW run is primarily balanced by a decrease in downward ground heat flux, G. At night the LH difference is small (Figure 4c) and the increase in downward SH (Figure 4a) in the NEW run is mostly compensated by the decrease of upward G (Figure 4e). The LWu difference between NEW and OLD runs varies from around −0.7 W/m2 during the day down to −1.8 W/m2 at night, corresponding to a nighttime cooling around 0.3K in the NEW run.

Figure 4.

July averaged diurnal cycle of all evergreen needleleaf grid cells averaged over the NLDAS domain. (a–e) Results from the NEW and OLD runs, while (f) only the difference is shown.

[17] During winter months in which the ground is fully covered by snow, LH is relatively small, and GVF does not directly affect G. Therefore, the overall impact of changing GVF on the surface energy balance is small during the winter (Figure 2). During the snowmelt period (Feb.–Apr.) LH increases and snow becomes patchy. At this point GVF can directly affect G and the overall impact of GVF on the surface energy balance becomes important again.

[18] To further show the impact of an increased GVF on energy fluxes during the snowmelt period, a grid box with evergreen needleleaf vegetation cover in the Bighorn Mountains of north-central Wyoming was chosen. Figure 5 shows the diurnal variations of energy fluxes in the last ten days of April, which correspond to the end of the snow season at this location. When snow depth (Figure 5e) is less than 0.08 m in locations of forested vegetation cover, snow fraction is estimated to be less than unity in the Noah model and GVF significantly reduces the magnitude of G (Figure 5c), as mentioned earlier.

Figure 5.

Last 10 days of the snowmelt period (April) for an evergreen needleleaf-dominated grid cell in the Bighorn Mountains of Wyoming.

[19] When the ground temperature is above freezing (or LWu is above about 315 W/m2), LH is significantly increased during the day in the NEW run (Figure 5b) because of the increase in TR and Ec allowed by higher GVF. At nighttime the G difference is dominant and the upward G is decreased in the NEW run (Figure 5c). SH in the NEW run could be higher or lower (Figure 5a), depending on the compensating effects of LH versus G. The LWu difference is small overall (Figure 5d).

[20] The new GVF data represent an improvement over the old data, because of the use of vegetation-type-dependent (rather than constant) Nv in equation (1) and the overall better quality of the MODIS NDVI data over the AVHRR data. However, the MODIS NDVI data from Moody et al. [2005] still have deficiencies for deciduous and mixed forests, grasslands, and croplands in winter. No attempt has been made to demonstrate that the new data actually improve the performance of Noah. In general, such a demonstration of improvement requires additional adjustment of the Noah model, which will be the task of future research. Also planned for future study is the derivation of multi-year GVF data using all MODIS NDVI data.


[21] This work was supported by NOAA (NA03NES4400013) and NASA (NNG06GA24G). Two anonymous reviewers are thanked for their constructive comments and suggestions.