Objective 1: In situ measurements of land cover, LAI, fAPAR, and NPP.

The plots have already been established at the orignal four BigFoot sites. For newer sites, we have established the plots using two Ashtech GG-24 Surveyor units. Plots are surveyed to within 50 cm horizontal root mean square error and plot centers are marked with a stake. Vegetation cover, LAI, fAPAR, and NPP is measured at several subplots within each plot for at least two years (see sampling design diagram and field activity schedule). LAI is measured using standard direct and optical methods at each site (Gower et al. 1999). Direct measurement approaches include periodic area harvest for non-forest sites and application of allometric equations to tree diameter data for forest sites. LAI and fAPAR are also estimated indirectly using the Li-Cor LAI-2000 Plant Canopy Analyzers (Fassnacht et al. 1994, Chen et al. 1997, Gower et al. 1997, Gower et al. 1999). Therefore, the number of LAI/ fAPAR measurement campaigns must vary among sites, as phenology of LAI development varies among biomes, among ecosystems within each biome, and between years for a given ecosystem. Consequently, we measure LAI and fAPAR three times each year at the forest sites and four to six times at other sites.

fAPAR is estimated two ways: from the DIFFN variable provided by the Li-Cor LAI-2000 Plant Canopy Analyzer (Gower et al. 1999) and from a continuous PAR tram system. We designed and successfully deployed a PAR tram at NOBS in 2000. The PAR tram measures incident and reflected PAR both above and below the canopy at small increments along a 30 m track. We will be installing a tram at AGRO, SEVI, TAPA, HARV, NOBS, and METL in 2002 for several reasons. First, the fraction of direct to diffuse PAR influences LUE (Gower et al. 1999) and this relationship varies with canopy structure. Furthermore, the continuous measurements provide more complete characterization of daily and seasonal patterns of fAPAR.

NPP (the sum of the annual biomass production of each tissue for all vegetation strata) is measured for a minimum of two years (see field activity schedule) at approximately 50 plots at all five newer sites. NPP is defined as equal to NPPW + NPPF + NPPCR + NPPFR + NPPU + NPPGC, where W = aboveground wood (e.g., stem + branches), F = foliage, CR = coarse roots, FR = fine roots, U = understory, and GC = ground cover (e.g., mosses and sphagnum). This equation is appropriate for any terrestrial ecosystem, but the field methods used to estimate each component vary among ecosystems (Gower et al. 1999). Aboveground woody biomass (e.g., stem and branch) and coarse root biomass is also estimated from allometric equations that correlate component biomass to stem diameter at breast height (1.3 m). Woody biomass increment is determined from radial growth, measured using increment cores. As tropical trees do not produce reliable annual growth rings, we are using rust-resistant dendrometer bands (Walker and Whiteaker 1988) to measure annual diameter growth. The number of tree species and size classes will be determined during the 2001 reconnaiance trip. Numerous abiotic and biotic factors have been shown to influence the allometric coefficients for new foliage biomass; therefore we are estimating new foliage production from annual leaf litterfall detritus production for forests where site- and species-specific allometric equations are not available (Gower et al. 1999). This approach assumes the canopy biomass is in steady state. Total foliage biomass and leaf area equations are from the literature. Where appropriate, biomass and leaf area data for harvested trees of the same species, but different sites, are composited and a generalized regression equation is used. NPPA of the shrub and herbaceous layers is quantified using clip plots. We also use clip plots throughout the growing season to quantify biomass production at the non-forest sites. NPPA of bryophytes, lichens, etc. are estimated using crank wires and ingrowth mesh plots (Gower et al. 1997, Bisbee et al. 2000). Fine root net primary production and mortality is estimated using minirhizotrons (Steele et al. 1997). Measurements on NPPB are restricted to the two dominant vegetation types within a site because of the large costs associated with obtaining and processing these data. Minirhizotrons are installed in each ecosystem and fine root growth is measured for two years. In the forest ecosystems, coarse root NPP is estimated from allometric equations.

Gower has experience in measuring components of the carbon budgets in grasslands and agriculture crops (Brye et al. 2000), tropical forests (Gower 1987, Gower and Vitousek 1989) and are adapting relative methodology to the new ecosystems.


Objective 2: Development of land cover and LAI surfaces.

To develop land cover, LAI, and fAPAR surfaces at any given site we are applying both general and specific sets of methods. Landsat ETM+ serves as the backbone of our remote sensing analyses. Each ETM+ image is radiometrically normalized and georeferenced. A 7 x 7 km area is extracted, linear statistical transformations and mapping decisions are developed and applied, and an error characterization performed. Multiple dates of Level 1G imagery from a given year are used and radiometric normalization commence by applying to each image the COST atmospheric correction algorithm of Chavez (1996), which converts digital counts to reflectance. The COST model is based on a simple but effective use of the dark object subtraction technique that accounts for both additive scattering and multiplicative transmmitance effects. COST uses the cosine of the solar zenith angle to approximate atmospheric transmittance and has been shown to be as accurate as models that use in situ (i.e., surface-based) atmospheric measurements and more rigorous radiative transfer code (Chavez 1996). For single-year, multi-image normalization at a given site, the image closest in date to maximum LAI is chosen as a reference and all other dates relatively normalized to it using a technique that locates the “ridge” of the two dimensional histogram formed by plotting a given ETM+ band from a subject date against that same band from the reference date. This method (conceived by R. Kennedy and W. Cohen) has been used quite effectively by Song et al. (in press). The ridge is located statistically (with the assistance of visual image inspection) and defines those pixels that have not undergone surface change. This results in a more robust relative normalization control set than is commonly obtained by selecting just a handful of pixels from the bright and dark ends of the brightness range of a given image (e.g., Hall et al. 1991). The ridge for each band of a given image pair (subject and reference) is subject to a regression analysis to calculate the normalization coefficients, which are then applied to the subject image to complete the normalization. This is done for each subject image from a given year for a given site.

Georeferencing is accomplished using the best source of reference data available. For the seven sites in the USA, the positional accuracy of the Level 1G-processed image is assessed by direct comparison with USGS digital orthophoto quadrangles (DOQs) in a 9 x 9 km area centered on the site. We found that for our original BigFoot sites, which are relatively small and free of significant topography, a small (<200 m) systematic shift in the x and y directions has been sufficient to provide a high-quality georeferencing of ETM+ to DOQs. After shifting the image into position, it is resampled to 25 m using the cubic convolution algorithm, and then clipped to a 7 x 7 km area centered on the site. The 7 x 7 km area provides a buffer that allows for subsequent alignment with a 5 x 5 km area of MODIS products. For the non-USA sites the same process is used, but IKONOS 1 m images that have been georeferenced with GPS are used instead of DOQs. This has worked well for NOBS and we expect similar results using the image being purchased via the Science Data Purchase Program for TAPA. Georeferencing in this way is done for the reference radiometric normalization image, and all images from other dates within the same year are shifted to match the reference image.

We continue mapping at all original four BigFoot sites in addition the five newer BigFoot sites. Land cover mapping relies on a combination of unsupervised classification, regression analysis, mixture modeling, and other techniques applied to the multi-date image set within a given year. Unsupervised classification is used to first stratify the scene into a single vegetation and several non-vegetation classes (e.g., water, barren, urban/built), but the process after that point is specific to the site. For example, at AGRO, we use a supervised classification to separate corn and soybean, which tend to be spectrally distinct, especially when seasonal development is captured via multi-date imagery. For forested sites, we also use an unsupervised classification to separate a forest class from non-forest classes. Then within the forested class, regression analysis is used to model percent tree cover and, if relevant, percent conifer versus hardwood. At HARV, because we use leaf-off and leaf-on imagery, we are able to identify conifer in the understory. The new sites each present a unique challenge, and our methods are tailored to the specific data sets available and information required for ecological modeling. Our work in the area of forest characterization under high LAI and biomass conditions, such as at TAPA, and in agricultural systems is current and extensive (e.g., Cohen et al. 1990, Cohen et al. in press, Lefsky et al. in press, Oetter et al. in press). For sites where low temporal frequency change is the norm, such as at the forested sites, subsequent years of land cover mapping rely on change detection. First, we determine if changes have occurred, then we label those areas that changed but carry the original label forward for those areas that have not changed. We used this procedure effectively in the Greater Yellowstone Ecosystem, which is a mix of forest, range, agriculture, and urban land use.

To map LAI and fAPAR, our primary concern is characterizing the seasonal maximum, which we do using regression analysis. Again, here we take advantage of seasonal development of spectral properties in relationship to maximum LAI. For this we rely on canonical correlation analysis (CCA, Seal 1964), which is an optimal alignment of the seasonal spectral data with an axis of LAI/fAPAR from low to high. An additional advantage of CCA is that it accomodates linear calibration, a technique that minimizes (the often significant) bias in regression model predictions when the true response variable (i.e., spectral data) is used as the independent variable in the model (Curran and Hay 1986, Snedecor and Cochran 1989). Unlike regression analysis, there can only be one independent variable in linear calibration. A vegetation index such as NDVI from a single date could be used, but the first CCA axis is superior as it weights all bands from all dates according to their contributions in predicting maximum LAI/fAPAR. We have used this method effectively at BigFoot sites and in other, independent in-progress studies. The combination of CCA and linear calibration is also used in the land cover mapping for BigFoot whenever we rely on regression modeling to provide unbiased continuous estimates of vegetation properties (e.g., percent forest cover).

Characterization of errors in our land cover, LAI, and fAPAR surfaces is critical if they are to serve as validation for MODIS. To this end, we use our reference data (field data and other ancillary information such as aerial photos) in combination with a method called cross-validation, which is similar to bootstrapping and jackknifing (Efron and Gong 1983). We have the option of collecting more reference data and will do so where feasible, but reference data are expensive to collect and process. Having data from 100 plots, we could set some proportion aside explicitly for accuracy assessment, but as the primary consideration is the development of maps of the highest possible quality, we choose to use all data to develop the maps. Cross-validation is a statistical solution to this problem (Neter et al. 1999), in that 100 separate models are developed, each time with data from 99 plots. Each model is tested on the plot that was left out, providing a nearly unbiased estimator of prediction error (Efron and Gong 1983).


Objective 3: Modeling NPP over the 5 x 5 km BigFoot footprint.

A description of the BigFoot scaling approach for NPP and its rationale are found in Reich et al. (1999). Briefly, we use ecosystem process models as our principal scaling tool. Inputs include the land cover and LAI surfaces previously described, soil data if available, and climatic variables. Model parameterization is cover-type specific (e.g. White et al. 2000). To derive an NPP surface for a given year, the model is run in each of 1600, 25 x 25 m grid cells with daily or annual outputs, which are temporally aggregated and the surfaces saved as needed. The daily climate drivers are derived from half-hourly observations at the flux towers and extrapolated to each cell if needed to account for the effects of elevation, slope and aspect (e.g., at KONZ, steep south facing slopes receive over 20% more solar radiation than the north facing slopes). Daily GPPs, at either the flux tower or spatially aggregated in the vicinity of the tower, are compared against flux tower-based GPPs. The NPP products for specific grid cells are compared with our field-measured NPP values.

In BigFoot, we use two different ecosystem models to compare water and carbon fluxes from the flux tower, and to estimate NPP for the 5 x 5 km MODLand footprint. The two models are Biome-BGC (Running and Hunt 1993) and IBIS (Foley et al. 1996, Kucharik et al. 2000). We selected Biome-BGC because it was developed specifically for application in a spatially-distributed mode in combination with satellite data (e.g., Hunt et al. 1996). Biome-BGC was also used in the development of the light use efficiency factors for the MODIS GPP algorithm, hence it is helpful in interpreting differences between the MODLand products and the BigFoot products. In addition to providing an independent assessment of NPP, IBIS was selected for several reasons. First, IBIS is an integrated ecosystem model that simulates carbon and water fluxes for terrestrial ecosystems and the output has been validated for a variety of ecosystems (Kucharik et al. 2000). The model employs multiple time steps, including an hourly time step, which allows for tighter comparisons to hourly flux estimates from the towers. For Biome-BGC comparisons, tower GPP estimates are aggregated to the daily time step. IBIS also does a complete carbon budget, so that outputs are checked directly against tower NEE. Soil respiration measurements are being made at several of our sites which provide additional information of heterotrophic respiration. Gower is already using IBIS at NOBS and CHEQ. Jon Foley, the author of IBIS, is a BigFoot collaborator, and he is using IBIS to simulate carbon and water exchange within LBA. In BigFoot we originally used PnET (Reich et al. 1999) in addition to Biome-BGC, but we have decided not to continue use of this model because it is too similar to Biome-BGC.

For 2002 and 2003 at AGRO and KONZ we will rely on new ETM+ imagery with existing algorithms to map land cover, LAI, and fAPAR. These are used with tower meteorological data to model GPP only, as there will be no new NPP measurements for these sites during those years (see field activity schedule).

For the purposes of assessing LUE algorithms, such as that used by MODLand, BigFoot will produce daily 1 km resolution data layers for PAR (photosynthetic active radiation), fAPAR, APAR (absorbed PAR), and Varepsilong (GPP efficiency factor). In each case, these data layers are initially derived at the 25 m resolution and aggregated to 1 km. PAR comes from the tower meteorological observations and DEM-based interpolations. For fAPAR, we map its distribution with ETM+ imagery for multiple dates across the growing season (described earlier). Continuous measurements of transmittance and reflectance at the flux tower help with the interpolation between the dates for which clear-sky ETM+ imagery is available. We then create a daily Varepsilong surface by dividing model-based daily GPP (checked against tower based GPP) by the daily APAR (PAR*fAPAR) just described. Thus, we produce a continuous record of spatially and temporally varying PAR, fAPAR, and APAR which can be aggregated to the 1 km grid cells and 8-day averages needed for direct comparisons with the components of the MODLAND NPP algorithm. These data could also be used with other LUE algorithms.

We can also gain insights into the daily unstressed or maximum GPP efficiency (Varepsilong*) which is used in the MODLand NPP algorithm. Looked at for all days across the growing season, the scatter plot of APAR against GPP at the daily time scale indicates the variability in Varepsilong for that vegetation type. GPP in this case could be taken directly from the tower data so that model accuracy is not an issue. The slope of the line demarking the upper limit of the scatter is the maximum GPP efficiency at the time scale relevant to the satellite-based LUE algorithms. By examining the relationship of departures from this line on any given day, and environmental factors such as maximum air temperature and daily average VPD, we evaluate the scalars for stress effects typically used in operational LUE algorithms (e.g., Goetz et al. 1999). Ultimately, these observations could become the basis for a new biome-specific parameterization scheme for Varepsilong* and the stress factor scalars. This scheme can be tested for eventual wider application using the MODLand land cover, PAR, and fAPAR products. Whether for validation or parameterization, better understanding of how Varepsilong varies across biomes and varies over the growing season within a biome is needed (Goetz and Prince 1999).


Objective 4: Validation of MODLand land cover, LAI, fAPAR, and NPP products.

There are several ways in which we validate MODLand products. The simplest and most straight-forward is direct map-to-map comparisons, and we do this for each land cover, LAI, fAPAR, and NPP BigFoot surface created. Within one 1 km2 MODIS cell, there are 1600, 625 m2 cells, so for each cell the frequency distributions of fine-grain BigFoot values can be contrasted against the single cell value of a MODIS product. This is particularly informative for land cover, and at the very least we would expect the MODLand cover class call for a given cell to be the same as the mode of the fine-grained distribution. For the numerical surfaces (e.g., LAI) we calculate the mean of the fine-grained values and compare this against the value in the cell of the coincident MODLand product. Summarizing the data across a single site, we evaluate the 25 MODLand cells in relation to the BigFoot aggregated 1 km fine-grained modes and means. These same data can be compared across sites. The MODLand GPP product is produced each 8 days, so for the purposes of comparison an 8-day 1 km BigFoot GPP product will be created.

The most basic level of confirmatory validation is that the slope of the BigFoot vs. MODLand trends across sites for the numerical products is positive and close to 1.0. The next level is that the absolute values match, and then that there is good correspondance among cells of a given site. Undoubtedly, there will be discrepancies. The MODLand surfaces can be compared directly to the field data, but this is probably only valid in the central, flux tower cell where the density of ground plots is high. However, the accuracies of BigFoot maps (characterized from comparisons with the field data) serve as confidence estimates for the quality of BigFoot surfaces as validation media.

To the degree that there are errors in our surfaces, some of the differences observed between BigFoot and MODLand surfaces will be unexplainable. Given our protocols and our ability to tailor a mapping process to a given site, however, the quality of our maps should be high enough (e.g., Figure 8d) to provide strong insights into the discrepancies between the BigFoot and MODLand NPP products. As such, we conduct several scaling exercises that involve generalization of the BigFoot surfaces. These include translation of BigFoot classes into MODLand classes, as was demonstrated by Thomlinson et al. (1999). Another is to successively coarsen the grain size of the BigFoot input images up to 1 km (e.g., Milne and Cohen 1999). With each kind and level of generalization we rerun the process models and evaluate the change in NPP output (Reich et al. 1999). If, for example, use of the 17+ classes of IGBP land cover, provides significantly better NPP estimates than use of 6+ classes of the current MODLand NPP product, a perhaps simple but effective means of improving the MODLand product will be at hand. If 250 m surfaces are required to, on average, preserve biome-specific vegetation patterns, then more reliance on the MODIS 250 m bands may be in order (Turner et al. 2000).

Other factors contribute to differences between the MODLand and BigFoot NPP products. First, both approaches use PAR, Tmin (minimum temperature), Tmax (maximum temperature), and daily average VPD. S. Running will therefore compare the BigFoot meteorological time series data for each site (derived from the flux towers) with that delivered to the MODLand algorithm by the NASA Data Assimilation Office climate model. A second factor is the spatial and temporal variability in light use efficiency. We can determine Varepsilonn (NPP/APAR) for each cover type from our NPP measurements and APAR accounting. We then run a simple Varepsilonn type LUE algorithm (Ruimy et al. 1994) with variations in the spatial resolution (25 m to 1 km) and land cover generality. These comparisons are informative about the sensitivity of LUE algorithms to spatial resolution and land cover generalization (Turner et al., in preparation).


Objective 5: Facilitating the continued development of GTOS.

GTOS has the aim of improving the quality and coverage of terrestrial ecosystem data, and integrating them into a worldwide knowledge base that will help us manage our planet wisely for future generations. A priority activity within GTOS is the Global Terrestrial Observing Network (GT-Net), which is envisaged as a "system of networks," formed by linking existing monitoring sites and networks as well as planned satellite remote sensing systems, with the aim of better understanding global and regional change. The main objective of GT-Net is to encourage existing networks, with similar objectives and geographical coverage to become more efficient in making observations, share and exchange environmental data, define data and information access policy, develop metadata standards as well as local, regional, and global in situ datasets, and undertake demonstration projects. The first demonstration project concentrates on improving current estimates of global terrestrial NPP. The project adopts a hierarchical approach and uses models that combine both satellite data and in situ observations. A set of output products, which have NPP as their common foundation, will be produced. The NPP Demonstration Project has two primary goals: distribute global standard NPP products (e.g., MODLand) to regional networks for evaluation/validation, and translate this standard product to regionally specific crop, range, and forest yield maps for land-management applications.

BigFoot is in a position to faciltate advancement of GTOS goals. Although we have not explicitly requested funds to conduct any specific GTOS activities, we think it is important to integrate our work within the context or framework of GTOS. To this end, we are collaborating with Jim Gosz, the Chair of the GTOS Steering Committee, on this proposal. Through his leadership role in LTER, Gosz supported BigFoot by funding the 1996 workshop and is now funding a second workshop in 2001 for BigFoot to bring together scientists from the International LTER community to take the first solid step towards fullfilling the goals of the GTOS NPP project. BigFoot (Turner) organized a carbon flux scaling workshop at the 2000 LTER All Scientists Meeting (ASM). The 2001 workshop is a follow-on to the ASM workshop. At this workshop participants integrate their NPP field and related data with remote sensing and models to develop NPP surfaces using BigFoot protocols. These surfaces are then tested against MODLand NPP surfaces in the same way that BigFoot is comparing surfaces. Additionally, we are updating our field manual (Campbell et al. 1999, see Appendix) for further distribution throughout the GTOS network. The BigFoot conceptual design is also shared with other interested scientists via Gower's involvement in GCTE. Gower is a member of the GCTE Science Steering Committee and the development of a global terrestrial observing systems is recognized as an important goal by this organization (Canadell et al. 1999).

Project Overview BigFoot Project Home Field Sites Data and Products Publications & Presentations Project Personnel Related Links Contact