Science Objectives

Our research effort links three Objectives. The First Objective focuses on the biogeochemistry of terrestrial systems under the forcing of natural climate variability. Among the topics of interest are the natural pattern of variability in net primary production, respiration, nutrient availability, and the flux of trace gases between terrestrial (including seasonally flooded) ecosystems and the atmosphere. Soil moisture and inundation are also important components of this research because of their controls on trace gas dynamics. In Objective Two, we address the problem of modeling the transient dynamics of human-altered terrestrial ecosystems, including agricultural systems, stages of succession, and the combined forcing of various landuse and climate change patterns. The Third Objective both supports and synthesizes the preceding Objectives. It provides a Geographical Information System (GIS) framework for model development, evaluation, and application. Within the model evaluation theme we will exploit remote sensing, using AVHRR and when available, MODIS and MISR data from EOS AM-1 as well as a formal error analysis based upon spatial-temporal statistical techniques. We will use the linked model to explore impacts of landuse and climate change scenarios on the biogeochemistry of the region.

III.1. OBJECTIVE ONE.

To describe key basin-wide biogeochemical processes and resultant states in terrestrial ecosystems as forced by natural climate variability.

Task III.1.1 focuses on improving the current TEM/WBM in order to address issues specific to Amazonian biogeochemistry (phosphorus, deep rooted trees, the role of seasonally flooded terrestrial ecosystems). These improvements will also establish the foundation for simulating succession (see Objective Two) and fluxes of trace gases (CH4 and N2O) from natural systems (Task III.1.2). We will refine the spatial and temporal scale of our models: the spatial scale will be reduced from a 0.5o grid to a uniform 8 km grid; the temporal scale will be dropped from monthly means (with statistical climatology for submonthly dynamics) to daily means.

III.1.1. Task. Modify the Terrestrial Ecosystem Model (TEM) to improve the simulation of carbon, nutrient, and water dynamics in the Amazon Basin.

Phosphorus. Much of the Amazon Basin is covered with phosphorus- deficient soils16 yet TEM does not consider phosphorus dynamics and the potential limitation of NPP by phosphorus in the Amazon basin.17 In our past studies, this potential problem was partially overcome by our calibration technique during which parameter values were defined to correctly predict NPP under equilibrium conditions in an Amazon forest growing on phosphorus-deficient soils. However, more fertile soils do occur in the Amazon basin such that NPP might be underestimated in certain regions (e.g., Rondonia). To simulate phosphorus limitation of NPP, we will use an approach that is similar to that used by TEM to simulate nitrogen limitation of NPP.18 The model will also constrain changes in carbon stocks in vegetation to maintain C:P ratios that are representative for mature vegetation found in the Amazon Basin and include occlusion of phosphorus by the sesquioxide clays found in oxisols and ultisols (Fig. 8).

Deep Rooted Systems. It has been estimated that half of the closed tropical forests in Amazonia have deep roots (> 8 m) that may allow NPP to be sustained during seasonally dry periods.19 In a simple preliminary analysis, we found that annual NPP of tropical forests estimated by TEM increased when tropical trees were able to utilize water below the 2.5 meter rooting depth normally defined in the water balance submodel (WBM).20 In addition, we found that the simulated timing of seasonal net carbon exchanges between the atmosphere and tropical ecosystems changed by up to six months when deep roots were included. However, without deep roots, WBM yielded annual water budgets consistent with a 5-year time series of measured discharge upstream of Obidos9; thus a critical inconsistency between model and field experiments is apparent and needs to be addressed. To explore the potential influence of deep roots on carbon and nutrient dynamics in the Amazon Basin, WBM will be modified such that soil moisture will be represented by more than one reservoir and TEM will be modified to increase rooting depth. We will compare seasonal estimates of NEP by TEM to net ecosystem exchange (NEE) data collected at eddy flux towers at LBA intensive sites throughout the Amazon Basin to help assess the importance of deep roots on seasonal carbon dynamics.

Wet-TEM. Currently, TEM estimates fluxes and pools of C and N in floodplains and wetlands using parameterizations of a comparable upland ecosystem type. These estimates do not include the unique biogeochemical dynamics that occur under anaerobic conditions. In addition, the current soil water model does not simulate the conditions of saturation under which anaerobiosis develops. We propose to simulate three mechanisms by which waterlogging of soils can occur: from precipitation in excess of infiltration, from the redistribution of soil water to low-lying areas and from seasonal floodplain inundation.

The Water Balance and Transport Model (WBM/WTM)9,25,21 will be improved upon in support of the proposed terrestrial modeling effort. The model will be calibrated using parameters appropriate at the defined scale22 and verified at the basin and sub-basin scale using observed hydrographs. We will employ two broad approaches: using 1) station-based interpolations of climate forcings provided through LBA and 2) an aerological approach using convergence/divergence fields (e.g, derived from the CPTEC meso- scale model or NCEP reanalysis products). A key aim in this redundant approach is to highlight inconsistencies and develop error estimates on derived water budget estimates which then can be interpreted as potential errors in ecosystem model outputs. The improved regional model (R-WBM/WTM) will maintain an explicit daily timestep using LBA meteorological data (the current version uses monthly steps with a statistical weather generator). It will include deep roots, a physically-based evapotranspiration function considering both canopy and soil water fluxes,23 variable infiltration capacities, and computation of waterlogging.

The model will also simulate sub-grid scale wetting to better define the extent and temporal pattern of soil moisture and waterlogging, of direct importance in predicting trace gas emissions as well as the spatial distribution of wetland ecosystems. Famiglietti and Wood24 showed that including sub-grid variability in larger-scale modeling could be done by identifying dominant process controls on the water and energy balance and then representing the spatial variability of these important controls in a statistical- dynamical framework.25 They identified a threshold scale (1-2 km2) for runoff and energy balance modeling. At larger scales the statistical-dynamical model formulation yielded minimally-biased simulation results as compared to more detailed spatially- distributed simulations. Subsequently, they proposed a general mesoscale model formulation which aggregates a simplified soil- vegetation-atmosphere transfer scheme (SVATs) with respect to a statistical distribution of topographic and soil properties.26 We propose to use the resulting mesoscale hydrological model as a land surface parameterization that will help determine the distribution of wet habitats by predicting the downslope redistribution of soil water within each of our 8-km grid-cells (J. Famiglietti letter of support in Section J). We propose to use this approach at the basin scale using a 1 km sub-grid resolution digital elevation model (DEM)27 and, if available, finer resolution data in the LBA meso-scale study areas as test sites.

In the Amazon, floodplain inundation of lowland ecosystems can last for more than six months each year and at its maximum extent cover more than 150,000 km2 along the mainstem and major tributaries.28 The floodplain thus has a substantial impact on the nature of mainstem and tributary hydrographs.8,25,29 Because the water exchanging with flooplains can be generated from several 100's to potentially 1000's of km upstream we require a linked water budget and water transport model (WBM/WTM). Computed runoff from single grid elements of the WBM/WTM will be routed using a set of simultaneous differential equations organized through a river networking scheme. We will employ routines already developed by our IDS-funded Global Hydrological Archive and Analysis System (GHAAS). Based on its proven utility in the Amazon9 we will employ a quasi- linear cascade model for river flows that will include wetland inundation. The WTM will generate discharge at any point within a simulated drainage basin, and any water fluxes associated with explicit flooding in those river reaches associated with wetlands.30

Previous versions of this model applied a procedure to optimize the choice of inundation parameters. An alternative approach to identifying floodplains for diagnostic studies is to use information from passive microwave satellite sensors such as SMMR or SSM/I. Studies exploring the use of 25-km resolution Nimbus-7 SMMR 37-GHz polarization temperature differences showed that a time series of inundation in the central Amazon could be remotely sensed.31 A recently developed inundation area time series from 1978-89 for twelve reaches along the mainstem Amazon used mixing models that consider end-members of major landscape units (i.e. closed forest, open water, flooded wetlands) that have distinct microwave emission signatures.32 By using a DEM32 in conjunction with stage information33 it will be possible to determine volumes of entrained water. We are now analyzing these volumetric estimates for direct input into the fluvial transport module of WBM/WTM to predict the onset, duration, and intensity of floodplain storage. We propose to extend the analysis, in collaboration with our U. California at Santa Barbara (UCSB) colleagues, into the contemporary and LBA operational time frames using SSM/I sensor data, additional stage, topographic information, and wetland mapping products available during LBA.34

III.1.2. Task. To develop maps of CH4 and N2O fluxes from natural ecosystems in the Amazon basin.

Methane. Tropical wetlands (including seasonally flooded systems) contribute about half the global wetland flux of CH4 to the atmosphere.35 We will focus our efforts on quantifying fluxes from seasonally inundated floodplains by addressing the key uncertainties of area and duration of inundation. These calculations will come from the WBM/WTM (see Task III.1.1). We will calculate floodplain methane fluxes based on inundation area and TEM ecosystem productivity for the flooded ecosystem. A linear correlation between methane production and net ecosystem productivity (NEP) has been demonstrated,36 based on simultaneous, mid-day measurements. We will modify this relationship to use simulated productivity and decomposition as the driver of methane flux, based on published flux data and model simulations. An emission factor for open water will relate methane flux to water depth,37 which we will relate to river stage from WBM/WTM. Reported mean methane fluxes from Amazonian floodplains are ~200 mg CH4 m-2 d-1 from the grass mats, ~100 mg CH4 m-2 d-1 from the flooded forest, and ~50 mg CH4 m-2 d-1 from open water.40 Similar flux rates were found both during the rising floodplain water table (late wet season) and falling water table (mid-dry season) with no significant difference in floodplain flux rates across the basin.38 Highest methane fluxes were observed at low water levels for the open water and during rising and lowering water levels for the flooded forest and floating mat.39 Methane flux for non-floodplain wetlands in the basin will also be calculated using the productivity-flux relationship of Whiting and Chanton40 and wet- TEM productivity simulations.

We will model CH4 uptake as a maximum rate (based on published data from Amazonia and LBA field data) modified by a soil moisture scalar derived from the WBM/WTM. This approach is appropriate because methane uptake by upland soils40 is remarkably consistent across the landscape, with uniformly low rates of ~0-5 mg CH4 m-2 d- 1. The flux depends primarily on soil diffusivity, which correlates with soil moisture content (higher water content implies lower diffusivity).

We will include methane flux from fire in our regional analysis by assigning an emission factor to areas burned.41 Total CH4 flux per unit area from a forest fire is very roughly equivalent to the annual flux from a tropical wetland,42 thus fire emissions should not be ingnored. We will rely on remote sensing analysis from LBA and historical records for data on locations and areas burned.

Nitrous Oxide. To generate regional maps of N2O flux, we will use TEM-simulated N-mineralization and the empirical relationship between nitrogen cycling rate and nitrous oxide flux for tropical forest ecosystems.43 Highest fluxes are reported in pasture, then upland forest and clearings, and lowest in varzea (flooded forest),44 consistent with the cited finding on N mineralization. Wet season nitrous oxide fluxes in the cerrado is uniformly low, even following burning, and dry season fluxes are probably also low.45 The study by Nobre50 concluded that nitrification was the major source of N2O from the cerrado, that the soils were nitrogen-limited, and that only cultivated, fertilized agriculture would generate large N2O fluxes from this region. Maps developed in this manner (an approach to that used by others in global analysis of N2O fluxes46) should be adequate for undisturbed ecosystems, but not adequate for agricultural lands (see Task III.2.3).

Through our NASA EOS-IDS efforts, Changsheng Li is developing a forest version of the DNDC model with GIS capability. We will evaluate this forest version against published data and LBA field work, and apply the model in GIS mode using the data layers of landcover for the forested regions of LBA domain (Task III.3.1). DNDC12 N2O flux results will be compared to our empirical model and, along with LBA field studies, will enable identification of any other key factors (besides N-mineralization) that strongly influence N2O fluxes in regions without human disturbance.

III.2. OBJECTIVE TWO.

To develop a landuse/cover model that simulates spatial patterns and temporal dynamics of landuse and landcover change.

Task III.2.1 focuses on development of a landuse change model. The purpose of this effort is not to develop the definitive landuse model for Amazonia, but rather to establish the capability for running reasonable, coherent landuse scenarios as we explore the linked issues of biogeochemical cycling and sustainability through our ecosystem models. Landuse cannot be treated as a purely exogenous forcing since landuse patterns depend, in part, upon the changing patterns of ecosystem productivity, soil moisture, and nutrient availability. The coupled land-evaluation model, landuse model, and terrestrial models will make spatially-explicit estimates of 1) the rates of deforestation for agricultural (pasture and crop) and timber production, 2) abandonment of cropland and pasture, and 3) forest regrowth in Amazonia. For Task III.2.2 we will apply the DNDC model to pasture and cropland ecosystems to simulate carbon and nutrient biogeochemical and trace gas fluxes. In Task III.2.3 we will develop a successional TEM model to simulate succession following abandonment. Task III.2.4 explores an alternative approach to dynamic vegetation modeling of successional behavior based on some recent developments in the mathematical representations of ecosystem models.

To determine landuse in a grid cell, we will expand the landuse decision matrix used by GEOMOD to include price, income, trade, transport infrastructure and biogeochemical constraints. We will construct the socioeconomic sub-matrix from existing socioeconomic data at the county level. For the biogeochemical sub-matrix we will use annual production estimates for alternative landuses by the terrestrial ecosystem models: TEM or alternative approaches for natural ecosystems (mature and successional forests) and DNDC for cropland and pasture ecosystems. In addition to price and trade, the level of annual production in cropland and pasture will be used to determine when these areas are abandoned. We will also calculate a distance sub-matrix to describe proximity of a grid cell to: 1) cities and large towns, (access to major markets); 2) major roads (access to transportation); 3) surface waters (access to transportation and irrigation); and 4) existing cropland, pasture or timber sites (adjacency).

We hypothesize that the spatial patterns and temporal dynamics of landuse and landcover change are the result of a dynamic imbalance between supply and the demand of land for alternative uses. We will reconstruct the historical changes in the spatial distribution of landuse and landcover for the 1978-1990s using two levels of search and allocation of land for alternative uses. The first will be at the county (municipios) level. We will use county level tabular data to drive the landuse/cover model and derive spatial distribution of landuse/cover changes over time. The second will be at the country level, driven by country level tabular data. Iterative implementation of these two levels of analysis will allow the model to respond to a variety of driving forces (i.e. international, national and local policies) at different spatial- temporal scales.

We will use landcover data from 1978 derived from fine resolution satellite remote sensing47 as the baseline for actual landcover in Amazon. Landsat-derived landcover data in 1988 and in the 1990s will be used to evaluate the modeling effort.54 As a product we will make available a dynamic landcover data set with distinct age structure of landcover in a grid cell, usable as inputs for the biogeochemistry models to estimate the effect of landcover and landuse change on carbon and nutrient dynamics in Amazonia (Objective Three). The task will lay the foundation for addressing sustainability of landuse and landcover change in Amazonia in the context of future changes in climate and socioeconomic environment.

III.2.2. Task. Model the carbon and nutrient biogeochemistry and trace gases fluxes in human-disturbed ecosystems.

Conversion of forest to agriculture (pasture or cropland) will cause changes in soil carbon, nutrient stocks, and trace gas fluxes. In some cases, decreases in soil carbon content have been observed following conversion to pasture,48 while in another case soil carbon increased.56 Conversion to tilled agriculture generally results in a decrease in SOM.49 The net change in soil organic matter content will depend on the rates of loss of forest-derived SOM and net inputs of pasture- or crop-derived SOM, which depends in turn on decomposability of SOM and crop/pasture management practices (vegetation and litter types and amounts, fertilization rates, soil types etc.). Keller and Reiners50 found that the conversion of tropical forest (in Central America) to pasture led to an increase in N2O and methane emissions. For methane, the forests are sinks, while the pastures are weak sources, at least for some portion of the year, due to wetter soils. Wetter soils in pastures also lead to enhanced N2O fluxes, perhaps coupled with some increased N- mineralization and decreased N uptake soon after clearing. Keller and Reiners57 hypothesize that this increase lasts for ~10-15 years, based on a pasture chronosequence study. Neill et al.51 concluded from net nitrification measurements that enhanced N2O flux from Rondonia pastures might only last a few years.

We will use the DNDC model to simulate both soil carbon and nutrient biogeochemistry and N2O flux for cropland and pasture. DNDC has four primary submodels: soil climate, crop/vegetation growth, decomposition, and denitrification. It requires daily weather data, soil properties, crop type, and agricultural management as drivers. It is designed to simulate agroecosystems, and thus has algorithms in place for cultivation, fertilization, irrigation, and other agricultural practices. Tropical agriculture and pasture simulations with DNDC have been focused on Costa Rica (Fig. 5), and will be adapted to the LBA region. Crop parameters (e.g., biomass allometry, root and shoot C:N ratios, water use) have been developed for a number of crops common in the region, including maize, sugarcane, soybeans, and bananas. Parameterizations for other crops will be gathered from the literature.52 We will use vegetation classification maps and national and regional agricultural statistics to assign crops to cultivated agricultural land. We will also evaluate whether current pasture grass vegetation parameters in DNDC are appropriate for Amazonian pastures.

For land use change scenarios (e.g., forest to cropland/pasture to secondary forest), we will use TEM and DNDC. The two models differ in the details of their structure and state variables (i.e. TEM's Csoil and Nsoil pools are disaggregated in DNDC into a number of pools). However, these pools in DNDC can be initialized based on total soil organic matter and ecosystem type or history. Initialization of the more dynamic pools (e.g., nitrate and ammonium) has little impact on ecosystem behavior beyond the first month or so. A major challenge will be to generate realistic initializations of the plant litter pools left on site following deforestation. In simulations of conversion of forest to agriculture in Costa Rica, we found that matching N2O fluxes for the first month or so following deforestation was very difficult to achieve due to uncertainties and variability in the litter pools, but model and field results were in general agreement after this initial period.

Our approach will use the models in series upon conversion of a landscape element from forest to pasture/cropland. DNDC will receive the initial state of soil and litter carbon and nutrient pools from TEM. DNDC will then simulate soil and vegetation biogeochemistry while the land is in managed agriculture, and return the state of the soil, litter, and vegetation pools upon pasture abandonment, to successional TEM (see Task III.2.3). For prognostic simulations, the landuse model will be used to determine the timing and location of land conversions.

III.2.3. Task. Develop a successional TEM to simulate the regrowth of secondary vegetation after the abandonment of pastures and croplands in the Amazon Basin.

The abandonment of agricultural land is generally assumed to be caused by pest infestation or the reduction of soil fertility such that agricultural production cannot be sustained.53 Thus, fluxes and pools of carbon, nitrogen and phosphorus in secondary vegetation of these abandoned areas may or may not attain the same levels as found in "undisturbed" natural vegetation.54 The recovery of natural vegetation in abandoned areas depends upon the intensity and length of the agricultural activity and the amount of soil organic matter on the site at the time of abandonment.55

To simulate the biogeochemistry of secondary vegetation, we will continue our development of a successional TEM for application in the Amazon Basin. The successional TEM simulates biogeochemical cycles and vegetation dynamics simultaneously with information exchange occurring interactively among all the structural and functional components of an ecosystem. The model can simulate successional development of plants from either bare ground (primary succession) or sites that have been disturbed by human activities (secondary succession) based upon age-dependent patterns of NPP and biomass accumulation (Fig. 9). Simulated patterns of plant growth during secondary succession depend substantially on the status of nutrient pools inherited from the previous stage.

In this version of TEM, vegetation carbon has been divided into four pools: leaves, sapwood, heartwood and roots. The model uses a two-layer soil water model in which vegetation functional types compete for water based on their rooting depth and the allocation of biomass to roots. Plants experiencing water stress will alter the allocation of carbon to different pools such that more carbon will be allocated to roots. Processes such as reproduction, establishment and light competition have been added to the model and interact with the C, N and water cycles. Disturbance regimes such as fire are also incorporated into the model. Currently, only two functional types (temperate deciduous forests and grasslands) have been tested in the model (Fig. 10). We will expand the model parameterizations to include vegetation types found in the Amazon Basin as part of our EOS-IDS activity.

III.2.4. Task. Develop, test, and apply an alternate model to treat the successional dynamics of terrestrial systems in the Amazon Basin.

Scaling is a central problem in ecosystem ecology because large- scale patterns arise from complex dynamics at smaller scales.56 As part of this proposal, we plan to develop a dynamic vegetation model (DVM) that will bridge the gap between the scale of individual plants and the large spatial scale of highly aggregated ecosystem models (e.g., TEM). At the core of this new model framework will be a formal scaling of the performance of individual plants through the dynamics of interspecific competition on a heterogeneous landscape. Formally addressing the issue of scaling will allow us to use LBA data at a variety of spatial resolutions. In addition, it will facilitate error analysis (see Task III.3.3) by propagating distributions of errors on measured parameter values through the full model to obtain confidence limits on model predictions. Explicitly addressing the dynamics of plant competition and succession within sub-grid scale patches will make the model well suited to predict terrestrial carbon dynamics following disturbance and land use change. We propose to develop a new model with an arbitrary number of sub-grid scale patches and a continuum of biodiversity (distribution of functional types) within each grid cell to address sub-grid scale heterogeneity. The basic structure of the proposed DVM model for the dynamics of carbon in vegetation is a formal limit of a stochastic individual-based model of plant community dynamics in which plants compete for local resources. A first test of the preliminary DVM confirmed that a single simulation with DVM was equivalent to the ensemble mean behavior of many simulations with SORTIE57 using the same functional forms and parameter values (Fig. 11). Because the DVM approximation is much simpler than the corresponding full stochastic individual based model, it will scale-up to large spatial scales more readily than models that track each individual plant.

The model is a system of two partial differential equations (PDEs) for each model grid cell (Table 2). The first PDE describes the dynamics of the density of plants or carbon in plants of different types and sizes, in patches of different ages. The second PDE describes the dynamics of the density of patches of different ages. Boundary conditions include initial conditions, plant reproduction, and the creation of patches following disturbances (Table 2). The model can be extended to include multiple axes of biodiversity. It can also be extended to include qualitatively different categories of land, such as active areas of cultivation, heterogeneous soils, and topography. Kohyama has successfully used simple versions of models of this type in other regions.58 Using the same scaling algorithms as the SORTIE example above (Fig. 11), but more general functional forms for growth, mortality, fecundity, and how plants alter the resoure environment, the DVM again exhibits behavior consistent with our understanding of sucessional dynamics (Fig. 12).

A key to the success of model development along the biodiversity axis is the underlying simplicity in the strategic differences between broadly different functional types (such as the differences between grasses and trees). There are strong indications that a limited number of parameters are necessary to characterize species differences. For example, work on SORTIE illustrates that species cannot have any combination of parameter values (there are no super species and there are no nonviable species), but rather have parameters that lie on a lower dimensional surface. In another example, a principal component analysis of data on plants ranging from trees to shrubs illustrates that a single principle component explains 85% of the variation in height, root to shoot ratio, and stem wood percentage of those species.59

The development of the patch dynamics component of our model will rely heavily on disturbance, land use, and land use change data. These data will come from a variety of scales: ground based case studies to large scale remote sensing. We currently have abundant data with which to begin model development. For example, we have remote sensing data, environmental data, and ground-based data on plants and plant community structure spanning a precipitation gradient over a large region of NW Brazil (L Solarzano, personal communication). There are repeated census data60 and allometric data61 on thousands of Panamanian trees from more than 300 species. We also have ongoing efforts to build custom statistical estimators to parameterize sub-models of the life history components of individual plants. Finally, model fitting can also be used to estimate parameter values or compare functional forms for processes for which we do not currently have field data.

A key of feature of this new model is our ability to propagate measured parameter uncertainty through the model to construct confidence limits on model predictions (Task III.3.4). The parameters in our DVM will be measured (or measurable) on individual plants in the field. The uncertainty in these measurements can then be propagated through the model to give confidence estimates on model predictions.64 We plan to use remote sensing data and inversion studies obtained during LBA to test model predictions at the large scale (Task III.2.2). Data from LBA ecological field studies will be used to test model predictions at smaller scales.

III.3 OBJECTIVE THREE.

Synthesize modeling and data analysis approaches for prognostic simulations, including uncertainty estimates.

The final objective has five components: (1) link the terrestrial models into a distributed GIS framework, (2) evaluate the model performance using remote sensing, (3) perform error analysis using site data, (4) characterize the range of model uncertainty, and (5) design and execute scenarios that investigate the effects of changing climate and land-use on the biogeochemical cycles Amazonian ecosystems.

III.3.1. Task Develop and Implement a Web-based GIS-Model Integration and Synthesis Framework..

We will continue development of our GIS database structures (UNH EOS Explorer, Section II.3), investigating and implementing various GIS architectures, and giving particular consideration to emerging software developments. We will add LBA data to the system as it becomes available and cooperate fully to support the LBA "community GIS". Finally, we will link this internet-based GIS with a model simulation and evaluation framework. This linkage will facilitate scale analysis, as well as data and model evaluation. We believe that this approach, particularly when linked with other LBA data sets, will be valuable to the LBA community for planning and synthesis efforts. Our EOS Explorer has been demonstrated to the Oak Ridge National Laboratory (ORNL) DAAC, and we are currently investigating cooperative arrangements for supporting LBA data and information needs.

Focus 1. Evaluate the spatial and temporal patterns of TEM model predictions and sensitivities using remotely-sensed indices that are correlated with biogeochemical variables. We will use methods from spatial statistics and time series analysis to compare the model output (principally NPP) to empirical, satellite-based estimates from a production-efficiency model with biome-specific coefficients.62 First, we will compare the spatial patterns of our NPP predictions to NDVI-based estimates. Second, we will evaluate the temporal patterns of predicted NPP as well as the patterns of response of NPP to climate and other disturbance, including multi-year time lags.

We propose to use vegetation indices (VI) to continually evaluate patterns of model predictions by analyzing standardized differences between the modeled NPP and the RS-derived proxy. (Fig. 13) We will delineate spatial regions and temporal periods of inconsistencies between model and data. Explaining these inconsistencies will highlight areas for improvement in model calibration, process-level detail, or input data. Model calibration will be addressed specifically using remote sensing in the next focus.

We will further conduct statistical analyses to investigate the realism of the modeled response of NPP to climate. We will use spatial and spatio-temporal statistics to compare the ecosystem- level responses simulated by TEM to the responses inferred using satellite data19,63 (Fig. 7). This analysis will provide insight into the representation of model processes which can lead to lagged responses to perturbations, most notably, those mechanisms involving nutrient and water feedbacks to net carbon exchange.64

III.3.3. Task. Model Evaluation and Error Analysis.

We will conduct two kinds of model testing to examine the generality of parameters derived in the TEM calibration process. First we will apply the calibrated model to the same calibration sites using input data from 1997 and later compare TEM-derived estimates of seasonal carbon fluxes to the seasonal ecosystem exchange fluxes measured at the LBA tower sites during this period. Second, we will apply the calibrated model to input data collected at the new flux study sites in the eastern transect (Para) and the western transect (Rondonia) in the LBA study. We will compare the model-based estimates of seasonal carbon fluxes to the measured seasonal carbon fluxes of the primary forest sites in each of these two transects.

We will utilize quantitative model evaluation techniques, applicable to a linearized compartmental version of the model. In such a setting we can readily produce parameter estimates including mean prediction of fluxes and associated error bounds using nonlinear regression analysis.65 Figure 14 illustrates the steps required for calculating confidence regions for model parameters and predicted responses for a hypothetical example. Numerical algorithms and diagnostic graphical displays related to these estimation procedures are fully implemented in most comprehensive statistical software programs (S-PLUS).66 Comparing magnitudes of measured fluxes to temporal confidence bands around the model-predicted response (flux) is a powerful way of evaluating ecosystem model performance analogous to statistical process control in industrial settings. If model testing shows a reasonable agreement between the model estimates and the observed data, we will apply the calibrated models to estimate regional dynamics of seasonal carbon fluxes in Amazonia. If not, our activities will focus on additional model development.

III.3.4. Task. Uncertainty Analysis for Ecosystem Modeling.

There are two basic approaches to uncertainty propagation in prognostic analyses. First, the Monte Carlo procedure randomly samples values of the parameters from an a priori specified parameter distribution and evaluates the ranges of the responses.64 This method works in theory for complex models but becomes computationally demanding for high-dimensional parameter spaces and hence does not lend itself for regional and global scale predictions. Alternatively, a "first order variance" analysis quantifies errors during the calibration period and then extends these over the prediction period by linear approximation. First order variance analysis is analytically intractable for very complex models and typically produces bounds that are too large because observational error is accumulated in independent increments over the prediction period.67 The computational burden of sensitivity studies and uncertainty analyses associated with large scale ecosystem modeling calls for simplification in the model structure and the consideration of only the most uncertain parameters.68

Uncertainty analysis can be cast in a Bayesian statistical framework using prior distributions of the model parameters. Prior estimates can directly be obtained from the outputs of the nonlinear regression procedures (Task III.3.3) from the site specific model evaluations. By considering a simplified model (e.g., polynomial in the parameters) and defining mean output (e.g., NPP) as another parameter, its posterior distribution can be calculated for each time step using dynamic time series estimation69 or extended Kalman filtering.70 Recent advances in Bayesian computation, in particular Markov Chain Monte Carlo (MCMC) posterior calculation methods71 have promoted the application of complex Bayesian models that were previously beyond the reach of computational methods. We will use analytical and computational tools of this type to begin developing error fields associated with prospective modeling studies. These tools successfully avoid the computational burden of numerous Monte Carlo runs of the model by accounting for uncertainty in each time step via a Bayesian prior-posterior calculation. Furthermore MCMC calculations are a tool for combining sensitivities of model outputs with respect to individual parameters into a single measure of uncertainty. In addition the spatio-temporal analyses of present model-based NPP and of associated satellite observations (Section III.3.2) will provide supplemental information in the effort of prognostic uncertainty calculations at the larger spatial scale of the entire basin.