Large Scale Ecosystem Modeling Using Parameters Derived from Imaging Spectrometer Data

Carol A. Wessman1, Brian Curtiss1 and Susan L. Ustin2
1CSES/CIRES
University of Colorado
Boulder, CO 80309-0449
 
2Department of Land, Air, and Water Resources
University of California
Davis, CA  95616

Abstract

The capability to predict the response of ecosystems to change relies on our ability to understand and model the effective functioning of biotic processes at large scales and the transport functions of the atmospheric/hydrospheric processes. To successfully evaluate changes in ecological processes at the required spatial and temporal scales, remote sensing technology and ecosystem theory must be considered jointly. A review of developments in remote: sensing analysis using high spectral resolution sensors has led to the selection of a potential set of parameters to be used in ecosystem models. These parameters quantify the light interception properties that scale from leaf to landscape. Spectral mixture analysis forms a framework for the systematic separation of both vegetative and non-vegetative components at sub-pixel spatial resolution. The spectral concentrations of the vegetative components defined by the spectral mixture analysis are then used to drive canopy radiative transfer models from which the ecosystem parameters are inferred.

1. Introduction

Linkages between ecosystems via atmospheric and hydrologic transport cycles strongly affect ecosystem dynamics over time scales of decades to centuries. Predicting ecosystem behavior and transport processes is limited by our ability to l) extrapolate biotic functioning at large spatial scales and 2) measure and model transport processes1. Evaluation of changes in ecological processes at the required spatial and temporal scales necessitates a new conceptualization of the use of remotely sensed parameters as inputs into ecosystem models. The Earth Observing System (Eos) will collect remotely sensed data at a range of spatial scales and over a much greater region of the electromagnetic spectrum than current satellite sensors. New opportunities will continue to arise to aid us-in our approach to ecosystem modeling and the remotely sensed inputs that drive them, and to consider ways to improve the robustness of model predictions by using new and refined remotely sensed ecosystem parameters. We review developments in ecosystem models designed to incorporate remote sensing data and present a new approach to model parameterization using developments from high spectral resolution measurements of surface reflectance.

Interactions among climate, edaphic and biological properties explain much of the observed variation in terrestrial biogeochemical cycling. The type of model exemplified by CENTURY2 or Linkages3 include much of what is known about regulation of terrestrial biogeochemical cycles. These models assume a paradigm in which 1) abiotic and biogeochemical factors interactively constrain primary production and carbon storage, and 2) that water and nutrient limitations (e.g., N) are not independent but correlated at steady state conditions. A foundation is established from these models for the inclusion of the terrestrial biosphere into atmospheric/hydrologic system models.

2. Remote Sensing And Ecosystem Models:

Current remote sensing approaches to monitoring ecosystem processes were pioneered by C. J. Tucker4,5 who developed the use of vegetation indices based on the difference in reflectance between the red and near infrared wavelengths. Tucker et al.6 based their carbon cycle model on the fact that there is a strong relationship between net primary productivity and annual integrated light interception by the plant canopy7. The most commonly utilized vegetation index, the normalized difference vegetation index (NDVI), is well correlated with light interception, and its annual integral with annual light interception8,9,10. Fung et al.11 assumed that NPP could be calculated from the NDVI, which defines the phasing of photosynthetic uptake of CO2, and an empirically derived biome-specific factor defining the efficiency at which carbon is fixed in biomass. Running et al.10 used NDVI with a digital terrain model to measure spatial variation in vegetation cover, which was then input into an ecosystem model to predict daily to annual photosynthesis and transpiration.

The primary deficiency in existing ecosystem models using remotely sensed inputs is the feet that landscape processes affecting image data occur at a range of spatial scales. While the majority of existing ecosystem models are parameterized on leaf-scale properties, many important ecosystem process are controlled, at least in part, by canopy properties at much larger spatial scales12. It is only through the modeling of large spatial scale canopy properties that the small scale properties can be accurately inferred. For example, light interception by the canopy is controlled not only by small scale canopy properties (ea. LAI, leaf angle distribution, leaf thickness), but is also strongly influenced by the architectural properties of the canopy at a scale of meters to tens of meters13-16. Sellers8,17 found that instantaneous NDVI could be used to estimate photosynthetic capacity. However, the NDVI varies substantially with topography and canopy geometry due to the variation in illumination, canopy cover and structure, and canopy shading. As an alternative to the NDVI, a quantitative measure of APAR, combined with other information, e.g., net radiation and surface temperature, can be used to estimate photosynthetic activity and net primary productivity18,19.

In addition to APAR measurements, quantitative determination of vegetation canopy chemistry would provide the capability for dynamic spatial modeling of biogeochemical processes such as decomposition. Changes in ecosystem processes are often expressed in the foliar chemistry as a result of altered carbon allocation patterns, metabolic processes and nutrient availability. Many of the biochemical constituents in leaves, ea. water, cellulose and lignin, have unique absorption features in the reflected solar spectrum. Because high spectral resolution datasets (ea. AIS, AVIRIS and HIRIS) have several wavelength bands spanning these features, curve-fitting techniques can be utilized to precisely calculate the amount of light absorption associated with each of these canopy constituents.20,21 In these spectroscopic analyses an absorption model assuming known concentrations of the constituents is used to match the measured curve. But as is also the case in transmission spectroscopy, just knowing the absorption coefficients and the amount of light absorption is not sufficient to infer constituent concentrations; the path length through the absorbing medium must be known a well.

As the previous examples demonstrate, it is not possible to make quantitative determinations of canopy properties: from reflectance data without first understanding how the light is absorbed, transmitted and scattered within the canopy. In order to determine canopy light interception from remotely sensed data, hemispherical absorbed photosynthetically active radiation (APAR) must be determined. This can be achieved either by integrating bidirectional APAR over the full range of viewing angles, or through the modeling of canopy architecture from a single bidirectional observation. Obviously, because of the lack of multiple view angle data-sets, the latter would provide a more satisfactory solution.

A model for describing the optical properties of a canopy must span the scales from leaf-level interactions to landscape interactions, and must do so in a hierarchical fashion. By utilizing high spatial resolution sensors (30 meter pixels), canopy roughness at crown to landscape scales can be determined directly. These large scale canopy roughness parameters can then be used to parameterize a canopy model that infers the sub-pixel scale architectural properties of the canopy.

3. A New Approach To Ecosystem Modeling

We propose that the use of spectral un-mixing techniques for analyzing high resolution image data provides a better mechanism to develop input parameters for ecosystem models22-24. Un-mixing provides a strategy to stratify the image data into fractional proportions of vegetative and non-vegetative surface classes, to identify the classes using a spectral library, and to determine both the spatial and spectral degree of fit25. One additional component of the un-mixing analysis is that an image is obtained of the shade/shadow in the scene23,24. Curve fitting or other analytical tools can be hierarchically applied to the end-member spectra to define the physical parameters of interest (Table 1), including the light path length and canopy biochemistry. We propose that the light path length through the canopy can be determined using the spatial variance in the shadow end-member image to define the dominant canopy spacing (gap size and frequency) and canopy depth, when used in conjunction with a canopy radiative transfer model. The output expected from these analyses will permit calculating the canopy biochemicals in physical abundance units.

4. Summary

Sophisticated ecosystem models driven by remote observation would permit monitoring of ecosystem dynamics at local to global scales. A growing body of theory relates light interception by plant canopies to ecosystem functioning. The high spectral resolution sensors planned for Eos will be capable of providing considerable information about light interaction within canopies, over a greatly expanded region of the electromagnetic spectrum. We now need a conceptual framework that allows us to analyze ecosystem processes using models that rely on the new physical parameters that will be available from Eos instruments. Because current ecosystem models are parameterized to use traditional types of environmental and biometeorological data, this new conceptualization requires reevaluation of both the biological/climatological processes and the qualitative and quantitative inputs from remotely sensed measurements.

We propose that spectral mixture analysis is particularly suited to the multiband image data-sets that will be produced by the Eos HIRIS and other sensors. This analysis provides a framework for systematically defining both large and small scale features in the image data. We expect that this procedure can be used to sequentially model the smaller scale spectral features of plant canopies useful for quantification of ecosystem functioning. We have provided a table of expected biophysical products that will be produced by the Eos sensors that are particularly promising for use as input parameters in ecosystem process models. Remote sensing scientists need to interact closely with meteorologists and ecologists to evaluate the new Eos sensor products and their inclusion into appropriate biosphere models.

5. Acknowledgements:

These ideas have benefited from conversations with numerous colleagues, in particular, we wish to thank John Adams, Milton Smith and Alan Gillespie for their contribution in developing the unmixing analysis techniques, without which this work could not have been done.

6. References

List of Tables

Table 1. Expected biophysical products from the Eos High Spectral Resolution Imaging Spectrometer (HIRIS) used in conjunction with other data can be used as input parameters for ecosystem process models.
Biophysical Parameters Derived from HIRIS
Example Process Models
Fractional Cover
of Vegetation
Crown Spacing
and Variance
Crown
Height
Canopy Light
Path-Length
Vegetative
Absorbed PAR
Canopy Concentration
of Canopy Lignin
Other Data
Net Primary Production
 X
 
 
 X
X 
 
Net radiation, surface
temperature, precip.,
soil texture
Decomposition
 X
 
 
 X
 
 X
Surface temperature
Evapotranspiration
 X
 X
 X
 
 
 
Net radiation, surface temperature, precip.

1998, Center for Spatial Technologies and Remote Sensing (CSTARS)
University of California, Davis