Remote sensing data analysis, forest ecosystem dynamics modeling and GIS were used together to determine the possible growth and development of a northern forest in Maine, USA. Field measurements and airborne synthetic aperture radar (SAR) data were used to produce maps of forest cover type and above ground biomass. These maps along with a soils map were used in a GIS to determine the initial conditions for forest ecosystem model simulations. Using this information and model results allowed the development of predictive maps of forest development for 25 and 50 years in the future. The results obtained were consistent with observed forest conditions and expected successional trajectories. The study demonstrated that ecosystem models may be used in a spatial context when exercised with GIS data sets.
The circumpolar boreal forest is one of the earth's major vegetative ecosystems, accounting for nearly 20% of the terrestrial plant carbon and covering one-sixth of the Earth's land surface (Bolin 1986). The northern and southern margins are especially sensitive to climate change as evidenced by the northward migration of boreal species since the end of the Wisconsin Ice Age. There has also been a notable decline in the health and vigor of spruce-fir forests throughout the northern hemisphere during the past two decades (e.g., Van Deusen et al. 1991). In addition, increasing harvesting pressures coupled with the ever present impacts of forest fires are dramatically changing the face of the boreal forest. The nature and extent of the impacts of these changes, as well as the feedbacks on global climate are not well understood, but may be addressed through modeling the interactions of the vegetation, soil, and energy components of the boreal ecosystem (Pastor and Post 1988). The boreal forest is immense in extent and, although similar in structure across the biome, has significant regional and local variability (Larsen 1980). This is especially evident in the hydrology and soils of the biome due to the effects of glacial activity during the Wisconsin Ice Age (Rourke et al. 1978). The use of combined ecosystem and remote sensing models presents an especially efficient and tractable method to study regional and global environmental changes.
The Forest Ecosystem Dynamics (FED) Project at GSFC involves the development and integration of models to understand soil, vegetation, and radiation dynamics in northern forest ecosystems. Through the use of simulation models, remote sensing, field investigation, and GIS, the vegetation, soil, and energy components within northern forests are being investigated, and their responses to global change or other disturbance are being explored and quantified. Data collection and model development has focused on northern forests of North America. Intensive field measurements coinciding with aircraft and satellite overflights, as well as ancillary data, were obtained and incorporated into a GIS. The GIS is used to provide driving variables for the models, initialize model runs, validate model predictions, and identify areas in the field requiring more intensive study. The project also incorporates a significant remote sensing modeling component to simulate expected signature response for a variety of forest simulations (see Ranson et al., 1997). This paper discusses methods and results using remote sensing images and ancillary data in a GIS to initialize and exercise a coupled forest ecosystem model. In this study models of forest dynamics and soil processes along with remotely sensed data were used to examine the growth and dynamics of a northern forest over a 250 year period. Specifically, we describe the connections of a forest succession model with a soil physics model and the application of GIS data sets including soil maps and remote sensing derived maps of forest type and biomass to initialize and test the model. This paper is organized into four sections. Section 2 (Background) provides brief descriptions of the FED modeling components and the coupled model environment, the remote sensing and field data. In section 3 (Methods) the implementation of the forest succession model, the analysis of remote sensing data and the prediction of forest dynamics are described. The results are found in section 4.
Forest Dynamics Modeling
Mathematical models that simulate forest dynamics have gained widespread acceptance and use over the past two decades. The most successful models (in terms of general applicability to diverse forest types) are individual tree-based models called gap models (Shugart et al. 1992, Botkin 1993). The strength of these models lies in their versatility to predict qualitative successional patterns related to species composition and forest structure.
The gap model, ZELIG (Urban, 1990), is an individual tree simulator that simulates the establishment, annual diameter growth, and mortality of each tree on an array of model plots. Model states are recorded in a tally of all trees on a plot, with each tree labeled by species, size (diameter), height to base of live crowns, and vigor (based on recent growth history). The competitive environment of the plot is defined by the height, leaf area, and woody biomass of each individual tree determined by allometric relationships with diameter. Plot size is defined by the primary zone of influence of a single canopy-dominant tree. The plot is considered homogeneous horizontally, but vertical heterogeneity (canopy height and height to base of crown) is simulated in some detail. Adjacent cells interact through light interception at low sun angles. Establishment and annual diameter growth is first computed under optimal (nonlimiting) conditions, and then reduced based on the constraints of available light, soil moisture, soil fertility, and temperature. Climate effects are summed across simulated months. Seedling establishment, mortality, and regeneration are computed stochastically, while the growth stage is largely deterministic. Simulations can start or stop at any point within the life cycle of a forest.
Soil Process Modeling
The goal of simulating the soil system beneath the forest is to understand the controls and feedbacks that operate within the soil as well as between the soil and the rest of the forest environment. This includes physical, biological, chemical, and mineralogical characteristics and mechanisms that vary at short-, medium-, and long-term temporal scales within soils. The FroST model which includes the physical processes occurring within the soil was used with the FED modeling framework. FroST (Frozen Soil Temperatures) is a simulation model of soil properties which produces estimates of water content, matric potential, temperature, and ice content within each soil horizon. FroST was developed from the Residue model of Bidlake et al. (1992) which couples surface residue to the soil-atmosphere system, and uses network analysis to describe heat and moisture transfer, and phase changes in water. Short-wave and long-wave radiative transfer, changes in energy status, rainfall interception, infiltration, redistribution, evaporation, and drainage are all accounted for. Climate input requirements include global short-wave radiation, air temperatures, average wind speed, and precipitation. General site, canopy, and soil characteristics for individual horizons are also needed. Enhancements to the Residue model to produce FroST included algorithms for calculating surface runoff, transpiration, Penman demand, and a simple snow model. In FroST, surface residue from the Residue model is configured to simulate above ground characteristics of forested sites. Snow is simulated by changing the characteristics of the surface soil node from soil characteristics to snow characteristics. Precipitation increases the node's thickness, and a simple melt factor is used to melt off the snow. Once the snow has melted off, node characteristics are reset to that of soil (Levine et al., 1997).
Modeling Environment
To interactively use the forest succession and soil processes the models were incorporated within a modeling environment that supports interactive configuration, manages the transfer of variables among models, and dynamically displays results (Levine et al. 1993, Knox et al. 1996). The FED modeling environment, WISE (for Workbench for Interactive Simulation of Ecosystems) allows two or more process models to be coupled using a generic query-response system where parameter values from detailed models in one discipline can be provided to drive models of other disciplines. Models are encapsulated and then run synchronously from a common external clock. Database-like features added while encapsulating each model allow models to query one another while running. Each encapsulated model also has X windows panels defining a model-specific graphical "sub-interface" and a version of a configuration tool to check parameter values entered interactively against rule sets defining allowable combinations of values. (Example WISE panels may be viewed over the Internet via World Wide Web http://fedwww.gsfc.nasa.gov). Currently, several models are encapsulated including ZELIG and FroST. With this modeling tool, scaling parameters from detailed models can be derived to improve values used in simpler models for the same parameter.
Remote Sensing
The remote sensing work in the FED Project has two major objectives. The first is to understand the sensitivities of remote sensing observations to forest structural attributes such as successional stage, biomass, LAI and forest spatial characteristics. The second component is to use remote sensing models to examine the expected signatures from forest stands with a range of the above parameters. These objectives require detailed sets of remote sensing imagery and field data and robust canopy reflectance and scattering models. We are fortunate to have both requirements well-in-hand through our earlier FED and collaborative efforts (e.g., Sun and Ranson, 1995, Ranson et al., 1997). In this paper the work involved producing maps of forest type and biomass using synthetic aperture radar SAR data (and described in detail by Ranson and Sun 1994a, 1994b, and 1997)
GIS
The Forest Ecosystem Dynamics project has been conducting research and collecting data at the International Paper Northern Experimental Forest (NEF) near Howland, Maine, USA (Figure 1.). The site is located at approximately 45o 15' N latitude and 68o 45' W longitude. The area comprises approximately 7000 ha containing several intensive experimental sites, where detailed ecological and mensuration measurements have been obtained. It contains an assortment of small plantations, multi-generation clearings, and large natural boreal-northern hardwood transition forest stands consisting of hemlock-spruce-fir, aspen-birch, and hemlock-hardwood mixtures. Topographically, the region varies from flat to gently rolling, with a maximum elevation change of less than 135 m within a 10 by 10 km study area. Due to the region's glacial history, soil drainage classes within a small area may vary widely, from excessively drained to poorly drained. This site was the focus of a NASA Multi-sensor Aircraft Campaign (MAC) for the Forest Ecosystem Dynamics project at GSFC (e.g., Goward et al. 1994) and was a backup Supersite for the recent SIR-C/XSAR missions (Ranson et al. 1997). The FED project has collected a wide variety of imagery, map data, and point data at the NEF and surrounding areas within a Geographic Information System (GIS) for research purposes. These data include field, tower, aircraft and satellite based measurements and are described and distributed from the FED project's GIS database available via World Wide Web at http://fedwww.gsfc.nasa.gov.
In this paper we demonstrate that parameter maps developed from remotely sensed data can be used to initialize and test a forest succession model. Figure 2 outlines the methods used for this paper. First, remote sensing data were analyzed for forest type and biomass levels and maps were developed. Second, a forest model (ZELIG) was coupled to the soil physics model (FroST) and run for a range of soil conditions found at the site over a 250 year period (see Weishampel et al. 1997). Individual pixels from the forest type and biomass maps along with a soil type map were compared with the results of the forest model to determine the age of the forest represented by the pixel. These age and forest conditions were used for model initialization and simulation results for 25 and 50 years in the future were recorded The individual steps of the method are described in the following sections.
Remote Sensing Analysis
Image Registration - Image registration was required to use multidate AIRSAR images. Since our research area is quite flat (maximum change in elevation of 145 m over 10 km), and the flight directions and incidence angle range of the pair of images were similar, the registration was relatively easy. A linear interpolation with about 10 control points yielded results superior to a cubic polynomial interpolation with 20 control points. The conditions during the AIRSAR flights apparently were quite stable and the distortion was linear. The AIRSAR image acquired on April 15, 1994 was registered to the AIRSAR image acquired on October 7, 1994.
Forest Type Classification - Ranson and Sun (1994a) produced a forest type map of the Howland area using AIRSAR images. They found that combining summer and winter images produced better results than using data from a single date. They used principal components analysis to reduce the number of channels used with a maximum likelihood classifier. Here, we used all non-redundant channels from both dates with a supervised minimum distance classifier.
A parallelepiped classifier (Moik, 1980), which approximates the hyperellipsoid decision boundaries of Bayesian classifier by parallelepipeds, was used in this study (see Ranson and Sun 1997) Nine land cover classes were selected from a generalized cover type map provided by International Paper: water, bog, wetland, grassland, clearing, regeneration, mixed forest, hardwood forest, and softwood forest. The latter five classes represent the state of the forest stands in the area from harvest through regrowth (regeneration) to mature "pure" or mixed stands.
The classifier was trained for the nine classes by locating areas identified from forest cover maps, aerial photos and field observations on the SAR imagery. As described above the AIRSAR image data was acquired with 12 channels (C-, L- and P-band with HH, HV, VV, VH polarizations). The set of channels for this analysis used only one cross polarization channel (VH) for each frequency. The registered forest type map was placed in the GIS for further use.
Figure 2. Conceptualized diagram for using GIS data to initialize a forest succession model. Model can then be run to determine future forest type and biomass for a given pixel.
Above Ground Biomass Mapping - Forest stands, measured during 1992 and 1994 were located on AIRSAR images and 3X3 block of pixels were extracted from which the average backscatter was calculated. The field biomass data was acquired from stands of predominantly spruce and hemlock and mixtures of hemlock and hardwood species.
An earlier analysis of AIRSAR data over the Maine study area estimated above ground standing biomass from a combination of radar channels (Ranson and Sun 1994b). A combination of P-band HV and C-band HV (i.e. PHV-CHV dB) was found to have the best sensitivity to total above ground biomass. Briefly, the procedure involves developing a linear regression equation with biomass and SAR backscatter. We used a cube root transformation for the dependent variable (biomass) and combined SAR channels as the independent variable. The cube root transformation equalizes the variance and produces a normal distribution of the biomass data (e.g. Ranson et al., 1997). The relationships between SAR backscatter values and the cubic root of forest biomass were determined using linear regression. Measurements from 17 homogeneous stands large enough to provide representative radar signatures were used to develop the regression model. An additional 28 stands were used for testing. The equation using a combination of AIRSAR bands (i.e., PHV-CHV) was:
Forest Model Implementation
The forest model ZELIG (Urban 1990) adapted as described in Levine et al. (1993), was used to simulate the successional dynamics of the southern boreal/northern hardwood forest transition zone found at the NEF. Because soil moisture is considered to be of primary importance in determining the structure (e.g., biomass and species composition) of these forests (Bonan and Shugart 1989), waterlogging effects (adapted from Botkin 1993) were included. This required connecting the ZELIG model to a soil physics model that simulates depth to saturated soils (see Weishampel et al. 1997). Diameter at breast height (dbh), height, height to base of crown, and foliage density) were recorded for each individual tree in nine (10 x 10 m) ZELIG plots.
To implement the model, site parameters (e.g., soil fertility and monthly values of temperature and precipitation) and autecological parameters (e.g., height and diameter maxima and growth tolerances) were derived from empirical data and published sources (e.g., Pastor and Post 1985, Botkin 1993). Forest succession on ten soil types (see Table 1) found at the NEF were simulated starting from bare ground. Depth to water saturated soil and water holding capacity of the soil were derived using the FroST model (Levine et al. 1995). The spatial scale of the ZELIG simulations were performed at a resolution (30 x 30 m) to correspond to the scale of typical remotely sensed data. Model results were recorded at five year intervals up to 100 years and at 50 year intervals through 250 years. Because gap models possess underlying stochasticity in their regeneration, mortality, and weather routines, fifteen separate runs were performed to generate a range of stand responses from which stand averages were calculated.
Biomass for simulated trees was calculated from modeled dbh using allometric equations developed for central Maine, USA forests (Young et al., 1980). The average biomass was then determined for the simulated 30 m plots.
Remote Sensing Analysis
Forest type classification results for the AIRSAR data were very good when compared with field information (Ranson and Sun, 1997a). Briefly, all non-forest classed were 100% correctly identified. Forest type classification for softwood was 94% correct. Hardwood showed 86.5% correct classification and mixed forest had 84.0% correct classifications.
Biomass classification results were also consistent with ground observations. Comparing biomass predicted with Eq. 1 against the 28 field measurements resulted in the equation:
We found that the biomass method worked best for biomass values below 15 kg/m2.
Forest Modeling
Figure 3 presents the average biomass trajectories simulated from fifteen 30 x 30 m (900 m2) stands growing on three of the ten NEF soil types used in the simulation over the 250 years. The range of biomass simulations illustrate the importance of considering soil types in our study area. Generally, better drained or mesic soils produced higher biomass values in less time, whereas poorly drained soils required much more time to establish maximum biomass values. Because of the stochastic timing of tree birth and death replicate simulation can exhibit considerable variation about these averages. The simulated biomass trends and the underlying patterns of dominance by needle leaf (conifer) and broadleaf (deciduous) species were consistent with field measurements reported by Ranson and Sun (1994) and Levine et al. (1994). A total of 3450 simulations were conducted (10 soil types X 15 replications X 23 time steps).
Using the forest type, biomass and soil maps to initialize the forest succession model produced the images shown in Figure 4. The maps represent an area where detailed soils information was available in the study area, including the small isolated area to the east. Maps of forest type and biomass produced from model runs for 25 and 50 years beyond initialization are also shown in Figure 4. The initialization results are consistent with the remote sensing images (not shown) indicating that this is a viable method for simulating forest dynamics in this study area.
Figure 3. Mean biomass of forest stands growing over different soil types for a 250 year simulation.
Predictive Images
The predictive images can be used to assess forest dynamics as the change in forest type and biomass over time. Assuming that no areas of forest are harvested during the 50 year period (although this can included in the analysis) the forest can be expected to develop under current climate conditions as shown in Table 1. Briefly, average biomass increases slowly over the 50 year period with a gradual decrease in standard deviation. These results indicate that the forest is mostly mature slowly growing stands. The change in forest types indicate that the hardwood stands change into mixtures of hardwood and softwood or into softwood stands. This is consistent with observations of the forest that the forest types of newly disturbed areas generally start out as hardwood or mixtures then develop into softwoods. Of course, detailed analyses of individual stands are required to confirm that these observations are valid at the local level .
This paper describes a procedure to use forest type and biomass maps
developed from remote sensing data to initialize a forest growth model
for a northern forest in Maine, USA. The remotely sensed forest attributes
were used along with a soils map to determine the stand age at which to
begin model simulation runs. This approach enabled the development of predictive
maps of forest type and biomass for 25 and 50 years in the future. Longer
periods can be examined by extending the duration of the model runs.
| FOREST COVER | ||
| YEAR 0 | YEAR 25 | YEAR 50 |
 |
 |
 |
| BIOMASS | ||
| YEAR 0 | YEAR 25 | YEAR 50 |
 |
 |
 |
| Simulation Year 0 | Simulation Year 25 | Simulation Year 50 | |
| Average Biomass (kg/m2) | 13.1 (8.9) | 14.6 (7.6) | 15.6 (6.8) |
| Percent Forest Type | Simulation Year 0 | Simulation Year 25 | Simulation Year 50 |
| Softwood | 67.8% | 73.5% | 76.4% |
| Mixed Forest | 10.1% | 7.2% | 8.5% |
| Hardwood | 21.3% | 18.5% | 14.3% |
ACKNOWLEDGMENTS
This work is supported in part by NASA Mission to Planet Earth through the Terrestrial Ecology Program (Forest Ecosystems Dynamics) and Solid Earth and Natural Hazards Program (SIR-C/XSAR) at NASA Headquarters.
REFERENCES
Bidlake, W.R., G.S. Campbell, R.I. Papendick, and R.F. Cullum. (1992), Seed-zone temperature and moisture conditions under conventional and no-tillage in Alaska. Soil Sci. Soc. Am. J. 56:1904-1910.
Bolin, B. (1986), How much CO2 will remain in the atmosphere. In, B. Bolin, B.R. Doos, J. Jager and R.A Warrick (eds.) The Greenhouse Effect, Climate Change and Ecosystems. (SCOPE 29),
John Wiley, Chichester. Bonan, G.B. and H.H. Shugart. (1989), Environmental factors and ecological processes in boreal forests, Annual Reviews of Ecology and Systematics 20:1-28.
Botkin, D. B. (1993), Forest dynamics: an ecological model, Oxford Univ. Press, New York.
Friend, A.D., A.K. Stevens, R.G. Knox, and M.G.R. Cannell. (1996), A process-based, terrestrial biosphere model of ecosystem dynamics (Hybrid 3.0) Ecological Modeling (in press).
Goward, S.N., D.L. Williams and D.L. Peterson. (1994), NASA multisensor aircraft campaigns for the study of forest ecosystems. Remote Sensing of Environment 47:107-108.
Knox, R.G., V. Kalb, E.R. Levine, and U. Bindingnavle (1996), A framework for integrating models to simulate forest ecosystem dynamics, Proc. of Third International Conference/Workshop on Integrating GIS and Environmental Modeling.
Knox, R.G. and A.D. Friend. (1997), Vegetation dynamics from physiological and functional principles and a resulting paradox in mechanistic reduction. (in press).
Larsen, J.A. (1980). The Boreal Ecosystem, Academic Press, New York. 500 pp.
Levine, E. R., K. J. Ranson, J. A. Smith, D. L. Williams, R. G. Knox, H. H. Shugart, D. L. Urban, and W. T. Lawrence. (1993), Forest ecosystem dynamics: linking forest succession, soil process and radiation models, Ecological Modeling 65:199-219.
Levine, E. R., R. G. Knox, and W. Lawrence. (1994), Relationships between soil properties and vegetation at the Northern Experimental Forest, Howland, Maine. Remote Sensing of Environment, 47:231-241.
Levine, E. R. and R. G. Knox. (1997), Modeling soil temperature and snow dynamics in northern forests. Journal Geophysical Research (in press).
Moik, J. G. (1980), Digital processing of remotely sensed images, NASA SP-431, Science and Technical Information Branch, NASA, 327 pp.
Pastor, J. and W. M. Post. (1985), Development of a linked forest productivity-soil process model, ORNL/TM-9519, Oak Ridge National Laboratory, Oak Ridge, TN.
Pastor, J, and W. M. Post. (1988), Response of northern forests to CO2 induced climate change. Nature 334: 55-58.
Ranson, K.J. and G. Sun, (1994a), Northern forest classification using temporal multifrequency and multipolarimetric SAR images, Remote Sensing Environ. 47(2):142-153
Ranson, K.J. and G. Sun. (1994b), Mapping biomass for a northern forest using multifrequency SAR data, IEEE Trans. Geosci. Remote Sens. 32:388-396.
Ranson,K. J. and G. Sun. (1997) An evaluation of AIRSAR and SIR-C/X-SAR data for estimating northern forest attributes, Remote Sensing of Environment 59:203-222.
Ranson, K.J. , J.A. Smith, G. Sun, J.F. Weishampel and R.G. Knox. (1997), Forest biomass from combined optical and microwave modeling and measurements, Remote Sensing of Environment, 59:118-133.
Rourke, R.V., J.A. Ferwerda, and K.J. LaFlamme. 1978. The Soils of Maine. Life Sciences and Agric. Exp. Sta., Misc. Report 203, Univ. of Maine, Orono, ME.
Shugart, H. H., T. M. Smith, and W. M. Post. (1992), The potential for application of individual-based simulation models for assessing the effects of global change. Annual Review of Ecology and Systematics 23:15-38.
Sun, G. and K.J. Ranson, (1995), A three-dimensional radar backscatter model for forest canopies, IEEE Trans. Geosci. Remote Sens. 33:372-382.
Urban, D.L. (1990), A versatile model to simulate forest pattern: a users guide to ZELIG, Version 1.0. Environmental Sciences Dept., University of Virginia, Charlottesville, VA, 22903.
Van Duesen, P.C., G.A. Reams, and E.R. Cook. (1991). Possible red spruce decline. J. Forestry 89(1):20-24.
Weishampel, J.F, R.G. Knox and E.R. Levine. (1997), Modeling soil saturation effects on forest dynamics across a southern boreal landscape, Bull. Ecological Soc. Am (in press).
Young, H.E., Ribe, J. H., and Wainwright, K. (1980), Weight tables for tree and shrub species in maine, Miscellaneous Report 230, Life Sciences and Agriculture Experiment Station, University of Maine at Orono, September.