Vegetation Mapping of Forested Ecosystems in Interior Central Alaska

Susan L. Ustin¹, Lai-Han Szeto¹, Qing-Fu Xiao¹, and Quinn J. Hart¹, Eric S. Kasischke² ¹Department of Land, Air and Water Resources, University of California, Davis CA 95616-7600 USA ²School of Environment, Duke University, Durham NC Author for Correspondence:
Susan L. Ustin
Department of Land, Air, and Water Resources
University of California
Davis, CA 95616
Phone: (530) 752-0621
FAX: (530) 752-5262
email: slustin@ucdavis.edu

Abstract

A digital elevation map (DEM), digital soil survey map, and a 1991 Systeme Probatoire d'Observation de la Terre (SPOT) satellite image were used in a geographical information system (GIS) to identify 12 forest types at the Bonanza Creek Experimental Forest (BCEF) in central Alaska. A potential vegetation distribution map for the boreal forest ecosystem in the interior of Alaska was created using a hierarchical decision tree. The site was first stratified into montane, flood plain, and lowland zones based on topography. Within the montane zone, topographic information (elevation, aspect, and slope) used to define potential vegetation classes. SPOT band 3 (0.50-0.59 mm) data was used to identify the boundary of a 1983 wildfire within the zone. The forests in the flood plain and lowland zones were not strongly associated with topographic features due to the low relief, so the Normalized Difference Vegetation Index (NDVI) was the primary basis for defining forest types. Separating the flood plain from the lowland zone was done based on variance ranges in NDVI. This method has promise for mapping vegetation distributions in other boreal regions when little or no ground data is available for validation.

INTRODUCTION

Global boreal forests cover 12 million km² and only a small portion have been mapped in any detail. Even a low resolution vegetation map defining conifer, broadleaf and non-forest classes would provide information needed for assessment of the terrestrial carbon cycle and cycles of other trace gases. The total carbon storage, carbon source and sink terms (in both above ground and below ground compartments) and the rate dependence varies widely with ecosystem type and radiation regime (Bonan, 1990, 1991). Most empirical evidence indicates that broadleaf and conifer forests differ widely in nutrient use, turnover rates, and in fluxes of trace gases (Bonan et al., 1992). Changes in the distribution or abundance of major forest ecosystems could have significant global implications for the storage, turnover and release of greenhouse gases (Bonan, 1991, 1993; Kasischke et al., 1994; Pastor and Mladenoff, 1992).

Predictions from Global Climate Models (GCM) for higher temperatures in boreal regions under scenarios of increased CO₂ concentrations require urgency in the development of a reliable method for assessing the types and extent of boreal ecosystems and for monitoring changes in their distributions and characteristics over time. Remote sensing technology provides the most feasible means of acquiring repeated terrestrial information at detailed space and time scales. The goal of this research was to test how well topographic and satellite information could be used together to define boreal forest distributions and to differentiate among the major boreal forest ecosystems and their characteristic successional stages. We hypothesized that vegetation types could be accurately mapped using rules developed from a few ecologically based assumptions. We present an approach using a hierarchical decision tree that produces a forest cover map with a minimum of user intervention and with accuracies at or above current estimates of forest distributions. This type of approach has the potential for extension to boreal ecosystems generally.

Boreal Forest Ecosystems

The boreal forests of Alaska are floristically simple with only nine tree species dominating the zone (Takhtajan, 1986). The Bonanza Creek Experimental Forest (BCEF) contains the four late seral forests described by Van Cleve and Viereck (1981) which are: (1) flood plain white spruce, (2) upland white spruce, (3) upland black spruce and (4) lowland black spruce. In addition to these distinctive late seral forest ecosystems there are several mid-successional stages that include broadleaf forests (e.g., alder or balsam poplar) and conifer/hardwood mixtures. It is generally agreed that these forest ecosystems are associated with temperature and precipitation gradients, nutrient availability and quality, and fire history (Viereck, 1983; Van Cleve and Viereck, 1981; Van Cleve et al., 1983; Bonan et al., 1992). Topography, especially elevation and aspect, is critical in determining the potential distribution of these species (Bonan, 1989).

METHODOLOGY

The study site at the BCEF is located approximately 40 km southwest of Fairbanks, Alaska. The area is crossed from east to west by the Tanana River. Most of the area north of the river is dissected by rounded hills and ridges unglaciated Yukon-Tanana uplands of Precambrian Birch Creek shist overlain by loessal deposits from the Tanana River (USDA, 1977). The geology around and south of the river contrasts sharply, as the flood plain was formed by Pleistocene glaciers that left broad outwash plains and gravelly moraines. A DEM of the site was obtained from US Geological Survey. A vegetation map developed from interpretation of 1978 aerial photos and field surveys was obtained from J. Yarie (University of Alaska). SPOT multispectral image of the region was acquired on August 28, 1991. The NDV was computed using bands (3-2)/(3+2). This index has been widely applied in vegetation research and provides a continuous distribution of canopy "greenness" and is a surrogate measure of spatial productivity or foliage density. SPOT images were co-registered to the ground-based vegetation map.

Classification Procedure

The initial step in the decision tree was to divide the region into an uplands and lowlands region based on topography. The lowlands boundary was defined as the upper edge of the Tanana River alluvial plain at the transition where the elevation contours became closely spaced and the slope angle rapidly increased. For the montane region of the image, aspect and elevation thresholds were applied to define potential upland forest community distributions. The relation was developed using species distributions from the literature and from histograms using overlays of the topography and ground-based maps. For areas where both white and black spruce had equal potential to occupy a site, slope was added as a third classification variable. Black spruce communities were defined as occupying slopes >30o and white spruce <30o. Finally, in the fourth step, the NDVI value was used to partition the stands into "actual" vegetation classes from the topographically defined "potential" vegetation classes. This allowed a forest identification to vary when more than one vegetation type was possible at a particular combination of elevation, aspect, and slope. This sequence of rules permitted identification of several successional forest stages. The alluvial plain was distinguished from the flood plain zone by high variance in NDVI. Because topographic relief was minimal, the final vegetation classes for the two lowland zones were determined using separately defined NDVI thresholds. An accuracy assessment of the classification was performed.

Classification Accuracy Dependence on Resolution of Community Type

To evaluate the influence on the number of classes, the classification procedure was repeated for each of three levels of community description. In the most detailed description, 11 classes were identified (six forest classes dominated by a single species, four mixed classes, and a non-forest class). In the intermediate level, only six single species dominated forest classes were identified and the non-forest class. Finally, in the most consolidated grouping, only three classes were identified: deciduous, conifer, and non-forest classes.

The overall pixel accuracy of the predictions based on the ground-based map ranged from 35%, 48%, to 60% for the unsmoothed 11, seven, and three forest class models, respectively. However, the 11 class level of forest classification found good agreement in terms of overall forest patterns and of total area occupied by each forest type even though numbers of absolute pixel matching was low. The ratio of the predicted area/mapped area vary from 0.34 to 1.78 for each forest type with a mean of 0.87. The patterns observed for the uplands, lowlands, and flood plain zones are visually consistent with the surveyed vegetation map. Where differences between forest types exist, they are largely between related types. Some misidentified types are between single species forest types and the corresponding mixed forest type. Accuracy nearly doubles when pixels from related forest types (e.g. black spruce and black spruce mixtures) are combined. In other cases, the white and black spruce forests are confused or the broadleaf forest types (e.g., aspen and paper birch) are confused. The two forest types of greatest extent, white spruce and black spruce mixed forests, showed the highest individual accuracies, which ranged between 34-50%. Thus the results are promising in terms of total area predicted per forest type and in the overall accuracy of predictions. Classification errors might be further minimized if additional information layers could be included in the GIS.

Another obvious difference between the surveyed map and the predicted map is in the level of spatial detail. Clearly the predicted forest map exhibits much greater spatial patchiness then the surveyed map. This is partially due to the higher spatial resolution of the image based map.

DISCUSSION

Our analysis was strongly predictive of the four major upland and lowland late seral forest ecosystems described by Viereck and Van Cleve (1981) and for their most important mixed forest and broadleaf successional stages. The predicted forest distribution map indicates good agreement with the ground-based map, in terms of location and areal extent of forest distributions. These results are significant given the difficulty of mapping the large areal extent of forest distributions and the often inaccessible terrain. Knowledge of the spatial extent and location of these forest ecosystems is important for understanding climate biosphere feedbacks because they have considerably different patterns of carbon storage, exchanges of carbon dioxide and water, and in the cycling of other trace gases e.g., methane (Bonan, 1991).

The site is roughly divided 3:2 between lowland and upland forest ecosystems. The distribution of forest types varies widely. In both upland and lowland forests, forests in the black spruce successional sequence exhibit greater total area then white spruce seral sequences. The lowland black spruce forests occupy more than twice the area occupied by flood plain white spruce. Nonetheless, only a small fraction of the lowland black spruce forest is at a late seral stage while nearly half of the flood plain white spruce is in a late seral stage. The reverse is true for upland forests, with nearly half of the black spruce sequence at a late successional stage and less than 10% of upland white spruce forests are at late successional stages. Only a small fraction of the total area is classified as non-forest, also indicating that most of this landscape is at the relatively late seral stages (conifer or mixed conifer classes). This indicates only 1% and 4% of the land is at an early seral stage. Depending on the number of years a forest will remain in this initial phase of secondary succession, the actual turnover rates for these forests is probably more than 100 years. Possibly the differences in proportion of late seral stages between upland and lowland forests indicate a contrast in the types of fires and the extent of combustion in these forests.

We chose parameters for our GIS analysis that can be obtained for many remote areas so that the results could be extendible into interior boreal forests of similar types. In addition, minimal user intervention or a priori site knowledge is required for the analysis, in contrast to methods that require extensive "training sites." Our analysis largely adopted typical photo-interpretation techniques by defining the criteria for acceptance or elimination of forest types based initially on topographic or soils conditions, followed by interpretations based on spectral patterns and texture in the images. The essential difference between this and other methods has been to direct the analysis in a GIS, using a rule based approach. This difference is critical since photo-interpretation requires considerable time and analyst skill to produce vegetation maps covering large regions, and may be subject to user bias. The rules developed and tested in this paper are being applied to other sites in central Alaska (Kasischke et al., 1994b; Xiao et al., 1994).

Figure 1

Acknowledgments

This research was performed under NASA ERS-1 grant program, NAGW-2636 and EOS grant NAS5-31359 to the University of California, Davis, and NASA ERS-1 grant NAGW-2645 to the Environmental Research Institute of Michigan. The computer analyses were performed on Digital Equipment Corporation DEC 5000 and DEC Alpha computers provided under the Sequoia 2000 research program. We wish to recognize the DEM and ground-based vegetation maps and SPOT image used in this study were contributed by Dr. JoBea Way (Radar Sciences Group, Jet Propulsion Laboratory), Dr. Les Viereck (Institute of Northern Forestry, Fairbanks, AK) and Dr. J. Yarie (U. Alaska, Fairbanks, AK).

REFERENCES

Bonan, G.B., Shugart, H.H., and Urban, D.L. "The sensitivity of some high-latitude boreal forests to climatic parameters." Climatic Change. 16 (1990): 2-29.

Bonan, G.B. "Atmosphere-biosphere exchange of carbon dioxide in boreal forest." Journal Geophysical Research 96 (1991): 7301-7312.

Bonan, G.B., Pollard, D., and Thompson, S.L. "Effects of boreal forest vegetation on global climate," Nature 359 (1992): 716-718.

Bonan, G.B. "Importance of leaf area index and forest type when estimating photosynthesis in Boreal forests." Remote Sensing of Environment 43 (1993): 303-314 .

Kasischke, E.S., Christensen, N.L. Jr., and Stocks, B.J. "Fire, global warming and the mass balance of carbon in boreal forests." Ecological. Applications (1994a, .in review).

Kasischke, E.S., Ustin, S.L., Christensen, N.L. Jr., French, N.H.F., and Bourgeau-Chavez, L.L. "Estimating carbon storage and release during fires in Alaskan boreal forests using satellite imagery." Journal Geophysical Research (1994b,.in review),

Pastor, J. and Mladenoff, D.J. Southern boreal-northern hardwood forest border. In: Shugart, H.H., Leemans, R.L., and Bonan, G.B. A Systems Analysis of the Global Boreal Forest, New York: Cambridge University Press (1992) pp. 216-240.

Takhtajan, A. Floristic Regions of the World. Berkeley, CA: University of California Press (1986) pp. 522.

Van Cleve, K. and Viereck, L.A. Forest succession in relation to nutrient cycling in the boreal forest of Alaska. In: West, D.C., Shugart, H.H. and Botkin, D.B. (eds) Forest Succession: Concepts and Applications. New York: Springer-Verlag (1981) pp.185-211.

Van Cleve, K., Oliver, L., Schlentner, R., Viereck, L.A. and Dyrness, C.T. "Productivity and Nutrient Cycling in Taiga Forest Ecosystems." Canadian Journal Forest Research 13 (1983) 747-766.

Viereck, L.A. The effects of fire in black spruce ecosystems of Alaska and northern Canada. In: Wein, R.W and MacLean, D.A. (eds.), The Role of Fire in Northern Circumpolar Ecosystems. New York: Wiley (1983) pp 201-220,

Xiao, Q.-F., Ustin, S.L., and Wallender, W.W. "A Spatial and Temporal Continuous Surface Subsurface Hydrological Model." (1996)Journal of Geographical Research 101(D23):29565-29584.

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