0. ABSTRACT

1. INTRODUCTION

Currently the best tool for examining landscape structure is remote sensing, because remotely sensed data provide complete and repeatable coverage over landscapes in many climatic regimes. Many sensors, with a variety of spatial scales and temporal re peat cycles, are available. The Airborne Visible/Infrared Imaging Spectrometer (AVIRIS) is an aircraft mounted sensor with a large number of spectral bands (Vane et al., 1993). Ironically, despite the hyperspectral capability of this instrument, the bes t data product for the examination of landscape structure over a large number of sites are the much reduced, one-band "quicklooks." Although high spectral resolution is desirable for investigating qualities of different land cover classes on th e landscape, in terms of spatial structure, most bands provide roughly equivalent views of the geometric relationships (i.e. patch size, patch connectivity) for highly distinct spectral classes like vegetation, soil and water.

Quicklooks are one-band images, averaged every two lines and two samples, and available for every site AVIRIS imaged between 1992-1994 (and more previously and subsequently), which constitutes over 4000 scenes from over 100 different sites throughout N orth America. Quicklooks have several advantages for the survey of ecosystem structure. They are small (0.5 Mb) and easy to manipulate, yet retain a pixel size (40 m) which approximates the spatial resolution of current land observation satellite system s like Landsat -TM, -MSS and SPOT. Unlike panchromatic SPOT, they are free and easily available over the Internet; other free data (e.g. AVHRR) has much coarser spatial resolution. The single quicklook band is centered at approximately 700 nm (nominally 10 nm wide), which avoids major atmospheric absorbtions (Lillesand and Kiefer, 1994), yet is in a region where basic terrestrial materials, like water, soil/rock, vegetation, and snow/ice can generally be distinguished (Lillesand and Kiefer, 1994). Simp le image processing techniques (contrast stretching, level slicing) allow the scenes to be classified into spectrally distinct landscape components for analysis. Though only a crude approximation of the landscape, these methods allow each scene to be ana lyzed in terms of the same component classes while retaining the spatial heterogeneity of the landscape.

A broad range of ecosystems representing a variety of climatic, geological and ecological conditions are represented in the quick look dataset. The dataset spans locations from 18º N to 56º N latitude and from 68º W to 126º W longitude, from borea l forest to tropical uplands, from coastal British Columbia to Key West, Florida. Eighteen of the twenty-seven North American vegetation types identified by Barbour and Billings (1988) and eleven of the fourteen North American physiographic regions iden tified by Vankat (1979) are represented in the dataset, including several examples each of montane conifer forest, boreal forest, temperate deciduous forest, desert scrub, Mediterranean scrub, prairie grasslands and tidal marshes. Though the dataset cont ains sites outside the United States, for this study we include only sites within the continental United States where ancillary data for comparison are available. Nevertheless, a short list of sites includes Mt. Rainier, WA, Biscayne Bay, FL, Los Angeles , CA, Rocky Mountain National Park, CO, Dismal Swamp, VA, Organ Pipe, AZ, Indian Pines, IN, and Suncook, NH. For most archive locations, several scenes are available, and for a few locations, multiple dates as well. Each scene is considered an independ ent sampling of that landscape type.

Climate interacts with landscape structure by influencing the amount of vegetation cover and the distribution of physiognomic types across the landscape (Prentice et al., 1993). Different plant communities have distinct spatial arrangements which form particular structures. Climate may also influence landscape structure through landforming processes like erosion and fluvial geomorphology. For this study we represented climate using the climate water balance diagram which integrates the temperatu re and moisture requirements of plants (Stephenson, 1990). By examining the seasonal timing of water supply vs. evapotranspiratory demand, we calculate expressions for the average, annual water surplus and deficit.

Climate is not the only influence on landscape structure however, and in many cases not the dominant one. Human land use and topography often have dominant influences on landscape pattern. Human land use breaks up vegetation patterns, creates new ones (e.g. agriculture), and is relatively indifferent to climatic effects. Topographic variation interacts with regional climate to create local climatic zones which in turn influence vegetation distributions. Mountainous terrain dissects the landscap e, creating more complicated and variable landscape structures. River basins and channels influence the kind and structure of nearby vegetation. Inter-correlations between climate, land use and topography are subtle and complicated, yet important for un derstanding the large scale causes of landscape structure.

Because of the important influence of land use and topography, these variables have also been included in our analysis. Fortunately, for the contiguous United States, free, publicly available datasets are available over the Internet. We downloaded th e appropriate land use and topographic geographic information system (GIS) coverages and clipped out the portion of the coverage corresponding to the quicklook. For land use we summarized the percentage land use for several major categories: urban, agri culture, range, forest, wetland and barren. For topographic variation, we calculated mean and standard deviations of elevation, slope and aspect, as derived from 1:250,000 scale Digital Elevation Models (DEMs).

This paper describes the methodologies used to evaluate the landscape structure of quicklooks and generate corresponding datasets for climate, topography and land use. A brief discussion of preliminary results is included at the end. Since quicklooks correspond exactly to their parent AVIRIS scenes, the methods used to derive climate, topography and land use data should be applicable to any AVIRIS analysis.

2. METHODS

2.1 AVIRIS Quicklooks

The selected quicklooks were downloaded either as raw data files or gif format from the AVIRIS Internet homepage, then uploaded into IDL (Research Systems Inc., Boulder, CO) for manipulation and analysis. Each quicklook was given a numeric code, 1-109, and a longer name composed of the date of acquisition, and flight, run and scene number (e.g. 940815B06.02). Flightline names provided by the AVIRIS staff were also retained (e.g. Jasper Ridge).

Each scene was stored in both continuous and nominal representations. Remote sensing images are useful for landscape analysis because they provide, without modification, a continuous measure of the landscape structure, but can be easily classified into nominal measures as well. Landscape structure metrics typically apply either to continuous or nominal data, but not both. As presented, quicklooks have a continuous range of gray levels from 0 to 255. These represent non-atmospherically corrected , upwelling radiance from the ground surface. To classify the scenes, we used a simple density slice technique to roughly identify patches of soil and vegetation in the image. Typically vegetation is darker than soil in the quicklook band (~700 nm). We confirmed this pattern by taking five calibrated AVIRIS data cubes, for which quicklooks and field data were available. These five scenes (Jasper Ridge, CA; El Dorado National Forest, CA; Santa Monica Mountains, CA; Winters, CA; Petaluma Marsh, CA) were deliberately chosen from different climate types and vegetation communities in California. The data cubes were classified into vegetation, soil or water using the Spectral Angle Mapper algorithm implemented in ENVI (Research Systems Inc., Boulder, CO) a nd image derived endmembers for those classes. The AVIRIS scene classifications were qualitatively verified by field work conducted by the CSTARS Laboratory (Department of Land, Air and Water Resources, University of California, Davis) at these sites. T hese classified images were compared to their corresponding quicklooks to derive gray level thresholds which distinguished vegetation from soil, with a small intermediate class. Although initially we had also planned to classify open water, there was suf ficient variability in the quicklook band to make the identification of water unreliable. Because the quicklook classifications have not been rigorously confirmed by field reconnaissance, they are referred to as Class-V, Class-S and Class-I, so that it i s clear they are based only on a simple remote sensing interpretation. An example of a quicklook and its classified product are shown in Figures 1c and 1d.
 

2.2 Landscape Structure Metrics

2.3 Climatic Water Balance

We obtained mean monthly temperatures and total precipitation from the Global Historical Climatological Network (Petersen et al., 1997), available for free over the Internet. This database has over 6000 precipitation stations and over 4000 temperature stations, with a large concentration of stations in North America, each with a historical record of at least 10 years. For each weather station, we averaged their long term records, then found the five closest stations within 100 km of each site. These five stations were averaged to estimate monthly temperature and precipitation at the quicklook site. Stations were reviewed to remove potentially unrepresentative data from the means (e.g. a weather station on the other side of a mountain range from the quicklook site).

Potential evapotranspiration at each site was calculated using the method of Thornthwaite and Mather (1955), which requires knowledge of only the mean monthly temperature, latitude and available soil water capacity of the site. Though more sophisticat ed methods are available, it is unclear whether the results are significantly better for large extent studies like this one (Milly, 1994). Using this method and the average precipitation, we used mass balance to calculate actual evapotranspiration, annua l water deficit and surplus, given an estimate of the available soil water capacity.

In the past scientists have often assumed a uniform level of available soil water capacity for all sites (Eagleman, 1976; Major, 1977), but today it is possible to get a geographically specific estimate using the Natural Resources and Conservation Serv ice State Soil Geographic (STATSGO) database (USDA, 1994). The STATSGO database provides 1:250,000 scale maps of soils and soil properties for the continental United States with a minimum mapping unit of about 625 ha. By overlaying each quicklook polygo n on the appropriate soil coverage (each available for an entire state), we calculated an area-weighted average available water capacity for each site. Soil water capacity has a strong influence on the annual water deficits and surpluses because it deter mines both the amount of water stored in the soil and the proportion of actual to potential evapotranspiration, making estimations of this parameter an important step in accurate water balance calculations. A representative water balance diagram is shown in Figure 1b.
 

2.4 Topographic Data

2.5 Land Use Data

Land use coverages are identified by using an on-line map to find the appropriate 1:250,000 quadrangle. Each quadrangle has a corresponding EPA code which identifies the appropriate coverage to download and then import into ARC/INFO. The quickloo k area of interest is clipped from the LULC coverage and summarized by proportional area of the Level 1 land use/land cover type. The LULC data have been used for several past landscape structure studies (e.g. Riitters et al., 1995; Hunsaker et al., 1994 ), but provide a much different sense of the landscape than remotely sensed data because they have been interpreted into homogenous polygons. An example of the land use data is shown in Figure 1f.
 

2.6 Analysis of Landscape Structure and Ancillary Data

3. RESULTS AND DISCUSSION

3.1 Synopsis of the Landscape Structure and Ancillary Datasets

The proportions of classes V, M and S (representing vegetation, mixed and soil, approximately) varied dramatically across the quicklook datasets. The proportion of vegetated surface varied from 4-89% and the proportion of bare soil varied from 6-9 4%. The scale, as estimated from the autocorrelation function, varied from 1-46 pixels (40-2020 m), indicating that the quicklooks have an appropriate grain (40 m) for capturing variation of these landscapes. Scale was positively correlated with the pro portion of class S in a scene and negatively correlated to the number of patches.

Variation in texture measures (angular second moment and inverse difference moment) were strongly positively correlated with contagion, since these metrics measure the homogeneity of the landscape, though from the perspective of continuous and nomi nal data, respectively. These measures were positively correlated with the proportion of class V in a scene, and negatively correlated with classes M and S.

The climatic water balance annual totals show significant variation site to site. Annual water surpluses vary from 0 at many sites in the western United States to over 185 cm, and deficits from 0-170 cm. Interestingly, highest surpluses and defic its were both recorded in Washington. Estimates of annual actual evapotranspiration and deficit are in similar ranges to those calculated by Eagleman (1976) and range over values that Stephenson (1990) associated with vegetation types from deciduous fore st to desert scrub. Substantial variation in the available soil water capacity (from 0.07 – 28.79 cm) indicates how important this parameter is to the climatic water balance. Many earlier authors assumed constant available soil water capacities of 10 cm (Major, 1977) or 15.24 cm (Eagleman, 1976), which although good approximations can lead to serious errors.

Quicklook sites varied in elevation from sea level to over 3400 m. Slopes also varied from near flat to over 20%. The variation in slope and elevation within a quicklook scene (measured by the standard deviation of those variables) tended to increase with elevation, since most high elevation sites were in mountainous terrain.

Land use also varied widely across the sites. Agricultural, rangeland and forest land use types were the most prominent in the dataset overall, though urban and wetland land use/land cover types were locally dominant in some scenes. Agricultural land use is strongly, positively correlated with available soil water capacity.
 

3.2 Correlative Hypotheses between Landscape Structure and Climate, Topography and Land Use

For example, there is a strong positive correlation between annual actual evapotranspiration and the proportion of forestland in the scene, as predicted by Stephenson (1990). The proportion of forestland is in turn positively correlated with the propo rtion of Class V (roughly vegetation), and the proportion of Class V is correlated with homogenous, contagious landscapes as measured by angular second moment, inverse difference moment and contagion. Simultaneously, however, mean elevation is negatively correlated with both actual evapotranspiration and the proportion of Class V. Does elevation drive the pattern of actual evapotranspiration which then drives the amount of vegetation and the homogeneity of the landscape, or do actual evapotranspiration and elevation both act independently on the amount of vegetation? A similar set of hypotheses can be formulated for the relationship between elevation, rangeland, proportion of class S, and heterogeneous landscapes. Forest and range land uses may expres s climate derived landscape structure more clearly than other land uses because they are based on potential vegetation types which are largely climate driven.

Another set of hypotheses involves factors governing the compactness of patch shape. The available soil water capacity is positively correlated with the proportion of agricultural land use in a scene. However agriculture is negatively correlated with mean elevation, variation in elevation and mean slope. Agriculture is also positively correlated with the more compact shapes (indicated by perimeter-area ratio and fractal dimension) and negatively correlated with the fraction of small patches in the landscape.

 

4. CONCLUSIONS

5. ACKNOWLEDGEMENTS

6. REFERENCES