Elements of Change 1995: Session I White

Over the past two decades, scientists have developed an increasingly refined ability to quantify and monitor the surface of the Earth through the use of satellite remote sensing. In particular, as our understanding of humanity's role in altering global patterns of vegetation and weather has increased, so has our interest in the detection of anthropogenic changes in land cover.

The term land cover itself is open to multiple interpretations. Since the introduction of the first, and still widely used, comprehensive global land cover classification (Matthews, 1983), a proliferation of studies has led to considerable confusion as to the extent and definition of major biomes (see figure i.1,Townshend, 1991). Townshend concluded that the next step in land cover modeling should be a satellite-derived classification scheme.

Such a classification would help to fulfill a growing need for accurate, repeatable, and globally extensive land cover maps. While the number of desired land cover classes varies considerably by discipline, a biome level classification is often used for large-scale terrestrial ecosystem modeling. In fact, obtaining biome distribution is one of only four critical precursors to implementation of the current generation of large scale biogeochemical (BGC) computer simulation models. The second requirement is Leaf Area Index (LAI), a dimensionless variable representing the ratio of leaf to ground surface area. LAI can be approximated from remote sensing, most commonly as a function of the Normalized Difference Vegetation Index (NDVI) and biome type. Daily climate and topography are also required.

For individual biome types, such as grass, deciduous broadleaved, or evergreen needleleaved, White and colleagues have created detailed files which include factors such as respiration coefficients, carbon to nitrogen ratios, and other basic physiological parameters. This file, with the addition of LAI, topography and climate, is then used to execute one of the family of BGC models, ranging from stand to global levels. Outputs from BGC models include Net Primary Production (NPP), respiration, and evapotranspiration, among others.

Obtaining biome distribution is one of only four critical precursors to implementation of the current generation of large scale biogeochemical computer simulation models. The other requirements are leaf area index, daily climate and topography.

As topography and climate are relatively easy to obtain, and LAI is calculated from NDVI and biome type, representing biome distribution becomes extremely important in regional to global ecological modeling.

The purpose of White's presentation was to provide a basic outline of current satellite-driven land cover modeling activities at the Numerical Terradynamic Simulation Group (NTSG) at the University of Montana. Throughout development, the group's goal has been to produce a land cover classification scheme which is 1) based on clearly defined vegetation structure; 2) driven by remotely sensed data; and 3) useful to large-scale carbon models. Within this overall framework, three primary stages have been completed.

Running et al. (1994) presented the initial logical concept for a global vegetation land cover classification. A series of decisions is used to drive the implementation logic (Figure 21.1 ). First, permanence of above-ground woody biomass separates grasses and broad leaf crops from trees and shrubs. The length of the growing season, possibly defined as the length of time the NDVI curve stays above a threshold value (Figure 21.2 ), could be used to make this first decision; a longer duration above the threshold defines year-round biomass.

The addition of surface temperature and a seasonal trajectory component enhanced the operational capability of the original logic.

Second, leaf longevity is used to distinguish between vegetation types which maintain leaf cover for more than one year and those that do not. The amplitude of the NDVI curve is theorized to differentiate between evergreen (low amplitude) and deciduous (high amplitude) vegetation types. For grasses and broadleaf annuals, this is not an issue, since all the cover types are deciduous. For trees and shrubs, though, this is an important distinction. Third, leaf type, needleleaf or broadleaf, is defined. This last division differentiates between, for example, oak (deciduous broadleaf) and pine (evergreen needleleaf). For theoretical discussion of possible implementation of this methodology and associated problems, see Running (1994).

This logic defined the framework from which subsequent modifications were built. Using the 1991 bi-weekly composite NOAA/AVHRR dataset from the EROS Data Center, White and colleagues classified the biome cover of the conterminous U. S. (Nemani, in press). While similar to the logic outlined by Running, this classification included numerous modifications. Essentially, the addition of surface temperature and a seasonal trajectory component enhanced the operational capability of the original logic.

The first decision relies on surface temperature, calculated from thermal infrared (TIR) channels 4 and 5, to separate groups with typically high canopy temperatures (groups I and III, Figure 21.3 ) and groups with lower canopy temperatures (groups II and IV, Figure 21.3). A 35 degree threshold was used for this distinction, based on empirical evidence that well-watered closed canopies do not exceed 32 degrees C. It was assumed that even when water stressed, the aerodynamically rough canopies of forests will mix well with the atmosphere, and temperatures will usually not elevate by more than 2 to 3 degrees. Crop, grass, shrub, and barren areas, which are relatively smoother aerodynamically, will not mix well with the atmosphere and may reach canopy temperatures well above ambient levels.

Distinction of forest/non-forest was 90%. Forest type detection was also good, while shrubs and barren class detection was slightly lower. Grasses and crops were identified with the least success. In general, the classification captured the major distribution of significant biomes.

Second, an NDVI threshold of 0.4 was used to separate group I (barren, shrub, and grass) from group III (crops) and group II (wetlands and non-vegetated) from group IV (forest). The underlying theory for this threshold is the assumption that at an NDVI of 0.4, 75% of photosynthetically active radiation (PAR) has been absorbed. Forests and crops, then, were assumed to absorb more than 75% of PAR.

The classification into four groups is iterated for each composite period. Typically, most classes begin in Group II, energy limited, and then move toward another group. Whichever group besides Group II the pixel occupies a majority of the time becomes the final group classification.

Group I was further separated into barren, grass, and shrub by computing a growing season average NDVI, NDVIgs defined as the average of the NDVI composite periods with surface temperature above 5°C. Barren areas were assumed to have low NDVIgs, shrubs intermediate, and grass high. Group II, wetlands and non-vegetated, was also separated using NDVIgs. Group III was divided into deciduous and evergreen using NDVIgs. Further separation into broadleaf or needleleaf was achieved with a near infrared (NIR) threshold, based on empirical evidence that coniferous canopies, especially boreal forests, have low NIR reflectances. Crops required no further separation.

Results were validated by comparing the group's land cover image to a reclassification of Loveland's 1990 map. Distinction of forest/non-forest was 90%. Forest type detection was also good, while shrubs and barren class detection was slightly lower. Grasses and crops were identified with the least success, possibly because of common land cover mixing which is difficult to detect at 1 km resolution. In general, though, the classification captured the major distribution of significant biomes.

In areas where human land use has resulted in the removal of native forest vegetation for agriculture or grazing, there has been an evident reduction in LAI. Southeast Asia, western Europe, parts of Africa, and India, are areas where this pattern appears to dominate.

An initial analysis of the impacts of human land use practices on global biogeochemical cycles has been completed (Nemani, in press). This area of work is controversial because of the difficulty in establishing pre-agricultural land cover distribution. Nonetheless, a dataset based on long-term climate and soil types was used to derive biome distribution, grided to 0.5 degrees. Standard physiognomic plant optimization theory was used to calculate a maximum historical LAI. Biome BGC, an NTSG ecosystem simulation model, was iterated until LAI reached a stable value. Modern biome distribution was taken from a current atlas-based land cover map. Biome distribution and the 1985-1990 NOAA Global Vegetation Index (GVI) were used to calculate current global LAI. Maximum values from each of the six years were selected and then averaged to arrive at a single NDVI value in the hope of minimizing cloud contamination.

While all usual Biome BGC outputs were generated, the focus here is on changes in LAI from historical to current conditions. LAI is assumed to be a crude surrogate for biomass. Two types of land use change seem to have resulted in distinct patterns of vegetation change. In areas where human land use has resulted in the removal of native forest vegetation for agriculture or grazing, there has been an evident reduction in LAI. Southeast Asia, western Europe, parts of Africa, and India, are areas where this pattern appears to dominate. Conversely, other areas of the world, such as the southeast United States and eastern Australia have experienced increases in LAI as a result of agricultural development.

White and colleagues recognize that the NOAA GVI has problems in terms of cloud contamination, multiple satellite calibration, and inadequate atmospheric corrections. Nonetheless, their simple analysis seems to have captured known patterns of global land use. Currently, the group is working to analyze global land use impacts using a remote sensing derived land cover map. Herein lies the potential for global change monitoring: to create decadal, remote sensing derived global land cover maps, and thus to monitor the ongoing impacts of human land use, and ultimately to suggest political and economic strategies to ameliorate negative impacts.