Determining land surface fractional cover from NDVI and rainfall time series for a savanna ecosystem
Introduction
Savanna ecosystems are characterized by the coexistence of woody and herbaceous vegetation, the relative abundance of which defines important aspects of the biome such as fuel biomass and combustion factors for fire Hoffa et al., 1999, Shea et al., 1996, fauna habitat Doergeloh, 2000, Dean et al., 1999, nutrient cycling Belsky, 1994, Frost, 1984, and resources for human subsistence. Frequently stressed and sensitive to change (Guenther, Zimmerman, Greenberg, Scholes, & Scholes, 1996), savanna ecosystems are responsive to climate variability on relatively short time scales. Water is the main driving force in shaping the vegetation composition and distribution for these semiarid systems Rodriguez-Iturbe et al., 1999a, Rodriguez-Iturbe et al., 1999b, Smit & Rethman, 2000. The dynamic quality of the vegetation with respect to precipitation forcing can be monitored over a large spatial area with the aid of remote sensing and inferences about rainfall–vegetation processes in these savanna ecosystems can thus be developed. With the added benefit of sufficient temporal coverage, remote sensing can now be used to make predictions of the earth's vegetation dynamics with respect to future climate scenarios based upon analysis of these past observations. In this paper, we merge ground-based rainfall measurements with remotely sensed data to infer the surface cover components of a savanna system from the rainfall response properties of the individual components.
Several factors make the Kalahari Transect (KT) in southern Africa an ideal location for assessment of this concept. The KT is one in the global set of International Geosphere–Biosphere Programme (IGBP) transects (Koch, Scholes, Steffen, Vitousek, & Walker, 1995), spanning a north–south aridity gradient from Angola and Zambia, through Botswana, and into South Africa (Thomas & Shaw, 1993). A homogeneous aeolian sand formation underlies a large portion of the transect, providing a relatively uniform background reflectance. Savanna vegetation is optimal for the use of remote sensing (Palmer & van Rooyen, 1998) because, unlike with closed canopies, spectral saturation is not a problem, thereby enabling the recognition by a remote sensor of subtle differences in the true amount of green biomass. Previous applications of remote sensing in Africa have focused on the link between water availability and vegetation biomass (e.g., Farrar et al., 1994, Fuller & Prince, 1996, Nicholson & Farrar, 1994, Richard & Poccard, 1998). In southern Africa, there is observed to be a strong correspondence between the mean climatic distribution of precipitation and green biomass, as measured by mean annual normalized difference vegetation index (NDVI) (Goward & Prince, 1995), yet on shorter time scales the interannual variability in precipitation does not, in general, result in much variability in NDVI (Fuller, 1994). The exception to this was found to be in “marginal zones” (Goward & Prince, 1995), areas bounded by high and low annual precipitation such as the middle region of the KT, where the vegetation appears to respond strongly to year-to-year precipitation variability. This raises the question: What factors associated with the vegetation in these marginal zones cause the NDVI to be more dynamic in its response to precipitation? We hypothesize here that relative mixture of trees and grasses plays a primary role in defining this observed phenomenon.
In terms of strategies for using water, grasses are considered to be intensive exploiters while trees and shrubs are extensive exploiters (Burgess, 1995). With dense, shallow root systems, grasses make use of water that is ephemerally available in the upper layer of the soil while trees, which have root systems that penetrate both the shallow and deeper soil layers, have a more persistent supply of soil water. Relative to trees, grasses exhibit a greater areal expansion of biomass in response to rainfall in savanna ecosystems. Short-term greening of trees is restricted in areal extent by the standing woody biomass. Additionally, the photosynthetic pathways associated with trees and grasses differ such that grasses, in general, are able to synthesize more carbon per unit of water loss via transpiration than are trees (Ehleringer & Monson, 1993). This was supported by flux measurements taken along the KT (Scanlon & Albertson, submitted). All these factors contribute to greater expected NDVI response to precipitation by grasses relative to trees. This is the distinguishing property of the vegetation components that will be used to detect and model the dynamic vegetation composition of the savanna ecosystem along the KT, with wider applicability to savannas in general.
Subpixel-scale information about land cover classification or composition has often been obtained from spectral “unmixing” analysis, a richly developed and active area in the remote sensing and photogrammetry literature. This analysis is based on the principle that a spatially coarse spectral signal is a weighted function of the spectral contributions from the smaller-scale (i.e., subpixel) individual components. Linear unmixing models (e.g., DeFries et al., 2000, Settle & Drake, 1993, Smith et al., 1985) used to detect surface cover are by far the most common, and, although theoretically imperfect due to the omission of the effect of multiple scattering between cover types Myneni et al., 1995, Roberts et al., 1993, the errors associated with the linear assumptions have been found to be relatively minor (Kerdiles & Grondona, 1995). The output from the spectral unmixing applications generally falls into two categories: land cover classification types (i.e., forest, grassland, urban, etc.) or fractional-cover components (i.e., tree, grass, bare soil). A typical approach for determining either product, as outlined in DeFries et al. (2000), first involves the generation of numerous metrics from the multiple spectral bands (channels) that are present on remote sensing instruments. Often, the phenology of the vegetation is captured by selecting metrics that are related to the within-year variability measured by particular channels or combinations of channels. Next, the number of metrics is reduced to a smaller number of variables for use in the unmixing model by performing linear discriminant analysis. This procedure weights the metrics by maximizing the ratio of class means to within-class variance, thereby enabling maximum spectral separation between the cover types. Training sets, or spectral information from preclassified cover types, can then be used to define the spectral end members for the “pure” cover types for use in the unmixing model. Another way that spectral end members can be selected based on the metrics is through the use of principle component analysis Bateson & Curtiss, 1996, Smith et al., 1985, Van Der Meer & De Jong, 2000. Both of these methods rely fundamentally upon empirical relationships between the vegetation and the spectral reflectance to define the land surface cover.
In the method presented here for finding the fractional surface cover components, we use a 16-year time series of NOAA Advanced Very High Resolution Radiometer (AVHRR) data along with the relationship that is derived between rainfall and NDVI as a means for a process-based identification of the tree, grass, and bare soil cover components. As pointed out by DeFries et al. (2000), using a multiyear AVHRR record must be cautiously undertaken due to the artifacts that remain in the record as a result of factors such as changes in sensors (NOAA-7 to NOAA-9 to NOAA-11 to NOAA-14 satellites), the impacts of aerosols from volcanic eruptions, other uncorrected atmospheric effects, and drift in equatorial crossing time. These very real problems require recalibration of the spectral end members each year with traditional spectral unmixing applications. The sensitivity of the unmixing results to the recalibrated end members can lead to inferred land cover changes that did not actually occur (DeFries et al., 2000). By using the long term mean values of NDVI and the regressed sensitivity of the NDVI to rainfall along the KT, the net effect is that our method is insensitive to shorter time scale variability in NDVI that is not tied to variations in rainfall. As a consequence, we are able to largely circumvent this aforementioned problem.
The objectives of this research are to (1) determine the mean NDVI and sensitivity of NDVI to rainfall related to the individual components of the surface cover (tree, grass, bare soil) along the KT, (2) incorporate these properties in a linear unmixing model to derive fractional cover, and (3) develop a method to predict fractional cover in response to various future rainfall scenarios.
Section snippets
Data
Monthly NDVI data at 8-km resolution along the KT were acquired from the NASA/NOAA-sponsored AVHRR Land Pathfinder data set Agbu & James, 1994, James & Kalluri, 1994 for the years 1983–1998 in the area bounded by 12° to 26°S latitude and 20.4° to 24.7°E longitude. The north–south oriented swath of NDVI data has the dimensions of 196×61 pixels. The maximum NDVI value for each pixel during a given month was assigned as the monthly value in an effort to eliminate cloud cover contamination; the
Results
Fields of the 16-year wet season mean and interannual standard deviation of NDVI are shown in Fig. 2a and b, respectively. The mean NDVI shows a clear, general decrease from north to south over the KT. The standard deviation in the NDVI peaks near the middle of the transect, bounded by areas of lesser interannual fluctuation in the green biomass cover to the north and south. The slightly speckled appearance of the standard deviation field is caused by AVHRR sensor noise that was introduced at
Discussion
Although potential limitations associated with the analysis of multiyear AVHRR-NDVI data have been recognized (e.g., DeFries et al., 2000), the long-term temporal aspect of the data can be extremely beneficial for extracting information about land-cover processes in water-limited environments when carefully analyzed. Rather than basing our analysis on relative differences in NDVI between years, which would be problematic due to shifts in sensor-related instead of vegetation-related factors, we
Conclusions
There currently exists the need to establish estimates of fractional surface cover over large geographical areas for use in fuel load models of biomass burning as well as in land–atmosphere exchange models. In both cases, the compositional mixture of trees and grass plays a major role in defining the function of the system. Reliable multiyear estimates of fractional cover in savannas have proven to be elusive, however. Additionally, large-scale linkages between rainfall and vegetation cover
Acknowledgements
Funding for this research was provided by a NASA New Investigator Program in the Earth Sciences award (NAG5-8670). The authors appreciate the thoughtful discussion and input by Tim Newman. We also thank the three anonymous reviewers for their comments on an earlier version of this paper. Data used by the authors in this study include data produced through funding from the Earth Observing System Pathfinder Program of NASA's Mission to Planet Earth in cooperation with the National Oceanic and
References (43)
- et al.
A method for manual endmember selection and spectral unmixing
Remote Sensing of Environment
(1996) - et al.
Large trees, fertile islands, and birds in arid savanna
Journal of Arid Environments
(1999) - et al.
The influence of soil type on the relationships between NDVI, rainfall, and soil moisture in semiarid Botswana: II. NDVI response to soil moisture
Remote Sensing of Environment
(1994) - et al.
On the use of NDVI for estimating rainfall fields in the Kalahari of Botswana
Journal of Arid Environments
(1997) - et al.
Spectral response of a plant canopy with different soil backgrounds
Remote Sensing of Environment
(1985) - et al.
Optical remote sensing of vegetation: modeling, caveats, and algorithms
Remote Sensing of Environment
(1995) - et al.
The influence of soil type on the relationships between NDVI, rainfall, and soil moisture in semiarid Botswana: I. NDVI response to rainfall
Remote Sensing of Environment
(1994) - et al.
Detecting vegetation change in the southern Kalahari using Landsat TM data
Journal of Arid Environments
(1998) - et al.
Green vegetation, nonphotosynthetic vegetation, and soils in AVHIRIS data
Remote Sensing of Environment
(1993) - et al.
The influence of tree thinning on the soil water in a semi-arid savanna of southern Africa
Journal of Arid Environments
(2000)
The evolution and characteristics of the Kalahari, southern Africa
Journal of Arid Environments
The NOAA/NASA Pathfinder AVHRR land data set user's manual
Autogenic succession in a subtropical savanna: conversion of grassland to thorn woodland
Ecological Monographs
Influences of trees on savanna productivity: tests of shade, nutrients, and tree–grass competition
Ecology
Dessert grassland, mixed shrub savanna, shrub steppe, or semidesert scrub? The dilemma of coexisting growth forms
Global continuous fields of vegetation characteristics: a linear mixture applied to multi-year 8 km AVHRR data
International Journal of Remote Sensing
A new global 1-km dataset of percentage tree cover derived from remote sensing
Global Change Biology
Relative densities and habitat utilization of non-utilized, terrestrial gamebird populations in a natural savanna, South Africa
African Journal of Ecology
Evolutionary and ecological aspects of photosynthetic pathway variation
Annual Review of Ecology and Systematics
Organic matter and nutrient dynamics in a broadleafed African savanna
Cited by (0)
- 1
Now at Department of Civil and Environmental Engineering, Box 90287, Duke University, Durham, NC 27708-0287, USA.