Braswell and colleagues have demonstrated a method for extracting biophysical charac teristics over large regions using coarse resolution remote sensing data and inverse radiative transfer modeling. The objective of this approach is to obtain variables which can assist process modeling of terrestrial biogeochemistry and land surface energy balance. These variables are important for understanding how the terrestrial biosphere is coupled to the physical climate system. An example of this coupling is the effect of climate change/variability on net terrestrial carbon dioxide exchange. While the interactions between human activity and terrestrial vegetation can be studied with higher resolution data, the issues under investigation here are inherently low or course resolution problems.
In order to extract variables like fPAR (the fraction of absorbed photosynthetically active radiation) and albedo from remote sensing, a model is needed. One approach is to use empirical relationships between vegetation indices and a variable ( e. g., Leaf Area Index [LAI], fAPAR). The first requirement of this method is a good index that is invariant with respect to atmospheric contamination, soil variability, and other contaminating factors. Then, one must construct relationships based on field data. Another approach is to use a physical model, as is discussed here. In contrast, the physical model uses parameters describing the architectural and optical characteristics of vegetation components and produces reflectance that is a function of wave length and sun-sensor geometry, known as the bidirectional reflectance factor (BRF). In order to calculate the parameters, an optimization routine is used to determine the parameter set that minimizes the difference between the measured and modeled BRF.
This research utilized a modified version of the Scattering by Arbitrarily Inclined Leaves (SAIL) model (Verhoef, 1984). Modifications include a hot-spot parameterization by (Kuusk, 1991) and the inclusion of a second set of canopy elements (Qin, 1993) that allows for the distinction between leaf and non-leaf scattering and absorption of light. The hot-spot parameterization is necessary in order to account for the effect of self-shading by the vegetation, and the second (non-photosynthetically active) component allows for more accurate simulation of real canopies that contain significant amounts of stems or senescent vegetation. The two-component SAIL model (SAIL2) parameters include optical and structural characteristics for leaves and stems, background reflectance, and the relative abundance of the components (leaves, stems, and bare soil). There are 18 parameters in all.
Braswell and colleagues have performed an informal validation of the model by comparison with field data (a Maize canopy) from Qin (1993). Their model has roughly the same complexity as Qin's so that an almost identical parameterization was possible. SAIL2 was successful in simulating the canopy near infrared and visible reflectances under a number of illumination conditions. With one component only it was impossible to fit the data. Inclusion of a stem fraction in the canopy introduces more anisotropy (not the same in all directions) to the predicted BRF, especially by decreasing forward scatter transmittance. The model also reveals significant differences in computed Normalized Difference Vegetation Index (NDVI) from forward scatter to backscatter directions. This strong variation suggests that there are considerable problems interpreting NDVI observed from different angles.
Braswell presented an integrated algorithm for the use of Pathfinder 8x8 km AVHRR data to estimate vegetation parameters Figure 1.1). In order to obtain enough information to perform a parameter retrieval, he constructed a BRF by treating all pixels within some neighborhood (in this case a circle of radius 50 km) as if they were the same target. Because of the way the data is composited, adjacent pixels in a scene often are from observations on different days and thus may have very different view zenith and relative azimuth angles. This additional information enables the successful inversion of the model. Supporting the main flow of the algorithm (boxed portion of Figure 1.1) are a series of analyses involving a vegetation index. An NDVI climatology was constructed from three years of data (1986-88) and a principal component rotation was performed to obtain a continuous land surface characterization (LSC) based on the time trajectory of the index. The LSC is used in two ways, the first of which is to ensure spatial continuity of a neighborhood. All pixels which have a significantly different annual pattern of NDVI are treated as outliers and neglected. The second use of the LSC is to stratify a region of interest, for example, by applying parameters and constraints differently for different functional ecosystem types. Finally, Figure 1.1 shows that field measurements may be applied directly to the inversion process by defining fixed parameter values and constraints.
There are a number of advantages associated with using model inversion:
It allows for the explicit use of soil and leaf optical properties.
It allows for the incorporation of a priori knowledge of land-surface type.
It complements (and utilizes) broad-brush vegetation index approaches. Any factor that leads to uncertainty in biophysical estimates derived from Radiative Transfer Model inversion also results in uncertainty in results derived from vegetation index methods. In theory, however, inversion is a tool that can allow for reduction in the uncertainty through the use of ecological knowledge and field measurements.
An initial investigation was focused on a transect across the Central African Republic (Figure 1.2). This transect represents a steep ecological gradient from dry, seasonal grass land in the north near Chad to moist, evergreen forest in the south near Zaire. In between, there are savannas, woodland systems, and drought-deciduous forest. Each circle represents a field site where soil/litter background spectral measurements were taken as well as some LAI measurements. Each cross represents a "site" (neighborhood) where data was gathered from the AVHRR for the inversion analysis. Each of the 15 sites initially contained 121 8x8 km pixels. Through the imposed requirement of spatial continuity as discussed above using the LSC, as many as half of the pixels may have been rejected. Two time periods were focused on: January (dry season) and September (wet season) of 1988. The inversion was performed to estimate the canopy parameters of each patch and for both time periods. Then, a series of forward integrations were applied using those parameters to generate fAPAR and albedo estimates.
Figure 1.3 shows some preliminary results. The top panel is folar fAPAR for the transect (the x-axis is the distance from the northernmost point). It is evident that in the dry season there is a sharp gradient beginning with patch number nine (approximately midway), from the sparse woodlands to the evergreen forest. In the wet season, the fAPAR is more or less constant along the transect, with a slight decline. Both these results are reasonable qualitatively as well as quantitatively, based on knowledge of these systems, but it is not known whether the small scale features are real or artifacts. Specifically, the slight decline in wet-season fAPAR going south could be due to the fact that the same fixed parameters and the same mode of inversion were used along the whole transect. The albedo results in the bottom panel are more difficult to interpret, but show season-to-season differences that are reasonable.
The preliminary results presented here indicate that inversion techniques have the potential to provide accurate estimates of surface properties that are physically and ecologically based. Continued development of this approach will include investigation of how to apply fixed parameters and parameter constraints on a biome-by-biome basis. More careful validation of the SAIL2 model for both forward and inverse modes is also planned. Eventually, this technique will be mature enough to produce time varying, two-dimensional fields of these quantities for use in terrestrial biogeochemical and climate modeling studies.