The short answer is "yes and extremely well" (but see the provisos below). This is not surprising considering the displacement of community type class "signatures" in VISNIR space apparent in the scatterplots for selected training sites. Model fits to observations are rather good and in this study mean RMSEs were <=0.03 reflectance, with a small standard deviation : the phenomena were well-described. The question of whether the adjusted or modelled reflectances are valid in a physical sense depends partly on the quality of the model inversion; partly on the noise penalty incurred in modelling to geometries other than those at which the data were originally acquired (as measured by the weight of determination); and partly on the ability of the model to explain rather than merely describe BRDF. If the model can be taken to explain in some degree the processes governing BRDF, then it is clear that the choice of modelling geometry will be conditioned by promotion of those geometries which provide better discrimination of and information on surface features (in this application primarily the plant canopy), in addition to consideration of the angular sampling distribution at the time of data acquisition. This latter point is extremely important; see the provisos below. Modelling existing scenes using retrieved parameters and the acquisition view and illumination geometry demonstrates the value and limitations of this type of BRDF model; differences between the modelled and original scenes are not significant in view of other sources of uncertainty (calibration, atmospheric correction, contamination, misregistration); see these examples for 3rd August (PM) and 13th August (AM).
Differentiation of grassland community types is greatly improved with
BRDF-adjusted or modelled VIS and NIR reflectances compared with using uncorrected
AVHRR reflectances. The extent of the improvement can be quantified for different
adjustment and modelling scenarios by evaluating separability between grassland
type classes in VISNIR space, calculated here in terms of transformed divergence.
In addition, the separability afforded through use of the model's anisotropic
parameters alone as signatures shows that useful surface information may be
obtained; see this
summary of transformed divergence results.
The greater separability obtained when modelling to higher solar zeniths with nadir viewing and vice versa (the model is reciprocal) can be explained in simple physical terms. As the sun zenith increases a number of phenomena occur which promote information on vegetation structure and optical properties over information on the soil background. First, a smaller proportion of the soil background is illuminated leading to a reduced contribution to reflectance; soil "noise" has long been recognized as major problem in remote sensing of vegetation, especially in arid zones where soils can be bright in both the visible and near-infrared wavelengths (e.g. Huete et al, 1991, 1992). Second, the proportion of shadows in the sensor's instantaneous field-of-view increases, further reducing the contribution of the soil background to reflectance. Third, the interaction of incoming radiation with the leaves and stems of the canopy changes, with a greater proportion of leaf facets normal to the direction of incoming radiation (at least for erectophile canopies such as those of grasslands).
Further work is clearly required to explore other applications which may become
possible with data from the AVHRR when BRDF models are used to apply BRDF
adjustments, produce modelled reflectance values for optimal geometries and provide
useful surface information in their parameters. The methods outlined here require
further validation over different semi-arid biomes, although since the spatial,
optical and structural properties of the soils and vegetation of these biomes are
rather similar the results are expected to remain promising. In particular,
modelling bidirectional reflectance scenes may provide a better approach to the
provision of multitemporal series of observations than the widely-used
Maximum-Value Compositing criterion. Finally, new sensors with better angular
sampling characteristics than the AVHRR planned for launch over the next few years
will allow improved sampling of the surface BRDF, providing even better potential
for robust inversions of BRDF models of the kind discussed here. The possibilities
remain to be explored.