BRDF Applications in Semiarid Grassland Monitoring with the AVHRRs

Why not just use MVC with NDVI?

The widely-adopted maximum-value compositing procedure whereby the sample with the greatest NDVI is selected from a short series of observations has been shown to be problematic : Yang et al, 1996; Stoms et al, 1997; Moody and Strahler, 1994; Gallo and Huang, 1998; Zhu and Yang, 1996; Goetz, 1997; and (of course) Cihlar et al, 1994. In order to investigate the effects of applying the MVC_NDVI criterion to reflectances from the 17 NOAA-PM orbits used in the Inner Mongolia study, signatures were extracted from a MVC_NDVI composite for the ten training sites, as before. The transformed divergence values obtained were very high : a best average separability and a best minimum separability of 1,840 and 439 were recorded, respectively. Although not quite as high as the transformed divergence obtained through BRDF modelling, this would seem to indicate that MVC is a useful technique in processing NDVI from "1km" AVHRR datasets*.

However, an examination of the distribution of points in VIS-NIR space and of the view and solar angles resulting from the preferential selection of maximum NDVI samples show that the data are biased towards extreme geometries (note that where viewing zeniths are low, solar zeniths tend to be high). This is because less soil is seen and/or illuminated relative to vegetation at high view and/or sun zeniths and means that when plotted in VIS-NIR space, distinct classes are clustered into two regions.

The main conclusion is that quantitative use of AVHRR bidirectional reflectance data from composites derived through the MVC_NDVI criterion is compromised. However, it is not so straightforward : part of the aim of MVC_NDVI is to select near-nadir observations in order to maximize spatial resolution and minimize spatial distortion (although it doesn't consistently do this). Here we are presented with a dilemma : we can have either inconsistent data with a higher spatial quality (through some scheme for selecting near-nadir observations) or consistent data with a lower spatial quality (through inversion of BRDF models). If we want the former, the techniques are becoming rather sophisticated through the use of precise renavigation models; outside such techniques, there is no hope of obtaining image data with a ~1.1km resolution. With a scanline of 2,048 samples, only a small proportion will have a viewing zenith close to nadir (Earth curvature). It therefore seems preferable to aim for high consistency in the data dimension rather than high spatial resolution.


*of course the samples are almost never exactly 1km or even the 1.1km described in the NOAA Polar Orbiter Data Users' Guide - angular sampling means that much larger surface areas contribute to off-nadir observations.


Previous page | Next page | Index | Disclaimer
1