Bi-directional NDVI and atmosphere-coupled BRDF inversion

Abstract

The normalized difference vegetation index (NDVI) has been widely applied in optical remote sensing. It has been demonstrated that NDVI is still partially affected by atmospheric path scattering and bi-directional (illumination and viewing geometry) effects. We present a feature of using bi-directional NDVI. Based on the assumption that a clear day has a larger NDVI value than a dusty day (smaller atmospheric path scattering in near infrared band and larger atmospheric path scattering in red band), we used the square ratio of observed NDVI values and expected NDVI values as weights for the observations. The initial weights for each observation are calculated by the ratio of the observed NDVI value and mean NDVI value of all bi-directional observations. The inversion process will loop until all weights converge, while the expected NDVI values are calculated from the previous loop's model prediction. Our preliminary research on the early Terra/MODIS data using the semi-empirical kernel-driven bi-directional reflectance distribution function (BRDF) model (RossThick-LiTransit) shows that this new method can improve the inversion results when some of the cloudy pixels are not filtered out. As sub-pixel cloudiness is always a problem, this technique should still be very useful even as cloud detection and atmospheric correction get better.