Exploiting the multi-angularity of the MODIS temporal signal to identify spatially homogeneous vegetation cover: A demonstration for agricultural monitoring applications

AbstractMODIS has been providing daily imagery for retrieving land surface properties with a spatial resolution 250m since the year 2000. In many places, this pixel size is closer to that of individual landscape elements, such as managed forest stands or crop fields, than those of time series more typically used in vegetation analyses (with pixel sizes ranging from about 1km to 8km). With such spatial resolution, combined with its increasingly long archive, MODIS data offers great potential for vegetation monitoring applications in general, and crop growth monitoring in particular. However, due to its whiskbroom design, the observation geometry of the MODIS instrument, combined with the spatial uncertainty in the registration of the images, can result in different (albeit neighbouring) physical areas being mapped onto the same pixel depending on the view zenith angle (which varies from day to the next). Rather than considering this as an inconvenience, a method is here proposed to exploit this peculiarity to identify pixels corresponding to a homogeneous plant cover, in order to retrieve surface specific time series of satellite products. This method is based on quantifying the temporal signal-to-noise ratio (SNR, hereafter) of the daily MODIS NDVI time series, defined as the variance of smoothed temporal signal over the variance of the residues. If consecutive observations of the same pixel (which have thus sampled the spatial vicinity of that pixel) provide similar NDVI values, the resulting temporal signal is relatively stable and the SNR is high. Such stability can indicate that the signal comes from a spatially homogeneous surface, such as a single large field covered by the same crop with similar agro-management practices. On the contrary, a noisy time series indicates instead a transition zone between different land uses or between fields with different management practices. SNR maps therefore serve as a proxy for sub-pixel homogeneity from which time series originating from a specific land cover or land use can be retrieved. This approach is demonstrated over 12 contrasting agricultural landscapes across the globe from which clearly distinctive crop specific signals are extracted. Exploiting the full MODIS archive to derive surface specific information in this way should open new avenues for regional to global agricultural monitoring applications. Expanding this method to derive satellite products for specific land cover classes could also be useful for many other applications linked to dynamics of land cover and land use change