A cluster-based approach for global phenological monitoring

We propose and implement a new cluster-based approach for global phenological monitoring in which, instead of individual pixels or rectangular regions, phenologically and climatologically self-similar pixel clusters are monitored. We developed clusters based on the 1982-1999 global Pathfinder Advanced Very High Resolution Radiometer (AVHRR) Land (PAL) Normalized Difference Vegetation Index (NDVI) dataset, a global 10-minute resolution climatology, and the clustering approach developed by the Oak Ridge National Laboratories (ORNL). In the ORNL approach: (1) n cluster centers are defined based on the multi-dimensional NDVI/climate space; (2) pixel distances from the centroids are calculated; (3) pixels are assigned to the minimum distance cluster. While any number of clusters may be specified, we found that a global 500-cluster approach provided a satisfactory global distribution. In traditional rectangular approaches a group of pixels could contain desert, grassland, and tropical forest. Here, longitudinally extensive but latitudinally limited regions such as the Sahel exist as distinct groups. Thus, our approach avoids problems affecting single-pixel approaches (misregistration, cloud contamination) and rectangular approaches (mixed phenological signals). Based on the pixel clusters, we then used a spatial compositing approach based on a logistic regression filter and a version of the BISE algorithm to obtain a single time series for each cluster. We then extracted phenological metrics such as the onset and offset of greenness. Global phenological patterns and strategies for ground validation are presented and discussed.