Exploring spatiotemporal phenological patterns and trajectories using self-organizing maps

Consistent satellite image time series are increasingly accessible to geoscientists, allowing an effective monitoring of environmental phenomena. Specifically, the use of vegetation index time series has pushed forward the monitoring of large-scale vegetation phenology. Most of these studies derive key phenological metrics from the Normalized Difference Vegetation Index (NDVI) time series on a per-pixel basis. This paper demonstrates an approach to analyze synoptic spatiotemporal phenological patterns over large areas, rather than per pixel. The selected approach involves data mining using a self-organizing map (SOM) and Sammon's projection. To illustrate our approach, we trained a SOM using 13 years of ten-day NDVI composites from the Système Pour l'Observation de la Terre-VEGETATION over the Kruger National Park, South Africa. This resulted in a topologically ordered set of phenological synoptic states. The Sammon's projection was then used to create a simplified representation of the trained SOM that reflects the similarities among the synoptic states. Subsequently, we depicted phenological trajectories for each vegetation season to show how phenological development changes between years. This time series data mining approach provides a holistic characterization of the main regional phenological dynamics and effectively summarizes the information present in the time series, thus facilitating further interpretation.