Seminar: Claire Robin

Vegetation activity can vary over just a few meters, producing heterogeneous responses to climate extremes. The launch of Sentinel-2 and other decameter-resolution satellites with sub-weekly revisit times now allow to monitor this fine-scale variation. However, their short observational records limit the detection of extreme vegetation anomalies, as a few years of data do not capture the full range of variability.Here, we reduce this uncertainty using an eco-cluster approach, which groups locations with similar phenological patterns to increase the effective sample size for threshold calculation. We validate the method via cross-sensor comparison with MODIS, aligning Sentinel-2 data temporally and spatially. Results show that the eco-cluster approach halves median errors at the 10% percentile and increases cross-sensor co-detection rates by 43%.Applied at Sentinel-2 resolution, our method reveals that vegetation extremes are highly localized (75% occur in <30% of a MODIS pixel, ~0.02 km²), longer-lasting (median 48 vs 32 days), and more intense than observed at MODIS scale. By reducing threshold uncertainty, this approach enables reliable detection of vegetation extremes from recent higher resolution resolution satellites, revealing their spatial extent, intensity, and duration.