I recap the functioning of our fully self-supervised learning pipeline based on the spectral-temporal Barlow Twins. The SSL approach generates highly informative representations at 10m spatial resolution from cloud-corrupted optical time series. The resulting representations are well correlated with GEDI-derived relative height measurements so that an AGB model for vegetation/forest of up to 300-500 t/ha can be derived. I show that the model transfers well between years making it possible to train the model on (for example) one year of Sentinel-2 data together with the corresponding GEDI measurements, and applying the frozen model to Landsat data acquired in previous years. Bio: 2010-2023: Full Professor for Remote Sensing/Geomatics - since 2016 Lead of Mantle's research team
