Assessing the generalization of geospatial embeddings for temporaltransfer in land cover classification

Nowadays, remote sensing through Earth observation (EO) satellites is the preferred means of mapping land cover at various scales, from the regional to global level. While obtaining satellite data on most surface areas is more and more democratized as there are several publicly available sources, the ground truth (GT) data gap still prevents the frequent update of land cover maps or the production of historical maps in certain regions, notably in Sub-Saharan Africa. It is common for GT to be available only for a few specific years over study sites, and directly applying a model learned on a particular year to other mapping years often fails because of the data distribution shifts caused by different acquisition conditions and environmental or climatic factors. Unsupervised domain adaptation (UDA), a specific case of transfer learning, has been examined in recent years to address data distribution shifts, allowing the temporal transfer in land cover mapping of a model learned in a specific mapping year with GT data available to other mapping years without GT, using satellite image time series (SITS). In the meantime, the availability of massive unlabeled EO data and recent advances in self-supervised learning have enabled the development of geospatial foundation models and the release of geospatial embeddings such as those of AlphaEarth Foundations (AEF) and TESSERA. Although they offer a practical way to temporal transfer as they are devised to encode temporal dynamics and are fixed year-to-year vector representations, their potential for temporal transfer remains under explored. In this work, we proposed to assess their generalization for temporal transfer scenarios in land cover mapping, using a multi-year open dataset collected in Koumbia, Burkina Faso. First, we evaluated the AEF and TESSERA embeddings through a direct temporal transfer approach, and then through the UDA paradigm. Our results notably highlight the better fit of geospatial embeddings on Sentinel-2 SITS for temporal transfer tasks and the competitive performances obtained with TESSERA with respect to AEF embeddings.