Speaker: Ass. Prof. Konstantin Klemmer (University College London)
Earth observation data capture the natural and built environments of our planet. They come in various forms, ranging from satellite imagery to in-situ measurements. These data can support decision-making processes in both the private and public sectors, for example, by improving flood resilience or forecasting crop yields. However, analyzing these vast amounts of data is often challenging, as they are frequently unstructured, unlabeled, and multimodal. The presentation introduces a new class of representations specifically developed for Earth observation data: Earth Embeddings. Approaches to Earth Embeddings combine the scalability of neural networks with geospatial expertise and established knowledge from geodesy and geostatistics. Earth Embeddings are universal descriptors of location information that capture natural and physical characteristics. They provide powerful features for downstream modeling and can help address long-standing challenges in geospatial machine learning.