In this presentation, participants will discuss the above topics, with some emphasis on their applicability to early warning systems
Earth observation and climate data tend to be extremely large and detailed. Sometimes they are noisy but often also hard to interpret. Using data-driven algorithms to complement traditional processing schemes holds great promise to speed up the creation of actionable insights, speed up the development of novel applications, and improve the quality of the output compared to existing algorithms.
However, accounting for the data gravity of earth-observation data, data access and training schemes like distributed computing, federated learning, and general filtering and sharing of data across borders must be employed.
Furthermore, compared to the volume of available data itself, high-quality annotations are either expensive or not available in abundance.
The latter has led to the adoption of concepts used by large language models, but adapted to geospatial data, resulting in geospatial foundation models. Trained on very large volumes of earth observation data, these models can be adapted to various downstream tasks, e.g. for floods or wildfires, with a minimal number of labels, while generalizing across continents.
In this presentation, participants will discuss the above topics, with some emphasis on their applicability to early warning systems.