AI for EO & Earth Science
AI is one of the fastest developing domains, with an overwhelming majority of businesses looking to adopt AI and analytics to improve their operations. Machine Learning and Computer Vision are sub-domains that are gaining a lot of traction from a wide range of industries, with a huge weight put in by the automotive, communications and multimedia. In Earth Observation, there are now being explored alternatives to the computationally expensive and sometimes sub-optimal classical image information content retrieval methods, with the advent of 2-D and 3-D CNNs among other methods, with a clear technology push from multimedia. While EO is one of the main data sources for Earth System Science, the adoption in Earth Sciences is moving at a different pace. The complexity of the phenomena, involving interactions, teleconnections, processes that occur at different spatio-temporal scales, without a directly applicable scalability of physical models, makes it central that AI methods for Earth Science be designed in full agreement with and complemented by Earth System Models, integrating physical constraints. The challenges of interpretability, data preparation (for training as well as validation) and the 5Vs of Big geospatial data are common threads for EO as well as Earth Science challenges. This session will look into these topics, addressing the commonalities and specifics of the EO and Earth Sciences. The session is supported by The European Lab for Learning and Intelligent Systems (ELLIS), in particular the Machine and Climate Science Programme section. The session will be organized around key presentations that aim to push forward the fronteers of EO data exploitation in a data-intensive framework, to advance capability for better, faster and more reliable knowledge discovery for emerging Earth and Climate Science challenges: climate change, anthropogenic effects, geohazards, etc.
- Data-driven dynamic modelling and forecasting in Earth System Science (Markus Reichstein/ Miguel Mahecha)
- Spatio-temporal detection, prediction and attribution – Climate Anomalies and Extreme events (Markus Reichstein/ Miguel Mahecha)
- Data for AI, AI for Data – challenges and solutions to benchmarking datasets (XiaoXiang Zhu / Devis Tuia )
- Linking Physics and ML Models, Learning and Explaining feature representations for Earth System Science (Gustau Camps-Vals)
- The signal and the noise – dealing with errors and uncertainty in EO & ESS data and models (Volker Markl / Devis Tuia)
- Multimedia & EO – finding the sweet spot between domains (Fredrik Heintz)