to: 18 Jun 2021




The advantages of Machine Learning/Deep Learning (ML/DL) techniques have been seen in a wide range of applications such as image recognition, traffic prediction, self-driving vehicles, and medical diagnosis. These techniques have also gained popularity within the Earth System Observation and Prediction (ESOP) community due to their ability to improve our understanding and prediction capabilities on the Earth’s complex and wide-scale dynamics. Together with the increase in computing power, these techniques are valuable to automatically process and analyse a large scale of available data but they still present some limitations, as an example of DL methods that need large amounts of curated and labelled data. As a consequence, they have become a common language between academia and industry across several Earth Observation sectors. Therefore, one of the goals of the workshop is to share all domain experiences from the recent progress and synergies between ML/DL (combined with conventional tools) for satellite observations, weather and climate models, and post-processed model outputs.

Enhancing satellite observations has proved to be an ML/DL ability by fusing multiple resolutions (spatial, temporal, and spectral) with different sensor designs (satellite, drone, ground-based instruments). This super-resolution task improves satellite imagery precisions enabling detailed land cover classification, vegetation monitoring in complex landscapes, and comprehensive urban mapping.

Typical Data Assimilation (DA) approaches to solve optimisation problems and minimise model error estimations are often conceptually equivalent to modern ML/DL algorithms. In this case, the question is how to explore Hybrid models (i.e. DA combined with ML/DL) which are able to draw from the strengths of both technologies?

Geophysical Forecasting and Post-Processing are topics where ML/DL solutions have also been seen as a potential success, mainly in forecasting the future state of a system based on its current and past states. The statistical nature of ML/DL models enable them to represent probability distributions rather than individual instances. ML/DL techniques are useful to increase the quality of a non-expert system in a post-processing scheme. As an example, an algorithm can be trained to identify cases where a physical model is likely to be inaccurate and learn to correct its prediction.