ESA-ECMWF Workshop: "Machine Learning for Earth System Observation and Prediction"



Organisation: European Space Agency (ESA), European Centre for Medium-Range Weather Forecasts (ECMWF)

Workshop Motivation and Description

Machine/Deep Learning (ML/DL) techniques have revolutionized numerous fields and have proven to be particularly advantageous in various applications such as image recognition, traffic prediction, self-driving vehicles, and medical diagnosis. These techniques have garnered significant attention and adoption within the Earth System Observation and Prediction (ESOP) community due to their ability to enhance our understanding and prediction capabilities of the Earth's complex dynamics. In line with this, the workshop aims to delve into the progress and synergies between ML/DL and conventional tools in satellite observations, weather and climate models, and post-processed model outputs.

One prominent area where ML/DL techniques have proven invaluable is in the development of highly fidelity digital models of the Earth on a global scale. These models serve as comprehensive monitoring, simulation, and prediction systems that enable us to analyse and forecast the intricate interactions between natural phenomena and human activities. By providing a holistic understanding of the Earth's dynamics, these models contribute to the achievement of the European Commission's Green Deal and Digital Strategy goals towards a green & digital transition (Twin transition).

ML/DL solutions have also showcased promising advancements in weather forecasting and climate prediction. Algorithms can be trained to identify instances where physical models may exhibit inaccuracies and subsequently learn to correct their predictions accordingly. Moreover, AI-based models have the potential to create hybrid forecast models that combine the strengths of traditional, physics-based NWP/Climate prediction methodologies with the capabilities of ML/DL, ultimately enhancing the accuracy and reliability of predictions.

In recent years, novel computing schemes have emerged, encompassing High-Performance Computing (HPC), edge computing, and quantum computing (QC). Quantum and edge computing paradigms represent significant shifts in technology and hold immense potential for the field of Earth Observation (EO). HPC enables the execution of complex algorithms and simulations at high speeds, empowering scientists, and researchers to tackle large-scale challenges in EO. Complementing HPC, edge computing brings computational capabilities closer to the data sources, enabling faster processing and analysis of data, particularly in scenarios where real-time decision-making is crucial. Additionally, QC emerges as a disruptive force in the realm of computing, offering new opportunities for tackling complex optimization problems, enhancing machine learning algorithms, and unlocking novel insights from vast EO datasets. The advent of QC in EO holds promise for revolutionizing various aspects of data analysis and processing, further augmenting the capabilities of ML/DL techniques.

Researchers are actively exploring the use of mixed models that can effectively work with both text and images, including in the field of Earth Observation. These models demonstrate the capability to generate results by answering textual questions related to Earth Observation data, opening up possibilities for various applications and domains such as environmental monitoring, land use analysis, disaster management, and urban planning. By combining textual information and visual data from satellites and other Earth observation platforms, these models have the potential to enhance our understanding of the Earth's dynamics and support decision-making processes in areas such as climate change mitigation, natural resource management, and urban development.

Workshop Aims

The workshop aims to explore the fusion of traditional ESOP techniques with machine learning (ML) and deep learning (DL) methods. It seeks to showcase the impact achieved through this fusion, while also addressing the remaining challenges that need further exploration. The presenters will show their contributions to this field and engage the attendees in discussions to provide a comprehensive understanding of the subject. The workshop strongly welcomes industry to demonstrate their commercial lenses for ML4ESOP applications.

The workshop is planned to run over four days, during which the state-of-the-art in ML/DL techniques across various domains applied to ESOP will be covered. The oral presenters are expected not only to discuss their own work but also provided an overview of the subject. The poster presenters on the other hand, will deep dive into the topics of the thematic areas, having the opportunity to discuss one-to-one or one-to-many during this session. Following this, parallel working groups will be formed to delve deeper into the limitations of the status quo and propose strategies for advancing the fusion to extract more value from it. These parallel working groups play a crucial role in facilitating discussions focused on the main issues that will be identified during the workshop. These groups aim to investigate deeper the challenges, propose potential solutions or approaches, and determine future directions in ML4ESOP. The main workshop output is to produce a report which would serve as valuable contributions to the field.