7 - 10 May 2024| ESA-ESRIN
Thematic Areas
The Destination Earth (DestinE) is a flagship initiative of the European Commission (EC) to develop a highly accurate digital model of the Earth on a global scale. This model will monitor, simulate and predict the interaction between natural phenomena and human activities. It will contribute to achieving the objectives of the twin transition, green and digital as part of the EC’s Green Deal and Digital Strategy.
DestinE delivers a continuous flow of data along with tailored services supporting users in accessing and exploiting these data, deriving information, and implementing and operating their own services. DestinE initiative is jointly implemented by ESA, ECMWF, and EUMETSAT and is based on the following functional decomposition of the DestinE system:
• DestinE Digital Twin (DT) Engine (DTE): an engine capable of providing a common system approach to a unified orchestration of Earth-system simulations, delivering data from digital replicas of the Earth through the fusion of observations with models.
• DestinE Data Lake (DEDL): data access harmonisation of DT data and federated providers such as ESA, EUMETSAT, ECMWF, Copernicus, and many other sources. Big data processing capabilities provided to allow computing in proximity to the data.
• DestinE Service Platform (DESP): point of entry to DestinE. It provides users, applications and service providers with direct access to the data and functionalities provided by the other two DestinE components (DTE and DEDL).
The AI/ML contribution to DestinE cut across all domains; therefore, this TA welcomes for example, presentations on (i) data dimensionally reduction to speed up transfer and reliability, (ii) examples of operational benchmark training datasets for atmospheric sciences, and (iii) traceability tools applied on academic/policy repositories and social media to detect and assess the usage of open-source assets (e.g., [benchmark] datasets, AI/ML trained models, images, journal papers, etc).
In recent years, there has been a growing interest in using mixed machine learning models that combine text and image data. In the field of Earth Observation, researchers are exploring the integration of natural language processing (NLP) and deep learning techniques to analyse satellite images. By combining NLP with convolutional neural networks (CNNs), recurrent neural networks (RNNs), and CNN-LSTM architectures, researchers can pose textual questions and obtain answers using the information in satellite imagery. This approach enables the detection of land cover changes, tracking of wildfires, monitoring of urban growth, and identification of environmental changes. These mixed models allow for a comprehensive understanding of the Earth's dynamics, supporting decision-making in areas such as environmental management, disaster response, agriculture, and urban planning. By leveraging the advancements in mixed ML models, researchers can extract valuable insights from Earth Observation data, leading to more informed decision-making and a deeper understanding of our planet. This TA welcomes the ML4ESOP community to show recent advances in this field and to demonstrate how this approach can became a tool for decision makers.
We solicit contributions in the field of Machine Learning (ML) to replace or enhance components of the standard Numerical Weather Prediction (NWP) and Climate workflow. ML developments can replace parts of the workflow, such as forecast models and data assimilation, or provide end-to-end solutions from observations to forecasts on NWP and Climate timescales. ML-based forecast models, trained on historical data, aim to improve forecast accuracy by capturing complex relationships in weather and climate data. ML techniques applied to data assimilation can enhance the assimilation process by incorporating diverse sources of data. Additionally, researchers are exploring end-to-end ML solutions that integrate observations, data pre-processing, model training, and forecast generation. The goal is to advance the field by leveraging ML to improve forecast skill, enhance data assimilation, and streamline the NWP and Climate workflow. This TA welcomes contributions and aims to stimulate discussion between domain experts in data assimilation/NWP/Climate and computer/data scientists interested in the application of ML/DL to weather forecasting and climate prediction.
Hybrid approaches in NWP and climate prediction (e.g., hybrid DA-ML, hybrid forecast models) combine Machine Learning (ML) with traditional physics-based methodologies to enhance forecasting capabilities. These approaches integrate ML into data assimilation and forecast models, aiming to improve accuracy, capture complex processes, and address computational challenges. Challenges include reliability, interpretability, and data quality. Overall, hybrid approaches offer promising opportunities for advancing weather forecasting and climate projection. This TA aims to illustrate the state of the art in the interpretable application of ML/DL to specific components of the NWP and climate prediction workflow, highlight opportunities for further progress and discuss current challenges.