Forecasting the future: AI in early warning systems

Opportunities and challenges for AI-based warnings of complex climate risk 

Markus Reichstein, Vitus Benson

Climate change presents growing challenges for disaster management, necessitating robust Early Warning Systems (EWS) to protect human societies. These systems are essential for sustainable development and must overcome difficulties in hazard forecasting, risk communication, and decision-making efficiency. This introduction discusses the integration of Artificial Intelligence approaches, in developing Multi-Hazard Early Warning Systems (MHEWS). By combining meteorological and geospatial data, these AI-driven systems aim to enhance prediction accuracy, but also can help with intuitive interfaces and community feedback.  

Optimizing the WFP Humanitarian Aid Supply Chain 

Giulia Martini

Given the high unpredictability of WFP operations and global supply chains, it is important for WFP to take effective  decisions rapidly and assess all possible scenarios to address key questions at a global scale, such as where to strategically store commodities, when to procure food and how/when to best use limited resources to reach as many people in need as possible. 

SCOUT is an optimization system that helps answer these questions by consolidating WFP’s demand, supply and network data and providing quantitative insights on key planning decisions along the entire supply chain, from suppliers to recipient countries. 

In 2023, SCOUT supported the global planning of specialized foods that boost nutrition in children and women, and enable, for example, to avoid potential interruption of distribution to more than 170,000 people in need in Eastern and Western Africa. It is also instrumental in the planning the procurement of approximately 300,000 tons of commodities a year in West Africa, with a projected cost avoidance of US$ 574,000 in sorghum purchasing alone. In 2024, SCOUT will be used for all WFP operations and commodities, greatly amplifying the efficiency of WFP’s global supply chain.

Causality and Knowledge Integration for Understanding Complex, Dynamic Systems  

Joachim Denzler 

The interplay of causal understanding and integrated knowledge is crucial in comprehending and predicting the behavior of complex, dynamic systems. The presentation will cover state-of-the-art methodologies and applications that leverage causality and domain knowledge to enhance our understanding of dynamic systems, including those from ecology and climate science. We will discuss novel data-driven models that integrate physics-aware methodologies with machine learning techniques, such as Dynamic Mode Decomposition with Control (DMDc), to analyze ecosystem respiration dynamics under varying climatic conditions. Additionally, the talk will cover the impact of plant diversity on soil properties and its subsequent influence on ecosystem stability, focusing on the causality between biodiversity and soil thermal diffusivity using Physics-Informed Neural Networks. By merging causal inference with empirical data and knowledge, we not only want to predict system behaviors but also to inform sustainable management and policy decisions. Open future research directions are robust, scalable models that can handle the complexity and unpredictability of natural systems.

Cloud-Optimised, ML-based rainfall forecasting for Early Warning Systems over the Horn of Africa

Shruti Nath 

Effective early warning systems are paramount in any disaster risk management framework. The Strengthening Early Warning Systems for Anticipatory Actions (SEWAA) initiative,  funded by Google over multiple years, seeks to assess the skillfulness and sustainability of creating a cloud-based machine learning post-processing system. This system aims to enhance the accuracy and reliability of high-impact weather and climate forecasts. Our objective is to demonstrate the generation of forecasts utilizing machine learning methodologies and to explore their practical applications in anticipatory actions.

Geospatial vegetation forecasting for anticipatory action

Vitus Benson

High resolution satellite imagery allows monitoring and forecasting vegetation status at the scale of individual fields. Within the EarthNet initiative we train deep neural networks to forecast vegetation greenness into the future conditional on weather scenarios. This allows for a new type of  early warning system, focusing on the land surface impacts of droughts and heat waves instead of their atmospheric drivers. Now, we are exploring the extensions of geospatial vegetation forecasting to anticipatory action, for instance for pastoralist communities in the horn of Africa.

Disentangling the increasing Complexity of Extreme Heatwave Risks in a warming Climate

Kai Kornhuber

Driven by anthropogenic Climate change, heatwaves have become increasingly record breaking, occur simultaneously around the globe and sequentially over a specific region posing new and unexpected risk to natural and societal systems, such as global supply chains, disaster response and the global food system. In this talk I will provide an overview of recent advances in benchmarking state of the art climate models in reproducing observed trends and heatwave interrelationships. In addition, preliminary results and pathways on how to further our understanding of heatwave processes that govern record shattering extremes and their early detection using machine learning methods will be discussed. 

Optimizing the WFP Humanitarian Aid Supply Chain

Giulia Martini

Given the high unpredictability of WFP operations and global supply chains, it is important for WFP to take effective  decisions rapidly and assess all possible scenarios to address key questions at a global scale, such as where to strategically store commodities, when to procure food and how/when to best use limited resources to reach as many people in need as possible. SCOUT is an optimization system that helps answer these questions by consolidating WFP’s demand, supply and network data and providing quantitative insights on key planning decisions along the entire supply chain, from suppliers to recipient countries.

Cloud-Optimised, AI-based rainfall forecasting for Early Warning Systems over the Horn of Africa

Shruti Nath

Effective early warning systems are paramount in any disaster risk management framework. The Strengthening Early Warning Systems for Anticipatory Actions (SEWAA) project is a multi-year initiative with partners from UN World Food Programme, University of Oxford, the Horn of Africa climate prediction and assessment centre (ICPAC), and local meteorological departments in Kenya and Ethiopia (KMD and EMI respectively). The project seeks to assess the skillfulness and sustainability of setting up a cloud-optimised, AI-based forecasting system for use in operational forecasts over the Horn of Africa region. In particular, we seek to provide improved rainfall forecasts that can be continuously updated on a user-friendly interactive web platform. Ultimately, we hope that this will facilitate timely and accurate information dissemination for Early Warning Systems. In this talk, we will showcase the cloud infrastructure set up for training the AI model for operational forecasting use, we furthermore will compare the model developed in this project to other AI weather models (GraphCast and Fuxi), finally we will show the interactive web-platform hosted by ICPAC that provides AI based rainfall forecasts.