Leveraging Machine Learning for Crisis Propagation Forecasting and Early Warning of Stakeholders in Urban Areas

Kurzfassung

Rapid urbanization and escalating frequency of natural disasters pose significant challenges for urban environments, demanding timely and precise crisis response strategies. Current crisis management systems typically separate crisis forecasting from citizen warning systems, creating dependencies that hinder timely communication, as shown by the 2021 German floods. Furthermore, existing warning solutions are often generic and insufficiently personalized, reducing their effectiveness for individual users. Additionally, systems enabling two-way citizen communication frequently rely on unreliable crowdsourced data, such as social media, compromising crisis response accuracy and reliability. These shortcomings indicate a need for systems capable of real-time forecasting and personalized communication to protect citizens. To overcome these challenges, an integrated grid-based crisis warning system combining Machine Learning (ML) and Natural Language Processing (NLP) technologies is introduced. The framework is composed of three core stages: A Convolutional Neural Network (CNN)-based imputation model to handle incomplete data from regions without urban sensors and citizen reports, a Convolutional Long Short-Term Memory (ConvLSTM) forecasting model to predict crisis evolution, and a Retrieval-Augmented Generation (RAG) system to deliver personalized warnings. The crisis warning system was evaluated through extensive synthetic storm scenario analyses in Darmstadt. By transforming incomplete crisis patterns into a coherent spatiotemporal representation, the CNN estimates affected urban areas, enabling the ConvLSTM model to forecast crisis dynamics with top-1 accuracy reaching 86% in scenarios with sufficient historical context. The ConvLSTM’s performance significantly improves with increased historical context (i.e., more past timesteps), underscoring the importance of longer contexts for better forecasting. Subsequently, the RAG-based crisis advisor generates reliable, actionable advisories aligned explicitly with official guidelines from the Bundesamt für Bevölkerungsschutz und Katastrophenhilfe (BBK), minimizing misinformation risks and hallucination. The crisis advisor provides situational warnings and supports two-way communication, allowing citizens to interactively seek clarification and receive customized guidance relevant to their personal circumstances. Results for the crisis advisor demonstrated effective performance and high reliability in crisis communication for the storm scenarios (faithfulness ≥0.98) indicating high factual consistency of generated advisories. In summary, this thesis contributes to urban crisis management by integrating crisis forecasting with citizen-centric, context-specific warning and communication to enhance public safety.