Rapid identification of disaster hotspots by means of a geospatial information fusion from remote sensing and social media

Kurzfassung

Effective management of complex disaster scenarios relies on achieving comprehensive situational awareness. Recent disasters, such as the 2024 floods in Southern Germany, have highlighted the critical need for timely geoinformation to protect communities. During the response phase, it is vital to rapidly identify the most affected areas to guide emergency actions and allocate limited resources effectively. This process is typically iterative, incorporating continuous updates as new or improved information becomes available. Initial estimates, often based on incomplete or imprecise data, play a crucial role in forming an early situational overview before detailed damage assessments are conducted. Early-stage proxies, such as population distribution and hazard zones, can support planning data collection efforts, enhancing situational understanding and focusing response efforts efficiently. This study introduces a method for rapidly identifying disaster hotspots, particularly in scenarios where detailed damage assessments or very high-resolution satellite imagery are not (yet) available. The approach leverages the H3 discrete global grid system and employs a log-linear probability pooling method with an unsupervised hyperparameter optimization routine. It integrates flood hazard data from systematically acquiring high-resolution satellite imagery (Sentinel-1 and Sentinel-2), disaster-related information from X (formerly Twitter), and freely accessible geospatial data on exposed assets. The methods effectiveness is assessed by comparing its outputs to detailed damage assessments from five real-world flood events (USA August 2017, Mozambique 2019, Mexico November 2020, Germany July 2021, Pakistan September 2022). Results demonstrate that disaster hotspots can be identified using readily available proxy data. An extensive hyperparameter analysis revealed that while equal-weight methods offer simplicity and effectiveness, optimized pooling weights generally yield superior results. Context-specific tuning was shown to be critical for optimal performance in log-linear pooling. Notably, an unsupervised method minimizing the Kullback-Leibler divergence between input distributions and predictions outperformed supervised approaches, overcoming the limitations of training data. This method’s transparency and adaptability allow it to incorporate geospatial layers with varying resolutions and semantic relevance, making it particularly suitable application to other hazards (e.g., landslides, wildfires, earthquakes) or exposed assets (e.g., roads, railways, critical infrastructure).