DESIGN AND RESULTS OF AN AI-BASED FORECASTING OF AIR POLLUTANTS FOR SMART CITIES

This paper presents a modular approach for extracting disaster-relevant data from web news using machine learning, rules, and dictionaries. Extracted entities are geolocated and classified into a flood disaster cycle framework, enabling fine-grained spatiotemporal analysis in synergy with geospatial data. The West Germany flooding event 2021 serves as a case study. While rule-based limitations persist, the approach demonstrates strong potential for disaster management applications.

For Bavaria, Germany, we developed the first region-wide data-driven landslide susceptibility map, training a neural network on over 11,000 landslide events using terrain, geological, and land cover data. The model shows strong predictive performance, with geology and soil properties as the most influential features. Validation against newly recorded landslides shows that inventory incompleteness and selection bias translate directly into susceptibility underestimation in poorly mapped regions.