Conference Agenda

Heavy rainfall regularly leads to local flooding in Germany, urban areas with a lot of sealed surfaces are particularly at risk. In Berlin alone, heavy rainfall has caused damage of €174 million between 2002 and 2021 according to the German Insurance Association. To enhance preparedness and reduce losses, advanced methods for nowcasting are vital for disaster management and emergency response. Typical factors such as water depth and flow velocity give a first indication on the magnitude of an event. But how can this information be translated into impacts such as the expected number of affected households or potential damage to buildings?

For the rapid impact assessment in a real-time framework, we developed the “Flood Damage Estimation Tool” (FlooDEsT), a machine learning algorithm trained on survey data from past heavy rainfall events. Using recursive partitioning, the algorithm captures non-linear relations between variables addressing the hazard and potentially affected objects. In its application, the trained model only requires the forecasted maximum water depth and velocity. Further, building characteristics are needed for a specific damage estimation. If missing, these variables are sampled randomly from the survey data distributions, while providing an uncertainty range. With a short running time, the model is applicable in real-time. Regardless of the property’s size or market value, relative qualitative damage classes are provided on the building level. Validation with external insurance data showed a more precise damage assessment for heavy rainfall events with the new model than with typical river flood damage models. FlooDEsT offers a fast identification of potentially affected objects and associated impacts to complement real-time hazard information and predictions for the near future. Based on the information output, impact-based warnings can be issued for a better understanding of the situation-related consequences within the population.

Additionally, regional available data can be added to the results for further information. For example, evaluating the affectedness of critical infrastructures and points of interest (e.g., fire brigades, hospitals), which are important for emergency response teams. By masking population density with the affected areas, the number of potentially affected people can be estimated, too.

Apart from the real-time application the model can be used for scenario-based risk analysis. Data aggregation and intersection with additional data can show regional hot spots of vulnerability. The model thus allows a prospect of impacts for regional planning in hot spot regions and the potential for increased resilience towards heavy rain events. These outcomes can therefore support risk communication and raise awareness.

Neural networks have demonstrated the capability to surpass process-based hydrological models in runoff prediction, achieving remarkable computational efficiency. However, these models typically lack representation of sub-catchment hydrodynamics and lateral water transport processes, such as streamflow routing. Conversely, routed semi-distributed hydrological models enhance spatial representation by partitioning the catchment into finer hydrological response units (HRUs), enabling more accurate simulation of spatial variability in land use, soil characteristics, topography, and river routing.
Recent case studies have indicated the potential of neural networks for semi-distributed river routing. Nevertheless, these methodologies are currently not scalable to larger regions due to their computational inefficiency, or require advanced engineering and substantial computational resources. Our goal is to develop a relatively simple yet accurate semi-distributed streamflow model for hydrological Switzerland with explicit routing capabilities, leveraging deep learning techniques. This end-to-end system aims to improve local runoff predictions and routed streamflow simulations, facilitating real-time, spatially explicit forecasts that can enhance water resource management and risk mitigation strategies.

Reliable flood hazard modeling is crucial for enhancing disaster preparedness, supporting decision-making, and protecting lives and properties. The rising frequency and intensity of flood events, exacerbated by climate change, demands high-resolution predictions that incorporate building-level details, quantify uncertainties, and can be rapidly updated in response to dynamic environmental changes. Current flood hazard models often struggle with slow computation times and limited resolution, hindering their effectiveness in real-time applications. To overcome these limitations, we present a novel AI-driven flood hazard modeling methodology that dramatically accelerates computation speeds, offering highly accurate results thousands of times faster compared to conventional numerical methods.

The AI model is trained with data generated by validated high-fidelity, physics-constrained numerical algorithms. Generating our own training datasets ensures continuous feeding of statistically relevant cases for optimal fine-tuning of the model. The achieved speed and accuracy remove the bottleneck of the computational cost, enabling several key enhancements in the flood prediction methodology. The most significant advancement facilitated by the AI model is the ability to perform real-time nowcasting with high-resolution spatiotemporal data. Additionally, the model effectively addresses the inherent uncertainty in flood relevant input data, a challenge that conventional numerical models have difficulty coping with due to extreme computational requirements. Given that typical flood forcings like weather forecasts and event sets are probabilistic, treating them as distributions rather than fixed points provides a more reliable representation of reality. Our approach tackles this by running multiple hazard simulations accounting for the full range of possible input scenarios, mapping the probability of specific weather events to their potential impacts on the ground.

Our AI-based flood hazard model is designed for seamless integration into existing workflows across multiple disciplines, including meteorology, disaster management, and cartography. This integration fosters operational synergies throughout all phases of disaster management—monitoring, forecasting, early action, preparedness, response, and recovery. By significantly reducing computation time and enriching information provided, our approach allows stakeholders to make quicker, more informed decisions, enhancing the overall effectiveness of disaster mitigation efforts.

By integrating AI with traditional hydrological and meteorological models, we provide a scientifically robust and practically applicable tool. This approach transforms complex data streams and impact forecasts into actionable insights, supporting cross-organizational information management. Our AI-based hazard modeling approach represents a substantial advancement in disaster risk management, offering a powerful tool that can be readily integrated into existing systems to improve preparedness, response, and recovery efforts.

Mangrove forests are vital ecosystems in Thailand, providing essential ecological services such as coastal protection, carbon sequestration, and biodiversity preservation. However, these ecosystems are under threat due to urbanization, industrialization, and tourism, particularly in Chonburi Province, where mangrove degradation has led to significant environmental challenges. Accurate species classification is fundamental to the restoration and sustainable management of these forests. This study integrates high-resolution multispectral satellite imagery from Pleiades Neo with machine learning (ML) models, specifically Random Forest (RF) and Decision Tree (DT), to classify five mangrove species:
Rhizophora mucronata (Rm), Rhizophora apiculata (Ra), Avicennia marina (Am), Avicennia alba (Aa), and miscellaneous species (Ms).
The research methodology involved the acquisition of six-band Pleiades Neo imagery, field data collection, and the application of ML algorithms. Advanced processing tools like ArcGIS Pro, ENVI, and Python libraries were employed to preprocess data and train the ML models. RF and DT were compared based on their classification accuracy, precision, and recall, with RF demonstrating superior performance. The RF model achieved an overall accuracy of 85.23%, outperforming DT's 82.37%. These results underscore RF's robustness and precision in handling highdimensional and heterogeneous datasets typical of mangrove ecosystems.
A detailed analysis of the classified mangrove areas revealed that RF provided more reliable results, especially for species such as Avicennia alba and miscellaneous species, where DT struggled. Table 1 below summarizes the specieswise classification accuracy of both models.

The integration of ML with satellite imagery offers transformative potential for ecological research and resource management. By providing accurate, scalable, and cost-effective methods for monitoring mangrove forests, the findings of this study can significantly contribute to the development of targeted conservation strategies. Furthermore, the successful application of RF highlights its suitability for ecological applications, particularly in diverse landscapes requiring precise species differentiation.