Predicting Building Damage using Object-Based Fusion of Hydrological Forecast and Multi-Modal Remote Sensing

Kurzfassung

Floods rank among the most devastating natural hazards, yet current risk assessment frameworks remain fragmented, with hydrological forecasting and damage estimation largely treated as separate problems. While remote sensing has enabled accurate post-event damage mapping and hydrological models provide increasingly sophisticated hazard forecasts, the translation from predicted hazard to expected impact at the building scale remains a critical challenge. This study addresses this gap by developing a data-driven framework for forecasting flood-induced building damage through the integration for heterogeneous geospatial predictors. High-resolution exposure and vulnerability information derived from remote sensing are combined with coarse-scale hydrological forecast variables within a machine learning framework. A systematic ablation study is conducted to quantify the contribution of distinct predictor domains - hazard, exposure, and vulnerability - and to evaluate their interactions in determining damage outcomes. Gradient-boosted decision trees are employed to capture complex, non-linear relationships while retaining interpretability. Model performance is rigorously assessed using nested cross-validation to ensure robust generalization across spatially heterogeneous flood events. Feature attribution based on SHAP values enables transparent identification of dominant predictors and supports model simplification. The proposed approach achieves reliable building-level damage prediction, with a maximum F1-score of 0.80, demonstrating that meaningful impact forecasts can be derived despite substantial discrepancies in spatial resolution between predictor domains. Results consistently show that predictive skill depends on the joint integration of hydrological forcing, geomorphological context, and building-specific characteristics. Building height, terrain elevation, and runoff-related indicators emerge as the most influential predictors across model configurations. These findings demonstrate that data-driven models can effectively brigde the gap between hazard prediction and impact assessment. The presented framework provides a scalable foundation for anticipatory flood risk modelling, while highlighting key limitations related to data availability, class imbalance, and transferability that must be adressed to enable operational deployment.