STURM-FloodDepth: A deep learning pipeline for mapping urban flood depth using street-level and oblique aerial imagery

Flooding remains one of the most frequent and damaging natural disasters, intensified by climate change and urbanization. High-resolution real-time flood depth observations at the urban scale remain spatially sparse, thus alternative data sources are required to support risk assessment and emergency response. This study introduces STURM-FloodDepth, a deep learning pipeline to estimate and map urban flood depths using street-level and oblique aerial imagery. The workflow consists of two sequential modules: A. flood depth estimation, proceeding through vehicle detection (YOLO-World and SAHI), contextual cropping, super-resolution enhancement (EDSR), and flood level classification (fine-tuned ResNet-50); and B. georeferencing and mapping, proceeding through orthographic reference image construction, feature matching (SuperGlue), homography estimation (RANSAC), geospatial projection and mapping, conversion and export to GeoJSON. The classifier achieved AUC values ranging from 0.78 to 0.98 across all classes. Real-world qualitative validation confirmed its accuracy in operational conditions. STURM-FloodDepth is a modular, scalable, sensor-agnostic tool for urban flood monitoring, with applications for urban resilience, disaster management, and smart city. The framework is released as an open-source tool to foster further research and operational deployments.