AI-Driven Geospatial Data Assimilation for Urban Flood Forecasting

Thesis Abstract

Urban flooding remains a persistent risk in densely populated coastal cities, where rapid urbanization, climate variability, and aging drainage infrastructure interact to amplify inundation episodes. The study addresses the inadequacy of traditional flood forecasting systems, which often rely on standalone hydrological or hydraulic models and fail to assimilate diverse real-time data streams. The aim is to develop and validate an AI-driven geospatial data assimilation framework that fuses multi-source observations with high-resolution urban flood models to provide??, probabilistic forecasts at street-scale. Specific objectives include (1) to design a data assimilation architecture that integrates radar rainfall, satellite-derived soil moisture, crowd-sourced flood reports, in situ gauge measurements, and high-resolution digital elevation models within a unified geospatial platform; (2) to implement machine learning-enhanced parameter estimation and state augmentation for urban hydrodynamic models; (3) to evaluate forecast skill against observed inundation extents using metrics such as critical success index, Brier score, and Nash-Sutcliffe efficiency under multiple scenarios; (4) to assess robustness to data missingness and sensor outages through Monte Carlo uncertainty quantification; (5) to translate probabilistic forecasts into decision-support tools for emergency services and urban planners. The methodology adopts a mixed-methods design anchored in the Bayesian data assimilation paradigm, drawing on theories of operational geoinformatics and the theory of information entropy to justify adaptive weighting of heterogeneous data streams. The population comprises urban drainage networks, sensor arrays, and satellite/remote-sensing assets in three metropolitan areas with historical flood records. A stratified sampling approach will select 12 representative acquisition sites per city and 6 urban blocks with documented inundation events, yielding a dataset of approximately 216 sensor-day observations over two wet seasons. Data collection instruments include Doppler radar rainfall products (0.5 km spatial resolution, 5-minute updates), Sentinel-1 and Sentinel-2 imagery, ground-based rainfall and stage gauges (n?60 stations), crowd-sourced reports from municipal flood-reporting apps, and high-resolution LiDAR-derived digital elevation models (?1 m). The analytical workflow employs (i) a geospatial data assimilation engine that couples a 2D urban hydrodynamic model with an Ensemble Kalman Filter (EnKF) and a neural network surrogate for rapid state estimation, (ii) convolutional neural networks for feature extraction from radar and satellite data, (iii) Gaussian process regression for spatiotemporal interpolation of sparse observations, and (iv) a probabilistic post-processing layer to generate flood probability maps and time-to-flood forecasts. Validity and reliability will be ensured through cross-validation on historical flood events, sensitivity analyses of initialization and localization parameters, and benchmarking against baseline hydrodynamic models. The study expects to reveal that AI-integrated assimilation markedly improves lead times and spatial accuracy of flood forecasts, with expected improvements in critical success index by 15–25% and Brier score reductions of 0.05–0.10 on average across test cases, relative to conventional methods. It anticipates quantifying uncertainty via ensemble spread and validating decision-relevant metrics such as outage duration and exposed population estimates. Contributions to knowledge include (a) a scalable framework for real-time geospatial data assimilation in urban hydrology that leverages multimodal data streams and deep learning surrogates, (b) methodological advancement in integrating probabilistic forecasting with operational decision-support in municipal contexts, and (c) empirical evidence on the value of AI-enabled assimilation for resilience planning under climate-driven hydrological extremes. The study concludes that AI-driven geospatial data assimilation substantially enhances urban flood forecasting capability and provides actionable, uncertainty-aware guidance for emergency response, urban drainage design, and land-use planning. Recommendations include expanding sensor networks, developing standardized data-sharing protocols with municipal agencies, and establishing continuous model retraining pipelines to adapt to evolving urban landscapes and climate conditions.