Smart GIS-Based Urban Flood Risk Forecasting System with AI

Thesis Abstract

This study addresses the growing challenges of urban flood risk in rapidly urbanizing landscapes by integrating Geographic Information Systems (GIS) with artificial intelligence (AI) to deliver a real-time, decision-support system for forecasting flooding events and guiding resilience actions. The objective is to develop and validate a smart GIS-based urban flood risk forecasting system that leverages multi-source data, advanced machine learning, and spatiotemporal analytics to produce accurate short- to medium-term flood predictions and actionable risk maps for municipal planners and emergency responders. Specific objectives include (1) to assemble a multi-criteria dataset comprising high-resolution rainfall records, digital elevation models, land-use and impervious surface maps, storm surge indicators, historical flood extents, and sensor data from local hydrological stations; (2) to design a data fusion pipeline that harmonizes heterogeneous data streams for real-time processing; (3) to develop and compare AI-driven forecasting models, including Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTM) networks, and gradient boosting methods (XGBoost), for predicting flood inundation extents and probability of exceedance at the neighborhood scale; (4) to implement a GIS-driven visualization framework that generates dynamic risk maps and decision-support dashboards for alerting and resource allocation; and (5) to evaluate system performance under varied rainfall scenarios and hydraulic conditions using a rigorously defined validation protocol. The methodological approach adopts a mixed-methods, applied design anchored in the theoretical perspectives of the Ensemble Learning and Spatial-Temporal Modeling framework, alongside the Resource-Based View to justify organizational uptake of AI-enabled capabilities. The population comprises urban flood-prone municipalities with dense impervious surfaces and available historical flood data. A stratified random sample of 12 neighborhoods within the study city, representing diverse topographies and land-use patterns, will be selected to develop and test the forecasting models. Data collection will integrate (a) historical flood extents and inundation records from municipal archives (covering 15 years), (b) high-resolution rainfall data from meteorological stations and radar-derived rainfall estimates, (c) LiDAR-derived digital elevation models at 1-meter resolution, (d) recent land-use and imperviousness maps, and (e) sensor readings from a network of hydrological and rainfall gauges deployed or accessible within the city. Instruments include a data fusion pipeline, GIS-based visualization toolkits, and a suite of AI models implemented in Python using libraries such as TensorFlow/Keras, scikit-learn, and PyTorch. Validity and reliability will be established through cross-validation, back-testing with hold-out historical events, and sensitivity analyses of model hyperparameters. Data analysis will proceed in three stages (1) descriptive and exploratory analyses to characterize baseline flood risk patterns and data quality; (2) model development and evaluation through time-series forecasting and spatial-prediction tasks, using metrics such as Nash-Sutcliffe Efficiency, Root Mean Squared Error, AUC-ROC, and F1-score; and (3) scenario-based impact assessment to simulate rainfall intensities and urban drainage responses, followed by qualitative expert reviews of the decision-support outputs. The study will also implement feature importance analyses and partial dependence plots to interpret AI model decisions and ensure model transparency in line with explainable AI practices. Ethical considerations include data privacy for any sensitive urban infrastructure information and ensuring that the forecasting outputs support equitable risk communication among diverse communities. Expected findings anticipate that AI-augmented GIS forecasting will yield higher predictive accuracy for inundation extents and exceedance probabilities than traditional hydrological models, with improvements in lead times of 30–90 minutes under moderate to extreme rainfall events. The integration of real-time data streams is expected to produce near real-time risk maps with spatial resolutions at the neighborhood level, enabling more precise resource deployment and targeted evacuation guidance. The study will contribute to knowledge by demonstrating a replicable framework for smart city flood risk forecasting that combines GIS, AI, and multi-source data fusion, advancing the methodological integration of spatial analytics with machine learning in urban resilience. It will offer practical recommendations for municipal agencies on data governance, system deployment, and stakeholder engagement to enhance flood preparedness and response. The main conclusion posits that a GIS-centric AI forecasting system can substantially augment urban flood risk management, provided there is sustained investment in data infrastructure, inter-agency collaboration, and continuous model recalibration; recommendations include expanding the sensor network, establishing data-sharing protocols with neighboring jurisdictions, and integrating forecast outputs into emergency operation center (EOC) workflows and public communication channels.