Sea Grant coastal adaptation and resilience: AI-enabled urban flood monitoring systems

Abstract

Coastal cities across Alabama and Mississippi face mounting flood risks due to a convergence of climatic, geographic and infrastructural vulnerabilities. Urban centers such as Mobile, AL, and Gulfport and Biloxi, MS, are particularly exposed, with aging stormwater systems, rapid development and low-lying terrain intensifying their susceptibility to both extreme and nuisance flooding. These flood events are further exacerbated by increases in severe storms, placing vulnerable coastal populations at heightened risk. While traditional hydrologic and hydraulic (H&H) models have formed the backbone of urban flood analysis, their efficacy is limited in dense, heterogeneous environments. Despite advances in high-resolution topographic data and high-performance computing, urban flood simulations remain constrained by the absence of timely, spatially distributed validation data, which is essential for ensuring model accuracy and informing real-time decision-making.

To address this critical gap, the proposed project integrates advanced deep learning (DL) with stationary street-view camera feeds to develop a novel, automated visual analytics pipeline for flood detection and depth estimation. By leveraging the latest breakthroughs in computer vision and environmental feature extraction, our approach enables near-instantaneous detection of urban flooding conditions, improving emergency response times and reducing the analytical burden on human operators. This innovation transforms previously underutilized infrastructure into a near-real-time flood sensing network that captures hyperlocal flood dynamics, especially valuable in areas where traditional sensors are sparse or nonexistent. The pipeline will be further integrated with physical flood models to provide context-aware flood exposure analysis capabilities, enhancing both short-term response and long-term planning.

In parallel, the project includes a community-centered design process, developed in close collaboration with the City of Mobile’s engineering staff, to ensure outputs align with local needs and capacity. This approach emphasizes not only technical excellence but also practical usability, and long-term sustainability. The automated system will improve the visibility and documentation of "nuisance" flooding — frequent, low-level events that disrupt transportation, access to services, and local economies but often go unrecorded. By bringing these underreported impacts to light, the project empowers decision-makers and residents to better understand and address their day-to-day flood vulnerabilities.

This proposal directly aligns with the priorities outlined in the Mississippi-Alabama Sea Grant Consortium’s 2026–2027 Research Funding Opportunity. It contributes to the “Resilient Communities and Economies” focus area by assessing coastal residents’ risk awareness and identifying practical solutions to enhance flood resilience. It also supports the “Environmental Literacy and Workforce Development” focus area by translating cutting-edge flood detection research into applied community benefit. The research outcomes will include an operational prototype, a public-facing dashboard, and documentation to support adoption by local governments across the Gulf Coast and beyond. By merging artificial intelligence, near-real-time sensing, and community partnership, this project delivers a transformative model for urban flood monitoring and preparedness, helping coastal cities adapt to a future of increasing hydrologic uncertainty.

Objectives

This project seeks to establish an integrated, AI-enabled urban flood monitoring and advisory system by pursuing three interrelated objectives. First, it aims to develop and validate machine learning algorithms for near-real-time flood detection using image data from street-view cameras placed at high-risk locations across Mobile, AL. These deep learning models will be trained to recognize flood indicators such as water pooling and street submergence under varying urban, lighting and weather conditions, ensuring reliable detection with minimal false alarms. Validation will involve labeled image data and ground-truth verification by the City of Mobile’s engineering staff.

Second, the project will construct and calibrate a city-scale stormwater drainage model using EPA's Storm Water Management Model (SWMM), incorporating detailed infrastructure data provided by the city — such as drainage networks, topography and land use — along with high-resolution rainfall reanalysis and hydrodynamic boundary conditions from Mobile Bay. Calibration will rely on local hydro-meteorological data and expert input to simulate urban stormwater dynamics accurately. 

Third, the outputs of the machine learning detection system will be integrated with the SWMM model to form a unified decision-support framework. This fusion will enhance the timeliness and spatial precision of flood warnings, allowing city officials to convert raw data into actionable flood response measures. While initial deployment will be limited to select areas of Mobile due to camera availability, the framework is designed for future scalability and replication in other coastal cities (i.e. Biloxi and Gulfport, MS). The final deliverable will be a proof-of-concept system demonstrating how artificial intelligence can be embedded in municipal operations to improve flood risk mitigation and public safety.

Methodology

This project presents a comprehensive, multi-modal methodology that fuses machine learning (ML), high-frequency street-view camera imagery, and hydrologic and hydraulic (H&H) modeling to develop an AI-enabled urban flood monitoring and advisory system tailored for coastal urban settings like Mobile, AL. The approach consists of three integrated components. First, ML algorithms will be developed to detect flood onset and estimate water depth from live camera feeds positioned at flood-prone intersections, identified in collaboration with the City of Mobile. These deep learning models — including convolutional neural networks (CNNs) and transformer-based architectures — will be trained using a combination of historical flood footage and synthetically augmented images (via GANs and style transfer) to ensure robustness under variable weather, lighting and visibility conditions. Beyond water segmentation, the system will infer flood depth through 3D scene reconstruction, leveraging geometric cues and multi-object segmentation, validated against in-situ measurements and engineering reports. 

Second, the project will construct a physically-based stormwater drainage model using the U.S. EPA’s Storm Water Management Model (SWMM), informed by Mobile’s existing infrastructure datasets. Missing network elements will be reconstructed using slope-based interpolation and DEM-aligned drainage tracing. The model will be calibrated and validated using historical flood data, rainfall reanalysis products (e.g., AORC, ERA5), and estuarine hydrodynamic boundary conditions derived from PI Moftakhari’s previous estuarine dynamics models. Third, an integrated AI-driven framework will combine real-time ML detections with predictive SWMM simulations to support flood advisory generation. This system will incorporate an active learning loop, enabling expert-in-the-loop refinement of ambiguous flood images to enhance accuracy while minimizing manual labeling. Initial deployment will focus on high-risk areas in Mobile, with modular architecture designed for scalability and future replication in other coastal cities such as Biloxi and Gulfport, MS. All code, data pipelines, and models will be openly shared via PI-hosted repositories to maximize reproducibility, transparency, and community uptake.

Rationale

Urban flooding poses an escalating threat across the Gulf Coast, particularly in Alabama and Mississippi, where coastal cities like Mobile, Gulfport, and Biloxi face compounded risks from low-lying topography, outdated stormwater infrastructure, and socioeconomically vulnerable populations. As climate change increases the frequency of severe storms, these cities experience more frequent and disruptive flood events. Traditional hydrologic and hydraulic (H&H) models, though foundational, struggle in urban environments due to surface heterogeneity, data gaps, and limited real-time validation capacity. Even with high-resolution LiDAR and advanced computing, urban flood modeling remains constrained without timely, spatially distributed observational data. Existing validation methods—such as satellite imagery, UAVs, and crowdsourced photos—are hindered by technical, logistical, or labor-related barriers. To overcome these challenges, this project proposes an automated, deep learning-powered visual analytics pipeline that uses fixed street-view cameras to detect flood onset and estimate water depth in real time. This system significantly reduces reliance on manual analysis and provides emergency managers and city planners with rapid, reliable flood intelligence. By integrating ML-based flood detection with physical drainage models and community-driven design, the project offers a scalable, adaptive, and locally relevant flood monitoring framework. It directly supports Mississippi-Alabama Sea Grant Consortium (MASGC) goals by advancing environmental literacy and workforce development, improving community resilience, and addressing the vulnerabilities of at-risk populations. Specifically, it enhances risk awareness, provides actionable data to decision-makers, and elevates visibility of frequent but underreported “nuisance” flooding that disrupts daily life. Through cutting-edge AI technology and practical community engagement, this initiative represents a vital step toward strengthening flood preparedness and adaptive capacity across the Gulf Coast’s most flood-prone urban areas.