home.social

#spatialgeneralizability — Public Fediverse posts

Live and recent posts from across the Fediverse tagged #spatialgeneralizability, aggregated by home.social.

  1. Data-Driven Pluvial Flood Modeling For Urban Regions - An Investigation Of Spatial Generalizability
    --
    doi.org/10.1016/j.envsoft.2026 <-- shared paper
    --
    cbc.ca/news/canada/montreal/mo <-- shared media article
    --
    climateinstitute.ca/news/fact- <-- shared technical media article
    --
    H/T @Oveys Ziya | PhD Candidate at McGill University
    “[The authors have] developed a machine learning-based framework for rapid prediction of urban pluvial flood depth and velocity, with a particular focus on spatial generalizability. Using hydrodynamic simulations for Montreal [Canada], [they] demonstrated the framework's ability to provide accurate and computationally efficient predictions for unseen rainfall scenarios and locations.
    Importantly, the framework maintained strong predictive performance when transferred to an unseen city (Ottawa) without retraining.
    [Their] findings highlight the potential of data-driven approaches for real-time flood warning systems and urban resilience planning…”
    --
    “HIGHLIGHTS:
    • A framework is developed for fast pluvial flood depth and velocity prediction.
    • U-Net outperforms Random Forest and XGBoost across all test scenarios.
    • A physics-informed reference layer improves accuracy and spatial transferability.
    • The models generalize to spatially unseen urban regions without retraining.
    • The framework enables rapid and transferable prediction of urban flooding.
    ABSTRACT: Accurate and rapid flood prediction is essential for effective urban flood risk management, yet physics-based hydrodynamic models are computationally prohibitive for real-time or ensemble simulations. This study presents a machine learning-based surrogate modeling framework leveraging Random Forest, XGBoost, and U-Net to provide fast and reliable predictions of pluvial flood depth and velocity with generalizability to unseen hyetographs and locations. A key innovation is the introduction of a physically consistent reference depth/velocity input, which embeds hydraulic context and enhances transferability. Using hydrodynamic simulations for Montreal Island as training, validation, and testing data, the U-Net model consistently outperformed other models, achieving the lowest RMSEs for flood depth (0.093 m) and velocity (0.130 m/s) in test regions, and maintaining robust performance (RMSEs of 0.114 m and 0.151 m/s) when transferred to an unseen city (Ottawa). The presented framework offers significant potential for integration into real-time pluvial flood warning systems and urban resilience planning…”
    #Surrogatefloodmodeling #Urbanpluvialflooding #Machinelearning #Deeplearning #UrbanFlooding #SpatialGeneralizability #MachineLearning #DeepLearning #FloodModeling #flood #flooding #urban #pluvial #extremeweather #climatechange #Canada #Montreal #Ottawa #floodwarning #risk #hazard #cost #economics #humanimpacts #earlywarning #model #modeling #hydrodynamic #simulation #framework #floodprediction #riskmanagement

Share on Mastodon

Enter the server where you have an account.