Authors: Parvaneh Nikrou – University of Alabama
Title: Transferable Deep Learning Flood Inundation Mapping Trained on LISFLOOD-FP Simulations
Presentation Type: Poster Presentation
Abstract: Flood inundation mapping is essential for disaster preparedness, early warning systems, and risk management. While traditional hydrodynamic models like LISFLOOD-FP offer high accuracy, they are often computationally demanding and time-consuming. In recent years, machine learning—particularly deep learning models such as Convolutional Neural Networks (CNNs)—has gained traction as an alternative for efficient flood extent prediction. This study presents a deep learning-based approach in which a CNN model is trained using outputs from the LISFLOOD-FP model, high-resolution Digital Elevation Models (DEMs), and Land Use/Land Cover (LULC) data. Simulations were performed for three case studies using eight synthetic hydrographs per site, generated by scaling unit hydrographs with the 500-year peak discharge (Q500), allowing for variations in flood magnitude and shape. The trained model was subsequently evaluated on two unseen case studies to assess its ability to generalize. Results show that the CNN model accurately replicates LISFLOOD-FP flood extents, demonstrating strong spatial agreement. The key contribution of this work lies in the demonstrated transferability of the trained model across different flood events and geographic regions, underscoring its potential for scalable, near real-time flood mapping applications.