Authors: Tantu Mandal, Hongxing Liu, Sagy Cohen, Dan Tian – The University of Alabama; Lei Wang – Louisiana State University
Title: Improving River Channel Morphological and Hydraulic Representation within the OWP Hydrofabric using High-Resolution SAR, LiDAR, and Deep Learning
Presentation Type: Poster Presentation
Abstract: Accurate delineation of river channels and reliable estimation of their morphological and hydraulic properties are essential for developing hydrofabric datasets that support hydrologic modeling and flood forecasting. These capabilities also play an important role in sediment transport analysis and ecosystem monitoring. Although hydrographic data have advanced in recent years, many hydrofabric components still rely on static surveys. Such datasets often lack the spatial resolution and temporal variability required to represent channel geometry and flow dynamics in applications such as flood inundation mapping.
This study presents an automated framework for extracting river channels and associated geomorphic features from Synthetic Aperture Radar (SAR) imagery using deep learning and computer vision methods. We employ the Segment Anything Model 2 (SAM2) and fine-tune it using a curated training dataset developed through adaptive thresholding and rigorous quality control. The model is guided by point prompts and applied at the basin scale to segment river networks from Sentinel-1 imagery.
The initial segmentation outputs are refined through post-processing steps designed to reduce noise and improve boundary definition. The resulting binary masks are transformed into vector representations of channel banks and centerlines. Centerlines are derived using distance transforms and medial axis techniques and are subsequently smoothed to produce geometries suitable representation for hydrologic analysis. These representations enable the estimation of key channel attributes such as width curvature, sinuosity, and braiding indices through computer vision approaches.
The framework is applied to the Mobile River Basin and evaluated against manually delineated channels derived from high resolution PlanetScope imagery. Results demonstrate that the proposed approach produces consistent and detailed channel representations at large spatial scales. This work highlights the potential of integrating SAR data and deep learning into hydrofabric development workflows and supports improved hydraulic parameterization and flood modeling within OWP systems.