Assessment of vector hydrography derived from deep learning of remotely sensed data

Keywords: deep learning, U-net, National Hydrography Dataset, digital elevation model
Session Type: Virtual Paper Abstract
Day: Saturday
Session Start / End Time: 2/26/2022 02:00 PM (Eastern Time (US & Canada)) - 2/26/2022 03:20 PM (Eastern Time (US & Canada))
Room: Virtual 3

Abstract

Recent fire and flood hazard events highlight the importance of capabilities to monitor surface water features such as those included in the United States National Hydrography Dataset (NHD). Flow routing and other terrain analysis techniques are available to derive vector hydrography from elevation data, but invariably these methods require manual adjustments and validation through costly field surveys or tedious visual comparisons with high resolution imagery. Advanced applications of machine learning with elevation and optical image data have shown promise for automated extraction of detailed hydrographic features in raster form. Automated conversion of raster predictions to topologically correct vector features presents a complex problem and a topic for this research. We use a deep learning U-net neural network model on high-resolution elevation data (5-m in Alaska, and 1- or 3-m elsewhere) and optical image data (0.5- to 1-m) to predict hydrographic features based on training data from the best available vector hydrographic features. Lidar and interferometric synthetic aperture radar (IfSAR) data from the United States 3-D Elevation Program (3DEP), along with optical image data are used for U-net modelling. A series of operations convert the raster predictions to polygonal and flowline network features in vector form. We constrain several flow routing techniques for extracting drainage networks from elevation to follow U-net predictions, are compare these to reference data. Preliminary results indicate vector network extraction informed by deep learning improves on traditional network extraction and resolves connectivity issues common to raster hydrographic predictions.