Neural Network Model for Extraction and Validation of Hydrographic features in Alaska Using Ifsar Data

Abstract:

High-resolution digital elevation and optical image data, such as light detection and ranging (lidar) and interferometric synthetic aperture radar (ifsar) data collected by the U.S. Geological Survey 3D Elevation Program, are becoming more readily available. These data enable more precise hydrologic modeling and collection of vector-based hydrographic features than what was possible with earlier, lower-resolution datasets. However, the extraction of hydrographic features by modeling flow accumulation with a high-resolution (1- to 3-meter cell size) digital elevation model (DEM) entails the challenges of properly conditioning DEMs and tailoring effective upstream area thresholds for stream formation. Recent work with machine learning has shown promising results for extraction of hydrography and other features from lidar point cloud and other remotely sensed data. In this work we test the capability of the U-Net convolutional neural network model to extract hydrographic waterbody polygons and drainage lines in Alaska using ifsar-derived elevation and intensity data. Preliminary results using 5-meter resolution ifsar data on fourteen HUC-12 watersheds in Northern Alaska, ranging from 55 to 147 square kilometers, show feature extraction accuracies from 59% to 89%, averaging 70% correct. Aside from the DEM, the U-Net model uses layers for curvature, topographic position index, openness, geomorphon terrain class, sky view factor, and shallow-water channel depth, all derived from the ifsar DEM. In subsequent work, we evaluate the U-Net model performance with additional layers for radar intensity, topographic wetness, flow accumulation, and the digital surface model. Methods include significance tests to estimate the probable contribution of each layer to the U-Net model. Efficient implementation of the U-Net model is expected to assist the validation of hydrographic features and improve the accuracy of hydrography databases.