Further Enhancement of Satellite DEM Resolution and Accuracy Using Machine Learning and Remote Sensing Data

Abstract

AbstractHaving a high-resolution and high-accuracy Digital Elevation Model (DEM) is essential for flood modelling to increase the reliability of the resulting flood analysis and flood maps. However, it is quite a challenge to acquire highly reliable DEM as it is often confidential, time consuming and/or costly. This paper presents the latest continuous effort to further enhance the satellite DEM using 0.25 m resolution RGB (Red, Green and Blue) color values from Google Earth in addition to the remote sensing data from Sentinel-2 multispectral imagery. Shuttle Radar Topography Mission (SRTM) satellite is selected to demonstrate the ongoing satellite DEM enhancement effort; also a very high surveyed spatial resolution of 0.25 m is used in this study. Artificial Neural Network (ANN) is trained with 0.25 m reference DEM which were obtained with Lidar (Light Detection and Ranging). ANN is first trained with reference DEM of a training area and then applied to other areas for validation. The validation result shows that ANN generates a much clearer view of the area and result in a significant error reduction (RMSE) of about 30%.KeywordsMachine learningDigital elevation modelRemote sensing