Examining CNN terrain model for TanDEM-X DEMs using ICESat-2 data in Southeastern United States
Eric Guenther, Lori A. Magruder, Amy Neuenschwander, Donald Maze-England, J. Dietrich
Abstract
Accurate large-area Digital Terrain Models (DTMs) are crucial for many science applications. Spaceborne Synthetic Aperture Radar (SAR) platforms are often used to create these DTMs as they provide an effective tool to collect surface elevations across a wide extent. However, SAR-derived digital elevation models (DEMs) cannot accurately measure ground elevations in the presence of forests. This work demonstrates an approach to estimate terrain elevations from 12 m TanDEM-X by using a convolutional neural network (CNN) trained with ground elevations from ICESat-2 – a spaceborne laser altimeter. This approach demonstrated the ability to estimate terrain elevations from TanDEM-X DEMs for the greater North Carolina area. The CNN estimated terrain saw an improvement in RMSE from 11.28 m to 4.42 m within the entire area of interest, and a focused improvement in RMSE from 12.78 m to 4.95 m in forested areas when compared to ICESat-2. The CNN model outperformed linear, random forest, and gradient boosted regression models using comparable model inputs. This work combines 12-m TanDEM-X data with ICESat-2 profiles, resulting in a new DTM product with accuracy approaching that of reference elevations obtained from satellite laser altimetry in the southeastern United States.