UAV-Based Terrain Modeling under Vegetation in the Chinese Loess Plateau: A Deep Learning and Terrain Correction Ensemble Framework

Abstract

Accurate topographic mapping is a critical task for various environmental applications because elevation affects hydrodynamics and vegetation distributions. UAV photogrammetry is popular in terrain modelling because of its lower cost compared to laser scanning. However, this method is restricted in vegetation area with a complex terrain, due to reduced ground visibility and lack of robust and automatic filtering algorithms. To solve this problem, this work proposed an ensemble method of deep learning and terrain correction. First, image matching point cloud was generated by UAV photogrammetry. Second, vegetation points were identified based on U-net deep learning network. After that, ground elevation was corrected by estimating vegetation height to generate the digital terrain model (DTM). Two scenarios, namely, discrete and continuous vegetation areas were considered. The vegetation points in the discrete area were directly removed and then interpolated, and terrain correction was applied for the points in the continuous areas. Case studies were conducted in three different landforms in the loess plateau of China, and accuracy assessment indicated that the overall accuracy of vegetation detection was 95.0%, and the MSE (Mean Square Error) of final DTM (Digital Terrain Model) was 0.024 m.