Correction of Low Vegetation Impact on UAV-Derived Point Cloud Heights With U-Net Networks

Abstract

This study presents an approach to the problem of minimizing the impact of low vegetation on the accuracy of a UAV-derived DEM, based on the use of a deep neural network (DNN). It is proposed to use the U-Net network to determine corrections to the height of the raw point cloud so that the processed data reflect the actual earth’s surface. The implemented solution is therefore based on regression, not classification. As a result of the proposed processing method, the expected value of the land surface height is determined for each point of the unified point cloud. In addition, a second U-Net network is trained, enabling the uncertainty of the corrected heights of the land surface to be determined for each point of the unified cloud. The training set includes data from different seasons, which makes the models more resistant and allows for assessment of the impact of the season and more generally the related vegetation status on the model accuracy. The processing results can be used in DEM generation, and also for determining the vertical displacements of the terrain surface associated with underground mining, as well as natural phenomena such as landslides. A key advantage of the proposed processing method is the ability to predict the uncertainty of the results.