Vegetation Removal From Airborne LiDAR Point Clouds Via Deep Learning Theory

Digital Elevation Model(DEM)is an important geographic information product for elevation expression,it expresses the elevation of ground points in a discrete digital manner at a given coordinate.Considering its convenience for surveying and mapping,it has been widely used in both scientific research and practical production,and its high-precision obtainment has always been a hot research topic for many photogrammetry and remote sensing researchers.Airborne light detection and ranging(LiDAR)provides helpful data source for the high-precision obtainment of Digital Surface Model(DSM)and DEM,as this active remote sensing technology can efficiently gain a great amount of high-precision and high-density 3D point clouds for models' establishment-DSM,which represents terrain surface(including vegetation and buildings),can be directly produced from airborne LiDAR point clouds.Nevertheless,for the production of DEM,those point clouds which represents ground objects' surface should be filtered out from massive raw airborne LiDAR point clouds in the first place.Accordingly;,the quantity and quality of ground points are the key to the DEM's accuracy.Unfortunately,in many mountain areas,the ground surface is covered with thick vegetation,and the laser beam of LiDAR can hardly penetrate the vegetation layer.That is to say,the obtained ground points for DEM production after filtering are extremely limited.It is this lack that makes a great trouble for high-precision DEM obtainment.Traditionally,in this situation,the acquisition of high-precision DEM mainly depends on both manual measurement and interpolation.Actually,despite of the desirable effectiveness of current interpolation algorithms,it still strongly depends on existing ground points.That means they may be useless and unfeasible when facing serious lack of ground points[9].On the other hand,manual supplement to ground points based on topography and empirical knowledge is too laborious and less objective.Hence,more automatic methods for DEM production are needed.Deep learning algorithms are powerful and popular methods for learning data's representative and discriminative features and have displayed a great improvement on learning algorithms' intelligence and automation,hence,they have already exhibited good performance in fields like object classification and signal processing.Considering that DSMs can be directly obtain from airborne LiDAR point clouds,to achieve the automation of vegetation removal,a supervised method for DEM production can be proposed.Here,a robust deep learning architecture can be obtained by continuously learning the correlations between DEM and its DSM.Finally,with the DSM data of a mountain forest area and the obtained architecture,the DEM of this area can be directly predicted.In terms of deep learning algorithms and their architecture,two typical deep learning architectures,Stacked Autoencoder(SAE)and Convolutional Neural Network(CNN),are brought in vegetation removal and production of high-precision DEM tasks for mountain forest areas.The main innovations of this paper are as follows:(1)We propose a SAE-based vegetation removal method from a signal processing perspective,and this method enables the production of high-precision DEM from airborne LiDAR point clouds in mountain forest areas.(2)We also propose a CNN-based vegetation removal method from image super-resolution perspective,and a symmetric marginal padding process is brought in this method to offset the border loss of input matrix caused by convolution.Experiments on Chinese Fujian and Hainan dataset validate the effectiveness and automation of our methods for high-precision DEM production.