The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences
Publications Copernicus
Articles | Volume XLVIII-1/W2-2023
13 Dec 2023
 | 13 Dec 2023


O. Grothum, A. Bienert, M. Bluemlein, and A. Eltner

Keywords: Soil erosion measurement, point cloud processing, vegetation filtering, deep learning, random forest, pointnet++

Abstract. Various soil erosion processes can be quantified using digital elevation models (DEMs) of difference. In this study, cameras were used to capture images of bare soil during artificial rainfall simulations. The photos were then used to generate dense 3D point clouds with millimeter resolution using Structure-from-Motion and Multiview-Stereo (SfM+MVS) techniques. However, the point clouds also captured some vegetation, such as agricultural plants, grass, and weeds, present at the soil surface. It had to be removed to accurately measure soil erosion processes. The removal can be done manually and is hence time-consuming. In this study, several methods have been tested and compared to perform (semi-)automatic vegetation filtering from point clouds of soil surfaces. First, the point clouds were labelled into vegetation and ground data to establish a basis set for the subsequent experiments. Then, three branches of algorithms were tested. The first branch considers knowledge-based thresholding. Thereby, unique features were considered, e.g., point color, height, and roughness within a specified neighborhood. For instance, a threshold was set in the color space to separate green vegetation from brown soil considering the green band. The second branch used a machine learning (ML) algorithm to classify each point as vegetation or ground by automatically finding thresholds. Thereby, again features such as point color were used. In addition, multi-scale features were computed for each point to characterize it in the context of its neighborhood. The calculated features were used afterwards with the random forest (RF). The third branch considered an end-to-end learning approach and thus avoiding the necessity to define features. The deep learning-based architecture PointNet++ was used. For the classification, an adapted model for soil surfaces and vegetation was trained. The performance of the different methods was compared, and an assessment of each method was provided. Overall, the study aims to find a (semi-)automatic method to remove vegetation from time series of soil surface point clouds to achieve an accurate measurement of soil surface changes and thus eventually erosion processes while minimizing manual effort and time consumption.