Development of a feature-based semantic segmentation algorithm for separating terrestrial and non-terrestrial surfaces

Recent advances in remote sensing technology make it possible to digitize the real world almost automatically. Airborne laser scan results are georeferenced data type. They provide detailed 3D information about objects and the environment. Automated classification and detection of objects obtained from lidar is necessary to minimize production costs. Although the optimization of traditional methods using rule-based algorithms has expanded geospatial applications, significant manual editing is still required to obtain a high quality data set. Unlike images, point arrays are unstructured, sparse, and have a non-standard data format. This creates a lot of challenges, but it also provides a huge opportunity to capture the details of scanned surfaces with millimeter accuracy. Classifying and separating non-ground points from ground points greatly reduces the amount of data required for consistent surface analysis, saving time and simplifying further analysis. The main idea of scientific research is to use deep learning as a section of machine learning to analyze an array of points. The paper presents a feature-based algorithm that classifies ground and non-ground points in airborne laser scanning cloud.