2022 年 32 巻 S 号 p. 165-169
Land cover classification is vital for ecosystem assessment, desertification monitoring and vegetation inventorying. In recent years new technologies have emerged that can be used for land classification. One of these technologies are Unmanned Aerial Vehicles (UAV), that can take images applicable for remote sensing, and which allow collection of data in places that may be difficult to access. The study site is located in Ali Faren (42°49ʹ48ʺ E, 11°21ʹ43ʺ) at the Province of Arta, Djibouti. It is intended to use the 3D cloud point data generated from RGB cameras, which land cover features were classified using NDVI calculated from data obtained by multispectral cameras in the same study area. This classified point cloud was used for training and validating a deep learning model, which was applied for classifying cloud point datasets in three sites. It was found that there is a high accuracy when identifying “Ground” and “High vegetation” in the point cloud datasets with only RGB. However, there is a low degree of accuracy for identifying other classes, such as low vegetation. This can be attributed to the fact that shrubs and grasslands (3 Low vegetation, and 4 middle vegetations) are very scarce in the study area, so a higher emphasis on its identification is required.
