3D POSITIONING ACCURACY and LAND COVER CLASSIFICATION PERFORMANCE of the UAVS: CASE STUDY of DJI PHANTOM IV MULTISPECTRAL RTK

Abstract

Unmanned air vehicle (UAV) has become an indispensable mobile mapping technology of remote sensing thanks to offering low cost and high resolution spatial data. Particularly, camera equipped optical UAVs are large in demand by land-related professions, including mapping, agriculture and forestry. Regarding the requirements, the technological level of the optical UAVs rises day by day by adding novel payloads. For instance, global navigation satellite system (GNSS) receivers with real-time kinematic (RTK) positioning capability were added to facilitate the fieldwork for ground control point (GCP) set up and measurements before UAV flights. Multispectral cameras were added to increase the automatic land cover classification potential of generated ortho-mosaics. At this point, the most significant question is the contribution level of these technological payloads. In this study, our research group evaluated the RTK GNSS positioning accuracy and automatic land cover classification potential of "DJI Phantom IV Multispectral RTK", which is one of the most common optical UAVs for scientific and commercial applications. For the evaluations, a study area that includes a large variety of land cover classes was selected. The UAV RTK GNSS positioning accuracy was calculated by comparing the RTK GNSS data obtained from the UAV with the measured GCPs in the study area. Furthermore, the land cover classification performance of Multispectral UAV was analysed by pixel and object-based classification techniques separately. For this purpose, while spectral angle mapper (SAM), minimum distance (MD) and maximum likelihood (ML) classifiers were applied to perform pixel-based classification, nearest neighbour (NN) classifier was employed to utilize object-based classification. The positioning accuracy results demonstrated that the root mean square error (RMSE) of UAV RTK GNSS is ±1.1 cm in X, ±2.7 cm in Y, and ±5.7 cm in Z. The classification results showed that the highest overall accuracy was estimated as 93.56% with ML classifier and its classification performance was found to be superior compared to those of SAM (73.46%) and MD (75.27%) classifiers. On the other hand, the overall accuracy was calculated as 90.09% for object-based classification and it was 3% lower than the pixel-based ML classification result. This could be the result of heterogeneity of the image objects created during the segmentation stage. Further studies are required to improve the object-based classification accuracy by applying different segmentation methods and quality measures. © 2022 Elsevier B.V., All rights reserved.