3D Classification of Urban Features Based on Integration of Structural and Spectral Information from UAV Imagery

Three-dimensional classification of urban features is one of the important tools for urban management and the basis of many analyzes in photogrammetry and remote sensing. Therefore, it is applied in many applications such as planning, urban management and disaster management. In this study, dense point clouds extracted from dense image matching is applied for classification in urban areas. Applied images are acquired using a Micasense RedEdge multispectral camera to increase the classification accuracy. The band to band registration is one of the existing challenges of multi-spectral camera, which the SIFT algorithm is used to extract the corresponding features of each band. One band selected as reference and other bands are transferred to the reference band by projective transformation. Finally, the bands are combined to create a color image from each three bands. So, two point clouds are generated using dense image matching techniques from two sets of images. To produce a multi-spectral point cloud, the two set of point clouds have been integrated using nearest neighbor interpolation. The multi-spectral point clouds are classified by using random forest algorithm, structural and multi-spectral features. This process composed of three parts as structural information, multi-spectral information, and integration of both. Finally, the results are shown a 25% improvement in the accuracy of the integration of multi-spectral and structural information compared to multi-spectral information and 32% improvement in the accuracy of the integration of multi-spectral and structural information compared to structural information. Classification using visible information (RGB) instead of multispectral information resulted in an accuracy drop by 5%.