Semantic object-based urban scene analysis for feature fusion of VHR imagery and Lidar DSM

More complexities in urban areas due to the existence of buildings with roofs made of various materials, roads and parking lots, diverse plant species and vehicles along with developed very high spatial resolution (VHR) remote sensing imageries lead to increase the difficulties of urban objects recognition. Defining human cognition semantic categories of land cover objects is proposed in this paper to improve the object-based scene analysis. After performing the segmentation algorithm, semantically feature fusion of the VHR imagery and Lidar DSM is performed in the context of object-based procedures. For this reason, high-level semantics composed of human cognition object categorization and visual–conceptual descriptions of land cover objects in the complex urban scene is used for low-level features measurement followed by classification rules generation. The pan-sharpened WorldView-2 satellite image and Lidar DSM over a complex urban area in Rio de Janeiro (Brazil) are used to evaluate the capabilities of the proposed semantic object-based scene analysis. The evaluation of the obtained object recognition results shows the 93.98% overall accuracy and 0.937 Kappa coefficient. Comparing the obtained semantic object recognition results with non-semantic object-based procedures indicates the improvements in overall accuracy for 20.68 and 2.18% in object-based spectral and object-based spectral–structural classifications, respectively. Increasing the true positive pixels of the spectrally and height similar object categories (such as buildings and road), and using threshold training without requiring high-dimensional training samples are some of the advantages of the proposed semantic object-based scene analysis method.