Multispectral Lidar-imagery Data Fusion and Multilevel Classification of Lidar Point Clouds for Urban Features Extraction

Abstract

The world is witnessing massive demographic residents transitions from rural to urban, resulting in alarming consequences. These shifts highlight the necessity of comprehensive urban land-use mapping for current assessment and future city planning and management. The integration of Light Detection and Ranging (LiDAR) and imagery data allows using their complementary geometric and radiometric characteristics, respectively, with a potential increase in urban mapping accuracies. However, airborne LiDAR data may not always come with onboard optical images collected during the same flight mission. Indirect geo-referencing can be adopted if ancillary imagery data are available. Nevertheless, automatic recognition of control primitives in LiDAR and imagery datasets becomes challenging when collected on different dates. This dissertation proposes a generic geo-registration mechanism using the Phase Congruency (PC) model and scene abstraction to overcome the stated challenges. The Root Mean Square Error (RMSE) is between one to two pixels. The proposed workflow is computationally efficient, especially with small datasets, and generic enough to apply to various urban morphologies. A point-based classification was then performed on a point cloud covering a residential area using machine learning. The data was geo-registered to an aerial image. The classification tried multiple combinations of the inherited image spectra and the height as the only geometric feature in the classifying spaces. The acquired radiometric features added to the height gave an overall accuracy above 97% using Multi-Layer Perceptron (MLP) neural networks, surpassing those achieved with height-based geometric-oriented feature spaces. However, misclassifications occurred among different classes due to the independent acquisition of the aerial and LiDAR data and shadow and orthorectification problems from the aerial images. The dissertation then moves on to modify the color-based segmentation algorithm for the object-based classification of the multispectral LiDAR version of the same point cloud. The dissertation proposes a multilevel classification through vectors of inherited and calculated height-based geometric and spectral features structured in different classification scenarios (combinations). An overall mapping accuracy higher than 98% was achieved using MLP neural networks, combining the entire feature vectors and their optimized output space projected by the Principal Component Analysis (PCA). This combination also eliminates some of the abovementioned misclassifications.