Fusion of Full Waveform LiDAR and Passive Remote Sensing for Improved Land-Cover Classification
Land-cover classification is a crucial step in interpreting remote sensing data, and the accuracy determines the reliability of the product for further downstream applications. Hyperspectral sensors have been widely utilized for classification because of the discrimination afforded by its rich spectral information and high resolution in both the spatial and spectral domains. On the other hand, LiDAR (Light Detection And Ranging) data has gained increasing interest for use in classification because it provides precise three-dimensional (3-D) data for large areas with precise 3-D location information, and therefore greatly expands the domain of available spatial information. Reflected laser energy from targets is also collected by LiDAR systems, and contains information regarding target backscattering properties. With the introduction of full waveform LiDAR (FWL), the possibility of using LiDAR for target discrimination has been enhanced due to the additional structural information acquired. The geometrical information and backscattering properties measured by FWL is complementary to the reflectance characteristics recorded within in Hyperspectral imagery (HI). Thus, the fusion of FWL and HI is highly desirable.
There has been a fair amount of research investigating the fusion of LiDAR and HI for target characterization and land-cover classification. However discrete-return LiDAR point clouds were more thoroughly investigated in this area than FWL because of their wider availability and easier interpretation. In those studies that utilized FWL, the application of waveform data was mainly limited as a reference data source to provide height information pertaining to observed targets. Furthermore, application of fused FWL and HI data for target identification has been mostly limited to selected objects, such as trees or buildings, while the subject of land cover classification has been investigated in only a few works.
This dissertation aims to build a framework for fusing FWL and HI and to demonstrate the application of the combined data set for land-cover classification without being limited to a small sample of objects. Feature extraction methods and classifier designs are proposed considering characteristics of both data sets, and performance of the proposed methods are evaluated using two data sets collected in complex scenes by the National Center for Airborne Laser Mapping (NCALM). Experimental results show that the proposed methods are successful in extracting features from reconstructed FWL data, and the proposed classification scheme effectively utilizes the combined FWL and HI features for separating ground cover features in both data sets with over 95% accuracy.