Advanced Methods for Change Detection in LiDAR Data and Hyperspectral Images

In the last years Remote Sensing technology has significantly improved and new sensors capable of acquiring data with high spatial and spectral resolution have been developed. Light Detection And Ranging (LiDAR) and Hyperspectral (HS) sensors acquire data that accurately characterize the 3-D structure and the spectral signature of the area of interest, respectively. With the upcoming generation of small sensors designed for Unmanned Aerial Vehicle (UAVs) and new spaceborne missions such data will be acquired more and more often increasing the availability of multitemporal datasets. This requires the development of methods capable of considering the time variable in the analysis of LiDAR point clouds and HS images. In this context, this thesis provides three main contributions related to: i) Change Detection (CD) in LiDAR data, ii) multiple CD in HS images and iii) fusion of bitemporal LiDAR point clouds. The first novel contribution presents a method for the detection of 3-D changes at the individual tree level in conifer forests using bitemporal LiDAR data. Unlike most of the literature techniques, the method performs an object-based CD to estimate both the vertical and horizontal growth of the individual tree-crown working directly in the point cloud domain to fully exploit the information content of the LiDAR data. Multiple CD in HS images is addressed in the second contribution. Differently from most of the existing methods in the literature, we focus on the information content of each spectral channel to define a novel efficient representation of the change information. This representation is used to automatically discriminate between the different kinds of change. The third contribution presents two methods for the fusion of bitemporal LiDAR point clouds aimed at improving the modeling of the individual tree-crown. One is a compound approach used to improve the detection of tree-tops of conifers by reducing false detections and recovering missed detections. It exploits the temporal correlation between the two LiDAR point clouds by modeling the different probabilities of transition from one date to the other and using the Bayes rule for minimum error to perform the decision process. The other fusion method exploits the richer information content of high density point clouds to improve the parameters estimation of individual conifers in low density data. For each tree, it uses a 3-D model to reconstruct the shape of the crown using the parameters estimated on the high density data to drive the estimation on the low density point cloud. The proposed methods have been tested on LiDAR point clouds and on simulated and real bitemporal HS datasets. Quantitative and qualitative experimental results confirm the effectiveness of the proposed automatic and unsupervised techniques, which show equal or better results compared to existing unsupervised and supervised techniques.