Multidimensional and temporal SAR data representation and processing based on binary partition trees

Abstract

This thesis deals with the processing of different types of multidimensional SAR data for distinct applications. Instead of handling the original pixels of the image, which correspond to very local information and are strongly contaminated by speckle noise, a region-based and multiscale data abstraction is defined, the Binary Partition Tree (BPT). In this representation, each region stands for an homogeneous area of the data, grouping pixels with similar properties and making easier its interpretation and processing. The work presented in this thesis concerns the definition of the BPT structures for Polarimetric SAR (PolSAR) images and also for temporal series of SAR acquisitions. It covers the description of the corresponding data models and the algorithms for BPT construction and its exploitation. Particular attention has been paid to the speckle filtering application. The proposed technique has proven to achieve arbitrarily large regions over homogeneous areas while also preserving the spatial resolution and the small details of the original data. As a consequence, this approach has demonstrated an improvement in the performance of the target response estimation with respect to other speckle filtering techniques. Moreover, due to the flexibility and convenience of this representation, it has been employed for other applications as scene segmentation and classification. The processing of SAR time series has also been addressed, proposing different approaches for dealing with the temporal information of the data, resulting into distinct BPT abstractions. These representations have allowed the development of speckle filtering techniques in the spatial and temporal domains and also the improvement and the definition of additional methods for classification and temporal change detection and characterization.