Advanced Methods for High Resolution SAR information extraction: Data and User-driven evaluation approaches for Image Information Mining

Kurzfassung

We are concerned in this thesis by the problem of Image Information Mining (IIM) for exploitation and understanding of high resolution Synthetic Aperture Radar (SAR) data. Advances in this field of research contribute to the elaboration of tools for interactive exploration and extraction of the image content. In this context, analyzing and evaluating adequate image models and image information extraction methods according to user conjectures, constitute challenging issues. Our work contributes with solutions to high resolution SAR modeling and content estimation with a data-driven evaluation approach, and with the design of image mining scenarios by involving the user and his conjectures, achieved through an user-driven evaluation approach. To represent the data and to allow extracting the information, we focus on the properties of high resolution SAR images and how the stochastic models can represent and characterize the image content through a parameter estimation step. We perform a datadriven evaluation and validation of automatic information extraction methods for high resolution SAR scenes based on Gibbs Random Field (GRF) models. Specifically, Gauss Markov Random Field (GMRF) and Auto-binomial (ABM) models are implemented in information extraction methods following the two levels of Bayesian inference: model fitting and model selection. Both methods provide as results the speckle-free image and its structure parameters. In order to assess the quality of these methods, we perform detection tests on classes such as cities, vegetation, and water bodies; using specific qualitative metrics to quantify the quality of speckle removal. The accuracy of modelling and characterization of the image content are determined using both supervised and unsupervised classifications, and confusion matrices. We conclude that both methods enhance the image during the despeckling process. The GMRF model is more suitable for natural scenes and the ABM model for man-made structures. However, evaluating the information extraction methods is not enough for a complete validation of IIM systems, as we need to adapt to the user conjectures by designing validation scenarios and assessing the user satisfaction degree as well as the effectiveness of the retrieval process. We design and generate two study cases, which reflect the user needs in solving rapid mapping applications. The end-user is included in the loop in the user-driven evaluation approach by creating the two evaluation scenarios in the framework of disaster monitoring: oil spill and flood detection. The scenarios are carried out using ScanSAR and High Resolution Spotlight TerraSAR-X products, respectively. Quantitative metrics such precision and recall are used as figures of merit. In order to have measurements about the user satisfaction degree, a group of evaluators are asked to qualitatively rank the retrieved results. We conclude that the effectiveness of the retrieval process is more than 80 percent and the degree of user satisfaction is good for both scenarios.