This talk aims to present two methodological approaches recently considered to analyze multidimensional Synthetic Aperture Radar images. First, we will consider the use of robust statistics to tackle the problem of change detection in a series of high-resolution images. Traditionally, when the data is multidimensional such as for PolSAR images, the detection is performed through a hypothesis testing of covariance homogeneity over time. Such an approach relies on a Gaussian assumption to model the data which is inaccurate when for high-resolution images. To tackle this issue, we consider the use of robust statistics which introduce the model of Compound-Gaussian distributions better suited to tackle the data. A new hypothesis testing statistic is then proposed using this assumption which demonstrates better robustness to heterogeneous data. This methodology has shown an increased performance for change detection for real-world data. In a second time, we will consider the use of Riemannian geometry to consider the problem of classifying when the feature constitutes covariance matrices. This novel approach takes into consideration the underlying geometry of the SPD metric space which is a non-euclidean one. The principle between this approach as well as the useful tools will be presented to show how a traditional classifier in Euclidean space can be adapted to this Riemannian framework. Experiments in a pedestrian detection problem as well SAR clustering problem show the interest of this approach.