Advanced machine learning methods for synthetic aperture radar analysis and fusion

The steady increase in the availability of Earth observation data, particularly from Synthetic Aperture Radar (SAR) and Polarimetric SAR (PolSAR) missions, has led to new opportunities in remote sensing applications. These data provide unique advantages such as day-and-night imaging, cloud penetration, and more generally, a distinctive independence to atmospheric conditions. However, its analysis presents significant inherent challenges, including speckle noise and the presence of complex values in the scattering matrix. Moreover, in many practical scenarios, limited labeled training samples remain a critical issue, emphasizing the need for methods that remain effective with minimal supervision. This thesis addresses these challenges by developing novel methods for three critical problems in remote sensing: (i) contextual classification of PolSAR images, (ii) feature learning for multi-sensor and multi-temporal PolSAR image series, and (iii) heterogeneous change detection (CD) between hyperspectral (HSI) and PolSAR data. First, a complex-valued Support Vector Machine (SVM) classifier combined with a Markov Random Field (MRF) model is introduced for contextual PolSAR classification. A novel symmetric kernel is proposed to handle complex-valued data while ensuring Mercer's condition, and spatial-contextual information is incorporated through a hybrid SVM-MRF framework. This method achieves state-of-the-art classification accuracy and even outperforms deep learning techniques in scenarios with limited training data. Second, a feature learning framework is developed to extract information-dense features from multi-sensor, multi-frequency PolSAR time series. A combination of 1D Convolutional Neural Networks (1D-CNNs) and Stacked Autoencoders (SAEs) enables the modeling of temporal dependencies and compression of high-dimensional data. The learned representations are validated on unsupervised classification tasks, demonstrating their effectiveness in handling heterogeneous PolSAR datasets. Finally, a novel approach for heterogeneous CD is presented to fuse hyperspectral data with PolSAR imagery. The method integrates dimensionality reduction using Uniform Manifold Approximation and Projection (UMAP) and image-to-image translation using a dual CNN architecture (X-Net). Experimental results on semi-simulated and real datasets show that the approach successfully detects ground changes, outperforming similar heterogeneous CD techniques. The findings of this thesis demonstrate the importance of combining machine learning, deep learning, and sensor fusion techniques to address the challenges of PolSAR data analysis. The proposed methods balance computational efficiency and accuracy, providing solutions for applications where annotated datasets are scarce, and timely information is critical.