The processing of remotely sensed data is innately an inverse problem where properties of spatial processes are inferred from the observations based on a generative model. Meaningful data inversion relies on well-defined generative models that capture key factors in the relationship between the underlying physical process and the measurements.

Unfortunately, as two mainstream data processing techniques, both mixture models and latent variables models (LVM) are inadequate in describing the complex relationship between the spatial process and the remote sensing data. Consequently, mixture models, such as K-Means, Gaussian Mixture Model (GMM), Linear Discriminant Analysis (LDA) and Quadratic Discriminant Analysis (QDA), characterize a class by statistics in the original space, ignoring the fact that a class can be better represented by discriminative signals in the hidden/latent feature space, while LVMs, such as Principal Component Analysis (PCA), Independent Component Analysis (ICA) and Sparse Representation (SR), seek representational signals in the whole image scene that involves multiple spatial processes, neglecting the fact that signal discovery for individual processes is more efficient.

Although the combined use of mixture model and LVMs is required for remote sensing data analysis, there is still a lack of systematic exploration on this important topic in remote sensing literature. Driven by the above considerations, this thesis therefore introduces a mixture of LVM (MLVM) framework for combining the mixture models and LVMs, under which three models are developed in order to address different aspects of remote sensing data processing:​ (1) a mixture of probabilistic SR (MPSR) is proposed for supervised classification of hyperspectral remote sensing imagery, considering that SR is an emerging and powerful technique for feature extraction and data representation;​ (2) a mixture model of K “Purified” means (K-P-Means) is proposed for addressing the spectral endmember estimation, which is a fundamental issue in remote sensing data analysis;​ (3) and a clustering-based PCA model is introduced for SAR image denoising. Under a unified optimization scheme, all models are solved via Expectation and Maximization (EM) algorithm, by iteratively estimating the two groups of parameters, i.e., the labels of pixels and the latent variables. Experiments on simulated data and real remote sensing data demonstrate the advantages of the proposed models in the respective applications.