Tensor Representation And Manifold Learning Methods For Remote Sensing Images

It is the main purpose of earth observation to extract the interested information and knowledge from remote sensing (RS) images quickly and accurately. With the development of RS technology, RS images with very high resolution and hyperspectral channels have been able to provide a large amount of information, we have much more multispectral, high spatial resolution, and temporal resolution RS data than before. As a result of that, it is inefficient or even impossible to interpret these data by human beings. Thus, with the development of computer hardware, we need to explore some intelligent algorithms to process such mass RS data quickly and accurately. This thesis aims to propose some RS image information extraction algorithms based on the latest methods in machine learning area. In particular, we adopts the manifold learning technology as the mainline and combins the regularization theory, tensor method, sparse learning and transfer learning into our algorithms. The main contributions of this thesis are as follows:(1) A patch alignment framework is introduced to unify the conventional dimension reduction (DR) methods such as PCA&LDA, and some representative manifold learning algorithms snch as LLE, ISOMAP, LE, LTSA&HLLE into an optimization. This framework reveals the basic principle of manifold learning algorithms, that is1) algorithms are intrinsically different in the patch optimization stage and2) all algorithms share an almost identical whole alignment stage. As an application of this framework, we could develop new dimensionality reduction algorithms by modify the local alignment matrix.(2) The tensor based manifold learning methods in RS image analysis are extensively discussed. In this paper, we propose amethod for the dimensionality reduction of spectral-spatial features in RS images, under the umbrella of multilinear algebra, i.e., the algebra of tensors. The proposed approach is a tensor extension of conventional supervised manifold learning based DR. In particular, we define a tensor organization scheme for representing a pixel’s spectral-spatial feature and develop tensor discriminative locality alignment (TDLA) for removing redundant information for subsequent classification.(3) Two types of multiple feature combining (MFC) algorithms are proposed to deal with RS image classification tasks with multiple features as input. In these two methods, we use the LE and SNE to bulid the local alignment matrix, respectively. Then, the proposed adaptive manifold learning MFC algorithms combine the input multiple features linearly in the optimal way and obtain a unified low-dimensional representation of these multiple features for subsequent classification. Each feature has its particular contribution to the unified representation determined by simultaneously optimizing the weights in the objective function.(4) Two types of regularized discriminative manifold learning algorithms are proposed for hyperspectral target detection, named sparse transfer manifold embedding (STME) and supervised metric learning (SML). Technically speaking, these methods are particularly designed for hyperspectral target detection by introducing the multiple constraints into discriminative manifold learning framework. In STME, a sparse formulation and a transfer regularization are adopted, while in SML, a similarity propagation constraint and a manifold smoothness regularization are enforced. Both of the algorithms have showed the outstanding detection performance.(5) A general multilinear data analysis framework for tensor inputs is investigated for RS image classification. This framework generalizes the current classifiers which only accept vectors as inputs into multilinear condition. Based on this framework, we further propose some algorithms, e.g., support tensor machine, multiclass support tensor machine, and proximal support tensor machine. These methods show the effectiveness of tensor representation and analysis approaches in RS image information extraction with a small number of training samples.