The locally linear embedding (LLE) algorithm is hypothetically able to find a lower dimensional space than a linear method for preserving a data manifold originally embedded in a high dimensional space. However, uneven sampling over the manifold in real-world data ultimately causes LLE to suffer from the disconnected-neighborhood problem. Consequently, the final dimensionality required for the data manifold is multiplied by the number of disjoint groups in the complete data representation. In addition, LLE as an unsupervised method is unable to suppress between-class connections. This means that samples from different classes are mixed during reconstruction. This study presents CLLE, a classification-oriented LLE method that uses class label information from training samples to guide unsupervised LLE. The criterion for neighbor selection is redesigned using class-conditional likelihood as well as Euclidean distance. This algorithm largely eliminates fractured classes and lowers the incidence of connections between classes. Also, a reconnection technique is proposed as a supporting method for ensuring a fully connected neighborhood graph, so that CLLE is able to extract the fewest features. Experiments with simulated and real data show that CLLE exceeds the performance of linear methods. Comparable classification performance can be achieved by CLLE using fewer features. In comparison with LLE, CLLE demonstrates a higher aptitude for and flexibility towards classification.

Recently, nonlinear feature extraction algorithms based on a so-called kernel trick have appeared to reduce the limitations of linear feature extraction methods with respect to class discrimination. This study presents a new kernel function that integrates the discriminative information from class labels and spatial contexts into the basic radial basis function (RBF). We represent the mutual closeness of samples in terms of the average class membership probability and explore contextual information by means of Markov random field models. By fusing additional discriminative information into the kernel feature space, the proposed kernel function outperforms the basic RBF kernel function. A more compact set of features have shown equivalent effectiveness. Experiments also demonstrate that using spatial contextual information during feature extraction can be more efficient than using the information during the classification stage.

Most of change detection algorithms for multi-temporal images are performed in unit of pixels. Due to the limit of spatial resolution, a pixel is, in many cases, a mixed pixel that contains more than one ground cover types. In order to explore more information from images, we have reviewed several spectral unmixing algorithms and used them to accomplish subpixel change detection. In an application of landslide monitoring, we demonstrated the use of subpixel change detection for detection of landslide spreading. We used spectral unmixing algorithms to extract the abundance information from multispectral images. For the particular characteristic of landslides, we incorporated the slope feature into process. Our preliminary result shows that the subpixel change detection method can provide more detailed information about landslide spread than pixel-based change detection algorithms.

A parametric linear feature extraction method is proposed for multiclass classification. The skeleton of the proposed method consists of two types of schemes that are complementary to each other with regard to the discriminant information used. The approximate pairwise accuracy criterion (aPAC) and the common-mean feature extraction (CMFE) are chosen to exploit the discriminant information about class mean and about class covariance, respectively. Choosing aPAC rather than the linear discriminant analysis (LDA) can also resolve the problem of overemphasized large distances introduced by LDA, while maintaining other decent properties of LDA. To alleviate the suboptimum problem caused by a direct cascading of the two different types of schemes, there should be a mechanism for sorting and merging features based on their effectiveness. Usage of a sample-based classification error estimation for evaluation of effectiveness of features usually costs a lot of computational time. Therefore, we develop a fast spanning-tree-based parametric classification accuracy estimator as an intermediary for the aPAC and CMFE combination. The entire framework is parametric-based. This avoids paying a costly price in computation, which normally happens to the sample-based approach. Our experiments have shown that the proposed method can achieve a satisfactory performance on real data as well as simulated data.

It is observed in remote sensing that a finer spatial resolution does not necessarily improve the classification performance. These observations have been understood by using the conceptual explanation that "boundary effect" and "within-class variability" work against one another. Though easily understood, this conceptual explanation cannot be readily used for a quantitative investigation. The authors design a simulation scheme to evaluate systematically the impacts of various parameters on the classification accuracy. The authors employ a model for the class spectral covariance of pure pixels and a linear mixing model for the spectral responses of mixed pixels. Based on these models, the authors derive the statistical characteristics for mixed pixels and assess the corresponding classification errors. As the ratio of ground sampling distance to field size decreases, the classification error associated with pure pixels tends to increase, whereas the classification error associated with mixed pixels tends to decrease from the smaller area of mixed pixels. The simulation results show that the overall classification error first decreases with decreasing ratio of ground sampling distance to field width, reaches a minimum value, and then may increase with further decreasing ratio. The study on the classification error may help the development of classification schemes for high spatial resolution imagery.