Feature Extraction And Classification For Hyperspectral Remote Sensing Images

Recent advances in sensor technology have led to an increased availability of hyper-spectral remote sensing data at very high both spectral and spatial resolutions. Manytechniques are developed to explore the spectral information and the spatial informationof these data. In particular, feature extraction (FE) aimed at reducing the dimensionalityof hyperspectral data while keeping as much spectral information as possible is one ofmethods to preserve the spectral information, while morphological profile analysis is themost popular methods used to explore the spatial information.Hyperspectral sensors collect information as a set of images represented by hun-dreds of spectral bands. While ofering much richer spectral information than regularRGB and multispectral images, the high dimensional hyperspectal data creates also achallenge for traditional spectral data processing techniques. Conventional classifica-tion methods perform poorly on hyperspectral data due to the curse of dimensionality(i.e. the Hughes phenomenon: for a limited number of training samples, the classifica-tion accuracy decreases as the dimension increases). Classification techniques in patternrecognition typically assume that there are enough training samples available to obtainreasonably accurate class descriptions in quantitative form. However, the assumptionthat enough training samples are available to accurately estimate the class description isfrequently not satisfied for hyperspectral remote sensing data classification, because thecost of collecting ground-truth of observed data can be considerably difcult and expen-sive. In contrast, techniques making accurate estimation by using only small trainingsamples can save time and cost considerably. The small sample size problem thereforebecomes a very important issue for hyperspectral image classification.Very high-resolution remotely sensed images from urban areas have recently becomeavailable. The classification of such images is challenging because urban areas oftencomprise a large number of diferent surface materials, and consequently the heterogeneityof urban images is relatively high. Moreover, diferent information classes can be made upof spectrally similar surface materials. Therefore, it is important to combine spectral andspatial information to improve the classification accuracy. In particular, morphological profile analysis is one of the most popular methods to explore the spatial informationof the high resolution remote sensing data. When using morphological profiles (MPs)to explore the spatial information for the classification of hyperspectral data, one shouldconsider three important issues. Firstly, classical morphological openings and closingsdegrade the object boundaries and deform the object shapes, while the morphologicalprofile by reconstruction leads to some unexpected and undesirable results (e.g. over-reconstruction). Secondly, the generated MPs produce high-dimensional data, which maycontain redundant information and create a new challenge for conventional classificationmethods, especially for the classifiers which are not robust to the Hughes phenomenon.Last but not least, linear features, which are used to construct MPs, lose too muchspectral information when extracted from the original hyperspectral data.In order to overcome these problems and improve the classification results, we de-velop efective feature extraction algorithms and combine morphological features for theclassification of hyperspectral remote sensing data. The contributions of this thesis areas follows.1. As the first contribution of this thesis, a novel semi-supervised local discriminantanalysis (SELD) method is proposed for feature extraction in hyperspectral remotesensing imagery, with improved performance in both ill-posed and poor-posed condi-tions. The proposed method combines unsupervised methods (Local Linear FeatureExtraction Methods (LLFE)) and supervised method (Linear Discriminant Analy-sis (LDA)) in a novel framework without any free parameters. The underlying ideais to design an optimal projection matrix, which preserves the local neighborhoodinformation inferred from unlabeled samples, while simultaneously maximizing theclass discrimination of the data inferred from the labeled samples.2. Our second contribution is the application of morphological profiles with partialreconstruction to explore the spatial information in hyperspectral remote sensingdata from the urban areas. Classical morphological openings and closings degradethe object boundaries and deform the object shapes. Morphological openings andclosings by reconstruction can avoid this problem, but this process leads to someundesirable efects. Objects expected to disappear at a certain scale remain presentwhen using morphological openings and closings by reconstruction, which meansthat object size is often incorrectly represented. Morphological profiles with partialreconstruction improve upon both classical MPs and MPs with reconstruction. Theshapes of objects are better preserved than classical MPs and the size information is preserved better than in reconstruction MPs.3. A novel semi-supervised feature extraction framework for dimension reduction ofgenerated morphological profiles is the third contribution of this thesis. The mor-phological profiles (MPs) with diferent structuring elements and a range of in-creasing sizes of morphological operators produce high-dimensional data. Thesehigh-dimensional data may contain redundant information and create a new chal-lenge for conventional classification methods, especially for the classifiers which arenot robust to the Hughes phenomenon. To the best of our knowledge the use ofsemi-supervised feature extraction methods for the generated morphological profileshas not been investigated yet. The proposed generalized semi-supervised local dis-criminant analysis (GSELD) is an extension of SELD with a data-driven parameter.4. In our fourth contribution, we propose a fast iterative kernel principal componentanalysis (FIKPCA) to extract features from hyperspectral images. In many ap-plications, linear FE methods, which depend on linear projection, can result inloss of nonlinear properties of the original data after reduction of dimensional-ity. Traditional nonlinear methods will cause some problems on storage resourcesand computational load. The proposed method is a kernel version of the CandidCovariance-Free Incremental Principal Component Analysis, which estimates theeigenvectors through iteration. Without performing eigen decomposition on theGram matrix, our approach can reduce the space complexity and time complexitygreatly.5. Our last contribution constructs MPs with partial reconstruction on nonlinear fea-tures. Traditional linear features, on which the morphological profiles usually arebuilt, lose too much spectral information. Nonlinear features are more suitable todescribe higher order complex and nonlinear distributions. In particular, kernelprincipal components are among the nonlinear features we used to built MPs withpartial reconstruction, which led to significant improvement in terms of classifica-tion accuracies.The experimental analysis performed with the novel techniques developed in thisthesis demonstrates an improvement in terms of accuracies in diferent fields of applicationwhen compared to other state of the art methods.