Spatial And Spectral Information Based Hyperspectral Image Classification

This thesis is about classification of high resolution hyperspectral images using spatial and spectral information.Hyperspectral images can be acquired from airborne or space-borne platforms,and can be used for the identification of different materials on ground as well as achieving the geometrical properties of different objects like buildings,water,roads,agriculture fields,shadows etc.However in order to develop practical applications with hyperspectral data,it is necessary to study automatic techniques for an effective and efficient analysis of data.This thesis focuses on the hyperspectral image(HSI)classification,which is the basis for most of the applications related to environmental monitoring and target detection.Image classification translates features that represent the information presented in an image using different pattern recognition techniques.However,huge volume of data associated with hyperspectral images creates a challenging task for classification and the existing techniques are still inadequate for them.Many spatial-spectral techniques are presented to extract the useful information from hyperspectral images to improve the classification accuracy.In this thesis,some novel spatial-spectral classification approaches are proposed,which can automatically extract the useful information from hyperspectral images for different applications.Furthermore,limited availability of training samples and observed noise are common problems that many scientists encounter in practical remote sensing(RS),which are also concerned in the newly proposed classification approaches.The main studies and contributions in this thesis are as follows:1.Systematic comparison study of linear feature extraction methods for classification of hyperspectral images with noises.Hyperspectral image processing is usually time consuming,due to its huge data size.Nowadays hyperspectral imaging is used in many fields requiring real-time solutions.A systemic comparison study of linear feature extraction(LFE)methods for classification of hyperspectral images with various types of noises is carried out in this research.The performance of different LFE methods for classification and their computation cost reduction are compared.In practice,hyperspectral images are often contaminated by different types of noise,as the atmosphere around hyperspectral cameras may change all the time.To make it more realistic,different types of noise,including Salt-and-Pepper noise,Gaussian noise,Speckle noise and their mixtures,are artificially imposed on the hyperspectral image.Support Vector Machine(SVM)based classification is employed for classification performance comparison.The experimental results are very helpful for selecting linear feature extraction methods for classification of hyperspectral images,which are usually affected by noise.The experimental results show that:(1)Feature extraction(FE)before classification with Partial Least Square and Partial Least Square-SB performs best no matter what positive value of signal to noise ratio(SNR)is chosen.(2)FE before classification with Principal Component Analysis(PCA)performs best when SNR is very low,i.e.negative value.The processing time for all FEs studied is much shorter than those without FEs.The one with Maximum Noise Fraction is the fastest.It is also observed that classification maps are improved when LFE methods are used on noisy hyperspectral images.2.Extended morphological profiles with duality for hyperspectral image classification.In this research,the Extended Morphological Profile with duality(EMPD)is proposed for hyperspectral image classification.The shape noise can be greatly reduced and thus improved classification accuracy is obtained compared to the conventional Extended Morphological Profile technique.Furthermore,linear filtering technique is combined with SVM based classifier,and used for classification of the hyperspectral images,which further improves the classification accuracy on urban data sets.Spectral and spatial information is examined individually and combined using concatenated vector.For the dimensionality and redundancy issues,two non-parametric supervised feature extraction techniques(NSFETs),Decision Boundary Feature Extraction(DBFE)and Nonparametric Weighted Feature Extraction(NWFE),are investigated on hyperspectral data sets and EMPD.The classification maps are also improved after filtering,in particular where data sets are more congested.The shadows are much clearer in classification maps after filtering.Apart from this the effect of noise on NSFETs is studied and it is concluded that they should be avoided in harsh environment.3.Marker selection using SVM over-fitting and marker selection using skeletonization for very low training sample analysis of hyperspectral image classification.This research presents two novel methods named marker selection using over-fitting and the marker selection using skeletonization.Markers are the most reliable pixels that represent a particular class.Both methods are analysed for very low training sample analysis(VLTSA)of classification,as low as one training sample per class of hyperspectral image.Furthermore,it is shown that by using the spatial and spectral information with nonparametric supervised feature extraction methods,better classification accuracy can be achieved even in case with very low training samples.Only one to ten training samples per class are examined and it is concluded that by using Nonparametric Weighted Feature Extraction,better classification accuracy can be obtained.It is also investigated that reasonably fine classification maps can also be obtained(even when the training samples are very low),using VLTSA with the proposed methods.4.Classification using Iterative support vector machine with spatial-spectral information.In this research,a novel classification approach is proposed to improve the classification accuracy of hyperspectral images in the cases when very limited training samples are available and when misclassification occurs due to random training samples.The proposed approach is based on the Iterative Support Vector Machine(ISVM)and spatial-spectral information.To improve the classification accuracy of ISVM,the Major Voting(MV)and the marker map correction techniques are used to correct the training samples in each iteration of ISVM.Experiments on practical hyperspectral images including AVIRIS Indian Pine Image are conducted and the results show that the proposed approach works better than ISVM and other classifiers such as SVM-RBF,Linear-SVM and K-NN in cases with small training sample set.