Advanced Techniques For Object-Based Classification Of Hyperspectral Remote Sensing

Land-Use Land-Cover(LULC)mapping is an important issue in agriculture management.The demand for improved and more accurate LULC mapping has led to a methodology known as object-based image analysis(OBIA).The focus of this thesis is about the design and implementation of new methods for object-based(OB)classification of hyperspectral remote sensing images.Airborne earth observation instruments collect these images.Particularly,in this thesis we investigate the integration of methods for object-based hyperspectral classification in synergic way.Hyperspectral images mapping is a crucial step in environmental monitoring,disaster management,agriculture management and military applications.The recent researches to improved and accurate hyperspectral mapping have led to key methodology known as Object-based classification of remotely sensed images.The core intention of the OB images analysis is to group-of-pixels named "objects" through segmentation into a specific class.Object-oriented image analysis segments the data and constructs hierarchical network of homogeneous objects.Object-based enable the analysis of aggregated sets of pixels,exploit shape-related variation,as well as spectral characteristics.Traditional pixel-based classification methods always produce mixed pixel problems due to the independency of neighbor pixels.This problem motivates researchers to combine segmentation,color and many other parameters to illuminate the wrongly classified or mixed pixels into their relevant classes.While the new methodology solved problems and improved accuracy,but it also raises new challenges such as the loss of accuracy in terms of less abundant,but potentially important.The key challenges in hyperspectral images classification techniques are the high dimensionality of data,limited number of training samples,and combination of spatial and spectral information.In recent years,availability of new remotely sensed data with high dimensions,spectral and spatial resolution research attempts have shifted from conventional pixel-based to object-based hyperspectral image classification is a hot topic to overcome aforementioned challenges.The non-parametric classifiers included decision tree(DT),support vector machines(SVM)some ensemble learning algorithms(e.g.random forest(RF),bagging and boosting)in object-based mapping has been used for mapping of high-resolution images.This thesis work explored new methods to improve OB hyperspectral images classification performance by eliminating mixed-pixel problems due to the independency of neighbor pixels.In the literature review few studies applied for pixel-based classification on low or middle resolution images.Nevertheless,in the literature the RoF has not been utilizing for Object-oriented hyper spectral images.To improve the classification accuracy,first time we investigated an ensemble principle named Rotation-Based object-oriented classification of hyperspectral images(RoBOO).It is the combination of segmentation with support vector machine and nearest neighborhood algorithm.It uses random features selection and data transformation principle component analysis(PCA),technique to improve accuracy and diversity.We exposed improved accuracy using two key player's Multi-resolution segmentation(MRS)and supervised classification is used to obtain the objects by tuning the parameters included compactness,scale and shape.Furthermore,multiple classification results are produced by proposed method and results were integrated by using majority-voting rule to release results.Mostly used data transformation technique PCA was used.An experiential study on one hyperspectral dataset indicates that the suggested RoBOO slightly surpasses the single SVM and KNN.The rotation-matrix with the MRS improved the performance.The effect of parameters on accuracy of RoBOO(different training sets,amount of features in each subset,compactness,scale parameters,shape)is examined as well in this document.We deduced that the integrations of SVM and NN with MRS were impressive in the object-oriented hyperspectral mapping.Overall,this thesis explores solutions to two key problems of an object-based classification RoBOO with MRS system:accuracy and efficiency.Our proposed algorithm RoBOO improved the efficiency and with MRS integration it demonstrated improvements in accuracy.The outcome of this thesis advances the state-of-the-art by exploring novel methodologies for accurate object-based hyperspectral image classification;where the results obtained by experimentation on various real hyperspectral data sets confirmed their effectiveness.Concluding the thesis,concrete remarks to the aforementioned issues are discussed.