Classification Methods For Hyperspectral Image By Multi-Classifier Ensemble And Spectral-Spatial Feature Combination

Hyperspectral remote sensing image can simultaneously represent the characterization of the surface both in spatial domain and spectral domain. Image-spectrum merging speciality make hyperspectral remote sensing image have more advantage than conventional remote sensing image. However, high dimensional feature and the limitations of samples present a lot of challenges for hyperspectral image processing. Building classification model with higher precision and stability is in urgent, as samples of hyperspectral image are limited and spectral bands are high. Classification method for hyperspectral image by multi-classifier ensemble and spectral-spatial feature combination aggregates advantages in both spectral domain and spatial domain. The methods makes full use of the spatial information between adjacent pixel and correlation contexture, which makes up for the inadequacy of the classification based only on spectral information. Multi-classifier ensembles method improves the difference and complementarity between members of base classifiers, which obtained the higher classification accuracy, and preserve more detailed characteristics of the imagery.This paper systematically studies classification methods Hyperspectral Image by Multi-Classifier Ensemble and Spectral-Spatial feature combination according to theoretical research and analysis on spatial feature extraction and ensemble learning. Ensemble classification hyperspectral remote sensing image based on morphological attribute profiles is launched. In addition, Adaptive morphological attribute profiles based on ensemble method for hyperspectral image classification is raised. The two methods above are both using morphological attribute profiles for generating multiple spatial geometry features of remote sensing image. And they are belong to synchronous-process for spectral-spatial classification method. Moreover, post-process for spectral-spatial classification for hyperspectral image based ensemble method is also put forward-classification merged with clustering for hyperspectral imagery based multiple classifier ensemble. Image objects deriving from clustering map are used as elementary unit for classification. Results with less noise and strong robustness are obtained. Experiment on AVIRIS data set, ROSIS data set and GRSS-DFC-2013 data set respectively illustrate excellent performance of methods this paper proposed.The main research contents and the conclusions are depicted as follows:1) Morphological attribute profiles is utilized to express multi-spatial information of remote sensing image. Extreme Learning Machine (ELM), rotation forest (RoF), support vector machines (SVM) are employed to classification. Results of each base classifier are integrated by the way of majority voting. Differ from existing ensemble method, the study in this article takes morphological attribute profiles deriving from different feature dimension and different classification methods in consideration to form a multi-level ensemble. It improves the results of the classification accuracy and generalization ability.2) An adaptive morphological attribute profiles is put forward based on the concept of ensemble learning. Define a range for parameter first of all, and then choose parameters to construct morphological attribute profiles. We input the feature generated by morphological attribute profiles into classifiers to form base classification result. By loop iteration the process of the above many times, multiple base classification results will be produced. Finally, majority voting is used for multi-classifier ensembles. This method overcomes the problem of parameter selection for morphological attribute profiles. Morphological attribute profiles are used to get rich spatial detail characteristics of remote sensing image, in the meanwhile, ensemble learning method could enhance the stability of the classification results.3) Different clustering methods (K-Means, ISODATA, FCM and K-Mediods) are used to obtain clustering map containing context features. The regions in the clustering map are labeled by using a four-connected neighborhood labeling method to generate image objects, and a majority voting method is used to classify the objects based on the initial classification map derived pixel-wised method. Finally, a Chamfer neighborhood filtering technique is used to regularize the classification map, which partially reduces the noise. This method utilizing spatial information and context features from clustering takes advantage of supervised classification and unsupervisedclassification to gain noise reduction, high-accuracy and high homogeneity, which makes up for the inadequacy of the classification based only on spectral information.Methods of this paper are all in the framework spectral-spatial classification, considering the principle of multiple classifiers system. The relationship of them is parallel. The first two are belong to synchronous processing strategy of spectral-spatial classification, the third one belongs to the post-processing strategy.In this paper, the advantages of combination of spectral features and spatial features and ensemble learning are put in consideration. A study of classification methods for hyperspectral image by multi-Classifier Ensemble and spectral-spatial feature combination is conducted. Rich detailed characteristics are retained in the classification results. This method improves the accuracy and enhanced the stability for the classification map.