Hyperspectral Image Classification Based On Weighted Extended Multi-attribute Profiles And Extreme Learning Machine

Hyperspectral image contains rich spectral information and spatial information.It can describe ground objects features more delicately,so it is widely used in agriculture,environmental inspection,military,geology and other fields.However,according to recent studies,there are still many challenges in the research of hyperspectral image classification.For example,the data structure of each pixel of hyperspectral is very complicated,and the data dimension is very large.It is very difficult to achieve high-precision classification with less time consumption in the case of limited samples;hyperspectral image has a wealth of image information,but the current research has not fully utilized the different spatial information of hyperspectral image;the extracted image features cannot be well fused;the extended multi-attribute profiles(EMAPs)of hyperspectral image feature will introduce noise in the hyperspectral image when connecting multi-attribute features.These problems will have a great impact on the accuracy of hyperspectral image classification.This paper proposes a new framework for joint decision fusion and feature fusion for hyperspectral image classification.(1)Extended multi-attribute profiles(EMAPs)have been applied to feature extraction and classification of remote sensing images because of their good performance.Since the extended multi-attribute profiles connects multiple attribute features without considering the pixel-based hyperspectral image classification,the homogenous region of the hyperspectral image may become unsmooth due to the noise.In order to solve this problem,this dissertation uses weighted mean filters(WMFs)to reduce noise and smooth the homogenous regions in hyperspectral images,and proposes a weighted extended multi-attribute profiles based on weighted mean filters(WMFs)(weighted extended multi-attribute profiles,WEMAPs).(2)In order to obtain the characteristics of the hyperspectral image,this paper proposes feature fusion(FF)based on weighted mean filters(WMFs)and weighted extended multi-attribute profiles(WEMAPs)to improve the discriminative ability of the classifier.It will have some similar areas but not totally the same,because the hyperspectral image has a very complex structure.However,its extended multi-attribute profiles may ignore some distinguishing information of the hyperspectral image.In order to capture different spatial structures to more effectively model the discriminative information of hyperspectral images,In this dissertation,a multi-scale method is adopted to generate multi-scale hyperspectral image features,so as to extract different spatial structures of hyperspectral images and lead to better classification results,in which the classification results of each scale are combined into the last optimal result.(3)In addition,an extreme learning machine(ELM)is used as a classifier to perform classification.In order to achieve hyperspectral image classification,different classes in the hyperspectral image may need to be projected into different feature spaces.Therefore,in order to distinguish different categories,this dissertation proposes a revised extreme learning machine(RELM)to handle different categories with different predictions.Then,the classification results of three different versions of extreme learning machines: the revised extreme learning machine(RELM)proposed in this dissertation,generalized extreme learning machine(GELM),and kernel extreme learning machine(KELM)are combined together,and finally the decision fusion method(decision fusion,DF)is used to obtain the optimal classification result.This is a new framework(joint decision fusion and feature fusion,JDFFF)proposed in this dissertation.Experiments conducted on two publicly available hyperspectral datasets(i.e.,the Indian pine dataset and the Pavia University dataset)show that our proposed algorithm is significantly better than many of the latest hyperspectral image classification algorithms.