Ensemble Learning Method And Optimization Of Remote Sensing Image Classification Problem

In recent decades,remote sensing classification has been one of the important topics in the field of remote sensing and patern recognition.The ensemble classification algorithm is also a hot topic in the field of remote sensing classification in recent years.It has always been the focus of research about how to combine the characteristics of remote sensing image data,and improve the accuracy of ensemble algorithm effectively.In this paper,remote sensing data is taken as the research background,the ensemble algorithm is the main research object.The remote sensing image classification problem is studied from different perspectives,such as feature extraction of training samples,diversity of base classifiers and local model ensemble algorithm.The research work of this paper mainly includes the following aspects:(1)In order to solve the problem of remote sensing image classification by using Rotation Forest ensemble classification algorithm,Rotation Forest based on Non-negative Matrix Factor(NMF)and Rotation Forest base on Auto Encoder(AE)are proposed respectively.By imporving the algorithm,more effective image features can be obtained,and a training feature set with differences can be constructed to improve the basic classifier.Diversity will ultimately improve the overall classification accuracy.In the process of feature extraction,considering the non-negative characteristics of remote sensing images,it is easy to get negative values using Pincipal Cmponent Analysis(PCA),which is unable to maintain the non-negativity of the original data and affects the overall classification effect.The non-negative features can be accurately obtained using NMF.It keeps non-negativity consisitent with the original data,which is suitable for remote sensing image data.At the same time,there is a nonlinear relationship in the processing of complex data in remote sensing image data.PCA is a linear transformation method,which can not deal with the nonlinear relationship very well.AE network can represent both linear and nonlinear transformation.It can extract complex features are used for simulation experiments,and the experimental results show that the two algorithm have achieved good classification results.(2)From the view of diversity among base classifiers of ensemble algorithm,aiming at the problem of how to obtain a group of base classifiers with large differences,this paper proposes two methods of remote sensing image classification,selective ensemble based on mutual information and selective ensemble based on Q statistics.Some classical ensemble algorithms do not consider the relationship between base classifiers.In this paper,we select some base classifiers with large differences to integrate.From the theoretical and experimental view,the classification effect of the selected classifier group is better than that of all the classifiers.At the same time,from the number of base classifiers,it can be seen that the selective ensemble algorithm can greatly save the overall operational efficiency.In the process of selecting base classifier,considering that remote sensing image data has multi-features,high dimensions and contains high order statistical relationship.Mutual information theory has a good effect in dealing with high order statistics.Therefore,this paper uses mutual information theory to measure the differences.The classifiers with large difference are as the final classifier.In addition,considerring the complexity of the ensemble algorithm,Q statistic is a pairwise measurement method,which has the characteristics of simple calculation and easy to understand.Using Q statistic to measure the difference between base classifiers can save operation efficiency while ensuring classification accuracy.The experimental results show that the proposed two selective ensemble algorithms can effectively improve the classification accuracy and have good generalization performance.(3)Considering that hyperspectral remote sensing images have complex spectral characteristics,the classification algorithms of hyperspectral remote sensing images based on multiple reduced kernel extreme learning machine and deep learning extreme learning machine are proposed to solve the problem of multi-feature and different characteristics.In multiple reduced kernel extreme learning machine algorithm,by using multiple kernel model and using different kernel models for different features in hyperspectral remote sensing images,multiple kernel mixture has more complex dynamic feature representation ability,and can get better results when dealing with the classification of hyperspectral image data.In addition,the deep learning network model in combination with extreme learning machine to process the hyperspectral image data,it has many characteristics and high dimension characteristics of deep network has the function of automatic traing characteristics in the process.It can save the efficiency,through the adaptive parameters.And it can be effective to dimension reduction of hyperspectral image.After combined with extreme learning machine for classification,it can achieve the desired effect.The simulation results hyperspectral images show the proposed algorithms achieve good classification effects when dealing with hyperspectral image data,and have good pertinence and adaptability.