Research On Remote Sensing Image Classification Based On Multi-feature Fusion And Deep Beilef Network

Remote sensing image classification is to distinguish all kinds of ground objects in images by remote sensing classification technology.it is an important link of remote sensing image processing and is of great significance to national defense security construction,urban planning,disaster monitoring,landscape analysis and so on.The commonly used remote sensing image classification methods are supervised classification,unsupervised classification and so on.These classification methods can solve the problem of classification and recognition to a certain extent.With the progress of remote sensing image acquisition technology and the continuous improvement of image resolution,The image contains rich texture information,many features and complex distribution of ground objects,so it is difficult to classify and recognize.More and more scholars carry on the problem of high resolution image classification.Research,how to use this information has become the focus of research.Krizhevsky's classification accuracy when applying deep learning to image recognition is significantly higher than that of traditional classification methods.Since then,many scholars have begun to study in-depth learning,which has achieved good results in image,speech recognition and other fields.The success of in-deep learning in natural image recognition provides a basis for its application in remote sensing images.At present,how to apply in-deep learning to high-resolution remote sensing images has become a trend.In this paper,Matlab (R2017b) is used as the experimental platform,and the data are UC Merced_LandUse dataset and WHU-RS19 datase.The application of deep belief network in theclassification of high spatial resolution remote sensing image is studied in this paper.The main contents of the study are as follows:(1)Aiming at the high-resolution remote sensing image features,using a single feature to describe the classification accuracy is not high,a classification method combining multiple features and Deep Belief Network(DBN)is proposed to explore different image features.The relationship of classification accuracy.Firstly,seven characteristics of color,texture,locality and shape are extracted respectively,and then the fusion results are obtained according to the feature arrangement and fusion.Finally,the optimal feature combination is obtained by comparison of classification experiments.The analysis results show that the higher the number of feature combination categories,the better the classification accuracy.This method is used in the UC Merced_LandUse data sets and WHU-RS19.The experiment on the data set shows that the optimal feature combination results will lay the foundation for the classification research in the following paper.(2)In order to solve the problem of low classification accuracy of remote sensing images caused by feature redundancy in multi-feature fusion,ReliefF algorithm is introduced to optimize the features of the optimal feature combination.The weights of each feature variable are obtained by ReliefF algorithm,and different feature subsets are obtained by setting different threshold values.The optimal selection results are obtained by comparing the classification experiments between the feature subset and the full set of features.The experimental results show that the feature subset after removing redundant features by ReliefF algorithm can effectively improve the classification accuracy,and the classification accuracy of feature subset after feature optimization in two data sets is higher than that of feature full set.(3)In order to further improve the accuracy of image classification,In this paper,a linear discriminant analysis(LDA)and multi-feature optimization DBN high-altitude resolution remote sensing classificationmethod is proposed.According to the optimal feature subset and LDA feature,the feature fusion is performed to obtain a new feature vector,and then it is input into the DBN model as the total feature for image classification.The results show that this method can effectively improve the classification accuracy of UC Merced_LandUse data sets and WHU-RS19 data sets.The overall accuracy of LDA and DBN classification method based on multi-feature optimization is 89% on UC Merced_LandUse dataset and 87% on WHU-RS19 dataset,which is significantly higher than that of multi-feature optimal selection classification.Increased by 7% in UC Merced_LandUse dataset and raised on WHU-RS19 dataset It's 9% higher.