A Research On Dataaugmentation Technologies For Remote Sensing Image Classification

After decades of development,remote sensing technology has become a comprehensive science and technology,which focuses on the latest achievements of computer,electronics,optics and other disciplines.The application scope of remote sensing technology is also expanding day by day,which has involved many aspects such as national daily life,national defense and security.With remote sensing technology becoming an integral part of contemporary high and new technology,remote sensing image classification technology has also made rapid progress.How to effectively improve the classification accuracy of remote sensing images is widely concerned by many researchers.In the process of remote sensing image classification,limited by the small number of labeled samples and low quality of remote sensing image data,it often leads to the problems of unsatisfactory classification model and low image classification accuracy.Aiming at high-resolution remote sensing images and hyperspectral remote sensing images,this paper focuses on effectively increasing the number and quality of remote sensing image samples to improve the image classification accuracy.The main research contents are as follows.Firstly,in view of the limited size of labeled samples in high-resolution remote sensing images,resulting in poor generalization ability and low classification accuracy of classifier model,based on the idea of labeling unlabeled samples and screening effective labeled samples,so as to expand the samples,a improved data enhancement algorithm based on DBSCAN is proposed.In the stage of image classification,according to the characteristics of high-resolution remote sensing images,vgg-16 is selected as the classifier model to extract deeper image features,which is conducive to the improvement of classification accuracy;In the data enhancement stage,the DBSCAN clustering algorithm is improved by using information entropy,which can label the unlabeled samples and effectively screen the samples with large amount of information as new labeled samples,so as to obtain the effect of data enhancement and improve the final classification accuracy.Secondly,aiming at the problems of large amount of data,high-dimensional features and less labeled samples in hyperspectral remote sensing images,which lead to large amount of classification calculation of hyperspectral remote sensing images,low model operation efficiency and low classification accuracy,a data enhancement algorithm based on improved SMOTE is proposed based on the idea of sample expansion due to the labeled samples of a few classes at the synthetic boundary.In the stage of image classification,the neighbor support vector machine improved by kernel principal component analysis is used as the classifier model,which is helpful to improve the efficiency of image classification;In the data enhancement stage,GA-FCM clustering algorithm is used to improve Borderline SMOTE,shorten the running time to a certain extent,and achieve the data enhancement effect by synthesizing a few boundary samples,so as to improve the image classification accuracy.Finally,on different open source data sets used by researchers,the improved data enhancement algorithm based on DBSCAN and the data enhancement algorithm based on improved SMOTE are experimentally verified respectively,and compared with other similar data enhancement algorithms proposed recently.