Effects Of Compression Distortion In Remote Sensing Image On Classification Accuracy

The development of remote sensing technology,especially modern sensor acquisition technology,has increased the radiometric,spatial,spectral and temporal resolution of remote sensing image to some extent.Remote sensing has also played an important role in the fields of target detection,weather forecasting,fire detection,military surveillance,and agricultural practice evaluation.Wide coverage of remote sensing data and improvement of remote sensing image resolution make the amount of data collected by multi-channel sensors difficult to estimate.Massive remote sensing data requires sufficient bandwidth to transmit the data,enough space to store the data,and abundant computing resources to process the data.Under the existing actual conditions,these requirements are difficult to be fully satisfied at the same time.To resolve the contradiction between the increasing amount of remote sensing data and the extremely limited hardware storage and transmission conditions,to compress the remote sensing data is a means worth considering.The lossy compression on remote sensing image sacrifices some valuable ground information in order to obtain a larger compression ratio and a smaller amount of data.Therefore,lossy compression will inevitably introduce compression distortion in remote sensing image,which will cause image distortion and quality loss.The terminal end-users of remote sensing image are various types of remote sensing applications.The decline in image quality will directly affect the subsequent application value of remote sensing image.Remote sensing image classification is an important part of remote sensing image processing and one of the most widely used fields in remote sensing.It is very meaningful to study the effects of compression distortion in remote sensing image on classification accuracy.In view of the above problems,this paper quantitatively studies and analyzes the effects of compression distortion introduced by JPEG2000 in remote sensing image on classification accuracy,and proposes a prediction model for classification accuracy of remote sensing image based on fractal analysis.In this paper,GF-2,Landsat 8 and ZY-3 satellite images are used to establish a classification-oriented remote sensing compressed image data set.The JPEG2000 algorithm is used to compress the remote sensing images.The maximum-likelihood and SVM classification algorithms are used to classify the remote sensing images in the database and classification accuracy metrics such as Kappa coefficient,overall accuracy and average accuracy are calculated.Then,according to the characteristics of compression distortion introduced by JPEG2000 and analysis of the variation trend of classification accuracy with the degree of compression distortion,multiscale feature extraction is performed on the remote sensing image based on fractal analysis from two perspectives,i.e.,local and global structural feature description.In terms of local structure,the fractal dimension of the remote sensing image is calculated pixel by pixel based on the single fractal analysis to obtain the fractal dimension matrix,and then the corresponding statistical features are extracted as the local structural feature description based on the fractal dimension matrix.In terms of global structure,the multifractal spectrum of the remote sensing image is calculated,and the width of the multifractal spectrum is used as the global structural feature description.After the multiscale feature extraction is completed,a no-reference prediction model for remote sensing image classification accuracy is established by using the multiple kernel learning algorithm.Then the automatic prediction of the classification accuracy of remote sensing image can be realized.The experimental results show that the classification accuracy prediction model proposed in this paper performs excellently in predicting classification accuracy of remote sensing image and the prediction performance is outstanding,which prove the validity of the model.Besides,the prediction results are in good agreement with the ground truth.Compared with other reference comparison algorithms,the algorithm proposed in this paper has obtained optimal results on several prediction performance evaluation metrics such as PLCC,SROCC,KROCC,and RMSE.In repeated experiments,the data fluctuation of the prediction model in this paper is also the smallest,which further validates the model's effectiveness and superiority.Finally,the shortcomings of this paper and the further improvement directions are discussed.