Lossy Image Compression Quality Assessment For Remote Sensing Image Classification

Compression techniques are essential for reduce the volume needed for the storage and transmission.Lossy compression inevitably leads to image quality degradation,which is directly related to subsequent application.Generally,visual interpretation or quality index can be used to evaluate the compression image quality.Hovewer,it is a problem whether these methods can actually reflect the application ability.Classification is one of the most important fields of remote sensing and important part of application tasks.Lossy compression has not yet achieved global acceptance in remote sensing,mainly because the negative effects have not been quality fully investigated so far.JPEG 2000 is a normal compression standard.Thus,assessment of the effects of lossy compression for classification accuracy can be used to assessment the performance of classification algorithm.Besides,it is also related to whether lossy compression can be used in classification process.In light of these issues,we investigate the impact of lossy compression on image classification.As various compression ratios are considered,four classification methods are used on reconstruction images with same samples.The robustness of different methods and classifiers is tested.Besides,the influence of compression on the classification accuracy is analyzed.Then,we test the ability of two quality indexes on classification,and propose a prediction model by feature extraction.To summarize,our contributions are as follows:1)The robustness of different methods and classifiers is analyzed in this paper based on Landsat-8 and ZY-3 datasets.To be specific,classification methods include:pixel-based MLC,pixel-based SVM,object-based SVM and object-based RF.The results show that,the pixel-based classification accuracy drops with the increase compress ratio.Correspondingly,the object-based classification accuracy is also affected by compression,but not depends on the compression ratio.In some cases,the reconstructed image under the higher compression ratio can get better classification accuracy than the low compression ratio,which is because the compression makes the local smoothing.This phenomenon can improve the classification result to a certain extent.2)A method is proposed to predict the classification accuracy for lossy remote sensing imagery via extracting features by expressing spectral,texture and shape attributes.To be specific,the model etracts features from spectral,texture and shape aspect.Experimental results demonstrate the prediction index is more consistent with classification accuracy on two datasets,which can be used for classification accuracy prediction by the input of image features and classification method.The experiment shows that the mean feature can not be a good indication of the variation of classification accuracy for the classes greatly affected by compression.For the coarser texture classes(such as water),the texture features have little influence on the classification accuracy.For vegetation and forest,the features extracted from near infrared and red bands play an important role in the prediction of classification accuracy.