Research On Influence Factors Of Object-based Classification Of Medium Resolution Remote Sensing Data

With the development of sensors and platforms of remote sensing,the available remote sensing data are increasing,and their applications are becoming more and more extensive.As the basis of the application of remote sensing images,the classification of remote sensing images is of great significance.At present,the object-based method is an important research direction of remote sensing image classification.Many studies have found that this method is superior to the traditional method based on pixel.However,most studies using this approach have focused on high resolution images and lack of a comprehensive analysis of the factors that affect the performance of object-based classification.Choosing a suitable classification system,this paper makes a comprehensive analysis of the influence factors of a variety of medium resolution remote sensing data in object-based classification.The main work and results are as follows:1.On the basis of eCognition software,a set of rules is set up which can be used to carry out the classification with different parameter settings and the accuracy evaluation at the same time.On this basis,the object-based classification of Landsat8-OLI data,Landsat5-TM data and Gaofen-1 data are analyzed.2.The analysis of Landsat8-OLI data shows that:?1?The contribution rates of spectral,geometric and textural features to classification are 71.4%,11.8%and 16.8%,respectively.?2?The classification accuracy begins to be steady when the number of samples is 2-3 times of the feature quantity.?3?To segmentation parameters,the optimal values of scale are between 20 and 40.With the increasing of the scale,the shape decrease and the compactness present unobvious change.In addition,as the scale increase,the ratios of maximum objects quantity and minimum objects quantity vary in the range 7-11.?4?The parameter settings of the classifier show that to decision tree?DT?,the optimal values of the maximum tree depth vary from 5 to 8.To support vector machine?SVM?,the optimal values of Gamma are less than 10^-2 and C greater than 10².To random forest?RF?,the optimal values of active variables?Av?vary in the range 1-4 and max tree number?Mtn?in the range 30-100.In addition,SVM has the highest accuracy,followed by RT,Bayes,and DT.3.The analysis of Landsat5-TM data shows that:?1?the contribution rates of spectral,geometric and textural features to classification are 67.2%,11.5%and 21.3%,respectively.?2?When the number of samples is 2-4 times of the feature quantity,the classification accuracy can reach the stable state.?3?To segmentation parameters,the optimal values of scale vary from 4 to 6,and the values of shape and compactness are mostly less than 0.5.Moreover,with the increasing of scale,the ratios of maximum objects quantity and minimum objects quantity have a significant downward trend.?4?The parameter settings of the classifier display that the optimal values of the maximum tree depth within DT range from 3 to 8.To SVM,the optimal values of Gamma are less than 10^-2 and C greater than 10.To RF,the optimal values of Av vary in the range of 2 to7 and Mtn in the range of 30 to 100.In general,SVM has the largest accuracy,followed by RT,Bayes,and DT.4.The analysis of Gaofen-1 data shows that:?1?the contribution rates of spectral,geometric and textural features to classification are 65.5%,12.1%and 22.4%respectively.?2?When the number of samples is 2-4 times of the feature quantity,the classification results tend to be stable.?3?To segmentation parameters,the optimal values of scale vary in the range 15-25.For the shape and compactness,most shapes of SVM,RF,and Bayes are less than 0.5,and most compactnesses of SVM and RF are less than 0.5,while most compactnesses of Bayes higher than 0.5.To DT,all shapes are higher than 0.6 and the compactnesses present no obvious rule.Furthermore,with different scales,the ratios of maximum objects quantity and minimum objects quantity vary in the range 4-6.?4?The parameter settings of the classifier display that the optimal values of the maximum tree depth range from 3 to 6for DT.For SVM,the optimal values of Gamma are less than 10^-1and C greater than10.For RF,the most optimal values of Av range from 1 to 7 and range from 30 to 100for Mtn.On the whole,SVM has the largest accuracy,followed by RT,Bayes,and DT.5.The analysis of validation area demonstrates that the findings of object-based classification with medium resolution remote sensing data can be applied to images in different regions and phases.