Font Size: a A A

A Comparative Study Of Object-oriented Forest Classification Methods Considering Feature Types And Feature Selection

Posted on:2020-10-29Degree:MasterType:Thesis
Country:ChinaCandidate:H Y YangFull Text:PDF
GTID:2433330602451105Subject:Cartography and Geographic Information System
Abstract/Summary:PDF Full Text Request
Forest classification is the foundation and key link in the forest resource monitoring system.It has great significance for mastering the status of forest resources,improving the level of forest management and promoting the sustainable development of forestry to achieve forest classification efficiently,accurately.Remote sensing technology is widely used in forest classification and area extraction because of the advantage of wide range,high timeliness and low cost.However,at present,there are many problems in the forest classification,such as insufficient attention to new data sources in China,insufficient classification category,insufficient feature types selection,and insufficient information mining,et al.In view of the above deficiencies,this paper which used on GF-1 WFV images and combined with object-oriented image analysis methods and three classifiers including CART,SVM,RF research on the influence of different feature types,different feature selection methods on different classifiers'classification accuracy,and then selects the best results by the optimization of classification methods and classification strategies to achieve the fine forests classification of Deqing County.The main contents and conclusions of this paper are as follows.(1)Research on object-oriented multi-scale segmentation parameters.This study determines the band weight by OIF band selection method,and combines the multiple segmentation test and ESP scale evaluation tool to obtain the optimal scale parameters.The results show that when scale,shape factor,compactness factor is respectively set to 78,0.1,and 0.5,the segmentation effect is the best;combined with the OIF band selection method and ESP tool can effectively realize the selection of segmentation parameters and optimize the segmentation effect,which is of great significance for further classification research.(2)This paper combines different feature types such as multi-temporal features,spectral features,thematic index features,texture features,color features,shape features and topographic features and sets up eight sets of feature combinations.Then,under CART,SVM and RF three classification methods,it is used to explore the influence of different feature types on the classification accuracy and single land class accuracy,analyze the importance of different feature types to forest classification,and compare the classification performance of three classification methods.The results show that:1)Under all features combination,the overall accuracy of the three classification methods is that the RF classification is larger than the SVM classification than the CART decision tree classification,and the RF classification method has better recognition ability for forest types.2)For the three classification methods,multi-temporal features,spectral features,vegetation index features,HIS color features and topographic features can effectively improve the overall accuracy of classification,and the multi-temporal features are the most significant for the overall accuracy's improvement.The addition of texture features and shape features can't improve the classification accuracy,which in turn leads to a reduction in classification accuracy for some classifiers.3)Different feature types have different influences on different classifiers.In practical applications,we should combine with specific situations to select feature types.(3)Under the above three classification methods,this paper develops forest classification research through two feature selection methods including the features importance of random forest and Boruta algorithm.The conclusions are shown as follows:1)The importance of different feature types and different phase characteristics to forest classification is different.The importance degree of each feature type is:color feature>topographic feature>spectral feature>exponential feature>texture feature>shape feature;the importance degree of each phase is:August>December>February>October.2)From the perspective of the change of accuracy before and after feature selection,feature selection is the most effective for improving the classification accuracy of CART decision tree.SVM is second,and RF classification accuracy is least affected by feature selection.In addition,SVM classifier is most sensitive to changes of feature quantity,but RF classifier shows better stability.3)For the three classification methods,the combination of multiple feature types and feature selection can effectively improve the accuracy of forest classification.Compared with the feature selection method based on Boruta algorithm,the feature selection method based on feature importance of random forest is more significant for classification efficiency and classification accuracy.(4)By comparing the accuracy of classification results between different classification strategies and different classification methods,it is found that the classification accuracy is the highest when using RF as the classifier,integrating all feature types and using the feature importance of random forest as the feature selection method for forest classification.The overall accuracy and Kappa coefficients are respectively 88.75%and 0.8694.Compared with the survey data,the total forest area extraction accuracy is 98.63%,and the accuracy of most forest types is higher than 80%,indicating the effectiveness of the method.
Keywords/Search Tags:GF-1, Object-oriented, Forestmeticulous classification, Feature type, Feature selection
PDF Full Text Request
Related items