Research On Extraction Of Bamboo Forest Area In ZY-3 Remote Sensing Images

Bamboo forest resources are an significant natural resource in China,They not only occupy an important position in the natural landscape,but also bring huge social and economic benefits.Extracting bamboo forest resource information accurately provides a valuable reference for rational space allocation and optimization and adjustment of resource.At present,the classical pixel-based classification method and object-based classification method are widely used in the classification of features,but the current classification method can't meet the user's precision requirements.Although the machine learning method is continuously improved,the automatic extraction algorithm with simple rules and single structure shows great limitations in the ease of complex terrain distribution restricted by large differences in features,fragmentation of the land,and metamorphism,etc.Therefore,improving the extraction accuracy has become the key to the application of remote sensing technology in the forestry field.This paper makes a theoretical analysis and research on extraction of the bamboo forest in Yifeng County,Jiangxi Province,and puts forward an proposed classification algorithm to improve the accuracy of bamboo forest information extraction.The research work contains three aspects as follow:(1)Study of feature extraction and feature sets construction:On the basis of ZY-3 remote sensing images correction,fusion and features extraction,the importance of features should be assessed based on random forest.Then,traversing to select different features and building combinations,the extraction results as the source of evidence,using the weight coefficient to evaluate them.According to above results,some important features are selected and recombined to build different types of feature sets,such as spectral set,texture set and joint set.(2)Study of Gradient Boost-Random Forest Fusion Classification Model:There are many factors affecting the classification of bamboo forests.Excessive factors will not only cause data redundancy,but also reduce the accuracy of bamboo forest classification.Using gradient boost machine learning method to eliminate the useless factors and retain highly relevant factors.In addition,in order to inmprove the diversity and classification performance of the system,Using random forest machine learning method to select the optimal features among the candidate features,and construct a variety of different sub-decision trees.The final output is obtained by voting on the judgment result of the sub-decision tree.(3)Accuracy comparison verification:The bamboo classification results obtained by the improved classification algorithm,namely the gradient boost-random forest fusion model classification method,are compared with the classical classification results based on the pixel classification method and the object-oriented classification method.The main comparative analysis indicators are:overall classification accuracy,Kappa coefficient,bamboo forest producer accuracy,bamboo forest user accuracy(4)Analysis of bamboo forest area change:In 2017,the area of bamboo forest in Yifeng County of Jiangxi Province was 573.44 km2,accounting for 52.47%of the total forest area.The bamboo forest area is mainly distributed in the hilly areas of the southwest,northeast and west of Yifeng County.Compared with 2009,the area of bamboo forest in the study area showed a growth period during the study period,with a change range of 11.42%.The results of the study showed that based on the ZY-3 remote sensing images of Yifeng County,Jiangxi Province,a new classification algorithm,namely gradient boost-random forest fusion model classification method,was proposed to improve the accuracy of bamboo forest classification in the study area.Gradient boost is used as a technical means to extract bamboo forest features and build feature sets,and the random forest classification method is used to evaluate the importance of features and assign different weights.Finally,a variety of sub-decision trees are constructed,and the final output result is obtained through the voting result of the decision tree.The overall accuracy and Kappa coefficient of the newly developed algorithm for the extraction of bamboo forest information in the study area are 93.25%and 0.9239,the user accuracy and producer accuracy of bamboo forest extraction are 96.06%and 94.57%.Compared with the traditional,single-structured pixel-based and object-oriented classification methods,the combined machine learning classification method has a great improvement in the accuracy of each index.In summary,the newly developed classification method can extract bamboo forest features and construct feature sets,which can effectively improve the classification accuracy of bamboo forest classification in the study area.