Extraction Natural Forests Information In Yuanling Based On A Hierarchical Strategy

As an indispensable part of the ecosystem,the forest plays an irreplaceable role in human life and production.The natural forest is the main body of forest resources,and the degree of attention and protection of natural forests is also gradually increasing.To accurately grasp the stand composition and growth status of natural forests is the premise of quality evaluation and protection of the natural forest.However,limited by the topography and traffic conditions,the traditional natural forest survey methods take time and effort.In recent years,spatial information technology based on satellite remote sensing has provided technical support for rapidly acquiring forest information in a large area.However,due to the complex topographic conditions and stand structure of the natural forest,the accurate extraction of information on different forest types in natural forests is seriously restricted.In order to improve the extraction accuracy of natural forest information under complex topography,according to the difference in the spectral response of different forest types,this paper selects the natural forest of seven villages and towns in the northeast of Yuanling County,Huaihua City,Hunan Province as the research object,takes the multi-temporal Sentinel-2 image as the data source,combined with the sample data of different forest types of natural forest ground survey,using the maximum likelihood algorithm(ML),the artificial neural network algorithm(ANN),the support vector machine algorithm(SVM)and the random forests algorithm(RF)as basic classifiers.Based on the hierarchical strategy,an ensemble learning(EL)algorithm based on voting is constructed and applied to the information extraction of natural forests.The experimental results show that the hierarchical strategy and ensemble learning algorithm can fully use the differences in remote sensing features in different forest types and improve the accuracy and reliability of natural forest information extraction.The main results are as follows:(1)The voting-based ensemble learning algorithm improves the classification accuracy of natural forests by combining the results of a single classifier.According to the natural forest classification results of a single classifier and integrated learning algorithm,the natural forest classification result of the ensemble learning algorithm is the best.Its overall accuracy and Kappa coefficient is 83.36%and 0.78,respectively.Compared with the four single classifiers,the overall accuracy of the integrated learning classifier is improved by 0.31-5.19%.The experimental results show that the integrated learning algorithm can better integrate the advantages of every single classifier and improve the classification accuracy of natural forests.(2)Through the stratification strategy,it is found that there are differences in the ability of different remote sensing features to distinguish natural forest types.According to the spectral reflectance curve of the multi-temporal band,the classification nodes with the best separability are selected,and all land types are divided into three levels:the first layer is woodland and non-woodland;the second layer divides woodland into the bamboo forest and other forests;the third layer divides other forests into the broad-leaved forest and coniferous forest.The importance of random forest is used to rank the remote sensing features at each classification node to determine each node’s best classification feature.The results show that the best classification feature of forest land and non-forest land is B2,the best classification feature of bamboo forest and other woodland is B12,and the best classification feature of broad-leaved forest and the coniferous forest is B1.(3)The optimal set of features for identifying different forest types was obtained by fusing hierarchical strategies and integrated learning algorithms to improve natural forests’ classification accuracy and reliability.With the rank of feature importance and five classification algorithms to classify each node,the results show that the integrated learning algorithm achieves optimal classification results at each node,with 98.62%accuracy and the Kappa coefficient of 0.96 for the node of forest and non-forest;96.70%accuracy and Kappa coefficient of 0.86 for the node of bamboo forest and other forests;the accuracy of the broadleaf and coniferous forest nodes reached 86.71%with the Kappa coefficient of 0.74.After fusing the classification results of all nodes,the overall classification accuracy reached 87.18%with a Kappa coefficient of 0.82.The results demonstrate that combining hierarchical strategy and integrated learning can effectively improve the accuracy and reliability of natural forest information extraction.