Application Of Gaofen No.2 Remote Sensing Imaging In The Classification Of Forest Land In The Transition Belt Between Sichuan Plain And Mountain Area

Forests play an important role in maintaining the balance of terrestrial ecosystems and promoting social and economic development.It is significant to carry out forest resources investigation and monitoring,grasp its changes and forecast its development trend for promoting forest resources and promote environmental protection and even social and economic development.The traditional methods of forest resource regulation and monitoring can not meet the forest resource monitoring.The development of remote sense technology provides a very important technical mean for forest resource investigation and monitoring.Extracting forest land or forest type information by remote sensing image classification is an important method for forest resource investigation and monitoring.In recent years,with the launch of high-resolution satellites such as G-F and ZY,data support and service guarantee have been provided in the fields of land,agriculture,environment,disaster prevention and mitigation.However,due to the launch of high-resolution satellites just in few years,domestic scholars have relatively little research on the extraction and classification of forest information from high-resolution data,and there are many technical problems to be solved,especially in the areas between Sichuan Plateau and the Sichuan Basin.This research based on the remote sense image of GF-2 in Danjingshan Town,Pengzhou County,Sichuan Province,used the maximum likelihood classification with pixels,decision tree classification and object-oriented K-nearest neighbor classification to study the remote sense classification of forest land.On the basis of the accuracy evaluation of the three classification results,the best one will be determined.The good classification scheme was applied to Xinxing Town of Pengzhou County to verify.The purpose of this study is to explore the techniques and methods of extracting and classifying forest land information from GF-2 satellite images in the transition zone between Sichuan Plateau and the Sichuan Basin,meet the needs of forest resources investigation,monitoring and management in this region.The main results are as follows.(1)The best band combination of GF-2 remote sense image in Danjingshan Town of Pengzhou County was studied.In single-band statistical analysis,the order of mean and standard deviation of each band of GF-2 from large to small is as followed: band 1 > band 3 > band 2 > band 4,which indicates that band 1 and band 3 have relatively rich spectral information.In multiband statistical analysis,band 1,band 2 and band 3 have strong independence.In the analysis of OIF index,the best index(OIF)value of 123 band combination is the largest and contains the largest amount of information.This shows that band 123 is the best three-band combination of GF-2 remote sense images in Danjingshan Town,Pengzhou County.It has abundant information and obvious layers of ground objects,which is conducive to image classification.(2)In the image segmentation of GF-2 in Danjingshan Town,Pengzhou County,the thresholds of segmentation scale are set to 10,20,30,40,50,60,70 and 80,and the merging thresholds are 10,30,50,70 and 90.Through the analysis of the size of segmentation and merging objects and the discrimination between classes.It is found that when the segmentation threshold is 40,the merging threshold is 70,the image segmentation effect is the best.This shows that the segmentation scale 40 and the merging scale 70 are better scales for forest land classification of GF-2 remote sense images of Danjingshan Town,Pengzhou County,which can better distinguish different types and meet the forestry production.(3)Supported by ENVI5.3 remote sense processing software,forest land remote sense classification of Danjingshan Town in Pengzhou County is carried out by three classification methods:maximum likelihood classification,CART decision tree classification and K-nearest neighbor classification,and the classification result map are obtained.The overall classification accuracy of maximum likelihood classification,CART decision tree classification and K-nearest neighbor classification are 80.93%,83.59% and 91.62%,respectively.The Kappa coefficients are 0.7509,0.7659 and 0.8807,respectively.The overall accuracy of the three methods are over 80%,and the Kappa coefficient is over 0.7,which meets the requirements of forest land classification.Among them,the object-oriented K-nearest neighbor method has the highest overall classification accuracy and Kappa coefficient.Compared with the results of forest resource planning and design survey,the relative errors of K-nearest neighbor method in the area of Danjingshan Town are generally smaller,which shows that the object-oriented K-nearest neighbor method is the best method to classify forest land from the GF-2remote sense image of Danjingshan Town,which can meet the requirements of forest land classification.(4)The best classification case identified(the K-nearest neighbor method with a partition scale of40 and a combined scale of 70 in the 123-band combination of GF-2 remote sensing images)was applied to the forest land classification of Xinxing Town,Pengzhou County.The overall accuracy of the classification results was 84.58%,and the Kappa coefficient was 0.8137.Compared with the forest resources planning and design survey results of Xinxing Town in Pengzhou County,only the cultivated land area was wrong,and the differences are more than 10%.The areas errors of other land types are small,and the overall classification effects are better.In summary,the object-oriented K-nearest neighbor method,with a segmentation scale of 40 and a merging scale of 70,is suitable for remote sense classification of forest land in the transition zone between Sichuan Plateau and the Sichuan Basin under the condition of 123-band combination of GF-2remote sensing images,and can meet the requirements of forestry production.