The Study On Classification Of Forest Fuels Based On Multi-temporal Remote Sensing Images

Based on the 2017 Landsat8 remote sensing image data,this paper determines the forest fuels category in the F Miaofeng Mountain Forest Farm based on the field survey data and the map data of the Miaofeng Mountain Forest site map and the small class map.Comparative analysis of the main conifer species such as Pinus tabulaeformis,larch forest,and side berlin,the main broad-leaved tree species such as eucalyptus forest,eucalyptus forest,and Wujiao Fenglin,and the spectral characteristics of several forest fuel types such as the spiraea-based shrub forest.Then use the vector machine(SVM)algorithm in EnMap-Box,random forest(RF)and CART-based decision tree method to classify,select the classification method with the highest overall classification accuracy and apply it to image classification,and finally burn the forest.The categories are divided into eight categories:Pinus tabulaeformis,larch forest,side Berlin,eucalyptus forest,eucalyptus forest,pentagonal maple forest,shrub forest and others(including roads,buildings,etc.).At the same time,two non-fire-proof periods(May 7th,September 28th),two fire-proof periods(November 15th,December 17th),a total of four phase diagrams for spatial registration,and based on feature combinations Change detection,research explores differences in results based on single-phase and multi-temporal fuel classifications.Research indicates:(1)Establish the classification criteria of fuels by spectral characteristics,and compare and analyze the spectral characteristics of Pinus tabulaeformis,larch forest,side Berlin,eucalyptus forest,eucalyptus forest,pentagonal maple forest,shrub forest and other eight forest fuels.It was found that the distinction between coniferous forest and broad-leaved forest and shrub forest was obvious,while that of coniferous forest such as Pinus tabulaeformis and broad-leaved forest such as Quercus variabilis forest was more obvious.(2)This paper compares and analyzes three classification methods:support vector machine(SVM)algorithm,random forest(RF)and decision tree based decision tree(CART).In the non-fireproof period image(May 7,September 28),the overall accuracy of the support vector machine(SVM)method was found to be higher than the other two.Obtaining the penalty parameter(C)of 1000 and the kernel parameter(g)of 10 makes the SVM classification model optimal.The overall classification accuracy is 83.34%,the kappa coefficient is 0.78,and the accuracy is increased by 3.69%and 11.29%with respect to RF and CART,respectively..In the fireproof period image(November 15th,December 17th),the random forest classification method is often more accurate.(3)Studies have shown that multi-temporal images can better distinguish fuels from single-temporal images,and can reflect changes caused by seasonal alternation,especially in broad-leaved forests,with precision far exceeding that of single-phase images.Among them,based on the classification method of random forests,in the multi-temporal image forest fuels classification,the fuels categories of broad-leaved forests such as eucalyptus forests can be better classified.