Font Size: a A A

Region Identification Of Dendrolimus Superans Pest Based On Multispectral Remote Sensing Images

Posted on:2022-05-12Degree:MasterType:Thesis
Country:ChinaCandidate:W J ChenFull Text:PDF
GTID:2493306314494634Subject:Forest Engineering
Abstract/Summary:PDF Full Text Request
Forests have always played a vital role in the global ecosystem and keeping abreast of the health of forests is of paramount importance to maintain ecosystem stability.Forest insect pests are a very important factor in reducing the quality of forest health,so the study of forest insect pests has been closely followed by many experts and scholars for many years.In recent years,with the development of remote sensing technology,remote sensing technology has been widely applied to various aspects of forestry research,effectively solving some of the problems encountered in forestry.This paper focuses on the use of certain image features extracted from multispectral remote sensing imagery to identify areas of pest occurrence in forestry.In this study,the multispectral remote sensing images from Landsat-8 and Sentinel-2A satellites were used as data sources to extract band spectral features,spectral index features and texture features from the images as the basis features for pest area identification.The important features were selected using Analysis of Variance and XGBoost classifier feature importance analysis methods,and the selected features were used to identify pest areas in the study area using two classifiers,the random forest classifier and XGBoost classifier,to compare and evaluate the identification results.The results of this study showed the following.(1)Sentinel-2A satellite images are more suitable than Landsat-8 satellite images for the identification of larch caterpillar infestation areas in northeast China,and the infestation areas identified by them remain largely consistent with the actual survey situation.The overall recognition accuracy of Sentinel-2A images was 92%,with a kappa coefficient of 0.81.(2)Both Analysis of Variance and XGBoost classifier have significant effects on feature selection.In this study,12 features that were extremely insensitive to pest monitoring were filtered out through Analysis of Variance,reducing the number of feature variables from 36 to 24.Afterwards,the feature importance analysis using XGBoost classifier was used to rank the importance of the feature variables,which laid the data foundation for the next study of the feature importance-based pest zone identification model.(3)The XGBoost classifier recognition model with the top 14 features in terms of feature importance was used as the final study area pest recognition model based on Sentinel-2A imagery.The random forest was used to differentiate the study area between infested and healthy areas according to feature importance,with the top 10 features having an accuracy of 0.934 and the subsequent addition of features having less impact on the accuracy.The XGBoost classifier was used to distinguish between infested and healthy areas according to the importance of the features,with the top 14 features having an accuracy of 0.945 and the subsequent addition of features having less impact on the accuracy.
Keywords/Search Tags:Dendrolimus superans, multispectral remote sensing image, variance analysis, feature selection, pest zone identification
PDF Full Text Request
Related items