| Forest ecosystem is an important part of biodiversity and ecosystem stability on earth,and forest pests are one of the important factors leading to the deterioration of forest health,so timely and accurate understanding of the occurrence and evolution trend of forest pests is of great significance to protect forest ecosystem.In recent years,with the development and popularity of remote sensing technology,it has been widely used in forestry research,which can effectively solve the problems that are difficult to be covered by traditional monitoring means,such as large spatial range and high temporal resolution.In this study,taking bark beetle and aspen leaf miner as the research objects and using multispectral remote sensing Sentinel-2 image data,the regional detection method of forest pests is studied,aiming to improve the accuracy and efficiency of forest pest monitoring,provide support for forest pest investigation and management,and promote the sustainable development of forest ecosystems.The main research contents are as follows:(1)To study the pest area identification method of multispectral satellite images based on random forest.First,based on the calculation methods of multiple vegetation indices,the vegetation index features of the original Sentinel-2 images are calculated,and the original spectral features and vegetation index features are extracted to construct the band feature dataset.Then,each band factor is used as input to identify forest pest areas based on random forest algorithm.Finally,the importance score of each band is analyzed and the influence of each band factor is discussed.The experiments illustrate that the vegetation index feature information can improve the effectiveness of the random forest-based pest area identification method.Further comparison experiments are conducted to analyze the influence of red edges and vegetation index features derived from red edges on the recognition effect,which illustrates the importance of red edge information for detecting pest areas.(2)To study a multispectral satellite image forest pest region identification method based on improved UNet++.Firstly,the Sentinel-2 images and the corresponding raster form of pest region labels are cropped in the same regular grid way to construct a semantic segmentation dataset of forest pest regions.The common semantic segmentation models are compared and experimented,and then UNet++ is improved by using Res Ne St101 as the feature extraction backbone network and adding the sc SE attention mechanism module in the decoding part to construct the forest pest region segmentation model.The segmentation model is trained and tested using the dataset and compared with other common semantic segmentation models.The results show that the improved model can segment the forest pest region more accurately and achieve the best segmentation results.Finally,ablation experiments were conducted to separately quantify the effects of sc SE module and Res Ne St on the improved model,and the experimental results show that both have positive effects on the performance of the model.(3)To analyze the effects of vegetation index and multispectral data on the improved UNet++ model,and to study the band optimization method based on the importance of random forest band features.Based on the original RGB three-channel data,the native bands and vegetation index bands are added as the input data of the network,the improved UNet++ model is used for pest area segmentation experiments,the results are compared and analyzed after using different bands for training.The experimental results show that the vegetation index and multispectral data can improve the segmentation accuracy of the model compared with RGB data.The experiments show the importance of multispectral data and vegetation indices for deep learning-based forest pest detection models.In addition,based on the band feature importance score of random forest,a band optimization method is investigated to optimize the band and further improve the segmentation accuracy. |
PDF Full Text Request