Research On The Detection Of The Tree Infected With Pine Wilt Disease Ultilizing Satellite Remote Sensing

The pine tree is a symbol of fortitude,fidelity and longevity,and has great medicinal,ecological and economic value.Known as the "cancer of pine trees" and the "SARS" of the plant,pine wilt disease has caused the death of more than 600 million pine trees in China and economic losses of hundreds of billions of RMB.Satellite remote sensing is the primary methodology for census and monitoring of pine wilt disease in China.However,at present it can only identify the pine wilt disease that occur at the stand scale,which is a serious constraint on control efforts.In this context,this study combines satellite remote sensing technology,machine learning theory and other multidisciplinary knowledge to explore strategies for identifying pine wilt diseases occurring at the tree scale on a mega scope,with obvious theoretical depth,practical value and significance for forest ecological conservation.This paper is based on the pre-processed TripleSat satellite remote sensing data provided by the National Forestry and Grassland Administration.Firstly,the data were secondary processed and marked for the pine wilt disease occurring at the tree scale in the remote sensing images by ground verification quarantine and visual quarantine.Then,based on the results of the secondary processing of the data,the reasons for the discrepancies between the images captured by the satellites at different moments are analysed and solutions to this issue are explored.On this basis the processing results of the Wallis filtering algorithm and the histogram matching algorithm are compared.The spectral and spatial properties of diseased pine trees were in subsequent exploration by measuring the reflectance,shape and size of the trees.Afterwards,the relationship between the ability of deep learning models to process low-resolution images and the network architecture is explored.It also compares the recognition effectiveness of multiple deep learning models on satellite image slices of pine wilt disease.And latter,the upgraded HRnetV2 model and the improved HRnetV2 model were constructed by combining the advanced results of deep learning methods for processing low-resolution images and the characteristics of diseased pine trees.Accordingly,the corresponding pine wilt disease detection methods based on these two models are proposed.Finally,the detection method is applied to satellite images to verify its effectiveness in practice.Processing the data using the above method gives the following results.The wallis filtering algorithm can attenuate the variance between images better than the histogram matching method.In terms of mean values of reflectance,there are differences between the measured pine wilt diseased plants and other features.Infected pine trees have a regular oval crown with an average crown width of about 5.6 metres.The results of the experiments on the identification of pine wilt disease slices showed that the HRnetV2 was more effective than other models.Pine wilt disease detection experiments have shown that the improved HRnetV2 network-based method is more effective.The methodology is effective in identifying pine wilt diseases occurring at the tree scale on satellite images that meet the precision of quantitative processing.However,there were instances where some features with similar spectral composition in the forest area were misidentified as pine wilt disease-infected pine trees.The above results support the following conclusions.For one,the identification of pine nematodes at the tree scale requires data support from high-resolution remote sensing satellites.The radiometric and spatial information of the image is of great relevance to the recognition process.Secondly,relative radiation correction methods such as Wallis filtering and histogram matching,in addition to reference images,do not meet the requirements for quantitative processing of satellite images.Thirdly,the continuous downsampling process can seriously affect the ability of deep learning models to process low-resolution images.The less downsampling process of the designed model and the introduction of an attention mechanism from three dimensions of the image using the SE module and the spatial attention module were crucial to its ability to identify pine wilt diseases occurring at the tree scale.