Recognition Method Of Vegetation Pest Area Based On Airborne Multispectral Images

Inner Mongolia is the largest and most functional ecological protection area in northern China.In recent years,due to global climate change,the ecological environment in Inner Mongolia has become more and more fragile,and the phenomenon of vegetation being attacked by diseases and pests has become more and more serious.At the same time,the decline of vegetation growth environment quality provides convenient conditions for the spread of diseases and insect pests,and speeds up the invasion of diseases and insect pests.How to prevent the invasion of pests and diseases and ensure the healthy growth of vegetation has been an important topic studied by scholars at home and abroad.With the development of spectral imaging technology and the rise of unmanned aerial vehicle(UAV)low-altitude remote sensing platform,it is possible to use airborne multispectral remote sensing to complete the identification of vegetation diseases and insect pests.In this thesis,through investigation and field exploration,the multispectral data of many areas in the city and its surroundings were collected,mainly trees and farmland,and these data were taken as the research object to carry out the research on regional identification of vegetation diseases and insect pests.The specific work is as follows:(1)According to the characteristic of multispectral image data,the collected data are fully preprocessed,including data correction,image registration and orthophoto mosaic,etc.,to form a global remote sensing image that can visually display the collected area.In view of the small number of bands of multispectral images,a multi vegetation index model is proposed to expand the data set,and the supervised classification methods such as support vector machine(SVM),extreme learning machine(ELM)and multi-layer perceptron(MLP)are used to carry out control experiments.Experimental results show that the effectiveness of the proposed model.(2)In order to make full use of the spatial and spectral information of remote sensing images and further improve the classification accuracy,a classification model based on spacespectrum joint convolutional neural network and parallel heterogeneous extreme learning machine(SSCNN-PELM)is proposed.SSCNN is composed of two channels,2D-CNN and1D-CNN.2D-CNN extracts spatial information and partial spectral information,while 1D-CNN extracts all spectral information.Parallel heterogeneous extreme learning machine(PELM)completes feature fusion and classification.The experimental results show that the disease category recall rate of SSCNN-PELM model is 98.41% on the collected multispectral data set,and the overall classification accuracy on the Indian pines and Pavia university data sets are more than 99%.Compared with a variety of traditional methods,the SSCNN-PELM model improves the speed and accuracy of classification.(3)In order to extract the features of multispectral images more effectively and further improve the classification accuracy of multispectral images,a deep residual module combined with dual attention mechanism is proposed.Res Net network was improved by replacing the 2D convolution kernel of residual structure with the decomposed 3D convolution kernel,so as to extract the space-spectrum joint features and reduce the computation.Channel attention module and spatial attention module are combined in series to form a dual attention mechanism,and combined with the residual module,D3 DRes Net-CBAM network is proposed.Experiments show that the overall accuracy of the optimized network is 99.76% under the condition of 10%training samples,and the recall rate of diseased vegetation is 98.90%.Through ablation experiments,the effectiveness of each component of the model is verified.Compared with Res Net and SSCNN-PELM networks,the recognition accuracy was further improved,especially in the case of small samples.