Research On Automatic Extraction Of Tea Plantation Area Based On Hyperspectral Remote Sensing

With the rapid development of various hyperspectral remote sensing technologies,there are more and more ways to obtain a large amount of high-quality hyperspectral remote sensing image data.Hyperspectral imaging technology has been widely used in various earth science applications.Hyperspectral remote sensing not only has the characteristics of all-weather and strong immediacy of general remote sensing technology,but also can improve the identification ability of land cover by using its rich spectral information and the characteristics of image shape,texture and geometric structure.Tea garden is a kind of ground object category with regular shape and regular distribution.It can be used as the research object of target ground object recognition method based on hyperspectral remote sensing technology.Learning algorithms can perform deeper processing and analysis on hyperspectral remote sensing images.Therefore,it is feasible to use hyperspectral remote sensing technology to automatically extract the distribution of tea plantation areas.This research mainly studies the method of using deep learning model to automatically extract tea plantation areas mainly in Yongxing Town,Meitan County,Zunyi City,Guizhou Province based on the "Zhuhai No.1" hyperspectral satellite remote sensing image data,and explores the application of deep learning algorithms in the field of hyperspectral remote sensing.Advantages of target feature extraction.The main findings of this study are as follows:(1)Aiming at the problem that there is a large amount of redundant information between adjacent bands usually accompanies the research on high-dimensional data target recognition of hyperspectral remote sensing images,and the spectral features and texture features of target samples cannot be fully utilized for identification and extraction,a method based on The characteristics of the tea garden samples combined with the band selection method were used to select the feature band model method for extracting the tea garden planting area.Firstly,the 32-band data is preliminarily divided into four subspace sets by the automatic subspace division method,and then the spectral separability features of tea gardens in each subspace are selected from the perspectives of spectral separability and texture-specific features of tea garden samples.Representative bands(bands 1,4,17,28)and energy,homogeneity,and contrast texture feature representative bands(bands 3,16,and 28),and finally use the joint information entropy method to select the fourth most informative,17,28 band combination.The above method effectively and quickly improves the efficiency of selecting feature bands based on different feature information of ground object samples from high-dimensional data of hyperspectral remote sensing images.(2)Based on the knowledge system of transfer learning,according to the DeepLabV3+ model,it has the practical advantages of better performance in the task of semantic segmentation of various remote sensing images of fine features.The variable learning rate and the fixed learning rate(0.001,0.0001 and 0.00001)were trained by the control variable method,a total of four learning rates,and DeepLabV3+convolutional neural network models with 16,32 and 64 small batch samples,and the extraction accuracy and results of the tea garden distribution were evaluated.The results show that under the optimal model parameter settings of a fixed learning rate of 0.0001,the number of mini-batch samples is 16,and the rest of the training parameters are default,the recall rate,precision and IOU value of the DeepLabV3+model for the tea garden extraction in the study area reached 72.58%,67.64% and 0.5387.(3)The DeepLabV3+ model is compared with the traditional SVM and the minimum distance method for tea garden extraction.The results show that the model has the best extraction effect on the tea garden planting area.The difference between the maximum degree and IOU value is 12.71%,28.12%,26.68% and 0.2270;From the comparative analysis of the tea garden extraction results of the DeepLabV3+ model and the classic deep learning segmentation algorithms of Segnet and U-net,the results show that from the perspective of migrating the classic deep learning model,the DeepLabV3+ model has stronger advantages in tea garden extraction;Secondly,the DeepLabV3+ model is used to extract the tea garden planting areas in different sample areas.The accuracy evaluation shows that the overall classification accuracy of the model including the tea garden planting areas to different degrees has reached more than 85%,and the recall rate,precision and IOU value of tea garden extraction are the largest.The difference is 3.7%,2.08% and 0.0156,showing a good extraction effect in tea garden planting area as a whole;Finally,the tea garden extraction experiment is carried out on the same research area under different feature channels.The results show that the28 th channel training sample data,which is the representative band of spectral separability in the fourth subspace,and the representative band of contrast texture features,are used in the DeepLabV3+ model to evaluate the tea garden.The best performance in planting area extraction,the overall classification accuracy and tea garden extraction recall rate,precision and IOU value of the study area under this channel reached86.27%,76.73%,69.15% and 0.5716,respectively.The comprehensive evaluation results show that the DeepLabV3+model has a good performance in the automatic extraction of tea garden planting areas in hyperspectral remote sensing images,and has a wide range of comprehensive applications.