Classification Of Desert Grassland Grassland Based On Convolutional Neural Network

Grassland degradation has become one of the major ecological problems in China.China's grasslands are mainly distributed in the central and eastern regions of Inner Mongolia.They are important animal husbandry production bases and an important ecological protection barrier in China.In recent years,due to changes in human activities and the natural environment,the grassland environment has deteriorated,desert grassland environment has become frequent,and grassland degradation has been severe.Grassland degradation is not only reflected in the reduction of pasture production capacity,but also in the changes in vegetation community structure and surface soil properties.This project will focus on the Wulanchabu grassland,which has the most serious desertification in Inner Mongolia.The specific experimental site is the desertified grassland in the Siziwangqi area.In order to accurately obtain the degradation information of the grassland in real time,the ground hyperspectral remote sensing technology is adopted.Hyperspectral remote sensing images contain a lot of information,and can be combined into a map,which can be classified by hyperspectral imagery.However,due to the large amount of hyperspectral image data,the classification of grassland grass species is difficult,and it is difficult to obtain good classification results using traditional classification methods.Therefore,the use of powerful machine learning methods to solve the classification of hyperspectral images is the main research direction of this topic.In order to solve the problem of hyperspectral image classification,this topic will solve the problem of this topic from two technical directions:deep learning and traditional machine learning.The specific work is as follows:Firstly,for the problem of large amount of hyperspectral remote sensing image data,large band redundancy and strong correlation between bands,the sub-space segmentation based hyperspectral image band selection method is adopted to select the band of hyperspectral image and reduce the data.the amount.The automatic subspace division method is used to obtain the correlation coefficient matrix according to the correlation coefficient between each band.According to the correlation coefficient size,all the bands are divided into several subspaces,and then calculated according to the adaptive band selection method to obtain the subspaces.The maximum exponential band,and finally select the corresponding threshold to complete the band selection.Second,after the data is preprocessed,it is entered into deep learning and machine learning.In the deep learning training process,the data is first standardized by standard deviation and then input into the convolutional neural network.The feature is extracted from the image by the three-dimensional convolution kernel of the convolution layer,and then the downsampling is performed by the pooling layer to obtain the maximum value of the feature map,and finally the result value is output via the fully connected layer.In the process of machine learning training,two kinds of data input methods are adopted,one is to directly input images,the other is to extract texture features of images by using gray level co-occurrence matrix,and to classify data by using support vector machine(SVM)method.Finally,after adjusting the parameters to obtain the optimal convolutional neural network and SVM model,the quality of the two models is compared.The classification results of the two models show that the convolutional neural network has a higher advantage in image classification.