Forest Recognition And Classification Based On Hyperspectral Remote Sensing Image

The lucid waters and lush mountains are invaluable assets is proposed by President Xi.The forest sources are important part of the lucid waters and lush mountains.They are related to the sustainable development of human living environment.Therefore,it is necessary to use advanced methods to protect and manage forest resources.The hyperspectral remote sensing image record the spatial and spectral information of the ground object,which is an important remote sensing data,and they provide a new and effective way for the protection and management of forest resources.The forest recognition and classification is one of the most basic problems in hyperspectral remote sensing image data analysis.Dimensionality reduction,feature extraction,recognition and classification for hyperspectral remote sensing image using artificial intelligence and machine learning methods are the research hotspot in the field of remote sensing and some research results have been obtained.However,the realization of systematic,accurate and reliable forest recognition and classification is still an urgent problem to be solved at present,which is affected by high data dimension,large data redundancy,insufficient spatial resolution and limited labled samples.In this paper,three forestry related hyperspectral remote sensing image datasets are selected:Indian Pines by AVIRIS(Airborne Visible Infrared Imaging Spectrometer),Botswana by NASA(National Aeronautics and Space Administration)’s EO-1 and Yichun Liangshui forest farm,Heilongjiang province.Aiming at the problems existing in forest recognition and classification of hyperspectral remote sensing image,the following research work have been carried out:(1)In practical applications,when the feature dimension increased to a certain critical point,the performance of classifier will become worse if the feature dimension continues to increase,which is called "Hughes phenomenon".In order to solve the problem,we proposed a method for dimension reduction method for hyperspectral remote sensing image base on band combination(2D)2PCA,which achieved the purpose of reduce data redundancy and eliminate the "Hughes phenomenon",it lays a foundation for the subsequent feature extraction and classification work.(2)At present,most of the forest recognition and classification for hyperspectral remote sensing image are based on spectral features,and the relation between different pixels in hyperspectral remote sensing image is ignored,so the classification result is not well.In order to solve the problem,the HSI spatial spectral feature extraction model based on two-channel convolutional neural network is proposed.The model can effectively improve the efficiency of forest identification and classification by combining the spectral features and spatial characteristics contained in hyper-spectral images.(3)To solve the problem of weak generalization ability of classifier.The HSI forest recognition and classification model based on two-channel CNN-SVM fusion is proposed.The model makes full use of the powerful image feature extraction capability of convolutional neural network,the generalization ability of SVM is maximized,and the classification accuracy of the model is improved at the same time.The model is applied to the classification of hyperspectral remote sensing image.(4)In practical applications,only one part of hyperspectral remote sensing image analysis is researched,and lack of complete classification model.In order to solve the problem,the HSI forest recognition and classification model based on BC2S-Net’s spatial spectral feature extraction is proposed.The model first reduced the dimension of original hyperspectral remote sensing image on the premise of ensuring the accuracy of classification.Then the spatial spectral features of hyperspectral remote sensing image are extracted by using the two-channel convolutional neural network.Finally,the spatial spectral features are input into the classifier for classification,and gets the category for the corresponding pixels.(5)Aiming at data labeling is a problem of large workload,which is affected by human subjective factors.A unsupervised forest recognition and classification model based on fine-tuning is proposed,the model achieved the automatic labeling of unlabeled samples by clustering unlabeled samples and selecting reliable feature samples.By repeated iteration,the reliability of the samples from the center of the cluster is enhanced,the parameters of BC2S-Net are continuously optimized,and the performance of the whole model is getting better and better.In conclusion,the objective of this dissertation is the forest recognition and classification of hyperspectral remote sensing image,the hyperspectral remote sensing image is analyzed.The transform dimension reduction algorithm and non-transform dimension reduction algorithm are combined,the advantages of information extraction and feature expression of convolutional neural network and the advantages of SVM algorithm in classification are comprehensively utilized,a complete forest recognition and classification model of hyperspectral remote sensing image is constructed and extended to unsupervised application field.Some achievements have been made in dimensionality reduction,feature extraction,construction of classifier and supervised or unsupervised classification.The method proposed in this dissertation is applied to forest coverage survey,forest resource protection survey and forest classification,and the propsed methods have certain theoretical significance and the application value.