Research Of Hyperspectral Image Classification Based On Feature Extraction

In order to reduce the redundant information in hyperspectral image and noise interference and make full use of three-dimensional characteristics of hyperspectral image to do a good job classification,the article will with Indian Pines and the KSC two public HSI on the experimental data of data sets,mainly around the hyperspectral image feature extraction and classification task for a systematic and in-depth research,after many experiments in this thesis,a preliminary to the following conclusion:1.In order to reduce the information redundancy of hyperspectral images,a hyperspectral feature extraction method based on feature importance was proposed.Firstly,the Random Forest(RF)model obtained from the training of Bayesian optimization is used to evaluate the importance of numerous bands of hyperspectral remote sensing images,and then an appropriate number of bands of hyperspectral images are selected according to the evaluation results,so that the model can output new training samples.In order to verify the effectiveness of the evaluation model,the classical machine learning model Support Vector Machine(SVM)was introduced as a classifier,and its classification results were used as the evaluation criteria of the feature importance evaluation model proposed in this thesis.Experimental results: For Indian Pines: The proposed feature selection extraction method improves the overall accuracy by 6.48 percentage points,and 1.09 percentage points more than other traditional methods,such as Principal Components Analysis(PCA).The Kappa coefficient,however,increases by at least 2.77 more than the other methods.For KSC dataset: the proposed feature extraction method improves the overall accuracy of classification results by 9.04 percentage points,at least 1.23 percentage points more than PCA and other methods.Regarding the Kappa coefficient,Fi still performed well,increasing it by9.04.At least 1.23 more than K-means and other methods.The above data fully illustrate the effectiveness of the FI model.2.To enhance the efficiency and accuracy of hyperspectral image classification,this thesis introduces compared with ordinary convolution neural network(convolutional neural networks,CNN)model is more suitable for hyperspectral image related tasks of 3 d convolution neural network classifier,combining it with in this thesis,based on the characteristics of the importance of combining feature extraction model to build a deep learning classification framework based on feature importance.The experimental results are summarized as follows: for each data set,the FI-3D CNN combination model proposed in this thesis has completed the best classification,with the highest accuracy reaching 98.03% and 97.26%,respectively.Kappa,on the other hand,achieved 97.84 and 96.98.Compared with the basic control group,the classification accuracy of CNN improved by 12.76.The Kappa coefficient is increased by 14.53.The experimental results show that the combined model can well adapt to the task of hyperspectral image classification,and the classification accuracy and efficiency are better than most traditional algorithms.