Hyperspectral Image Classification Based On Nonoverlapping Sampling

Since the 20th century,modern science and technology have developed rapidly,and the field of remote sensing is also one of them.Hyperspectral remote sensing has been applied in many fields,such as geological prospecting,environmental monitoring,medical and health fields,etc.Hyperspectral remote sensing technology has many research directions,such as image classification,image segmentation,feature extraction,feature fusion,pattern recognition,etc.The main research direction of this paper is the field of image classification.Some basic classification methods only use spectral information and cannot achieve high classification accuracy.Later,the researchers introduced spatial information into the algorithm,that is,the spectral-spatial classification method,which greatly improved the classification accuracy.In the study of hyperspectral image classification,the sampling method of training samples is generally random sampling.The research of hyperspectral image classification is generally carried out on a hyperspectral image,so when random sampling is used in the spectral-spatial classification method,there will be information overlap between the training set and the test set,so it is impossible to Reflects the true performance of the classification method,because it is impossible to determine whether it is due to the spectral-spatial classification method or the increase in classification accuracy caused by sample overlap.In order to show the performance of the classifier more objectively,the samples can be sampled through a nonoverlapping sampling method,which can reduce the impact of sample overlap.In order to study the classification results of hyperspectral images in non-overlapping sampling mode,this paper proposes several classification algorithms based on this sampling mode.The main tasks are as follows:1)Propose a classification algorithm that introduces spatial information and combines the idea of Co-teaching.The algorithm is based on a convolutional neural network,and two neural networks with the same structure but different initialization parameters are constructed.The other settings of the two networks are exactly the same.In the iterative training method,extract the more reliable part of the test set of the two networks,add it to the training set of the other network,and then complete the iterative training of the two networks.As the training progresses,each network The number of training samples is increasing,and the information of the test samples is gradually introduced through spatial information,so that the generalization of the network is gradually improved.The reason for using two networks is that the network itself has a strong correlation with its own classification results.If the test samples classified by it are added to the training set again,the incorrectly classified samples will seriously affect the performance of the classifier.2)Propose a classification method combining neural network and image segmentation technology.Before performing the classification task,perform image segmentation on the hyperspectral image to obtain a certain number of regional blocks.The data is classified through the neural network,and the predicted label of the test sample will be obtained.For each area of the image segmentation,count the more reliable parts of the current test set and the number of categories in the initial training set,through a majority vote method Get the labels of all the test samples,and then add the test samples to the training set in the same proportion of each type,and perform tasks through the idea of self-training until the predetermined algebra is reached or the classification accuracy stabilizes.3)A self-training classification algorithm based on sample selection based on distance judgment is proposed.The test set is classified through the network,and then the Euclidean distance between each test sample and the feature center of the initial training set of all categories is measured.If a test sample is predicted by the neural network,the label,and its neighborhood and the current of a certain category If the smallest type of distance in the training set is consistent,then this test sample can be regarded as a sample with higher confidence.All the extracted samples are added to the training set in the same proportion of each type,and the training is carried out in the form of self-training.The number of training samples gradually increases,and at the same time,the feature information overlap between the training set and the test set is enhanced,which makes The performance of the classifier is gradually improving.