A Sample Pre-selection Method For Hyperspectral Image Classification

In recent years,researchers have paid much attention to hyperspectral image classification,among which the introduction of deep learning method has injected vitality into hyperspectral image classification.In this paper,we discuss the application of the convolutional neural network in hyperspectral image classification.The validation of hyperspectral image classification is based on randomly generated training sets.Because randomly generated training sets may have local nearest neighbor distribution in space,the recognition accuracy of different random training sets is different,which brings challenges to the objectivity of model comparison and the difficulty of reproduction.In this study,the unsupervised clustering algorithm is used to obtain the pseudo labels of each pixel,and then the regional information entropy of a pixel and its surrounding region is calculated by using the number of various pseudo labels,and then the representative pixels are selected as the training set in the arrangement of regional information entropy.A large number of experimental results show that this training set selection method has higher and more stable recognition accuracy than the randomly selected training set.In the process of sample selection under specific rules,the randomness is greatly reduced,and it is easy to reproduce the performance of recognition accuracy.Higher recognition accuracy stability makes the model comparison more objective.In addition,the improvement of recognition accuracy also has significant significance.In hyperspectral image classification,using the early stop strategy based on the loss of validation set is a common method,which often requires additional labeled samples.This paper proposes a method based on the loss of training set,that is,selecting the model with the minimum loss of training set as the final model,removing the validation set,while using fewer labeled samples In this case,the recognition accuracy is improved.In hyperspectral image classification,researchers often need to limit the number of training samples for each class of pixels,which is not conducive to the actual production practice.In view of this,this study proposes a sample pre-selection method based on the range of pseudo labels.Instead of using the information on range assignment of each type of real labels,it uses the information of each type of pseudo labels.This is a good way to practice.After obtaining a pair of hyperspectral images,we can get a pair of unlabeled hyperspectral images without using any prior information to give the pixels that need to be labeled.Through experiments,compared with using the real label information on range assignment,the overall recognition accuracy is only slightly reduced.Compared with the latest active learning method,under the same number of labeled samples,the proposed method has significant advantages in recognition accuracy.