Research On Problems Of Classification And Performance Evaluation Of Hyperspectral Images With Small Samples

Hyperspectral images(HSIs)contain continuous hundreds of spectral bands.HSIs can provide abundant spectral and spatial information to classify and identify objects more accurately.The characteristic of HSIs with high spectral resolution that can obtain both spectral information and spatial information at the same time makes it play an extremely important role in the fields of military and national defense,precision medicine,and environmental monitoring.Therefore,developing advanced hyperspectral image processing and analysis technologies meets the major strategic needs of our country for national defense security,national economy,and ecological protection.However,in remote sensing and medical scenes,insufficient number of labeled samples restricts the performance of some traditional classification algorithms,and limit the engineering application of hyperspectral image classification techniques in practical problems.In addition,the annotation of samples in ground truth has an impact on the evaluation performance of existing accuracy indexes when objectively measuring the feasibility and effectiveness of classification algorithms.Based on the existing machine learning theory,combined with the characteristic of HSIs and the prior knowledge of actual application scenarios,the thesis proposes a variety of effective feature extraction and sample augmentation methods for the problem of insufficient number of labeled samples in the data set,which significantly improve the classification accuracy in the case of small samples.In order to verify the effectiveness of the proposed methods,we set up sufficient experiments to measure the performance of the proposed methods on real hyperspectral images of various scenes.In addition,the thesis analyzes the effect of ground truth on performance evaluation of HSI classification algorithms in the field of remote sensing for the first time.The specific research contents are as follows:(1)In view of the problem that the traditional single-scale spectral-spatial feature extraction methods are difficult to achieve accurate classification in the case of small samples,the thesis introduces the multi-scale strategy to construct the discriminant features of different scales,and proposes a segmented principal component analysis and Gaussian pyramid decomposition based multi-scale feature fusion method for classification.The proposed method mainly includes the following steps: segmented principal component analysis based data dimensionality reduction,image pyramid decomposition based multiscale spectral-spatial feature extraction,and multiscale feature fusion.This method not only takes into account the correlation between the spectral bands of HSIs in the process of dimensionality reduction,but also effectively extracts and fuses the multi-scale Gaussian pyramid features of the images.This method enhances the separability of pixels.Experimental results show that the proposed method can guarantee high accuracy and efficiency of classification when the number of known category samples used for training is sufficient and limited.(2)For the problem that when the number of training samples is small,the classifier is not sufficiently trained and may generate over-smoothed classification maps,the thesis realizes image smoothing and denoising by implicitly minimizing curvature regularization,which can maintain the edge and structure of the images.A curvature filter and superpixel segmentation based multiscale feature extraction method is proposed.The proposed method mainly includes the following steps: curvature filter based image pyramid decomposition,superpixel segmentation based local multiscale spatial feature extraction,local and global multiscale feature fusion.This method can describe the spectral-spatial information in images at multiple levels and scales.The proposed method effectively combines the local multiscale spatial information and global multiscale spatial information of the images,and improves the classification performance of multiple classifiers in the case of small samples.Experimental results show that when the number of known category samples used for training in the data set is insufficient,the classification performance of the proposed method is still competitive,and the edges of the generated classification maps are more consistent with the original hyperspectral images.(3)Aiming at the problem that the traditional single label sample annotation method requires a lot of manpower and resources and it is difficult to annotate a large number of single label samples manually,the thesis introduces a novel multilabel training sample annotation method.The proposed annotation method only needs to label a small number of single label samples accurately,and simultaneously annotates a large number of pixels in certain regions by giving them multiple labels.In addition,this thesis proposes a multilabel sample augmentation method based on recursive filtering and superpixel segmentation to achieve label transformation effectively,which mainly includes the following steps: recursive filtering based feature extraction,superpixel based image segmentation and spatial-spectral similarity based mislabeled sample removal.Experimental results show that the sample augmentation method can make full use of the multilabel training samples,effectively improve the classification accuracy of multiple classifiers and reduce the demand of classifiers for accurate single label samples.(4)In view of the problem that it is usually difficult to collect a large number of tumor samples in medical scenes due to cost and other factors,the thesis makes full use of the advantages of different deep learning networks,and proposes a human brain HSI glioblastoma recognition technology based on the fusion of multiple deep learning models.The main steps of the proposed method include: spectral phasor analysis and data over-sampling,one-dimensional deep neural network based spectral feature extraction and classification,two-dimensional convolutional neural network based spectral-spatial feature extraction and classification,fusion and optimization of classification results,full convolution network based human brain parenchyma region and background segmentation.Experimental results show that our method can extract more accurate and robust distinguishing features,achieve high precision human brain tissue classification,and perform well in the recognition of Glioblastoma.(5)Aiming at the problem that the commonly used accuracy indexes may lead to over optimistic performance evaluation when the number of labeled samples in the ground truth is insufficient,the thesis constructs an impact analysis framework to analyze the influence of ground truth on the evaluation ability of accuracy indexes.This thesis discusses whether common objective indexes can provide fair and reliable evaluation results when the number of labeled pixels in ground truth is insufficient.In order to measure the robustness of different classification methods to the ground truth with limited labeled pixels during objective evaluation,four correlation and error indexes are introduced to objectively analyze the correlation between the accuracy index scores calculated based on different ground truths.Experimental results show that the ground truth with insufficient number of labeled pixels will limit the evaluation ability of the existing objective accuracy indexes,and may lead to over optimistic algorithm performance evaluation.Therefore,it is necessary to design new and more suitable objective accuracy indexes to provide reliable objective evaluation results for the ground truth with insufficient labeled samples.