Research On Remote Sensing Image Registration And Classification Based On Deep Representation Learning

The increasing development of remote sensing technology is infiltrating all areas and corners of our lives.Among them,when using remote sensing technology for efficient ground feature analysis,the registration of remote sensing images and the classification of hyperspectral images have been the research topics that have been the focus of attention.On the one hand,we can obtain more and more complex remote sensing images from various advanced remote sensing platforms,and these remote sensing images have complex structures,which usually have the characteristics of information distortion,noise impact,large amount of surface information,and few labeled samples,traditional image processing methods have been difficult to meet the effective analysis of remote sensing images;On the other hand,deep learning has developed rapidly and has achieved great success in the field of computer vision.Compared with manual features,it has powerful feature representation capabilities.Therefore,in this thesis,we consider the characteristics of remote sensing images and the success of deep learning,an effective remote sensing image registration method and a hyperspectral remote sensing image classification method are designed respectively.The details are as follows:(1)Robust feature point matching is a key process in remote sensing image registration based on point features.Aiming at the characteristics of remote sensing images,this thesis proposes a point feature-based registration method.During the registration process of this method,the focus is on the removal of outliers and the increase of inliers.This method includes using a two-branch network to judge the similarity of the local features(image patches)of the hypothetical matching point pair to remove the outliers,and introducing local range constraints strategy to increase the inliers.A self-learning method is introduced when selecting training samples to solve the problem of small and single training samples.Experimental results show the effectiveness of the method and can obtain more correct matching correspondences.(2)Aiming at the problem of high-dimensional spectral data of hyperspectral images and many unlabeled samples,this thesis proposes a novel semi-supervised hyperspectral image classification framework.The framework includes two processes,the feature representation process and the self-training process.In the feature representation process,a local feature(image patch)representation of the sample is performed using a convolutionalnetwork.During the self-training process,a small number of labeled samples were used for semantic information constraints to perform over-clustering,and unlabeled samples were gradually assigned high-confidence pseudo-labels,and a spatial constraint strategy was introduced to use spatial consistency within the image to correct false-allocation error label.In the process of self-training,high-confidence samples are gradually increased and added to the corresponding semantic classes,so that the semantic constraints are gradually enhanced.At the same time,the increase of high-confidence markers also contributes to the regional consistency within the hyperspectral image,which highlights the role of spatial constraints and improves the efficiency of hyperspectral image classification.Experimental verification proves the effectiveness,robustness and high accuracy of our method.