Research On Dimensionality Reduction And Classification Of Hyperspectral Remote Sensing Images Under Weakly Supervised Learning Framework

Hyperspectral images contain rich spatial information of ground objects and fine spectral bands,making it has obvious advantages in landscape identification.The classification of hyperspectral images is a key technology to achieve hyperspectral data understanding and earth observation.However,the high dimensionality and scarcity of labeled samples in practical applications often make it encounter the problem of "curse of dimensionality" in classification process,resulting in the degradation of classification accuracy.Therefore,in view of the high dimensionality and the scarcity of labeled samples in hyperspectral images,the classification accuracy of hyperspectral images can be improved by effectively reducing the hyperspectral dimension and making full use of the limited label information.It can meet the development frontier of data mining and intelligent processing of hyperspectral images and has strong research significance and application value.Weakly supervised learning is a paradigm for learning based on partially supervised information,which can greatly alleviate the requirement of large amounts of accurately labeled data in the training process of supervised models and thus has a broader prospect in practical applications.Meanwhile,with the development of hyperspectral remote sensing technology,the improvement of spatial resolution leads to the increase of spectral variation within the same ground object and the decrease of spectral variation between classes,i.e.,strong spectral variability,which increases the difficulty of classification tasks.Moreover,hyperspectral images contain spectral features and spatial distributions reflecting various land cover types and their combination relationships.The ground object distribution is essentially regionally similar and homogeneous,i.e.,high spatial correlation,and its intrinsic spectral and spatial discriminability guarantees the reliable classification results,thus extracting and utilizing its intrinsic separability is the prerequisite to achieve efficient hyperspectral remote sensing classification.Consequently,this thesis establishes a set of hyperspectral remote sensing image dimensionality reduction and classification models using advanced methods of machine learning under weakly supervised learning framework by combining the characteristics of hyperspectral images.The key works in this thesis can be briefly summarized as follows:(1)A spectral-spatial genetic algorithm-based unsupervised band selection method is proposed for hyperspectral image,which can select a subset of bands with low redundancy and high discriminative power from a large number of bands to alleviate the "curse of dimensionality" problem and improve the classification accuracy.In this method,the new crossover and mutation operations are designed to select a desired number of bands with low redundancy by restraining the search space.And an unsupervised fitness function that can comprehensively use spatial and spectral information is constructed by combining superpixel segmentation and Fisher's criterion to ensure that the selected subset of bands has high discriminability.(2)A superpixel-level global and local similarity-based clustering(i.e.unsupervised classification)method is proposed,which can quickly classify a large hyperspectral imagery without any label information and with higher accuracy.This method takes the superpixel as a sample,greatly reduces the time and space complexity in the process of large image classification.At the same time,the superpixel-level global and local similarity graph proposed in this method can better model the relationship among ground objects in the image,which guarantees the performance of the proposed algorithm.(3)A spectral-spatial superpixel label propagation-based hyperspectral image semi-supervised classification algorithm is proposed.This algorithm extends the previous clustering method to semi-supervised pattern,which greatly improves the classification accuracy.The superpixel-level spectralspatial graph designed in this method is able to model the structural information of the ground objects in the image,and compute spatial information at different scales.The method has the advantage of achieving high classification accuracy with only a very small number of labeled samples.(4)A multi-scale spectral-spatial convolution network-based semi-supervised model is proposed for hyperspectral image classification.This model further improves the above proposed spectral-spatial superpixel label propagation method,making the presented model both high-performance and highly scalable.In this model,a multi-scale spatial-spectral convolutional neural network with higher discriminative power is designed.Meanwhile,based on the full analysis of the differences between hyperspectral images and natural images,a new pseudo-label generation strategy is designed to ensure the reliability of semi-supervised learning,which further improves the classification performance.To verify the effectiveness of the approaches proposed in this thesis,typical hyperspectral remote sensing datasets with different spatial resolutions are applied to the above four methods.The experimental results indicate that these methods have more significant advantages in terms of classification accuracy and efficiency with extremely insufficient labeled samples.