Research On Deep Spectral-spatial Feature Extraction Algorithms For Classification Of Hyperspectral Image

Hyperspectral image(HSI)can simultaneously provide both detailed spectral information and spatial distribution information of land covers,and play a more and more important role in the field of remote sensing.However,hyperspectral image not only provides rich information,but also poses great challenges to conventional feature extraction algorithms duo to its characteristics of complex spatial variability of spectral signatures,high dimensionality,huge data and nonlinear separability.In recent years,deep learning has shown powerful expression ability and excellent generalization performance in feature extraction,providing new opportunities for feature extraction and classification of hyperspectral image.This dissertation proposes several deep spectral-spatial feature extraction frameworks,based on exploring the characteristics of hyperspectral image,to improve the classification performance.The specific research work is summarized as follows:1.A spectral-spatial feature extraction method based on deep similarity network is proposed.The spatial variability of spectral signatures in hyperspectral image is complicated,which leads to the phenomenon that the same material may have quite different spectral signatures,whereas different materials may share similar spectral information.Traditional handcrafted feature extraction methods extract only low-level features,such as edge and texture.Deep learning-based methods can extract more abstract deep features from hyperspectral data,but require complicated deep network,which could restrict its potential due to limited labeled samples in hyperspectral image.To overcome this barrier,we propose a novel deep similarity network(DSN),which can significantly increase the number of training samples.In addition,we introduce a maximum-margin-based loss to enhance the discrimination and alleviate the challenge caused by spatial variability of hyperspectral image.Moreover,in order to mitigate the workload of deep network in feature extraction and reduce the complexity of network,we combine the DSN and the extended multi-attribute profile to construct a feature extraction and classification framework for hyperspectral image.Experimental results and statistical tests demonstrate that the proposed method can achieve excellent performance on hyperspectral image classification.2.A deep structure-preserving spectral-spatial feature extraction algorithm is presented.Most existing deep feature extraction methods only exploit label information of land classes to supervise the training of deep network,and the feature extraction process is totally automatic with no any prior information embedded into this process.Manifold learning-based methods assumes that the hyperspectral data are lying on a low-dimensional manifold,and can obtain the intrinsic representation of hyperspectral image through constraining the extracted feature to keep the manifold structure of original data as much as possible.However,these methods only explored the manifold structure of the hyperspectral data with shallow models.Based on this,we introduce manifold learning into deep network framework,and propose a structure-preserving spectral-spatial network(SPSSN)in order to make the deep features maintain the manifold structure of original hyperspectral image and enhance the representative ability.The structure-preserving constraint can be converted into a structure-preserving loss,which is combined with cross-entropy loss as the final optimization objective function.To guarantee the feasibility,a simple and effective algorithm is devised to implement the proposed structure-preserving loss.Experimental results and statistical tests verify that the SPSSN shows significant superiority in terms of the classification accuracy over other state-of-the-art methods.3.A spectral-spatial feature extraction method based on weighted salience loss and attention network is built.Convolutional neural network shows obvious superiority to extract deep feature from the HSI cube with regular spatial size.However,not all pixels in an HSI cube are from same class of the center pixel,and these pixels,named interfering pixels,will interfere the feature extraction process.In order to suppress the influence of interfering pixels,we introduce the attention mechanism and propose an attention network to improve the discrimination.In addition,aiming at the unbalanced sample problem and hard samples,we improve the conventional cross-entropy loss,and propose a weighted salience loss with the intention to explore more information from the hard samples and the classes with limited labeled samples.The results of both experiments and statistical tests illustrate that the proposed method can enhance the representative ability of deep features and improve the HSI classification accuracy.