Research On Deep Learning And Adversarial Defense Methods For Hyperspectral Remote Sensing Image Classification

Hyperspectral image(HSI)classification is an important research task for many hyperspectral remote sensing applications.As the imaging spectrometry technology gradually matures,hyperspectral remote sensing is developing rapidly towards the direction of high spectral resolution,high spatial resolution,and high temporal resolution.Meanwhile,with the successful launch of more and more hyperspectral remote sensing satellites,the data volume of HSIs further presents an explosive growth trend.These challenges make the existing classification methods gradually unable to meet the needs of rapid,accurate,and automatic image interpretation for HSIs.In recent years,artificial intelligence(AI)technology has achieved vigorous development in the computer vision field and has become the frontier of science and technology.As an important research branch of AI technology,deep learning aims to simulate a hierarchical thinking process of the human brain and provides the machine with the ability of self-directed learning.It enables the machine to extract hierarchical feature representation from the massive data to carry out accurate recognition on texts,images,sounds,and other data,just like the human brain does.Therefore,developing deep learning algorithms may open up a new way for automatic and intelligent remote sensing image processing,considering its strong feature extraction and data fitting ability.Based on the above research background,this thesis systematically studies deep learning-based HSI classification from four aspects: supervised learning,semi-supervised learning,unsupervised learning,and adversarial defense.The main contents and contributions of this thesis include:(1)This thesis systematically summarizes the existing research theories and methods for the HSI classification task and makes analyses and discussions about the shortcomings of the existing deep learning-based HSI classification methods in detail.(2)The existing deep learning-based approaches for HSI classification often separate the spectral feature extraction,spatial feature extraction,and classifier training.Under this circumstance,the feature learning and classifier training cannot share the same objective function.To this end,a novel spectral-spatial unified network(SSUN)is proposed as a supervised learning framework,where the spectral feature extraction,spatial feature extraction,and classifier training are integrated into an end-to-end learning framework.In the implementation of the framework,a band grouping-based long short-term memory(LSTM)model is proposed for spectral feature extraction,considering the spectral continuity between adjacent bands of hyperspectral data.In the spatial branch of the network,a multi-scale convolutional neural network(MSCNN)is further proposed to learn spatial feature representation at different scales.(3)A robust self-ensembling network(RSEN)is proposed for semi-supervised feature learning of HSIs to address the insufficiency of labeled samples.The proposed RSEN consists of two subnetworks including a base network and an ensemble network.With the constraint of both the supervised loss from the labeled data and the unsupervised loss from the unlabeled data,the base network and the ensemble network can learn from each other,achieving the self-ensembling mechanism.To further improve the stability of semi-supervised learning,a consistency filter is proposed to filter unlabeled samples with high confidence.(4)Most existing deep learning-based methods suffer from the problem of high complexity and long training time.To achieve a simple yet effective deep learning-based framework,a novel random patches network(RPNet)is proposed for unsupervised learning.The proposed RPNet extracts random patches from the original HSI and regards them as prefixed convolution kernels without any training process.By combining both shallow and deep convolutional features,RPNet has the advantage of multi-scale,which possesses a better adaption for HSI classification,where different objects tend to have different scales.The experiments show that RPNet can achieve high classification accuracy with a relatively low time cost.(5)Existing deep learning-based models are usually vulnerable towards adversarial examples,which may also bring about a threat for the safety-critical HSI classification task.To address this issue,a novel self-attention context network(SACNet)is further proposed.With the help of self-attention learning and context encoding,the spatial dependence among pixels in HSIs is constructed and the global context features are extracted,which can significantly improve the robustness of deep neural networks when confronted with adversarial attacks.In the adversarial hyperspectral dataset,the proposed SACNet can still maintain superior classification accuracy,which better meets the security and reliability requirements of remote sensing tasks compared to existing approaches.The characteristics of hyperspectral adversarial examples and their influence on deep neural networks are also analyzed in detail.