Research On High-accuracy Classification For Hyperspectral Remote Sensing Image With Few Labeled Samples

Hyperspectral remote sensing is an advanced technology that organically combines spectroscopy and imaging theory.Hyperspectral image(HSI)obtains high-dimensional image data within the range of visible,short-wave infrared and even mid-infrared and thermal infrared spectrum,which reaches nanometer scale,and contains wealthy geometric,radiation and spectral information.It has been a research hotspot in the remote sensing community since the 1980 s.Nowadays,HSI has been widely applied in earth observation tasks such as ecological monitoring,and deep-space exploration missions such as Moon and Mars mineral analysis.Particularly,China’s Earth observation Gaofen-5,Moon exploration project Chang’e-1,and Mars exploration Tianwen-1 are equipped with hyperspectral imagers.It proves that hyperspectral remote sensing is not only an academic frontier but also has application value,and HSI plays a crucial strategic role in promoting the scientific and technological development and enhancing the economic construction of our country.HSI classification is one of the crucial research topics of hyperspectral interpretation.It aims to distinguish each pixel and assign a specific label to it to realize the pixel-level semantic parsing of the scene.Due to a large number of bands,the feature dimension of the HSI is high,which easily leads to the curse of dimensionality.Additionally,affected by imaging mechanism and conditions,HSI presents the spectral uncertainty of "spectral mixing","same material with various spectrums",and "different materials with a similar spectrum",which may cause the misclassifications.Furthermore,HSI labeling is laborious,expensive,and time-consuming in practice.Thus,the number of available labeled samples is too small to effectively extract discriminative and representative features,and it is difficult to obtain accurate classification results through training the classifier.To address these issues and deal with the greater challenges of high-accuracy classification for HSIs with few labeled samples,the paper studies the classification models for HSIs with high feature dimension and spectral uncertainty in the condition of small sample size,which are fully evaluated through real-world datasets and demonstrate substantial improvements.The main contributions are summarized as follows:Firstly,the paper proposes a deep prototypical network with a hybrid residual attention module(DPN-HRA)to address the problem that few labeled samples are more likely to trigger the curse of dimensionality.In particular,the model devises a discriminant distancebased cross-entropy loss function.Compared to the cross-entropy loss function that is commonly used in the deep learning-based multi-class classifiers,it is more effective in reorganizing and discovering features that reflect the distinction of different categories.To be specific,the loss function promotes the divisibility of high-dimensional features mapped to low-dimensional embedding space,i.e.,intra-class compact and inter-class separated.In this context,the decision boundary is more apparent.Furthermore,taking the image-spectrum merged structure of the HSI into account,the model proposes a hybrid residual attention network comprising a residual band attention module and a residual spatial attention module,which enhances the crucial spectral and spatial information.The ablation experiments prove that the devised loss function and hybrid attention mechanism can optimize the classification performance.The superiority of the proposed algorithm is demonstrated by comparison with several methods under different numbers of training samples.For instance,the DPN-HRA achieves 90.53% average accuracy on the Houston University 2013 benchmark dataset.Secondly,the study proposes a multi-direction network with attentional spectral prior(MDNASP),which presents a multi-direction neighborhood sample construction scheme to augment the diversity of the small samples.Through this strategy,MDN-ASP can alleviate that few labeled samples make it more difficult to extract representative spectral-spatial features.Besides,the model designs successive spectral and spatial three dimensional residual squeeze-and-excitation networks for thoroughly investigating the spectral-spatial structure features of the central pixel.Meanwhile,the spectral-prior information is utilized to establish the attention,which emphasizes the nontrivial neighborhood samples and suppresses the unimportant ones,facilitating adaptive fusion of multi-direction depth features with physical interpretability.The experiments demonstrate that the framework can improve the classification results on the edge of the land covers with obvious spectral uncertainty.For example,it acquires 93.97% average accuracy on the Pavia University benchmark dataset.Thirdly,this work proposes a multiscale context-aware efficient network(MSCEN)to further address the problem that few labeled samples make it more difficult to extract representative spectral-spatial features.Particularly,the MSCEN clusters the pixel-level features of the HSI into adaptive neighborhood samples to exploit the spectral-spatial structural information by multi-level superpixel segmentation.It then generates various spectral-spatial features through the multiscale context-aware perception module.Afterward,these features and the original spectral curves are fed into the ensemble deep kernel extreme learning machine to complete feature extraction and fusion,achieving high-precision and efficient classification performance.For example,compared to other methods,the model achieves better classification results on the Yancheng dataset acquired by the Gaofen-5 satellite,i.e.,96.37% overall accuracy with 15 training samples from each class.Fourthly,the paper proposes an efficient semi-supervised symmetric graph metric learning framework(SGML),which significantly improves the classification performance by leveraging the information of abundant unlabeled samples.First of all,the model utilizes deep metric learning and involves a metric loss function in the network so that the class boundaries in the low-dimensional space can be adaptively adjusted to the typical representation(center of the category)of the samples,improving the discrimination of the embedding features.Moreover,the model employs the adaptive multi-level superpixel patterns to establish the topological graph.Then,by interactive operation between multi-layer graph convolution and the selfchannel enhanced convolution,the potential spectral-spatial structure is explored to obtain the representative superpixel graph node features.Notably,the volume of the superpixel pattern is compressed compared to the original pixel-level features,dramatically improving the efficiency of the graph learning process.Finally,the accurate pixel-level classification results are obtained by symmetrically reprojecting the superpixel features into pixel level and integrating the multiscale outputs.Experimental results show that the proposed algorithm achieves superior classification performance than existing methods in the case of few labeled samples.For instance,the SGML obtains 99.02% overall accuracy with 30 labeled samples from each class on KSC dataset.In conclusion,the research solves the bottleneck problem of low accuracy in HSI classification with few labeled samples.The achievements can provide optimization theory and techniques for HSI classification in mineral exploration,intelligent agriculture,urban planning,deep space exploration,and other significant practical applications.