Facing the very high-resolution( VHR) image classification problem,a feature extraction and fusion framework is presented for VHR panchromatic and multispectral image classification based on deep learning techniques. ...Facing the very high-resolution( VHR) image classification problem,a feature extraction and fusion framework is presented for VHR panchromatic and multispectral image classification based on deep learning techniques. The proposed approach combines spectral and spatial information based on the fusion of features extracted from panchromatic( PAN) and multispectral( MS) images using sparse autoencoder and its deep version. There are three steps in the proposed method,the first one is to extract spatial information of PAN image,and the second one is to describe spectral information of MS image. Finally,in the third step,the features obtained from PAN and MS images are concatenated directly as a simple fusion feature. The classification is performed using the support vector machine( SVM) and the experiments carried out on two datasets with very high spatial resolution. MS and PAN images from WorldView-2 satellite indicate that the classifier provides an efficient solution and demonstrate that the fusion of the features extracted by deep learning techniques from PAN and MS images performs better than that when these techniques are used separately. In addition,this framework shows that deep learning models can extract and fuse spatial and spectral information greatly,and have huge potential to achieve higher accuracy for classification of multispectral and panchromatic images.展开更多
A novel fusion method of multispectral image and panchromatic image based on nonsubsampled contourlet transform（NSCT） and non-negative matrix factorization（NMF） is presented,the aim of which is to preserve both sp...A novel fusion method of multispectral image and panchromatic image based on nonsubsampled contourlet transform（NSCT） and non-negative matrix factorization（NMF） is presented,the aim of which is to preserve both spectral and spatial information simultaneously in fused image.NMF is a matrix factorization method,which can extract the local feature by choosing suitable dimension of the feature subspace.Firstly the multispectral image was represented in intensity hue saturation（IHS） system.Then the I component and panchromatic image were decomposed by NSCT.Next we used NMF to learn the feature of both multispectral and panchromatic images＇ low-frequency subbands,and the selection principle of the other coefficients was absolute maximum criterion.Finally the new coefficients were reconstructed to get the fused image.Experiments are carried out and the results are compared with some other methods,which show that the new method performs better in improving the spatial resolution and preserving the feature information than the other existing relative methods.展开更多
为发挥遥感图像在国防军事、公共安全、环境监测等领域的重要作用,如何融合已配准的高分辨率全色图像与低分辨率多光谱图像的互补信息成为当前研究的重点。尽管近年来全色锐化方法已取得较大进步,但大多数方法仍受到以下限制:一方面,利...为发挥遥感图像在国防军事、公共安全、环境监测等领域的重要作用,如何融合已配准的高分辨率全色图像与低分辨率多光谱图像的互补信息成为当前研究的重点。尽管近年来全色锐化方法已取得较大进步,但大多数方法仍受到以下限制:一方面,利用Wald协议退化生成不同尺寸图像时会造成信息损失;另一方面,受到网络结构和单一注意力的限制,无法同时利用全局和局部特征。为解决以上问题,本文提出了基于联合注意力的渐进式网络(Pan-sharpening based on multi-attention progressive network),称为MAPNet。在该网络中,首先采用多阶段训练以减小尺寸变化带来的光谱和细节损失。其次设计联合注意力模块,将自注意力、空间注意力和通道注意力结合,实现对全局特征和局部特征、空间特征和通道特征的多模态分析,进一步提高MAPNet对纹理细节的保留能力。在高分二号卫星上进行大量对比实验和消融实验,定性和定量结果表明,本文方法融合效果优于其他10种方法,能够改善光谱失真和细节纹理丢失等问题。展开更多
Pan-sharpening is a process of obtaining a high spatial and spectral multispectral image(HMS)by combining a low-resolution multispectral image(LMS)with a high-resolution panchromatic image(PAN).In this paper,a pan-sha...Pan-sharpening is a process of obtaining a high spatial and spectral multispectral image(HMS)by combining a low-resolution multispectral image(LMS)with a high-resolution panchromatic image(PAN).In this paper,a pan-sharpening method called PAIHS is proposed,which is based on adaptive intensity-hue-saturation(AIHS)transformation,variational pan-sharpening framework and the two fidelity hypotheses.The suitable objective function is established and optimized by adopting particle swarm optimization(PSO)to obtain the optimal control parameters and minimum value.This value corresponds to the best pan-sharpening quality.The experimental results show that the proposed method has high efficiency and reliability,and the obtained performance index is superior to the four mainstream pan-sharpening methods.展开更多
基金Supported by the National Natural Science Foundation of China(No.61472103,61772158,U.1711265)
文摘Facing the very high-resolution( VHR) image classification problem,a feature extraction and fusion framework is presented for VHR panchromatic and multispectral image classification based on deep learning techniques. The proposed approach combines spectral and spatial information based on the fusion of features extracted from panchromatic( PAN) and multispectral( MS) images using sparse autoencoder and its deep version. There are three steps in the proposed method,the first one is to extract spatial information of PAN image,and the second one is to describe spectral information of MS image. Finally,in the third step,the features obtained from PAN and MS images are concatenated directly as a simple fusion feature. The classification is performed using the support vector machine( SVM) and the experiments carried out on two datasets with very high spatial resolution. MS and PAN images from WorldView-2 satellite indicate that the classifier provides an efficient solution and demonstrate that the fusion of the features extracted by deep learning techniques from PAN and MS images performs better than that when these techniques are used separately. In addition,this framework shows that deep learning models can extract and fuse spatial and spectral information greatly,and have huge potential to achieve higher accuracy for classification of multispectral and panchromatic images.
基金Supported by the National Natural Science Foundation of China(60872065)
文摘A novel fusion method of multispectral image and panchromatic image based on nonsubsampled contourlet transform（NSCT） and non-negative matrix factorization（NMF） is presented,the aim of which is to preserve both spectral and spatial information simultaneously in fused image.NMF is a matrix factorization method,which can extract the local feature by choosing suitable dimension of the feature subspace.Firstly the multispectral image was represented in intensity hue saturation（IHS） system.Then the I component and panchromatic image were decomposed by NSCT.Next we used NMF to learn the feature of both multispectral and panchromatic images＇ low-frequency subbands,and the selection principle of the other coefficients was absolute maximum criterion.Finally the new coefficients were reconstructed to get the fused image.Experiments are carried out and the results are compared with some other methods,which show that the new method performs better in improving the spatial resolution and preserving the feature information than the other existing relative methods.
文摘为发挥遥感图像在国防军事、公共安全、环境监测等领域的重要作用,如何融合已配准的高分辨率全色图像与低分辨率多光谱图像的互补信息成为当前研究的重点。尽管近年来全色锐化方法已取得较大进步,但大多数方法仍受到以下限制:一方面,利用Wald协议退化生成不同尺寸图像时会造成信息损失;另一方面,受到网络结构和单一注意力的限制,无法同时利用全局和局部特征。为解决以上问题,本文提出了基于联合注意力的渐进式网络(Pan-sharpening based on multi-attention progressive network),称为MAPNet。在该网络中,首先采用多阶段训练以减小尺寸变化带来的光谱和细节损失。其次设计联合注意力模块,将自注意力、空间注意力和通道注意力结合,实现对全局特征和局部特征、空间特征和通道特征的多模态分析,进一步提高MAPNet对纹理细节的保留能力。在高分二号卫星上进行大量对比实验和消融实验,定性和定量结果表明,本文方法融合效果优于其他10种方法,能够改善光谱失真和细节纹理丢失等问题。
基金National Natural Science Foundation of China(No.61703278)。
文摘Pan-sharpening is a process of obtaining a high spatial and spectral multispectral image(HMS)by combining a low-resolution multispectral image(LMS)with a high-resolution panchromatic image(PAN).In this paper,a pan-sharpening method called PAIHS is proposed,which is based on adaptive intensity-hue-saturation(AIHS)transformation,variational pan-sharpening framework and the two fidelity hypotheses.The suitable objective function is established and optimized by adopting particle swarm optimization(PSO)to obtain the optimal control parameters and minimum value.This value corresponds to the best pan-sharpening quality.The experimental results show that the proposed method has high efficiency and reliability,and the obtained performance index is superior to the four mainstream pan-sharpening methods.
