Hyperspectral Image Reconstruction Based on Improved Residual Dense Network

Yong, 李勇 Li; Qiuyu, 金秋雨 Jin; Huaici, 赵怀慈 Zhao; Bo, 李博 Li

doi:10.3788/aos202141.0730001

Cited by 6 publications

(2 citation statements)

References 0 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Zhong et al [7] constructed a three-dimensional residual network (3D-ResNet) through residual connections to alleviate the impact of the accuracy drop caused by 3D convolution and network deepening. Li et al [8] proposed a classification method based on an improved residual dense network, which also achieved high classification accuracy. When the above-mentioned ResNet-based network model extracts the spatialspectral features, the spectral features extracted first will inevitably be affected by the spatial features extracted later, and the method based on the dual-channel spatial-spectral feature extraction can effectively alleviate this part of the impact.…”

Section: Introductionmentioning

confidence: 99%

Dual-channel mixed spectral-spatial features combined with second-order pooling for hyperspectral image classification

Wang

Zhang²

2022

Fifth International Conference on Mechatronics and Computer Technology Engineering (MCTE 2022)

View full text Add to dashboard Cite

In order to effectively improve the classification accuracy of hyperspectral images, we propose a classification method based on the combination of dual-channel mixed spectral-spatial features and second-order pooling. Firstly, we use PCA to reduce the dimensionality of the hyperspectral image; secondly, the fused spectral-spatial features are extracted by stacking three dual-channel mixed spectral-spatial residual modules; finally, the second-order pooling is used to fully mine the fused spectral-spatial features, and through the fully connected layer to realize classification. We conduct comparative experiments on Indian Pines and Pavia University datasets, and the overall accuracy of the proposed method respectively reaches 99.17% and 99.74%, which are higher than other classification methods. This verifies the effectiveness of the proposed method classification function.

show abstract

Section: Introductionmentioning

confidence: 99%

Dual-channel mixed spectral-spatial features combined with second-order pooling for hyperspectral image classification

Wang

Zhang²

2022

Fifth International Conference on Mechatronics and Computer Technology Engineering (MCTE 2022)

View full text Add to dashboard Cite

show abstract

“…Zhao [10] et al have proposed a multilevel regression network HRNet, using Pixel Shuffle as a method of layer interaction, using convolution operation to reconstruct the hyperspectral image. The adaptive weighted attention network AWAN was proposed by Li [11] et al, and the reconstruction accuracy of the hyperspectral image is improved a lot; Liu Pengfei [12] et al designed a residual structure as the basic module for generating adversarial networks, Learning high-frequency signals using guided models based on expansion convolution and multi-scale pyramids, Further improve the reconstruction accuracy; [13]…”

mentioning

confidence: 99%

Reconstructing hyperspectral images from RGB images fused with multi-scale features

Zhao¹,

Song²,

Zhao³

et al. 2023

Conference on Infrared, Millimeter, Terahertz Waves and Applications (IMT2022)

View full text Add to dashboard Cite

The Hyperspectral image (HSI) can provide spectral information on the surface of objects, and has been widely used in remote sensing imaging, medicine and other fields. Due to the high cost of hyperspectral imaging systems, reconstruction of hyperspectral images from RGB images has become a research hotspot in recent years, but the reconstruction accuracy of existing algorithms is relatively limited. Based on the Transformer network, this paper designs an MC-Transformer model that integrates multi-channel feature extraction, uses a multi-channel encoder based on pixel scrubbing to extract spectral and spatial dimension features of different scales, through designs a multi-scale fusion strategy to fuse features, and finally analyzes high-dimensional features information and reconstruction hyperspectral images by decoder. Experimental results show that the MRAE of the reconstruction results of this method can reach 0.1795, which is 39.1% and 23.6% higher than the HRNet and AWAN models, respectively.

show abstract

结合ncep再分析资料优化的6s模式大气校正方法

Wang Xuedan,

Wang Yufeng,

Liu Kai

et al. 2024

光学学报

View full text Add to dashboard Cite

Objective Unlike panchromatic and multispectral images, hyperspectral images have a high spectral resolution that causes more difficult atmospheric correction. Additionally, the atmospheric correction methods and correction accuracy based on the 6S model have been catching attention. China has launched several hyperspectral imaging sensor payloads, such as Environment 1 satellite, Gaofen 5, and resource satellites. Meanwhile, these satellites play an increasingly important position in agricultural remote sensing, forestry remote sensing, environmental monitoring and other fields, and the atmospheric correction technology for these hyperspectral remote sensing satellites is constantly developing. Taking China s hyperspectral remote sensing satellite Environment 1 satellite as an example, we carry out an optimized 6S atmospheric correction method combined with NCEP reanalysis data to improve 6S atmospheric correction accuracy.Methods First, considering the lack of standard reflectance products for hyperspectral images, the hyperspectral reflectance curve is constructed by the optimization estimation method and is regarded as the standard curve to verify the atmospheric correction results. Secondly, based on 6S atmospheric correction theory, we carry out sensitivity analysis and determine the most sensitive factors of aerosol optical thickness and the sensitivity of the aerosol model, atmospheric model, and atmospheric temperature and humidity to atmospheric correction coefficient. On this basis, an optimized 6S atmospheric correction method combined with NCEP reanalysis data is proposed. The aerosol optical thickness at 550 nm, atmospheric temperature and humidity profiles, and other data provided by NCEP are adopted to optimize the input parameters of the 6S model. Meanwhile, accurate atmospheric correction coefficients X a , X b , and X c can be obtained, and the reflectance spectral curves of different ground objects are thus obtained after optimized atmospheric correction. Finally, by choosing Xi an as the test area, the spectral curve of the water body is compared, and the accuracy of the correction results is evaluated via the standard curve. Results and DiscussionsThe reflectance results by the 6S model and NCEP are significantly better than those by the 6S model. Compared with the standard curve, they have the same trend in spectral reflectance, and the correlation coefficient between them can reach 0.8596 with a standard deviation lower than 0.0685 (Fig. 13). The average and standard deviations of pixelbypixel error of ground reflectance in each band are close to 0.02, which demonstrates that the optimized 6S model with NCEP data has obvious improvement on the atmospheric correction.Conclusions The absolute error of the reflectance curve and the standard curve obtained by the atmospheric correction of the 6S model optimized by the NCEP data is much lower than that of the 6S model, and the average absolute error of each band is also less than that of the 6S model. The correlation coefficients of the three c...

show abstract

Hyperspectral Image Reconstruction Based on Improved Residual Dense Network

Cited by 6 publications

References 0 publications

Dual-channel mixed spectral-spatial features combined with second-order pooling for hyperspectral image classification

Dual-channel mixed spectral-spatial features combined with second-order pooling for hyperspectral image classification

Reconstructing hyperspectral images from RGB images fused with multi-scale features

结合ncep再分析资料优化的6s模式大气校正方法

Contact Info

Product

Resources

About