18502 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 17, 2024 A Recurrent Feedback Hyperspectral Image Super-Resolution Reconstruction Method by Using Self-Attention-Based Pixel Awareness Ruyi Feng , Member, IEEE, Zhongyu Guo , and Xiaofeng Wang Abstract—Hyperspectral images (HSIs) contain abundant spectral information, while the spatial resolution is usually limited. To obtain high-spatial-resolution HSIs, various HSI super-resolution (SR) methods are proposed. Currently, deep-learning-based SR reconstruction methods are studied in depth, which take different measures to make full use of the spatial and spectral information of HSIs, and optimize the network with lots of trainings. Although they have achieved satisfying spatial resolution, the spectral consistency before and after reconstruction is difficult to guarantee. In this article, we proposed a self-attention-based recurrent feedback network for hyperspectral SR reconstruction, utilizing pixel-aware weights and pseudo three-dimensional convolution to enhance the spatial and spectral consistency during the reconstruction process. In addition, group reconstruction is used to reduce the redundancy of information. Spectral consistency regularization is proposed to ensure the spectral consistency before and after reconstruction. The effectiveness of the proposed method is tested on one set of natural images and three hyperspectral remote sensing image datasets. Index Terms—Feedback embedding, hyperspectral image (HSI), pixel awareness, recurrent network, super-resolution (SR). I. INTRODUCTION YPERSPECTRAL image (HSI) usually contains hundreds or thousands of bands, and each band corresponds to a specific wavelength range [1]. Compared with multispectral (MSI) and RGB images, HSI has higher spectral dimensions. Therefore, HSI is widely used in Earth observation [2], medical imaging [3], food quality control [4], military intelligence [5], and other fields. However, due to the technical limitations of imaging hardware, there is a tradeoff between spectral resolution and spatial resolution. To guarantee richer and more accurate spectral information, HSI often results in poor spatial resolution, which greatly affects its practical application. Hyperspectral super-resolution (SR) reconstruction, a key technique for overcoming the spatial-resolution limitations inherent in HSI, aims to enhance low spatial resolution HSI into H Received 11 June 2024; revised 25 August 2024; accepted 24 September 2024. Date of publication 9 October 2024; date of current version 21 October 2024. This work was supported in part by the National Natural Science Foundation of China under Grant 41925007 and Grant 42471430, in part by the Knowledge Innovation Program of Wuhan–Shuguang under Grant 2023010201020336, in part by Hubei Natural Science Foundation under Grant 2024AFB561, and in part by the Funds for International Cooperation and Exchange of the National Natural Science Foundation of China under Grant 42311530065. (Corresponding author: Xiaofeng Wang.) The authors are with the China University of Geosciences, Wuhan 430074, China (e-mail: fengry@cug.edu.cn; guozhongyu@cug.edu.cn; wxf@cug.edu. cn). Digital Object Identifier 10.1109/JSTARS.2024.3471899 high spatial resolution HSI while preserving the spatial and spectral information. This approach recovers high-frequency spatial details and maintains spectral fidelity, thereby significantly improving the utility of hyperspectral data across a wide range of applications. Recently, hyperspectral SR algorithms have been researched by more and more scholars, and many models are studied in depth. Among them, most of the models can be classified in two categories according to whether using the auxiliary image, single-image-based SR, and multiimage-based SR. Multi-image-based SR usually adopts RGB, multispectral, panchromatic images as the auxiliary information [6], [7], [8], [9] in the training process and afford sufficient reference information during reconstruction process. Although the multi-image-based SR shows excellent performance, how to acquire the appropriate auxiliary images is still a big challenge. While single-image-based does not require the assistance of auxiliary information. Additionally, single-image-based models typically show lower computational complexity and lower system requirements, making them suitable for scenarios with limited image resources. Therefore, single-image-based SR is more valuable and more practical. In this article, we will focus on recovering high-resolution HSI from a single low-resolution (LR) HSI without auxiliary images. In early study, various interpolation methods are proposed, i.e., adjacent interpolation [10], biquadratic interpolation [11], bicubic interpolation [12], etc. But these methods frequently yield distortions and artifacts attributable to loss detail, particularly at the edges and complex regions of the images. Since SR reconstruction is a typical ill-posed problem, more and more studies utilize prior knowledge as regularization terms and treat SR as a constrained optimization problem, such as sparse representation based SR approaches [13], [14], [15] and probabilistic analysis based SR approaches. Simultaneously, nonlocal similarity [16], low rank constraint [17], [18], [19], or other spatial or spectral correlations are usually adopted as prior knowledge. Nevertheless, due to the limitation of these methods, the quality of reconstructed images is not satisfying. With the development of deep learning, many scholars tried to simulate the inverse process of image sampling degradation by constructing different deep-learning models, and convolutional neural networks (CNNs) have been introduced and demonstrated remarkable effectiveness in hyperspectral SR reconstruction. However, CNN models still exhibit limitations in singleimage SR. HSIs possess abundant of spectral and spatial © 2024 The Authors. This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/ FENG et al.: RECURRENT FEEDBACK HYPERSPECTRAL IMAGE SUPER-RESOLUTION RECONSTRUCTION METHOD information. While current CNN methods [20], [21], [22], [23], [24], [25], [26] usually employ convolutional kernels with fixed size to extract features, which constrains the model’s receptive field, limiting its attention to local areas and making it difficult to capture distant spatial correlations. Meanwhile, CNN models also face challenges in extracting spectral correlations, resulting in poor consistency in spectral structure before and after reconstruction. To reduce the spectral distortion before and after reconstruction, enhance the spatial correlation for local and global image space, and further improve the quality of hyperspectral SR reconstruction, an improved single image SR reconstruction method based on deep learning is proposed, named pixel-aware recursive feedback network (PARF), which operates by grouping the HSI along the spectral dimension and utilizes a recurrent recursive network to reconstruct each group. PARF enhances spatial information using self-attention-based pixel awareness and achieving spectral consistency regularization through enhanced pseudothree-dimensional (3-D) convolutions. Compared to traditional methods, proposed method approach exhibits superior reconstruction performance. The main innovations are summarized as follows. 1) The pixel-aware weight based on self-attention is introduced, which can be acquired through the intermediate results to help reconstruct the missing spatial information. Meanwhile, the self-attention mechanism is used to expand the receptive field and enhance the local and global spatial correlations. 2) The optimized pseudo-3-D convolution operation is also used to ensure spatial and spectral consistency. The rest of this article is organized as follows. Section II provides the related works of single-image SR. In Section III, the proposed model architecture is introduced in detail. Then, extensive experiments and analyses are presented in Section IV. Finally, Section V concludes this article. II. RELATED WORKS A. Single-Image Super-Resolution Reconstruction Traditional single-image SR methods need to design a mapping function between LR and high-resolution HSI based on prior knowledge. While deep-learning-based SR reconstruction needs a large amount of training data to build models in an endto-end manner and learns the mapping relation autonomously. Notable models, such as very deep SR [27] and enhanced deep super-resolution (EDSR) [28], utilizing deep CNNs with residual learning and dense connections, can transmit information and gradients to learn deeper features more effectively and significantly enhance the reconstruction quality. Spatial– spectral prior for super-resolution (SSPSR) [26] uses spectral self-similarity to select image blocks and improves the effect of SR reconstruction. These methods learn the complex feature representation of images through common convolution, residuals, and dense connections, and achieve more satisfying results in spatial resolution than traditional methods. Recently, generative adversarial networks (GANs) and transfer learning have been widely employed to enhance the quality of super-resolved 18503 images. GANs, with their ability to generate photorealistic images, have shown promising results in image SR. DGAN [29], employing multiple generators within a GAN framework, effectively addresses the oversmoothing issue prevalent in traditional GAN-based image SR, resulting in significantly enhanced perceptual quality and accelerated training convergence. EEGAN [30] effectively addresses the challenge of recovering highfrequency edge details in noise-contaminated satellite images by combining an ultradense subnetwork for feature extraction and an edge-enhancement subnetwork for detailed refinement within a GAN framework. Meanwhile, transfer learning has been explored to mitigate the scarcity of training data by leveraging pretrained models, thereby accelerating the training process and improving performance. PSRGAN [31] addresses the challenge of limited infrared image training data by proposing a progressive SR GAN architecture that leverages information from available visible images through a multistage transfer learning strategy. However, these models perform poorly in the fidelity of the spectral structure before and after reconstruction due to ignoring the spectral relationships. In order to solve the spectral distortion, a grouped deep recursive residual network based on grouped embedding recursive module [32] is proposed, which fully considers the spectral correlation information and learns more effective mapping functions. To more effectively utilize the spectral and spatial information among consecutive LR bands, an internal fusion network [33] is proposed, which jointly learns spatial information, spectral information, and spectral difference information to further explore spatial and spectral information. For the purpose of better adapting to the high-dimensional image data of HSI, a 3-D fully CNN [34] is designed to learn the spatial and spectral correlation simultaneously through a 3-D convolution. Similar to this, a 3-D GAN is proposed by Li et al. [35] with band attention mechanism to alleviate the problem of spectral distortion. Despite these advancements, CNN-based methods still cannot overcome the drawback of fixed-size convolution, which is unable to obtain remote information associated with high-dimensional hyperspectral data. The limited receptive field also hinders the expressive abilities. Besides, traditional hyperspectral SR reconstruction models require a large HSIs for training, while the collection of sufficient HSIs poses challenges, and the vast training data burden the model. Particularly, HSIs exhibit strong correlations between adjacent bands, and it is crucial to exploit these correlations for hyperspectral highquality SR. Experiments made by Li et al. [32] confirmed this viewpoint, demonstrating that redundant information between bands contributes to high-quality SR reconstruction. To take full advantage of spectral band correlations, inspired by SSPSR [31] and GDRRN [32], many scholars have proposed grouping-based model frameworks. Those models split HSI into distinct groups first, reconstruct spectral data within each group, respectively, and combine the outcomes of each group finally. It not only enhances the quality of reconstruction but also improves the overall computational efficiency. However, these models only consider spectral correlations within groups, and neglect the spectral continuity between different spectral groups. To expand on the concept of grouping, Wang et al. [36] introduced 18504 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 17, 2024 Fig. 1. Error map of the reconstruction results and the ground truths of the EDSR model in 50 and 100 stages, respectively. Fig. 2. the super-resolution via recurrent feedback embedding (RFSR) network, which employed a recurrent feedback mechanism to leverage similarities and complementary information within adjacent spectral groups, and enhanced the reconstruction details. Subsequently, Wang et al. [37] investigated complementary information within adjacent spectral groups to enhance detailed reconstruction. At the same time, a dual-stage network [38] was designed to leverage spatial–spectral similarity between neighboring bands including both coarse and fine stages. Those innovative approaches enable hyperspectral SR to reconstruct spectral information precisely through feedback from different groups, fully leverage the continuity between spectral groups, and enhance the final SR performance. These models effectively capture the correlation information between distant spectral bands, but the convolution dependence on fixed-size kernels restricts models utilizing only a fraction of the available spatial information. B. Pixel Aware Weights Inspired by the idea of loop fusion, the intermediate results can provide information for subsequent regularization. In pixel perception based self-supervised SR reconstruction methods [39], pixel-aware weights and intermediate results are usually used to optimize and guide the following SR reconstruction process. In Fig. 1, it can be observed that the reconstructed background pixels can reach the minimum value in the early stage of training and then keep constant. But the foreground pixels behave differently. This suggest that each pixel presents distinct reconstruction difficulties and the background pixels are easier to reconstruct than those in the foreground. Thus, the reconstruction result from two different stages can be leveraged to assess the reconstruction difficulty of individual pixels. By subtracting the reconstruction results of two different stages, the difference results reflect the change of pixels in different regions. Pixels in the foreground area that need to be further reconstructed exhibit a larger difference, while those in the background region demonstrate a smaller difference. Therefore, to measure difference between the reconstruction result of one stage and the previous one can guide us for the following reconstruction. Main structure of pseudo-3-D convolution. C. Self-Attention Mechanism Transformer model has been widely used in the field of deep learning recently. Self-attention, as an important part of transformer, determines the weight of each position by calculating the similarity between each position and all other positions, and better capture the information worthy of attention from the image. The specific process is as follows: QKT V (1) Attention(Q, K, V) = softmax √ dk where Q, K, and V represent the input vectors after linear transformation. dk represents vector dimension. In the proposed work, self-attention model is adopted to capture spatial attention information of HSI at each stage after grouping, obtain weights to guide the reconstruction for the next stage, and improve the spatial reconstruction effect. D. Pseudo-3-D Convolution Traditional 2-D convolution operation has to process MSI band by band, while 3-D convolution can maintain the original channel correlation and better extract spatial spectral features in 3-D space. Generally, to standardize the spatial–spectral structure of hyperspectral data, 3-D convolution should be used. Considering the significantly growth in parameters and computational complexity, an efficient pseudo-3-D convolution [40] is adopted in the proposed model. The pseudo-3-D convolution is implemented using two convolutional layers, as shown in Fig. 2, in which the first (spectral) convolution has a filter shape of (1, 1, k) and the second (spatial) convolution has a filter shape of (k, k, 1) (here, k is set as 3). By decomposing the 3-D convolution operation into spatial and spectral parts, pseudo-3-D convolution can explore the correlation between spatial and spectra with less computational overhead. III. PROPOSED METHOD In this section, we will introduce the overall architecture of the proposed model first. Then, a detailed explanation of each part will be made, including a description of the loss function. In this article, the SR HSI is denoted as H sr ∈ RH×W ×B , where FENG et al.: RECURRENT FEEDBACK HYPERSPECTRAL IMAGE SUPER-RESOLUTION RECONSTRUCTION METHOD Fig. 3. 18505 Main structure of our proposed PARF includes RF-net module, pixel-aware module, and spatial spectrum consistent module. H, W, and B represent its height, width, and the number of band, respectively. x ∈ Rh×w×B represents LR HSI, where h, w, and B represent its height, width, and band number, respectively (here,h H andw W ). The flowchart of the proposed method is shown in Fig. 3. As shown in Fig. 3, the proposed PARF mainly consists of three parts, i.e., the RF-Net module (in blue box), the pixelaware module (in orange box), and the spatial spectrum consistent part (in grey box). We will give an overview of the entire work, and then introduce the main network structure, i.e., PARF in Section III-A, the acquisition of self-attention-based pixel-aware weights in Section III-B, the optimized pseudo-3-D convolutional modules in Section III-C, and loss function in Section III-D. For the input LR HSI, x ∈ Rh×w×B , we divide the spectral lr lr segments into G groups first, e.g.,x = [elr 1 , . . . , eg , . . . , eG ]. Each of these groups contains c spectral bands, and they will be fed into the network for reconstruction. The grouping operation can better utilize the correlation between adjacent spectral bands. Meanwhile, the feature dimension can also be reduced in the network. To fully maintain the continuity between groups, PARF extracts continuous information from the hidden state si , which is an abstract feature containing the spectral information of previous groups. With the flow of information between groups, the spatial–spectral information of the previous group is transmitted to the current group as the form of hidden state, which constrains the spectrum of the current group and helps reconstruction. The key process is as follows: sr lr b (2) êsr g = RFNet sg−1 , êg−1 , eg , eg , wg−1 b where elr g represents the input LR-group, eg is the HSI sr after bicubic interpolation, and êg is the output after reconstruction.êsr g−1 is the result of the previous group of SR reconstruction. With the previous set of reconstruction results, a hidden state sg−1 is also added to capture the long-distance dependence of the continuity of the spectral sequence. Inspired by the network of Xie et al. [41], we add pixel-aware weights wg−1 to obtain the reconstruction difficulty of pixels, and passes it backward to help the reconstruct of spatial resolution. The orange box in Fig. 3 is the way to obtain the pixel-aware weights, where Ei represents the result of the reconstruction of the ith group, and Ei = êsr i .E0 is initialized with 0. sr eb1 is upsampled by bicubic from elr 1 , and s0 , ê0 , and w0 are all initialized with 0. With Formula 2, we obtain the reconstruction result êsr 1 , and update s0 to s1 as the hidden state of second group. In pixel-aware module, we obtain new weights w1 through E0 sr and current results êsr 1 . Then, we send the s1 , ê1 , and w1 into the second group and use the same method as the first group for reconstruction. Repeat previous operation until the end of last group of reconstruction. In this way, the high precision reconstruction can be guaranteed and enhanced through the information transfer. Finally, we combine all the reconstructed results group by group and obtain the reconstructed result, i.e., Ĥ sr = sr sr [êsr 1 , . . . , êg , . . . , êG ]. To ensure the spectral consistency, we add a deeper 3-D convolution operation to perform the spatial– spectral regularization to generate the final result. The expression is as follows: H sr = PP 3D+ Ĥ sr (3) 18506 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 17, 2024 upsampled image through the residual structure h b ˆh êsr g = fg + eg + Up fg Fig. 4. Main model structure of recurrent feedback network. where PP 3D+ (·) represents the pseudo-3-D module, which will be introduced in detail in Section III-D. H sr represents the final reconstruction result with high spatial–spectral resolution. A. Recurrent Feedback Network Based on Pixel-Aware Optimization PARF is dedicated to obtaining information from feedback information and complementing the information between groups. h×w×c is the gth Its specific structure is shown in Fig. 4. elr g ∈R group input LR image. The reconstruction results êsr g−1 of the , and the hidden state sg−1 previous group, the current group elr g on the channel are connected, and then obtain an abstract feature representation by a 3×3 convolution. The convolution process is as follows: lr . (4) fgl = C3×3 sg−1 , êsr g−1 , eg Among them, [·] represents the Catch operation on the channel, and C3×3 represents the 3×3 convolution operation. Considering the correlation between spectral data, RCAB modules [42] based channel attention mechanism is utilized, which is expressed as fˆgh = RRCABD · · · RRCABd · · · RRCAB1 (fgl ) · ·· · ·· . (5) The RCAB module consists of two convolutional layers and a channel attention module. The channel attention module contains a global pooling operation, two convolution operations, and a softmax function. The correlation information fˆgh , obtained by RCAB module, will be input for two branches. One is the 3×3 convolution operation, which is used to obtain long-distance dependence sg , acted as the feedback information of the next group. Subsequently, pixel perception is added based on the original RF-Net, functioned as a weight wg−1 , which is the weight to measure the difficulty of pixels reconstruction, obtained by calculating the difference between the previous two groups. Then, the generated weight is applied to the current group, and formulated as follows to get fgh , which is a feature containing the spatial weight information: (6) fgh = wg−1 Up fˆgh where U p(fˆgh ) represents the intermediate result after upsampling. Subsequently, we combine the obtained fgh and the (7) where fgh is the intermediate result obtained by formula (6), and ebg is the unsampled image with bicubic interpolation. Up(·) is denoted as function performed by PixelShuffle operator [43]. This module can fuse the current group’s information with the preceding group, and transmit to the next group, which enhances the intergroup information correlation, and greatly improves the spectral consistency. B. Self-Attention-Based Pixel-Aware Weights At present, transformer is widely used in various fields owing to the excellent performance of self-attention. Self-attention mechanism can capture more information from image itself. As mentioned before, the loss of background pixels can reach the minimum in the early training stage, and then remain unchanged, while the foreground pixels and details continue convergence. Here, self-attention assists to search for more foreground regions, making the model fit data better. Then, we combine the previous reconstruction group and the current one to obtain the pixel-aware weights for the following reconstruction process, which can be defined as (8) wg = Sigmoid SA (Eg−1 ) − SA êsr g where êsr g represents the reconstruction results of group g, Eg−1 represents the reconstruction results of the previous group, and SA(·)represents the self-attention mechanism. This module cannot only obtain larger receptive field feature representations through self-attention, but also distinguish foreground and background regions through the difference between two groups and transmit targeted learning to the next group, which further enhances the model’s performance especially in spatial domain. C. Pseudo-3-D+ Two layers of pseudo-3-D convolution has been used in RF-Net to construct a pseudo-3-D residual block. Although the spatial–spectral consistency can be maintained to a certain extent, the extraction and expression of features are limited. To address this issue, we increased the number of layers of 3-D convolution and added RELU and regularization in the middle, making the module more nonlinear, and better abstract features. Specifically, the final result is obtained by stacking multiple pseudo-3-D modules mid1×1×3 = RELU BN CSpe1×1×3 Ĥ sr PP 3D+1 Ĥ sr = RELU (BN (CSpa3×3×1 (mid1×1×3 ))) H sr = PP 3D+K · · ·PP 3D+k · · ·PP 3D+1 (Ĥ sr ) · ·· · ·· (9) where mid1×1×3 represent the middle result. CSpa3×3×1 and CSpe1×1×3 represent spatial convolution and spectral convolution, respectively. Ĥ sr and H sr represent the intermediate result and the final reconstruction result, respectively. By stacking k FENG et al.: RECURRENT FEEDBACK HYPERSPECTRAL IMAGE SUPER-RESOLUTION RECONSTRUCTION METHOD layers of pseudo-3-D convolutions, the reassembled reconstruction results can be further regularized, and ensure the spectral structure consistency. It has been verified after several experiments that a better result can be achieved when k is taken as 3. D. Loss Function Most current SR reconstruction methods use pixel-level loss to measure the distance between the reconstruction result SR and the real HR image. Due to the oversmoothing result of the L2 loss, we choose L1 as the context loss to measure the reconstruction accuracy of the network, which is defined as 18507 spectral information is inherently encoded within the gradient structure of the hyperspectral data. Preserving the gradient information ensures that the reconstructed image maintains the same underlying spectral variations as the ground truth. In summary, the final target loss is defined as Ltotal (Θ) = Lhr + Llr + λ1 Lspe + λ2 Lgra (15) where λ is a tradeoff parameter for different losses. Specifically, λ1 controls the weight of the spectral term and the gradient term is constrained by λ2, which controls the weight of the spatial and spectral structure. N Lhr = 1 Hnhr − Hnsr 1 . N n=1 (10) A. Datasets and Experimental Setup Among them, Hnhr and Hnsr are the nth of real HSI and reconstructed HSI, respectively. N is the number of images in a batch. Considering the inevitable information loss in the process of HR HSI downsampling to obtain LR HSI, we added the dual loss. For the reconstructed high-resolution image SR, the LR SR Lsr n is obtained through the same downsampling method, and adopt the L1 as spatial loss, which is specifically expressed as follows: N Llr = 1 Llr − Lsr n 1 N n=1 n (11) sr where Llr n and Ln are, respectively, the nth of LR HSI and results after SR downsampling. Apart from the context loss and dual loss, spectral measurement is also utilized in this model to constrain the spectral structure, which is expressed as N Hnhr · Hnsr 1 1 arccos . (12) Lspe = N n=1 π Hnhr 2 Hnsr 2 In addition, as gradient information can enhance the spatial relationship of SR images [44], [45], in this article, we also design a gradient map of the HSI by computing the difference between adjacent pixels, we calculate the gradient maps in the height (h), width (w), spectral directions (s) dim, and then add the three results through the L2 norm M (H) = ∇H2 s.t. ∇H = (∇h H, ∇w H, ∇s H). (13) Among them, M ( · )represents the operation of extracting the gradient map. Considering gradient maps can better reflect the spatial–spectral dependence of HSIs; to mitigate the spatial spectral distortion, a gradient constraint is utilized to provide additional supervision and set gradient loss for image reconstruction. The gradient loss is defined as the gradient map difference between the super-resolved image and the corresponding HR image N Lgra (Θ) = 1 M (Hnhr ) − M (Hnsr ) 1 . N n=1 IV. EXPERIMENTS AND RESULTS (14) By minimizing this loss, the network is encouraged to preserve the gradient structure of the original image, thereby reducing spectral distortions. The rationale behind this is that the The proposed method is tested on four extensive datasets, including a natural hyperspectral dataset CAVE [46] and three hyperspectral remote sensing datasets, i.e., Pavia Centre, Chikusei [47], and Washington DC datasets. Five methods, i.e., bicubic interpolation, EDSR [28], SSPSR [26], RFSR [36], and GELIN [37], are selected as comparison algorithms for analysis. Due to the lack of spectral grouping GAN-based methods for SR of HSI. Therefore, we have chosen to compare the proposed method with other state-of-the-art grouping-based methods in order to make a more direct and relevant comparison within this specific research area. As the baseline model, RFSR is a recurrent feedback network for HSI SR, focusing on exploiting spectral correlations through group-based processing and spatial–spectral regularization. Bicubic interpolation as a classical image SR method is chosen to compare. EDSR is a deep residual network optimized for image SR by removing redundant modules and expanding network size. SSPSR employs group convolutions and a spatial–spectral prior network to effectively extract features from hyperspectral data. GELIN is a grouping-based HSI SR method that employs spatial–spectral embedding learning and neighboring group integration to enhance feature extraction and reconstruction. Meanwhile, five reference indicators, including average PSNR, spectral angle mapper (SAM), structural similarity index index, relative global error (ERGAS), and root mean square error (RMSE), are adopted for quantitative evaluation. B. Experimental Results on CAVE The CAVE dataset contains 32 HR HSIs, and the size of each HSI is 512×512 with 31 bands. We randomly select 23 HSIs as training set, 2 images as validation set, and 6 HSIs as testing set. For each image in the dataset, a 64×64 region with a step size of 32 pixels is extracted to increase the number of images used for training. We use zero value to fill the spectral band to 32 bands and divide them into 8 groups during training procedure. The LR HSI is generated by bicubic downsampling of HR HSI. In the test phase, we divide the image into the same size as the training set and send it to the model for testing. The average value of each index is taken as the final data of the image. In the experiment, the scale factor is set as 4 and 8, and the results are shown in Table I. The bold represents the 18508 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 17, 2024 TABLE I QUANTITATIVE EVALUATION RESULTS OF DIFFERENT METHODS ON THE CAVE DATASET Fig. 5. Error maps of three test HSIs in the CAVE dataset. best result, and the underline represents the second-best result. Since bicubic only reconstructs values by interpolation, the results of spectral reconstruction are far inferior to other methods, which can be observed in Table I. As the MPNSR, deep-learning-based methods can reach more than 42 dB, but bicubic is only 40.2358 dB. Furthermore, due to the lack of spectral information, the SAM of bicubic reaches the lowest at 5.484. Differing from bicubic interpolation reconstruction, EDSR focuses on the use of spatial information, enhancing spatial relationship effectively, as MPSNR indicate that it has 1.8 dB higher than the bicubic. Similarly, it has better results on ERGAS and RMSE. However, due to lack of spectral consideration, spectral distortion is unavoidable, the SAM index also shows that the method leads to obvious spectral distortion. Through grouping strategy, SSPSR and RFSR not only consider spatial information, but also impose spectral similarity, and there is a significant improvement in SAM, which is about 0.3 higher than EDSR. GELIN excels in both spatial and spectral performance by associating spatial and spectral information from more groups. While compared with the proposed method, GELIN shows slightly inferior to the proposed method for both 4 and 8 scale in all indicators. Specifically, in terms of spatial resolution reconstruction, the proposed method surpasses the second-best approach by approximately 0.74 dB, which is attributed to the pixel perception weight and provides more reliable spatial and spectral information for reconstruction. And in terms of spectral fidelity, the application of optimized pseudo-3-D convolution operation provides more spectral information to the model, and the proposed method outperforms the second-best method by 0.172. FENG et al.: RECURRENT FEEDBACK HYPERSPECTRAL IMAGE SUPER-RESOLUTION RECONSTRUCTION METHOD Fig. 6. 18509 Error maps of two test HSIs in the CAVE dataset. TABLE II QUANTITATIVE EVALUATION RESULTS OF DIFFERENT METHODS ON THE PAVIA CENTRE DATASET To better compare the different SR reconstruction methods, we calculate the error between the reconstruction results and ground truths from both spatial and spectral aspects. Fig. 5 shows three examples of the difference from CAVE dataset on the 30th channel at the scale factor 4, and the PSNR and SAM values of the reconstructed image are labeled at the bottom for comparison. Fig. 6 is the reconstruction result at the scale factor 8. From Fig. 6, it can be noticed the proposed method inherits the advantages of the RFSR method in restoring the spatial information of HSIs, and performs better in the reconstruction of object edges, especially for the marked region compared with the EDSR and SSPSR. To further compare the RFSR and proposed methods, we also compare the reconstructed images of the two methods on the same image and use the error map. As shown in Fig. 7, the proposed method is significantly better in edge and object texture reconstruction, owing to the training of self-attention-based pixel-aware weights. Fig. 7. Error maps of balloons and flowers between RFSR and PARF. In addition, the proposed model has also achieved the best results in spectral restoration, and the SAM index reached a satisfying level. Fig. 8 shows the spectral error curves of the reconstructed image and the ground truths on each band. It 18510 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 17, 2024 divide the image into multiple 64×64 sizes by taking 32 pixels as the step size, and obtain the training set of downsampling 4-fold and 8-fold by means of bicubic interpolation. During training, to divide the whole spectral bands into 13 groups, 2 zero-value bands are constructed and put in the final group. In the test phase, we segment the image into the same size as training set and send it to the model for testing. The average value of each index is taken as the final data of the image. Table II shows the quantitative results of the five methods on the evaluation indicators. From the results, it can be observed the proposed method is obviously better than other methods on both scales, which shows that the proposed method is robustness in both natural images and remote sensing images. Fig. 9 also shows the overall error map of different test images at the scale factor 4. It can be noticed that the proposed model has smaller errors than other models in loss of roofs and river banks. This is mainly attributed to pixel-aware weight based on self-attention that can help reconstruct the missing spatial information through intermediate results. Fig. 10 displays the spectral error. Compared to CAVE dataset, the Pavia dataset exhibits a noticeable increase in the number of bands, which makes spectral consistency more challenge. It can be observed that the grouping-based methods (SSPSR, RFSR, GELIN, and proposed method) demonstrate a significant advantage in spectral consistency over bicubic interpolation and EDSR by effectively capturing the correlations between distant spectral bands. While incorporating multiple layers of pseudo-3-D modules for spectral regularization, the proposed method obtains smaller average differences and illustrates greater stability in this regard. In Fig. 11, we select the 60th, 31st, and 12th band to synthesize true color image, and visualize the reconstruction results with a scale factor of 4. From the figure, it can be observed EDSR and SSPSR are partially blurred, and the images have significantly spectral distortion. Although the baseline model RFSR has a relatively satisfying performance, it is not smooth in the reconstruction of edge information. The results obtained by the proposed method is better visually. Fig. 8. Mean spectral difference curve in the CAVE dataset at the scale factor 4. (a) Balloons.mat. (b) Flowers.mat. (c) Egyptian_statue.mat. can be found the loss of the proposed method on each band can mostly reach the minimum, mainly because the feedback information is added to the spectral information. Besides, the optimized pseudo-3-D convolution also provides the model with more nonlinearity by stacking multiple 3-D convolution, which further improves the results of spectral regularization. C. Experimental Results on Pavia The Pavia Centre dataset is a scene captured by the ROSIS sensor as it flew over Pavia, northern Italy, and contains 102 spectral bands. The Pavia Centre dataset is a 1096×1096 HSI. Since some areas in the original image contain limited information, we discard this part and finally obtain a 1096×715 image. First, we D. Experimental Results on Chikusei The Chikusei dataset contains HSI of Ibaraki, Japan, obtained by the Hyperspec–VNIR–CIRIS. The size is 2517 × 2335, and 128 bands in the spectral range of 363–1018 nm. We select the part of 2048 × 2048 to evaluate the proposed method. Similar to the above datasets, it will first be divided into 16 images with 512 × 512, and then be divided into 64 × 64 as the HR image, and perform 4-fold and 8-fold downsampling through bicubic interpolation to get the LR HSI, which contains 128 bands. Then, grouping is taken to divide these data into 16 spectral groups, and each group contains 8 spectral bands. For Pavia data, due to the high spatial resolution and abundant spectral information, EDSR exhibits inferior performance in spatial and spectral reconstruction, and also performs poorly on Chikusei dataset. Therefore, for the Chikusei dataset, we solely conduct comparative experiments with the bicubic, SSPSR, RFSR, and GELIN models. Table III presents the quantitative FENG et al.: RECURRENT FEEDBACK HYPERSPECTRAL IMAGE SUPER-RESOLUTION RECONSTRUCTION METHOD Fig. 9. 18511 Error maps of three test HSIs in the Pavia Centre dataset. TABLE III QUANTITATIVE EVALUATION RESULTS OF DIFFERENT METHODS ON THE CHIKUSEI DATASET Fig. 10. Mean spectral difference curve of Pavia Centre dataset at scale factor 4/8. (a) Spectral error under the scale factor of 4. (b) Spectral error under the scale factor of 8. results for the four methods under 4× and 8× reconstruction. Proposed method outperforms the second-best GELIN method by approximately 0.35 dB in MPSNR. As the increase of spectra, proposed method performs better in keeping the spectral structure, while bicubic is significantly worse than the other methods due to the fact that it only reconstructs values by interpolation. The other grouping-based methods typically achieve SAM values around 3.3 to 3.5, proposed method achieves 2.85, highlighting the advantages of pseudo-3-D+ in preserving spectral structure. Despite encountering challenges in 8× reconstruction ratio due to information loss during downsampling, proposed method still achieves the best results across all indices compared to the baseline methods. Fig. 12 shows the error map of three test images. The darker the color, the smaller the error. It can be clearly seen that errors of the proposed method are the smallest for large farmland and the junction of river water and river bank owing to the excellent feature extraction ability of the self-attention mechanism for 18512 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 17, 2024 Fig. 11. Reconstructed images of two test HSIs in the Pavia Centre dataset with spectral bands 60-31-12 as R-G-B with the scale factor 4. large patch areas. In Fig. 13, to better visualize the results, the 70th, 100th, and 36th bands are selected to synthesize true color image with scale factor of 4 and 8. It can be noticed other methods have a certain blur in the visual sense, while the proposed method is much clearer. Fig. 12. factor. Mean spectral difference curve in the Chikusei dataset at the scale E. Experimental Results on Washington DC Washington DC dataset is an aerial HSI acquired by HYDICE sensor, primarily focusing on Washington DC shopping center. The data comprise 191 bands ranging from 0.4 to 2.4 μm, covering both visible and near-infrared spectral regions. The size of image is 1208×307, capturing features such as roofs, streets, paved roads, grasslands, trees, water bodies, and shadows. Following the same methodology as before, we partitioned the dataset into 64×64 patches first, and then performed bicubic downsampling by factors of 4 and 8 to construct the training set. Differing from the previous approach, we selected images size of 256×256 for testing due to the limitations of original image size. In training process, we also use zero value to fill the spectral band to 192 bands and divide it into 24 spectral groups, and each group contains 8 spectral bands. The experimental results are presented in Table IV in detail. In Table IV, as the number of bands reaches 191, the proposed model still achieves 0.55 dB higher than the second-best method in MPSNR indicator. Concerning spectral metrics, RFSR faces challenges in capturing long-distance correlations between groups due to its restricted information transfer, which is limited to the preceding group and the subsequent one, and results in difficulties in maintaining spectral structure. While FENG et al.: RECURRENT FEEDBACK HYPERSPECTRAL IMAGE SUPER-RESOLUTION RECONSTRUCTION METHOD 18513 TABLE IV QUANTITATIVE EVALUATION RESULTS OF DIFFERENT METHODS ON THE WASHINGTON DC DATASET Fig. 13. Reconstructed images in the Chikusei dataset with spectral bands 70-100-36 as R-G-B with the scale factor 4 and 8. GELIN is adept at preserving spectral structure as it can continually associate information between the current group and its preceding and succeeding groups (three groups in total). The proposed method also achieves outstanding results by collecting spatial and spectral information from the first two groups during the transmission process. Fig. 14. Mean spectral difference curve in the Washington DC dataset. (a) Spectral error under the scale factor of 4. (b) Spectral error under the scale factor of 8. Fig. 14 illustrates the spectral errors of different methods on Washington DC. As previously analyzed, the spectral error of RFSR usually shows a relatively poorer performance when the number of bands is large, due to the limited information transfer between groups. In contrast, the proposed method can maintain the as much as spectral feature benefiting from enhanced information transmission ability. In Fig. 15, the 60th, 27th, and 17th bands are selected as RGB bands and synthesized true color image with scale factor 18514 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 17, 2024 TABLE V QUANTITATIVE PERFORMANCE OF DIFFERENT LOSS FUNCTIONS TABLE VI QUANTITATIVE PERFORMANCE OF DIFFERENT MODULE Fig. 15. Reconstructed images in the Washington DC dataset with spectral bands 60-27-17 as R-G-B. of 4 and 8. From Fig. 15, it can be observed the proposed method exhibits clearer and sharper textures due to the better capturing of spatial structure by pixel-aware weight based on self-attention. F. Ablation Study Within the proposed method, the pixel-aware weights, pseudo-3-D+ and the losses will influence the performance. To testify how they affect the performance of the proposed method, the following ablation studies were conducted on the Chikusei dataset. In ablation experiments, RFSR is used as the baseline model, and L1 loss and spectral loss are employed as constraints. Subsequent experiments are conducted by, respectively, incorporating the dual loss and gradient loss to validate the effectiveness of adding losses. The results are presented in Table V. The results in Table V show that the addition of dual loss and gradient loss can slightly improve the accuracy, but the effect is minimal. Only by combining the two losses, the accuracy of reconstruction will have a certain gain. Dual loss reduces the size of solution space of model, so that the model can be solved better. Gradient loss uses the gradient information of the spectrum to enhance the model’s constraints on the spectrum and avoid spectral distortion as much as possible. Further ablation experiments were conducted on the proposed self-attention-based pixel-aware weights and enhanced pseudo-3-D+ convolutions, Table VI shows the results. Without pixel-aware weights and pseudo-3-D+ module, the quantitative evaluation indexes reach the lowest level. After integrating the self-attention-based pixel-wise weights, the spatial performance has a significant enhancement, improved by 0.28 dB. Similarly, the introduction of pseudo-3-D+ regularization also optimize the spectral structure by applying regularization operations from both spatial and spectral perspectives. According to the results in Table VI, the proposed method could reconstruct high-quality HR estimation when using the two modules, indicating the effectiveness of the proposed modules. V. CONCLUSION In this article, to further improve the spatial resolution of HSI SR reconstruction and maintain spectral consistency, selfattention-based pixel-aware weights is proposed to add to the recurrent feedback SR reconstruction model. In addition, the grouping operation is also used for HSIs for spectral consistency, and the correlation between the spectra is fully utilized by intergroup information transmission and enhanced pseudo-3-D FENG et al.: RECURRENT FEEDBACK HYPERSPECTRAL IMAGE SUPER-RESOLUTION RECONSTRUCTION METHOD convolution. Meanwhile, the pixel perception weight is continuously updated through the reconstruction results of the previous group to help the reconstruction of the subsequent spectral groups. Besides, the structure of pseudo-3-D convolution is optimized, and the model can better regularize the spectrum. On the loss function, dual loss as well as gradient loss is used to enhance the SR reconstruction. Experiments with four datasets have achieved excellent results in both qualitative and quantitative aspects. While the proposed method achieves promising results, there is still room for improvement. The computational cost associated with self-attention and the dataset-specific nature of grouping strategies are limitations. Future article can focus on developing more efficient architectures and exploring adaptive grouping strategies. Moreover, the potential applications of our method in various domains, such as remote sensing and medical imaging, warrant further investigation. REFERENCES [1] W. Zhao and S. Du, “Spectral–spatial feature extraction for hyperspectral image classification: A dimension reduction and deep learning approach,” IEEE Trans. Geosci. Remote Sens., vol. 54, no. 8, pp. 4544–4554, Aug. 2016, doi: 10.1109/TGRS.2016.2543748. [2] J. Hu, T. Li, M. Zhao, and J. Ning, “Hyperspectral image superresolution via deep structure and texture interfusion,” IEEE J. Sel. Topics Appl. Earth Observ. Remote Sens., vol. 14, pp. 8665–8678, 2021, doi: 10.1109/JSTARS.2021.3107311. [3] S. Puustinen et al., “Towards clinical hyperspectral imaging (HSI) standards: Initial design for a microneurosurgical HSI database,” in Proc. IEEE 35th Int. Symp. Comput. Med. Syst., 2022, pp. 394–399, doi: 10.1109/CBMS55023.2022.00077. [4] C. Cevoli, L. Di Cecilia, L. Ferrari, A. Fabbri, and G. Molari, “Potential of in-field vis/NIR hyperspectral imaging to monitor quality parameters of alfalfa,” in Proc. IEEE Int. Workshop Metrol. Agriculture Forestry, 2021, pp. 341–345, doi: 10.1109/MetroAgriFor52389.2021.9628816. [5] D. Lorts, “Real-time embedded hyperspectral image compression for tactical military platforms,” in Proc. Appl. Imagery Pattern Recognit. Workshop, 2002, p. 140, doi: 10.1109/AIPR.2002.1182267. [6] R. Dian, S. Li, and L. Fang, “Learning a low tensor-train rank representation for hyperspectral image super-resolution,” IEEE Trans. Neural Netw. Learn. Syst., vol. 30, no. 9, pp. 2672–2683, Sep. 2019, doi: 10.1109/TNNLS.2018.2885616. [7] W. Dong et al., “Hyperspectral image super-resolution via non-negative structured sparse representation,” IEEE Trans. Image Process., vol. 25, no. 5, pp. 2337–2352, May 2016, doi: 10.1109/TIP.2016.2542360. [8] M. R. Vicinanza, R. Restaino, G. Vivone, M. Dalla Mura, and J. Chanussot, “A pansharpening method based on the sparse representation of injected details,” IEEE Geosci. Remote Sens. Lett., vol. 12, no. 1, pp. 180–184, Jan. 2015, doi: 10.1109/LGRS.2014.2331291. [9] R. Fei, J. Zhang, J. Liu, F. Du, P. Chang, and J. Hu, “Convolutional sparse representation of injected details for pansharpening,” IEEE Geosci. Remote Sens. Lett., vol. 16, no. 10, pp. 1595–1599, Oct. 2019, doi: 10.1109/LGRS.2019.2904526. [10] H. Chang, D.Y. Yeung, and Y. Xiong, “Super-resolution through neighbor embedding,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit., 2004, p. I, doi: 10.1109/cvpr.2004.1315043. [11] D. Zhou, Y. Liu, X. Li, and C. Zhang, “Single-image super-resolution based on local biquadratic spline with edge constraints and adaptive optimization in transform domain,” Vis. Comput., vol. 38, no. 1, pp. 119–134, Jan. 2022, doi: 10.1007/s00371-020-02007-z. [12] J. Liu, Z. Gan, and X. Zhu, “Directional bicubic interpolation—A new method of image super-resolution,” in Proc. 3rd Int. Conf. Multimedia Technol., 2013, pp. 470–477, doi: 10.2991/icmt-13.2013.57. [13] N. Akhtar, F. Shafait, and A. Mian, “Bayesian sparse representation for hyperspectral image super resolution,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., 2015, pp. 3631–3640, doi: 10.1109/CVPR.2015.7298986. [14] Y. Xu, Z. Wu, J. Chanussot, and Z. Wei, “Nonlocal patch tensor sparse representation for hyperspectral image super-resolution,” IEEE Trans. Image Process., vol. 28, no. 6, pp. 3034–3047, Jun. 2019, doi: 10.1109/TIP.2019.2893530. 18515 [15] X.-H. Han, B. Shi, and Y. Zheng, “Self-similarity constrained sparse representation for hyperspectral image super-resolution,” IEEE Trans. Image Process., vol. 27, no. 11, pp. 5625–5637, Nov. 2018, doi: 10.1109/TIP.2018.2855418. [16] R. Dian, L. Fang, and S. Li, “Hyperspectral image super-resolution via nonlocal sparse tensor factorization,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., 2017, pp. 3862–3871, doi: 10.1109/CVPR.2017.411. [17] R. Dian and S. Li, “Hyperspectral image super-resolution via subspace based low tensor multi-rank regularization,” IEEE Trans. Image Process., vol. 28, no. 10, pp. 5135–5146, Oct. 2019, doi: 10.1109/TIP.2019.2916734. [18] J. Xue, Y.-Q. Zhao, Y. Bu, W. Liao, J. C.-W. Chan, and W. Philips, “Spatialspectral structured sparse low-rank representation for hyperspectral image super-resolution,” IEEE Trans. Image Process., vol. 30, pp. 3084–3097, 2021, doi: 10.1109/TIP.2021.3058590. [19] S. H. Mahajan and V. K. Harpale, “Adaptive and non-adaptive image interpolation techniques,” in Proc. Int. Conf. Comput. Commun. Control Automat., 2015, pp. 772–775, doi: 10.1109/ICCUBEA.2015.154. [20] Y. Li, J. Hu, X. Zhao, W. Xie, and J. Li, “Hyperspectral image superresolution using deep convolutional neural network,” Neurocomputing, vol. 266, pp. 29–41, 2017, doi: 10.1016/j.neucom.2017.05.024. [21] Y. Li, L. Zhang, C. Dingl, W. Wei, and Y. Zhang, “Single hyperspectral image super-resolution with grouped deep recursive residual network,” in Proc. IEEE 4th Int. Conf. Multimedia Big Data, 2018, pp. 1–4, doi: 10.1109/BigMM.2018.8499097. [22] S. Mei, X. Yuan, J. Ji, Y. Zhang, S. Wan, and Q. Du, “Hyperspectral image spatial super-resolution via 3D full convolutional neural network,” Remote Sens., vol. 9, no. 11, 2017, Art. no. 1139, doi: 10.3390/rs9111139. [23] J. Yang, Y. Zhao, C. Chan, and L. Xiao, “A multi-scale wavelet 3D-CNN for hyperspectral image super-resolution,” Remote Sens., vol. 11, no. 13, 2019, Art. no. 1557, doi: 10.3390/rs11131557. [24] Q. Li, Q. Wang, and X. Li, “Mixed 2D/3D convolutional network for hyperspectral image super-resolution,” Remote Sens., vol. 12, no. 10, 2020, Art. no. 1660, doi: 10.3390/rs12101660. [25] Q. Wang, Q. Li, and X. Li, “Spatial-spectral residual network for hyperspectral image super-resolution,” 2020, arXiv:2001.04609. [26] J. Jiang, H. Sun, X. Liu, and J. Ma, “Learning spatial-spectral prior for super-resolution of hyperspectral imagery,” IEEE Trans. Comput. Imag., vol. 6, pp. 1082–1096, 2020, doi: 10.1109/TCI.2020.2996075. [27] J. Kim, J. K. Lee, and K. M. Lee, “Accurate image super-resolution using very deep convolutional networks,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit., 2016, pp. 1646–1654, doi: 10.1109/CVPR.2016.182. [28] B. Lim, S. Son, H. Kim, S. Nah, and K. M. Lee, “Enhanced deep residual networks for single image super-resolution,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. Workshops, 2017, pp. 1132–1140, doi: 10.1109/CVPRW.2017.151. [29] M. Zareapoor, M. E. Celebi, and J. Yang, “Diverse adversarial network for image super-resolution,” Signal Process.-Image Commun., vol. 74, pp. 191–200, 2019, doi: 10.1016/j.image.2019.02.008. [30] K. Jiang, Z. Wang, P. Yi, G. Wang, T. Lu, and J. Jiang, “Edgeenhanced GAN for remote sensing image superresolution,” IEEE Trans. Geosci. Remote Sens., vol. 57, no. 8, pp. 5799–5812, Aug. 2019, doi: 10.1109/TGRS.2019.2902431. [31] Y. Huang, Z. Jiang, R. Lan, S. Zhang, and K. Pi, “Infrared image superresolution via transfer learning and PSRGAN,” IEEE Signal Process. Lett., vol. 28, pp. 982–986, 2021, doi: 10.1109/LSP.2021.3077801. [32] Y. Li, L. Zhang, C. Dingl, W. Wei, and Y. Zhang, “Single hyperspectral image super-resolution with grouped deep recursive residual network,” in Proc. IEEE 4th Int. Conf. Multimedia Big Data, 2018, pp. 1–4, doi: 10.1109/BigMM.2018.8499097. [33] J. Hu, X. Jia, Y. Li, G. He, and M. Zhao, “Hyperspectral image superresolution via intrafusion network,” IEEE Trans. Geosci. Remote Sens., vol. 58, no. 10, pp. 7459–7471, Oct. 2020, doi: 10.1109/TGRS.2020.2982940. [34] S. Mei, X. Yuan, J. Ji, Y. Zhang, S. Wan, and Q. Du, “Hyperspectral image spatial super-resolution via 3D full convolutional neural network,” Remote Sens., vol. 9, no. 11, 2017, Art. no. 1139, doi: 10.3390/rs9111139. [35] J. Li et al., “Hyperspectral image super-resolution by band attention through adversarial learning,” IEEE Trans. Geosci. Remote Sens., vol. 58, no. 6, pp. 4304–4318, Jun. 2020, doi: 10.1109/TGRS.2019.2962713. [36] X. Wang, J. Ma, and J. Jiang, “Hyperspectral image super-resolution via recurrent feedback embedding and spatial–spectral consistency regularization,” IEEE Trans. Geosci. Remote Sens., vol. 60, 2022, Art. no. 5503113, doi: 10.1109/TGRS.2021.3064450. [37] X. Wang, Q. Hu, J. Jiang, and J. Ma, “A group-based embedding learning and integration network for hyperspectral image super-resolution,” IEEE Trans. Geosci. Remote Sens., vol. 60, 2022, Art. no. 5541416, doi: 10.1109/TGRS.2022.3217406. 18516 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 17, 2024 [38] Q. Li, Y. Yuan, X. Jia, and Q. Wang, “Dual-stage approach toward hyperspectral image super-resolution,” IEEE Trans. Image Process., vol. 31, pp. 7252–7263, 2022, doi: 10.1109/TIP.2022.3221287. [39] W. Wei, J. Nie, L. Zhang, and Y. Zhang, “Unsupervised recurrent hyperspectral imagery super-resolution using pixel-aware refinement,” IEEE Trans. Geosci. Remote Sens., vol. 60, 2022, Art. no. 5500315, doi: 10.1109/TGRS.2020.3039534. [40] Z. Qiu, T. Yao, and T. Mei, “Learning spatio-temporal representation with pseudo-3D residual networks,” in Proc. IEEE Int. Conf. Comput. Vis., 2017, pp. 5534–5542, doi: 10.1109/ICCV.2017.590. [41] S. Xie, S. Chen, H. Jonathan, Z. Tu, and K. P. Murphy, “Rethinking spatiotemporal feature learning: Speed-accuracy trade-offs in video classification,” in Proc. Eur. Conf. Comput. Vis., 2017, pp. 318–335, doi: 10.1007/978-3-030-01267-0_19. [42] C. Dong, C. C. Loy, K. He, and X. Tang, “Image superresolution using deep convolutional networks,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 38, no. 2, pp. 295–307, Feb. 2016, doi: 10.1109/TPAMI.2015.2439281. [43] W. Shi et al., “Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit., 2016, pp. 1874–1883, doi: 10.1109/CVPR.2016.207. [44] Q. Yan, Y. Xu, X. Yang, and T. Q. Nguyen, “Single image superresolution based on gradient profile sharpness,” IEEE Trans. Image Process., vol. 24, no. 10, pp. 3187–3202, Oct. 2015, doi: 10.1109/TIP.2015.2414877. [45] J. Sun, J. Sun, Z. Xu, and H.-Y. Shum, “Gradient profile prior and its applications in image super-resolution and enhancement,” IEEE Trans. Image Process., vol. 20, no. 6, pp. 1529–1542, Jun. 2011, doi: 10.1109/TIP.2010.2095871. [46] F. Yasuma, T. Mitsunaga, D. Iso, and S. K. Nayar, “Generalized assorted pixel camera: Postcapture control of resolution, dynamic range, and spectrum,” IEEE Trans. Image Process., vol. 19, no. 9, pp. 2241–2253, Sep. 2010, doi: 10.1109/TIP.2010.2046811. [47] N. Yokoya and A. Iwasaki, “Airborne hyperspectral data over chikusei,” Space Appl. Lab., Univ. Tokyo, Tokyo, Japan, Tech. Rep. SAL-2016-0527, 2016. Ruyi Feng (Member, IEEE) received the B.S. degree in geographic information system from the Hunan Normal University, Changsha, China, in 2011, and the M.S. degree in surveying and mapping engineering and the Ph.D. degree in photogrammetry and remote sensing from the Wuhan University, Wuhan, China, in 2013 and 2016, respectively. She has been with the School of Computer Science, China University of Geosciences, Wuhan, since 2016, and is currently an Associate Professor. Her research interests include sparse representation, deep learning, hyperspectral image analysis, high-resolution remote sensing understanding, and intelligent interpretation of remote sensing imagery. Zhongyu Guo received the B.E. degree from the Hubei University of Education, Wuhan, China, in 2021, and the M.E. degree in computer technology from the China University of Geosciences, Wuhan, in 2024. His research interests include object detection, super-resolution reconstruction, change detection, and deep learning. Xiaofeng Wang received the doctoral degree in structural geology from the China University of Geosciences, Wuhan, China, in 2015. He has been with the China University of Geosciences, since 2013, and engaged in teaching and research in the field of geosciences information engineering. His research interests include deep learning for remote sensing, semiautomated lineament detection, and geoscience information engineering.
0
You can add this document to your study collection(s)
Sign in Available only to authorized usersYou can add this document to your saved list
Sign in Available only to authorized users(For complaints, use another form )