Hyperspectral Imaging Method Research Articles

Exploring near-infrared spectroscopy and hyperspectral imaging as novel characterization methods for anaerobic gut fungi

Abstract Neocallimastigomycota are a phylum of anaerobic gut fungi (AGF) that inhabit the gastrointestinal tract of herbivores and play a pivotal role in plant matter degradation. Their identification and characterization with marker gene regions has long been hampered due to the high inter- and intraspecies length variability in the commonly used fungal marker gene region internal transcribed spacer (ITS). While recent research has improved methodology (i.e. switch to LSU D2 as marker region), molecular methods will always introduce bias through nucleic acid extraction or PCR amplification. Here, near-infrared spectroscopy (NIRS) and hyperspectral imaging (HSI) are introduced as two nucleic acid sequence-independent tools for the characterization and identification of AGF strains. We present a proof-of-concept for both, achieving an independent prediction accuracy of above 95% for models based on discriminant analysis trained with samples of three different genera. We further demonstrated the robustness of the NIRS model by testing it on cultures of different growth times. Overall, NIRS provides a simple, reliable, and nondestructive approach for AGF classification, independent of molecular approaches. The HSI method provides further advantages by requiring less biomass and adding spatial information, a valuable feature if this method is extended to mixed cultures or environmental samples in the future.

Dilated Spectral–Spatial Gaussian Transformer Net for Hyperspectral Image Classification

In recent years, deep learning-based classification methods for hyperspectral images (HSIs) have gained widespread popularity in fields such as agriculture, environmental monitoring, and geological exploration. This is owing to their ability to automatically extract features and deliver outstanding performance. This study provides a new Dilated Spectral–Spatial Gaussian Transformer Net (DSSGT) model. The DSSGT model incorporates dilated convolutions as shallow feature extraction units, which allows for an expanded receptive field while maintaining computational efficiency. We integrated transformer architecture to effectively capture feature relationships and generate deep fusion features, thereby enhancing classification accuracy. We used consecutive dilated convolutional layers to extract joint low-level spectral–spatial features. We then introduced Gaussian Weighted Pixel Embedding blocks, which leverage Gaussian weight matrices to transform the joint features into pixel-level vectors. By combining the features of each pixel with its neighbouring pixels, we obtained pixel-level representations that are more expressive and context-aware. The transformed vector matrix was fed into the transformer encoder module, enabling the capture of global dependencies within the input data and generating higher-level fusion features with improved expressiveness and discriminability. We evaluated the proposed DSSGT model using five hyperspectral image datasets through comparative experiments. Our results demonstrate the superior performance of our approach compared to those of current state-of-the-art methods, providing compelling evidence of the DSSGT model’s effectiveness.

SBHSR: single-band super-resolution method for hyperspectral images based on blind degradation and fusion of auxiliary band

Despite natural image super-resolution (SR) methods have achieved great success, super-resolution methods for hyperspectral image (HSI) with rich spectral features are still a very challenging task. Furthermore, due to the diversity of HSI captured by different cameras, their degradation conditions vary. Currently available HSI SR methods are mainly based on fixed degradation models, and their performance is severely affected when the actual degradation does not match the assumed degradation model. To address this issue, this paper proposes a single-band hyperspectral image super-resolution (SBHSR) method for HSI. This method assumes that the degradation Gaussian blur kernel parameters of the HSI are unknown, and adapts to various degradation conditions by conducting blind super-resolution on the image. It also combines another band with better spatial structure as an auxiliary band for feature fusion. To address the spectral differences between the super-resolution band and the auxiliary band, we use the Spatial Adaptation Module (SAM) to map the feature distribution, ensuring consistent spectral brightness between the auxiliary band and the super-resolution band. Due to the use of blind degradation super-resolution method, our method is more robust compared to previous HSI SR methods. Additionally, as it is specifically designed for single-band hyperspectral image super-resolution, it has the advantage of being faster and more efficient. Experimental validation on two hyperspectral datasets demonstrates the superiority of our method, as it improves the spatial resolution of the image while preserving the spectral features, and performs better than existing blind super-resolution methods.

Impact of Storage on Nondestructive Detectability of Codling Moth Infestation in Apples

Highlights Codling moth responds differently to low temperature conditions during the supply chain. The length of storage and temperature determine the sensitivity of the codling moth to detection. Hyperspectral imaging sensor detectability of codling moth increased with temperature. Abstract. Different conditions during cold storage of codling moth (CM)-infested apples lead to different infestation levels, which affect overall product quality. In this study, the effects of postharvest storage duration (up to 20 weeks) and temperature (0°C, 4°C, and 10°C) on the detectability of CM-infested apples were investigated using the near-infrared (NIR) hyperspectral imaging (HSI) method (900–1700 nm). Fresh organic Gala apples were obtained directly from a commercial market and stored in a controlled environmental chamber at three temperatures for 20 weeks in two groups: control and CM-infested samples. Every four weeks, NIR hyperspectral images in reflectance mode were acquired directly for each set of samples. Machine learning models for the classification of CM-infested apples were developed based on the HSI data. The results revealed that storage duration and temperature had a significant effect on the performance of the classification models in the detection of CM-infested and control apples. Overall, the best classification rates were obtained for apples stored for 16 weeks, with accuracies of 97%, 94%, and 100% at storage temperatures of 0°C, 4°C, and 10°C, respectively. This study is critical for determining the effectiveness of HSI as a nondestructive method for sorting apples into classes based on CM infestation when stored under different conditions and duration, as in this study. Keywords: Apples, Codling moth, Detectability, Machine learning, Nondestructive method.

Self-Supervised Feature Learning Method for Hyperspectral Images Based on Mixed Convolutional Networks

Self-Supervised Feature Learning Method for Hyperspectral Images Based on Mixed Convolutional Networks

Enhancing anthracnose detection in mango at early stages using hyperspectral imaging and machine learning

Anthracnose, caused by Colletotrichum sp. infections, poses a significant threat to mango production worldwide, resulting in substantial losses. This devastating disease is challenging to detect and control, primarily due to its ability to spread rapidly. The methods currently used to control anthracnose are primarily corrective, relying on disease detection in the late stages when the infection becomes visible. Hence, there is a need for tools to detect the infection at early stages, before symptoms appear. Hyperspectral imaging systems are promising for developing non-destructive solutions to assess and detect external and internal damage in fruit, including decay caused by anthracnose. These advanced imaging systems make early detection possible before the symptoms are visible, allowing for timely intervention and potentially more effective disease control. This work aims to evaluate the possibility of early detection of anthracnose in two mango cultivars using hyperspectral imaging and machine learning methods. Secondly, to establish correlations between specific wavelengths and the physicochemical symptoms associated with anthracnose. Lastly, to develop a robust model for the spectral detection of anthracnose on mango fruit. Mangoes were inoculated with spores of Colletotrichum gloeosporioides. Hyperspectral images of control and infected fruit were captured in the 450–970 nm spectral range. Five machine-learning models were used to obtain the method that best fits the spectral data. The best model achieved an accuracy = 0.961, recall = 0.961, specificity = 0.992, F1 = 0.961 and Matthews correlation coefficient (MCC) = 0.953 for 'Keitt', and an accuracy = 0.975, recall = 0.976, specificity = 0.995, F1 = 0.975 and MCC = 0.971 for 'Osteen', showing the feasibility to detect early anthracnose infection in mango fruit within 48 h after pathogen inoculation.

Weighted spectral correlation angle target detection method for land-based hyperspectral imaging

In land-based spectral imaging, the spectra of ground objects are inevitably affected by the imaging conditions (weather conditions, atmospheric conditions, light conditions, zenith and azimuth angle conditions) and spatial distribution of targets, leading to uncertainties featured by “same object different spectrum”. That is, the spectrum of a ground object may change within a certain range under different imaging conditions. Traditional target detection (TD) methods are mainly based on similarity measurements and do not fully account for the spectral uncertainties. These detection methods are prone to false detections or missed detections. Therefore, reducing the impact of spectral uncertainties on TD is an important research topic in hyperspectral imaging. In this paper, we first review traditional TD methods and compare their principles and characteristics. It is found that the spectral correlation angle (SCA) method has good adaptability in land-based imaging. The shortcoming of the SCA method that it cannot reflect the local spectrum characteristics, is also analyzed. As the effect of spectral uncertainties cannot be completely overcome by the SCA method, a new similarity measurement method, the weighted spectral correlation angle (WSCA) method, is proposed. It can reduce the influence of spectral uncertainties on TD by increasing the weight of particular bands. Finally, we use two sets of experiments to analyze the effect of the WSCA method on TD. Its performance in overcoming spectral uncertainties caused by variations in imaging conditions or uneven spatial distributions of targets is tested. The results show that the WSCA method can effectively reduce the influence of spectral uncertainties and obtain a good detection result.Graphical

Removal of environmental influences for estimating soil texture fractions based on ZY1 satellite hyperspectral images

Soil texture is one of the important factors affecting the physical and chemical properties of soil, and is generally divided into sand, silt, and clay. Understanding its spatial distribution is crucial for agricultural and environmental decision-making. Estimation of soil texture fractions (STFs) can be achieved through visible, near-infrared, and shortwave infrared (VNIR-SWIR) spectra, making hyperspectral imaging technology particularly promising for low-cost and large-scale monitoring of STFs. However, the process of image acquisition is inevitably affected by soil environmental factors. In this study, a soil texture fractions estimation method for hyperspectral images was proposed based on Orthogonal Signal Correction (OSC). Firstly, the OSC algorithm was applied to eliminate the impact of soil environmental factors on the image spectra. Then, the spectral feature was enhanced by Fractional Order Derivative (FOD). Subsequently, a Partial Least Squares Regression (PLSR) model was established based on the optimal bands selected by Competitive Adaptive Reweighted Sampling (CARS). A total of 107 soil samples were obtained from the agricultural area in Lishu County, Jilin Province, China. The hyperspectral images of the ZY1-02D and ZY1-02E satellites were acquired simultaneously in this area. The results showed that the OSC algorithm effectively reduced the influence of soil environmental factors on image spectra, and significantly improved the accuracy of STFs estimation. After applying OSC to the original image spectra, the estimation accuracies (R2) for sand, silt, and clay content were improved from 0.52, 0.51, 0.50 to 0.81, 0.65, and 0.70 respectively. Compared to External Parameter Orthogonalization (EPO) correction, the model accuracy and stability were higher with OSC correction. The selected results from CARS indicated that the characteristic spectral bands of soil texture are concentrated around 500 nm, 750 nm, 900 nm, and 2200 nm. Further analysis indicates that the distribution trends of the soil texture fractions in this study were consistent with previous studies, demonstrating the good applicability of the estimation method.

S2EFT: Spectral-Spatial-Elevation Fusion Transformer for hyperspectral image and LiDAR classification

With the innovation of sensors, communication, computer and other technologies, remote sensing (RS) imaging methods show a diversification trend. At present, relying on a variety of earth observation platforms, more multi-source RS data can be obtained, which overcomes the inadequacy and incompleteness of the new earth observation information of single source RS image. In this paper, a classification method of hyperspectral images (HSI) and LiDAR data based on spectral-spatial-elevation fusion Transformer (S2EFT) framework is proposed. The Transformer framework is introduced into the task of multi-source RS image classification. Two simple but effective modules, spatial information recognition module and sliding group spectral embedding module, are added to the proposed framework, and the Patch form commonly used in traditional convolutional neural networks (CNNs) is used as the input. It effectively solves the shortcomings of Transformer’s insufficient attention to local information, reduces redundant spatial information, enhances the transmission of information, and fully integrates multidimensional information of pixels in same position. Experiments on three real data sets are compared with existing methods, and the best results are obtained by the proposed methods. The source code can be downloaded from https://github.com/SYFYN0317/S2EFT.

Comparison of the Efficiency of Hyperspectral and Pulse Amplitude Modulation Imaging Methods in Pre-Symptomatic Virus Detection in Tobacco Plants.

Early detection of pathogens can significantly reduce yield losses and improve the quality of agricultural products. This study compares the efficiency of hyperspectral (HS) imaging and pulse amplitude modulation (PAM) fluorometry to detect pathogens in plants. Reflectance spectra, normalized indices, and fluorescence parameters were studied in healthy and infected areas of leaves. Potato virus X with GFP fluorescent protein was used to assess the spread of infection throughout the plant. The study found that infection increased the reflectance of leaves in certain wavelength ranges. Analysis of the normalized reflectance indices (NRIs) revealed indices that were sensitive and insensitive to infection. NRI700/850 was optimal for virus detection; significant differences were detected on the 4th day after the virus arrived in the leaf. Maximum (Fv/Fm) and effective quantum yields of photosystem II (ΦPSII) and non-photochemical fluorescence quenching (NPQ) were almost unchanged at the early stage of infection. ΦPSII and NPQ in the transition state (a short time after actinic light was switched on) showed high sensitivity to infection. The higher sensitivity of PAM compared to HS imaging may be due to the possibility of assessing the physiological changes earlier than changes in leaf structure.

Compressive hyperspectral imaging based on Images Structure Similarity and deep image prior

In this paper, the structural similarity between RGB image and spectral image is studied, and a non-iterative Images Structure Similarity(ISS) method for fast reconstruction of spectral image is proposed. At the same time, the input of the Deep Image Prior (DIP) method is optimized for the first time by using the initial spectral data reconstructed by ISS. It raises the starting value of the iteration. Specifically, we take the RGB image data as the base of the spectral data, and solve the base coefficient by the least square method to quickly estimate the initial hyperspectral image. The Gaussian noise data is replaced by the estimated initial spectral data to constrain the solution space of the network and reduce the number of iterations. Finally, the structural similarity between RGB image and spectral image is used. The RGB three-channel graph is used to filter the iterative results to improve the reconstruction quality. Experimental results show that compared with other hyperspectral imaging methods, the proposed method can improve the quality of reconstruction in both spectral resolution and spatial resolution. In addition, compared with other methods based on Deep Image Prior (DIP), our improvement greatly reduces the reconstruction time and is more suitable for actual snapshot spectral imaging.

Zero-Shot Hyperspectral Sharpening.

Fusing hyperspectral images (HSIs) with multispectral images (MSIs) of higher spatial resolution has become an effective way to sharpen HSIs. Recently, deep convolutional neural networks (CNNs) have achieved promising fusion performance. However, these methods often suffer from the lack of training data and limited generalization ability. To address the above problems, we present a zero-shot learning (ZSL) method for HSI sharpening. Specifically, we first propose a novel method to quantitatively estimate the spectral and spatial responses of imaging sensors with high accuracy. In the training procedure, we spatially subsample the MSI and HSI based on the estimated spatial response and use the downsampled HSI and MSI to infer the original HSI. In this way, we can not only exploit the inherent information in the HSI and MSI, but the trained CNN can also be well generalized to the test data. In addition, we take the dimension reduction on the HSI, which reduces the model size and storage usage without sacrificing fusion accuracy. Furthermore, we design an imaging model-based loss function for CNN, which further boosts the fusion performance. The experimental results show the significantly high efficiency and accuracy of our approach.

Radial grid reflectance correction for hyperspectral images of fruits with rounded surfaces

Over the last two decades, the viability of hyperspectral imaging (HSI) technology as a tool for determining food quality has been established. This technology has been successfully combined with chemometric tools and image-processing techniques to enhance its effectiveness. Despite the advances in imaging technology, significant challenges related to non-uniformity caused by object geometry still need to be addressed. The inappropriate processing of non-uniform food products can significantly affect subsequent calculations, including accurate measurements and precise analysis. A popular method used to face up non-uniformity is the Lambertian surface correction, which assumes spherical surfaces of food products; however, it presents some issues with non-spherical shapes. This work proposes a generalized reflectance correction method for HSI of rounded fruits that are not necessarily spherical. The method is based on the correlation between the pixel’s position and its reflectance. For proof of concept, a case study was designed considering the HSI of mango fruits (Mangifera indica L) Kent variety. For comparison, each image was corrected using both the Lambertian surface correction and the proposed method. Results show a strong correlation between the standard deviation of the spectral profiles of images and the roundness of the fruit, which is more evident in the case of the proposed radial grid correction method. Hence, the proposed radial correction method produced a more uniform texture in corrected hyperspectral images.

Graph Dual Adversarial Network for Hyperspectral Image Classification

An end-to-end unsupervised domain adaptation method for hyperspectral image (HSI) classification based on graph dual adversarial network is proposed in this paper. First, in order to extract the domain invariant features of the source and target domains, the rich spectral information and spatial position of HSI are used to construct a spectral-spatial nearest neighbor graph, which is input into the graph convolutional network. Then, a prototype adversarial strategy is proposed, which uses the labeled data of source domain to reliably calculate the feature prototypes of different classes. Through the prototype adversarial strategy, the distances between different prototypes are appropriately extended, so that the clusters of different classes are far away from their respective decision boundaries, and the discriminability of features are also enhanced. The dual adversarial strategy is composed of the prototype adversarial strategy and domain adversarial strategy. It is worth noting that the dual adversarial strategy does not require feature extractor and discriminator to work in turn, which can be implemented through a gradient reversal layer. Finally, on the basis of adapting the overall features of both domains via the domain adversarial strategy, the source and target domain features are further adapted by minimizing the correlation alignment loss of each class of samples. Experimental results on two real HSI datasets of Botswana and Kennedy Space Center show the effectiveness of our proposed method.

Redundant compressed single-pixel hyperspectral imaging system

As single-pixel imaging technology is highly sensitive and low-cost, a large number of hyperspectral imaging methods based on single-pixel have been proposed. Although the image of each band can be reconstructed independently, the parameters that need to be computed for reconstructing hyperspectral images will increase in parallel with the number of bands. Therefore, in this paper, we propose a redundant compressed single-pixel hyperspectral imaging system (RCSHI), which takes full advantage of the spectral redundancy of hyperspectral images by jointly reconstructing images from different bands and compressing the spectral dimension of the data. RCSHI reduces the amount of redundant hyperspectral information stacking and the number of parameters that need to be computed to reconstruct hyperspectral images. In addition, we use progressive fusion enhancement to recover the compressed spectral dimension and enhance the spatial dimension of the images. Experimental results show that RCSHI is able to acquire hyperspectral data with high spatial and spectral resolution and performs well on various datasets and actual scenarios, especially at lower compressed sampling rates.

Effect of denoising methods for hyperspectral images classification: DnCNN, NGM, CSF, BM3D and Wiener

Hyperspectral images are widely used for land use/cover analysis in remote sensing due to their rich spectral information. However, these data often suffer from noise caused by various factors such as random and systematic errors, making them less useful for end-users. In this study, denoising methods (i.e., DnCNN, NGM, CSF, BM3D, and Wiener) for hyperspectral images were compared using the Pavia University hyperspectral dataset with four different noise types: Gaussian, Salt & Pepper, Poisson, and Speckle. After denoising, the k-nearest neighbor method was used to classify the image, and statistical and visual performance comparisons were performed on the classified data. Six performance metrics -Accuracy, Sensitivity, Specificity, Precision, F-Score, and G-Mean- were employed to compare the outcomes qualitatively. The findings demonstrate that DnCNN and BM3D have the best outcome performance for all four noise types. Due to their lack of sensitivity and specificity, the CSF and Wiener approaches had low performance for particular noise sources. For all noise types, the NGM approach had the worst results. The validated instruments not provide effective results when it came to denoising Salt & Pepper noise, but they managed to produce outstanding results when it came to denoising Poisson noise. In order to enhance the quality and usability of hyperspectral images for land use/cover analysis, this study emphasizes the significance of choosing an effective denoising technique.

Oxygen saturation mapping during reconstructive surgery of human forehead flaps with hyperspectral imaging and spectral unmixing

BackgroundOptical spectroscopy is commonly used clinically to monitor oxygen saturation in tissue. The most commonly employed technique is pulse oximetry, which provides a point measurement of the arterial oxygen saturation and is commonly used for monitoring systemic hemodynamics, e.g. during anesthesia. Hyperspectral imaging (HSI) is an emerging technology that enables spatially resolved mapping of oxygen saturation in tissue (sO2), but needs to be further developed before implemented in clinical practice. The aim of this study is to demonstrate the applicability of HSI for mapping the sO2 in reconstructive surgery and demonstrate how spectral analysis can be used to obtain clinically relevant sO2 values. MethodsSpatial scanning HSI was performed on cutaneous forehead flaps, raised as part of a direct brow lift, in eight patients. Pixel-by-pixel spectral analysis, accounting for the absorption from multiple chromophores, was performed and compared to previous analysis techniques to assess sO2. ResultsSpectral unmixing using a broad spectral range, and accounting for the absorption of melanin, fat, collagen, and water, provided a more clinically relevant estimate of sO2 than conventional techniques, where typically only spectral features associated with absorption of oxygenated (HbO2) and deoxygenated (HbR) hemoglobin are considered. We demonstrate its clinical applicability by generating sO2 maps of partially excised forehead flaps showed a gradual decrease in sO2 along the length of the flap from 95 % at the flap base to 85 % at the flap tip. After being fully excised, sO2 in the entire flap decreased to 50 % within a few minutes. ConclusionsThe results demonstrate the capability of sO2 mapping in reconstructive surgery in patients using HSI. Spectral unmixing, accounting for multiple chromophores, provides sO2 values that are in accordance with physiological expectations in patients with normal functioning microvascularization. Our results suggest that HSI methods that yield reliable spectra are to be preferred, so that the analysis can produce results that are of clinical relevance.

Scanning-based compressive hyperspectral imaging via spectral-coded illumination.

In this Letter, we present a novel, to the best of our knowledge, scanning-based compressive hyperspectral imaging method via spectral-coded illumination. We achieve efficient and flexible spectral modulation by spectral coding of a dispersive light source while spatial information is obtained by point-wise scanning, which can be applied to optical scanning imaging systems such as lidar. In addition, we propose a new tensor-based joint hyperspectral image reconstruction algorithm that considers spectral correlation and spatial self-similarity to recover three-dimensional hyperspectral data from compressive sampled data. Both simulated and real experiments show that our method has superior performance in visual quality and quantitative analysis.

Spectral characterization of fouled railroad ballast using hyperspectral imaging

The presence of fouling contamination in railroad ballast is a growing concern that causes deterioration, poor drainage functionality, and early failure of railroad components. Evaluation and characterization of fouled ballast are crucial for determining and monitoring fouling content (FC) and moisture content (MC) in ballast. However, human inspection, destructive testing, field sampling, and most available nondestructive evaluation methods (NDE) for inspection purposes pose safety risks, undermine the stability of the railroad substructure, or provide limited information about the condition of the ballast. This study presents the results of the spectral characterization of laboratory ballast samples using hyperspectral imaging (HSI) sensor. Nineteen (19) clean, dry-fouled, and wet-fouled samples were prepared and evaluated using HSI. The reflectance of ballast samples with different levels of MC and FC was determined in Visible-Near infrared (VNIR) and Near-infrared (NIR) wavelength range from 400 nm to 1700 nm. Results showed that; the presence of trapped water in ballast pores and free water in ballast voids reduced the reflectance; an increase in the dry fouling content increased the reflectance, and MC in wet-fouled ballast showed a nonlinear relationship with the reflectance or absorbance property. The presence of water was prominent in wavelengths of 1350 nm to 1550 nm. A second-order polynomial fit was adopted for the absorbance and reflectance features plotted against the MC, and the coefficient of determination (R2) values were determined. The R2 for the absorbance and reflectance plots between 900 nm and 1700 nm were 0.7747 and 0.8336 in the sample with 5% FC and 0.8790 and 0.9278 with 15% FC, respectively. For the 1350 nm to 1500 nm wavelength range, the R2 increased by 27% and 13%, respectively, for 5% FC and then 5% and 0%, respectively, for 15% FC. This study shows that absorbance features are more prominent than the reflectance features for determining MC in fouled ballast, with high prospects of detecting, monitoring, and quantifying MC and FC contamination using the HSI method.

Hyperspectral Image Classification Based on Fusion of Convolutional Neural Network and Graph Network

Convolutional neural networks (CNNs) have attracted significant attention as a commonly used method for hyperspectral image (HSI) classification in recent years; however, CNNs can only be applied to Euclidean data and have difficulties in dealing with relationships due to their limitations of local feature extraction. Each pixel of a hyperspectral image contains a set of spectral bands that are correlated and interact with each other, and the methods used to process Euclidean data cannot effectively obtain these correlations. In contrast, the graph convolutional network (GCN) can be used in non-Euclidean data but usually leads to over-smoothing and ignores local detail features due to the need for superpixel segmentation processing to reduce computational effort. To overcome the above problems, we constructed a fusion network based on the GCN and CNN which contains two branches: a graph convolutional network based on superpixel segmentation and a convolutional network with an added attention mechanism. The graph convolutional branch can extract the structural features and capture the relationships between the nodes, and the convolutional branch can extract detailed features in the local fine region. Owing to the fact that the features extracted from the two branches are different, the classification performance can be improved by fusing the complementary features extracted from the two branches. To validate the proposed algorithm, experiments were conducted on three widely used datasets, namely Indian Pines, Pavia University, and Salinas. An overall accuracy of 98.78% was obtained in the Indian Pines dataset, and overall accuracies of 98.99% and 98.69% were obtained in the other two datasets. The results show that the proposed fusion network can obtain richer features and achieve a high classification accuracy.