Information Of Hyperspectral Image Research Articles

Hyperspectral Image Classification with IFormer Network Feature Extraction

Convolutional neural networks (CNNs) are widely used for hyperspectral image (HSI) classification due to their better ability to model the local details of HSI. However, CNNs tends to ignore the global information of HSI, and thus lack the ability to establish remote dependencies, which leads to computational cost consumption and remains challenging. To address this problem, we propose an end-to-end Inception Transformer network (IFormer) that can efficiently generate rich feature maps from HSI data and extract high- and low-frequency information from the feature maps. First, spectral features are extracted using batch normalization (BN) and 1D-CNN, while the Ghost Module generates more feature maps via low-cost operations to fully exploit the intrinsic information in HSI features, thus improving the computational speed. Second, the feature maps are transferred to Inception Transformer through a channel splitting mechanism, which effectively learns the combined features of high- and low-frequency information in the feature maps and allows for the flexible modeling of discriminative information scattered in different frequency ranges. Finally, the HSI features are classified via pooling and linear layers. The IFormer algorithm is compared with other mainstream algorithms in experiments on four publicly available hyperspectral datasets, and the results demonstrate that the proposed method algorithm is significantly competitive among the HSI classification algorithms.

Cloud Contaminated Multispectral Remote Sensing Image Enhancement Algorithm Based on MobileNet

Multispectral remote sensing images have shown unique advantages in many fields, including military and civilian use. Facing the difficulty in processing cloud contaminated remote sensing images, this paper proposes a multispectral remote sensing image enhancement algorithm. A model is constructed from the aspects of cloud detection and image enhancement. In the cloud detection stage, clouds are divided into thick clouds and thin clouds according to the cloud transmitability in multi-spectral images, and a multi-layer cloud detection model is established. From the perspective of traditional image processing, a bimodal pre-detection algorithm is constructed to achieve thick cloud extraction. From the perspective of deep learning, the MobileNet algorithm structure is improved to achieve thin cloud extraction. Faced with the problem of insufficient training samples, a self-supervised network is constructed to achieve training, so as to meet the requirements of high precision and high efficiency cloud detection under the condition of small samples. In the image enhancement stage, the area where the ground objects are located is determined first. Then, from the perspective of compressed sensing, the signal is analyzed from the perspective of time and frequency domains. Specifically, the inter-frame information of hyperspectral images is analyzed to construct a sparse representation model based on the principle of compressed sensing. Finally, image enhancement is achieved. The experimental comparison between our algorithm and other algorithms shows that the average Area Overlap Measure (AOM) of the proposed algorithm reaches 0.83 and the Average Gradient (AG) of the proposed algorithm reaches 12.7, which is better than the other seven algorithms by average AG 2.

Quantitative study on impact damage of yellow peach based on hyperspectral image information combined with spectral information

Yellow peaches are graded based on their degree of impact damage to reduce economic losses, and the values of mechanical parameters can be used to characterize the degree of impact damage of them. In this paper, the combined hyperspectral image features and spectral features method was proposed to predict the values of mechanical parameters of yellow peaches. A collision device based on the pendulum principle was built to acquire the mechanical parameters which were maximum force, absorbed energy and average pressure. A hyperspectral imaging system was used to acquire the images of the bruised yellow peaches and the spectral information and color characteristics of images were extracted, respectively. The damage areas of yellow peaches were obtained based on the combined principal component analysis (PCA) and threshold segmentation method. Texture features of the characteristic wavelength images were extracted based on the grey-level co-generation matrix (GLCM). The PLSR models of mechanical parameters were built based on spectral information, image information, and spectral information combined with image information, respectively. The results show that the PLSR models based on spectral information combined with image information are best in prediction the values of damage area and maximum force. However, the PLSR models have not achieved satisfactory performance in prediction the values of absorbed energy and average pressure. To improve the performance of PLSR models of absorbed energy and average pressure, the feature wavelengths were selected by CARS, and the correlation analysis was used between image information and mechanical parameters. The PLSR models were built based on feature spectral information, image feature information, and feature spectral information combined with image feature information, respectively. The results show that the accuracy of all mechanical parameters is significantly improved by the PLSR models based on the feature spectral information combined with the image feature information. The prediction set correlation coefficient (RP) and the root mean square error (RMSEP) of the damaged area, absorbed energy, maximum force and average pressure are 0.928 and 86.632 mm2, 0.826 and 1.469 J, 0.924 and 61.765 N, 0.815 and 0.050 MPa, respectively. The results of the study not only provide a theoretical basis for postharvest grading of fruit, but also provide a reference for analyzing the mechanical properties of fruit.

An Efficient Hyperspectral Image Classification Method: Inter-Class Difference Correction and Spatial Spectral Redundancy Removal

Existing remote sensing images of ground objects are difficult to annotate, and building a hyperspectral dataset requires huge resources. To tackle these problems, this paper proposes a new method with low requirements for the scale of the dataset that involves correcting the inter-class differences of hyperspectral images and eliminating the redundant information of spatial–spectral features. Firstly, the algorithm introduces the spatial information of hyperspectral images into the classification task through the superpixel definition based on the entropy rate, which not only reduces the spatial redundancy information of hyperspectral images but also uses the similarity of spatial neighbors to increase the differences between classes and alleviate the differences within classes. Secondly, under the theoretical guidance of similar spectral fluctuation trends of similar objects, a feature extraction method based on fast Fourier transform is proposed to alleviate spectral feature redundancy and further eliminate the inter-class differences of the algorithm. Finally, to verify the improvement effect of the proposed idea on the traditional classification method, the idea was applied to the traditional SVM algorithm, and experiments were carried out based on the PaviaU and Indian Pines hyperspectral datasets. The simulation results show that the proposed (ERS–FFT–SVM) algorithm shows a significant improvement in classification accuracy when compared with the traditional classification algorithm and is able to perform the small sample classification of hyperspectral images.

Attention U-shaped network for hyperspectral image classification

To fully use the contextual information of hyperspectral images (HSIs), we propose a U-shaped network model combined with attention mechanism to achieve image-level HSI classification. First, the entire HSI is input into the network for end-to-end training, and the classification results of the entire scene are directly output. Then, the context information is used to improve the classification accuracy, while reducing many redundant calculations. Second, to improve the classification accuracy, considering two dimensions (i.e., space and channel), a hybrid attention module, mixing spatial and channel, is designed. Third, three datasets of the University of Pavia, Indian Pines, and Salinas are selected for the classification experiments. The experimental results show that, compared with other methods, the proposed method can obtain higher classification accuracy, and its training and testing efficiency is higher.

Multi-granularity vision transformer via semantic token for hyperspectral image classification

ABSTRACT The superior local context modelling capability of convolutional neural networks (CNNs) in representing features allows greatly enhanced performance in hyperspectral image (HSI) classification tasks by CNN-based methods. However, most of these methods suffer from a restricted receptive field and poor performance in the continuous data domain. To address these issues, we propose a multi-granularity vision transformer via semantic token (MSTViT) for HSI classification, which differs from the existing transformer view by modelling the HSI classification tasks as word embedding problems. Specifically, the MSTViT model extracts multi-level semantic features by a ladder feature extractor and applies a multi-granularity patch embedding module to embed these features simultaneously as different-scale tokens. Moreover, different-granularity tokens are fed to the vision transformer to capture the long-distance dependencies among the different tokens. A depth-wise separable convolution multi-layer perceptron is used to assist the attention mechanism for further excavation of the deep information of HSI. Finally, the performance of HSI classification is improved by fusing the coarse- and fine-granularity representations to generate stronger features. Experimental results on four standard datasets verify the marked improvement of the MSTViT over state-of-the-art CNN and transformer structures. The code of this work is available at https://github.com/zhaolin6/MSTViT for the sake of reproducibility.

Spectral-spatial classification of hyperspectral images based on non-linear principal component analysis and deep learning models

ABSTRACT Hyperspectral sensors, due to the acquisition of a large number of spectral bands, always have particular importance in monitoring the phenomena of the earth’s surface. The classification of hyperspectral images is the most crucial method of processing hyperspectral data, and many efforts have been made to increase its accuracy. In recent years, convolutional neural networks (CNNs) and spatial features have been a big part of how well hyperspectral images can be classified. CNNs, due to the automatic generation of features and reduction of parameters, compared to multilayer perceptron neural networks by sharing parameters in each layer, have received much attention from researchers in the field of pattern recognition. In previous research, much attention has not been paid to the simultaneous use of spatial feature extraction methods in CNNs. For this reason, in this paper, a new CNNs architecture has been introduced to classify hyperspectral images, which uses the spectral-spatial vector obtained from the morphological profiles method as the network’s input. Morphological profiles, which include a set of opening and closing filters, are generally applied to the components of the image that contain the most information. In this research, non-linear principal component analysis (non-linear PCA) implemented by neural networks was used to summarize hyperspectral image information. Therefore, the main innovation of this paper is to provide a three-stage framework for applying deep learning. The first step is to reduce the dimensions of the hyperspectral image using non-linear PCA, the second stage is to extract spatial features using the morphological profiles method, and the third step is to prepare the inputs and design the CNNs architecture. The proposed method was evaluated on three benchmark hyperspectral images of Pavia, Berlin and Telops. The results of the experiments show the superiority of the proposed method by 13, 11 and 17% in the overall accuracy parameter compared to the support vector machines (SVM) classification algorithm in the Pavia, Berlin and Telops images, respectively.

ES2FL: Ensemble Self-Supervised Feature Learning for Small Sample Classification of Hyperspectral Images

Classification with a few labeled samples has always been a longstanding problem in the field of hyperspectral image (HSI) processing and analysis. Aiming at the small sample characteristics of HSI classification, a novel ensemble self-supervised feature-learning (ES2FL) method is proposed in this paper. The proposed method can automatically learn deep features conducive to classification without any annotation information, significantly reducing the dependence of deep-learning models on massive labeled samples. Firstly, to utilize the spatial–spectral information in HSIs more fully and effectively, EfficientNet-B0 is introduced and used as the backbone to model input samples. Then, through constraining the cross-correlation matrix of different distortions of the same sample to the identity matrix, the designed model can extract the latent features of homogeneous samples gathering together and heterogeneous samples separating from each other in a self-supervised manner. In addition, two ensemble learning strategies, feature-level and view-level ensemble, are proposed to further improve the feature-learning ability and classification performance by jointly utilizing spatial contextual information at different scales and feature information at different bands. Finally, the concatenations of the learned features and the original spectral vectors are inputted into classifiers such as random forest or support vector machine to complete label prediction. Extensive experiments on three widely used HSI data sets show that the proposed ES2FL method can learn more discriminant deep features and achieve better classification performance than existing advanced methods in the case of small samples.

An effective global learning framework for hyperspectral image classification based on encoder-decoder architecture

ABSTRACT Most deep learning methods in hyperspectral image (HSI) classification use local learning methods, where overlapping areas between pixels can lead to spatial redundancy and higher computational cost. This paper proposes an efficient global learning (EGL) framework for HSI classification. The EGL framework was composed of universal global random stratification (UGSS) sampling strategy and a classification model BrsNet. The UGSS sampling strategy was used to solve the problem of insufficient gradient variance resulted from limited training samples. To fully extract and explore the most distinguishing feature representation, we used the modified linear bottleneck structure with spectral attention as a part of the BrsNet network to extract spectral spatial information. As a type of spectral attention, the shuffle spectral attention module screened important spectral features from the rich spectral information of HSI to improve the classification accuracy of the model. Meanwhile, we also designed a double branch structure in BrsNet that extracted more abundant spatial information from local and global perspectives to increase the performance of our classification framework. Experiments were conducted on three famous datasets, IP, PU, and SA. Compared with other classification methods, our proposed method produced competitive results in training time, while having a greater advantage in test time.

A Hyperspectral Image Classification Method Based on Adaptive Spectral Spatial Kernel Combined with Improved Vision Transformer

In recent years, methods based on deep convolutional neural networks (CNNs) have dominated the classification task of hyperspectral images. Although CNN-based HSI classification methods have the advantages of spatial feature extraction, HSI images are characterized by approximately continuous spectral information, usually containing hundreds of spectral bands. CNN cannot mine and represent the sequence properties of spectral features well, and the transformer model of attention mechanism proves its advantages in processing sequence data. This study proposes a new spectral spatial kernel combined with the improved Vision Transformer (ViT) to jointly extract spatial spectral features to complete classification task. First, the hyperspectral data are dimensionally reduced by PCA; then, the shallow features are extracted with an spectral spatial kernel, and the extracted features are input into the improved ViT model. The improved ViT introduces a re-attention mechanism and a local mechanism based on the original ViT. The re-attention mechanism can increase the diversity of attention maps at different levels. The local mechanism is introduced into ViT to make full use of the local and global information of the data to improve the classification accuracy. Finally, a multi-layer perceptron is used to obtain the classification result. Among them, the Focal Loss function is used to increase the loss weight of small-class samples and difficult-to-classify samples in HSI data samples and reduce the loss weight of easy-to-classify samples, so that the network can learn more useful hyperspectral image information. In addition, using the Apollo optimizer to train the HSI classification model to better update and compute network parameters that affect model training and model output, thereby minimizing the loss function. We evaluated the classification performance of the proposed method on four different datasets, and achieved good classification results on urban land object classification, crop classification and mineral classification, respectively. Compared with the state-of-the-art backbone network, the method achieves a significant improvement and achieves very good classification accuracy.

Fast spatial-spectral random forests for thick cloud removal of hyperspectral images

The long-term existence and extensive coverage of thick clouds have caused an amount of missing information in hyperspectral images (HSIs), hindering their applications greatly. Recently, with the continuous launch of new satellites, the increasing availability of HSIs provides more opportunities for using temporal information in cloud removal of HSIs. In this paper, a fast spatial-spectral random forests (FSSRF) method was proposed for removing clouds of HSIs. FSSRF is developed based on the advanced spatial-spectral random forests (SSRF) method that was developed to handle multispectral images. FSSRF greatly improves the computing efficiency while ensuring the accuracy of reconstruction. Experimental results on both GF-5 and EO-1 HSIs show that FSSRF is a much faster version than the original SSRF that uses directly all known bands of the temporally neighboring HSI with spatially complete coverage, and the accuracies of the two versions are similar. Compared with subSSRF (i.e., SSRF using several spectrally adjacent known bands for each cloudy band) and the popular MNSPI method, FSSRF can produce more accurate predictions. The evaluation from spectral dimension shows that FSSRF can recover the spectral characteristics of cloud pixels more satisfactorily. FSSRF has great potential in real-time cloud removal of HSIs due to its obvious advantages in balancing computational efficiency and accuracy.

AM³Net: Adaptive Mutual-Learning-Based Multimodal Data Fusion Network

Multimodal data fusion, e.g., hyperspectral image (HSI) and light detection and ranging (LiDAR) data fusion, plays an important role in object recognition and classification tasks. However, existing methods pay little attention to the specificity of HSI spectral channels and the complementarity of HSI and LiDAR spatial information. In addition, the utilized feature extraction modules tend to consider the feature transmission processes among different modalities independently. Therefore, a new data fusion network named AM <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</sup> Net is proposed for multimodal data classification; it includes three parts. First, an involution operator slides over the input HSI’s spectral channels, which can independently measure the contribution rate of the spectral channel of each pixel to the spectral feature tensor construction. Furthermore, the spatial information of HSI and LiDAR data is integrated and excavated in an adaptively fused, modality-oriented manner. Second, a <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">spectral-spatial mutual-guided module</i> is designed for the feature collaborative transmission among spectral features and spatial information, which can increase the semantic relatedness connection through adaptive, multiscale, and mutual-learning transmission. Finally, the fused spatial-spectral features are embedded into a classification module to obtain the final results, which determines whether to continue updating the network weights. Experimental evaluations on HSI-LiDAR datasets indicate that AM <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</sup> Net possesses a better feature representation ability than the state-of-the-art methods. Additionally, AM <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</sup> Net still maintains considerable performance when its input is replaced with multispectral and synthetic aperture radar data. The result indicates that the proposed data fusion framework is compatible with diversified data types.

Deep spatial-spectral prior with an adaptive dual attention network for single-pixel hyperspectral reconstruction.

Recently, single-pixel imaging has shown great promise in developing cost-effective imaging systems, where coding and reconstruction are the keys to success. However, it also brings challenges in capturing hyperspectral information accurately and instantly. Many works have attempted to improve reconstruction performance in single-pixel hyperspectral imaging by applying various hand-crafted priors, leading to sub-optimal solutions. In this paper, we present the deep spatial-spectral prior with adaptive dual attention network for single-pixel hyperspectral reconstruction. Specifically, the spindle structure of the parameter sharing method is developed to integrate information across spatial and spectral dimensions of HSI, which can synergistically and efficiently extract global and local prior information of hyperspectral images from both shallow and deep layers. Particularly, a sequential adaptive dual attention block (SADAB), i.e., spatial attention and spectral attention, are devised to adaptively rescale informative features of spatial locations and spectral channels simultaneously, which can effectively boost the reconstruction accuracy. Experiment results on public HSI datasets demonstrate that the proposed method significantly outperforms the state-of-the-art algorithm in terms of reconstruction accuracy and speed.

Fast robust fuzzy clustering based on bipartite graph for hyper‐spectral image classification

Hyper-spectral image (HSI) clustering has become a hot spot in remote sensing image research. However, due to the large amount and high dimension of HSI data, some traditional image clustering methods are no longer suitable, and even lead to poor clustering performance and long computing time. Therefore, this paper proposes a novel clustering method for HSI to make up for the shortcomings of traditional clustering methods. Firstly, a fuzzy similarity matrix is constructed by using the bipartite graph to obtain low-dimensional hyper-spectral data, which can reduce the complexity and the operation time of the algorithm. Secondly, fuzzy membership mapping is carried out through the constructed bipartite graph, and a non-negative regularization term is added to limit ill-conditioned problems. Thirdly, this paper uses the Geman-McClure function to optimize Euclidean distance. Finally, in this paper, the spectral information of HSI is considered while the neighbourhood information of HSI is added to achieve the unity of space spectrum, which enhances the robustness of the algorithm and improves the clustering performance of the algorithm. Compared with existing HSI clustering algorithms, the proposed algorithm has higher clustering accuracy and efficiency.

SpaSSA: Superpixelwise Adaptive SSA for Unsupervised Spatial-Spectral Feature Extraction in Hyperspectral Image.

Singular spectral analysis (SSA) has recently been successfully applied to feature extraction in hyperspectral image (HSI), including conventional (1-D) SSA in spectral domain and 2-D SSA in spatial domain. However, there are some drawbacks, such as sensitivity to the window size, high computational complexity under a large window, and failing to extract joint spectral-spatial features. To tackle these issues, in this article, we propose superpixelwise adaptive SSA (SpaSSA), that is superpixelwise adaptive SSA for exploiting local spatial information of HSI. The extraction of local (instead of global) features, particularly in HSI, can be more effective for characterizing the objects within an image. In SpaSSA, conventional SSA and 2-D SSA are combined and adaptively applied to each superpixel derived from an oversegmented HSI. According to the size of the derived superpixels, either SSA or 2-D singular spectrum analysis (2D-SSA) is adaptively applied for feature extraction, where the embedding window in 2D-SSA is also adaptive to the size of the superpixel. Experimental results on the three datasets have shown that the proposed SpaSSA outperforms both SSA and 2D-SSA in terms of classification accuracy and computational complexity. By combining SpaSSA with the principal component analysis (SpaSSA-PCA), the accuracy of land-cover analysis can be further improved, outperforming several state-of-the-art approaches.

Triple-branch ternary-attention mechanism network with deformable 3D convolution for hyperspectral image classification

ABSTRACT In recent years, the classification of hyperspectral images (HSI) has received extensive research attention. As compared with traditional HSI classification, which only uses spectral information, it is found that spatial information is also essential in HSI classification. To effectively utilize the spectral and spatial information of HSI, this paper proposes a triple-branch ternary-attention mechanism network with deformable 3D convolution (D3DTBTA). In D3DTBTA, three branches, i.e. the spectral, spatial-X, and spatial-Y branches, are combined with the attention mechanism in three directions, which can better capture the vector features of three dimensions in HSI. Furthermore, considering the adaptation of scale and receptive field size in the convolution operation, our method uses deformable convolution to enable D3DTBTA to enhance feature extraction. Our experimental results show that the framework outperforms existing algorithms on four hyperspectral datasets, especially when the training samples are limited.

Hyperspectral image information fusion-based detection of soluble solids content in red globe grapes

• Red globe grape are widely loved for their full grain, firm flesh, sweetness and high nutrient content. SSC is one of the important internal qualities of the fruit and is an important indicator of fruit ripeness. • The traditional method for testing SSC of fruits is to conduct destructive detecting, which is tedious and time-consuming, and only samples can be detected, and the experimental samples are completely damaged after detecting and cannot be sold and consumed. • In this paper, the hyperspectral images and spectral information of the samples were collected by hyperspectral imaging technology, and a PLSR model of SSC of red globe grape based on the fusion of spectral information, image information and hyperspectral image information was built. • After comparative analysis, an optimal detection model of SSC of red globe grape was determined, which provides a new method for the non-destructive detection of SSC of red globe grape. The Soluble Solids Content (SSC) of red globe grapes is an important indicator of internal quality. In this paper, 360 red globe grapes in the growing stage were collected as samples and the spectral information and images of the samples were extracted. The Raw spectral (RAW) information was extracted using the one-time dimensionality reduction algorithm (GA, CARS, SPA, UVE) and the combined dimensionality reduction algorithm (CARS-SPA, UVE-SPA) to build the PLSR model of the spectral information. The grey-scale co-occurrence matrix of the image was extracted as the texture feature information of the image and combined with the color information of the image (R, G, B, H, S, V, L, a, b) to form 19 image features to build the PLSR model of the image information. Thus, the PLSR model based on the fusion of hyperspectral image information was built by fusing the spectra extracted with the successive projection algorithm (SPA) feature wavelength and the image information after dimensionality reduction by the principal component analysis algorithm (PCA). The results showed that if only the spectral information was used for modelling, the SPA algorithm effectively extracted the characteristic wavelengths of the red globe grapes of SSC spectral information and improved the prediction performance of the model. If only image information was used for modelling, the PCA algorithm effectively improved the detection performance of the model by dimensionality reduction, but the improved performance was limited. The correlation coefficients of the calibration set and prediction set of the PLSR model were 0.9775 and 0.9762, and the detection effect and stability of the model were greatly improved compared with those built unilaterally based on spectral information or image information, and a new non-destructive detection method was found for the detection about SSC of red globe grapes.

Hyperspectral Image Mixed Noise Removal Using a Subspace Projection Attention and Residual Channel Attention Network

Although the existing deep-learning-based hyperspectral image (HSI) denoising methods have achieved tremendous success, recovering high-quality HSIs in complex scenes that contain mixed noise is still challenging. Besides, these methods have not fully explored the local and global spatial–spectral information of HSIs. To address the above issues, a novel HSI mixed noise removal network called subspace projection attention and residual channel attention network (SPARCA-Net) is proposed. Specifically, we propose an orthogonal subspace projection attention (OSPA) module to adaptively learn to generate bases of the signal subspace and project the input into such space to remove noise. By leveraging the local and global spatial relations, OSPA is able to reconstruct the local structure of the feature maps more precisely. We further propose a residual channel attention (RCA) module to emphasize the interdependence between feature maps and exploit the global channel correlation of them, which could enhance the channel-wise adaptive learning. In addition, multiscale joint spatial–spectral input and residual learning strategies are employed to capture multiscale spatial–spectral features and reduce the degradation problem, respectively. Synthetic and real HSI data experiments demonstrated that the proposed HSI denoising network outperforms many of the advanced methods in both quantitative and qualitative assessments.

Pyramidal and conditional convolution attention network for hyperspectral image classification using limited training samples

ABSTRACT Hyperspectral image classification methods based on convolutional neural network (CNN) and attention mechanism have obtained satisfactory results. However, the currently applied convolution operations treat all inputs equally, and the multiscale information extraction of hyperspectral data is not sufficient. The used attention mechanism is also one-dimensional, two-dimensional, or a simple combination of them. Alongside these, the high-precision classification models rely on many training samples. Based on these, we propose a pyramidal and conditional convolution attention network (PCAN), which can fully and effectively capture the information of hyperspectral input data through convolution kernels with different weights and different scales. And for the first time, the attention of three-dimensional weights is applied to hyperspectral data, and satisfactory results are achieved on minimal training sets. Specifically, PCAN is composed of spectral branch and spatial branch. The spectral branch uses conditionally parameterized convolutions (CondConv) to customize the convolution kernels according to the input features. It directly generates three-dimensional weights to quantify the importance of the entire feature maps. The spatial branch uses the pyramidal convolution (PyConv), which includes convolution kernels of different spatial sizes and depths and can capture different levels of detailed information in hyperspectral images. At the same time, competition and cooperation between different channels are captured through gating adaptation and channel normalization. Experiments on three well-known hyperspectral data sets show that our model achieves higher accuracy and less training time than other advanced methods, especially with fewer training samples.

Hyperspectral Image Classification Based on Spectral Multiscale Convolutional Neural Network

In recent years, convolutional neural networks (CNNs) have been widely used for hyperspectral image classification, which show good performance. Compared with using sufficient training samples for classification, the classification accuracy of hyperspectral images is easily affected by a small number of samples. Moreover, although CNNs can effectively classify hyperspectral images, due to the rich spatial and spectral information of hyperspectral images, the efficiency of feature extraction still needs to be further improved. In order to solve these problems, a spatial–spectral attention fusion network using four branch multiscale block (FBMB) to extract spectral features and 3D-Softpool to extract spatial features is proposed. The network consists of three main parts. These three parts are connected in turn to fully extract the features of hyperspectral images. In the first part, four different branches are used to fully extract spectral features. The convolution kernel size of each branch is different. Spectral attention block is adopted behind each branch. In the second part, the spectral features are reused through dense connection blocks, and then the spectral attention module is utilized to refine the extracted spectral features. In the third part, it mainly extracts spatial features. The DenseNet module and spatial attention block jointly extract spatial features. The spatial features are fused with the previously extracted spectral features. Experiments are carried out on four commonly used hyperspectral data sets. The experimental results show that the proposed method has better classification performance than some existing classification methods when using a small number of training samples.