Unsupervised hyperspectral greedy waveband selection strategy based on dual-frequency collaborative feature fusion

Abstract

The invention relates to an unsupervised hyperspectral greedy waveband selection strategy based on dual-frequency collaborative feature fusion, which is applied to the field of image classification, in particular to an unsupervised hyperspectral greedy waveband selection strategy. The objective of the invention is to solve the problems of data redundancy caused by low wave band selection accuracy of the existing method; and the problem that a traditional space attention module cannot effectively capture a long-distance or cross-region dependency relationship when capturing local context information, so that the global and local space dependency relationship cannot be effectively captured is solved. The method comprises the following steps: constructing a dual-frequency collaborative feature fusion and greedy waveband selection model; the model comprises a dual-frequency collaborative feature fusion module and a waveband selection module based on a greedy strategy. And acquiring hyperspectral image data to be measured, inputting the hyperspectral image data to be measured into the trained dual-frequency collaborative feature fusion and greedy wave band selection model, and outputting a selected wave band by the trained dual-frequency collaborative feature fusion and greedy wave band selection model.

Description

Technical Field

[0001] This invention is applied to the field of image classification, specifically involving an unsupervised hyperspectral greedy band selection strategy. Background Technology

[0002] Hyperspectral images (HSI) are a collection of images composed of multiple narrowband images, each representing a radiation signal in a different wavelength range. Each band reflects the same scene composed of objects imaged at different wavelengths [1] (C. Shi, D. Liao, T. Zhang and L. Wang, "Hyperspectral ImageClassification Based on Expansion Convolution Network," in IEEE Transactions on Geoscience and Remote Sensing, vol. 60, pp. 1-16, 2022.). In recent years, hyperspectral images have been widely used in various fields. The acquisition of scene HSI involves recording multiple spectral bands in a specific wavelength range. In the context of land cover classification, hyperspectral images, with their rich spectral bands, can accurately perceive and identify land cover [1]. However, as the data dimension increases, the model becomes too complex, leading to overfitting and the famous "Hughes phenomenon," also known as the curse of dimensionality [5] (L. Wei, H. Ma, Y. Yin and C. Geng, "Kmeans-CM AlgorithmWith Spectral Angle Mapper for Hyperspectral Image Classification," in IEEEAccess, vol. 11, pp. 26566-26576, 2023.). Therefore, dimensionality reduction has become an important step in processing hyperspectral image data.

[0003] Currently, dimensionality reduction technology is mainly achieved through two methods: one is feature extraction [6](Z. Chen, D. Hongand H. Gao, "Grid Network: Feature Extraction in Anisotropic Perspective for Hyperspectral Image Classification," in IEEE Geoscience and Remote Sensing Letters, vol. 20, pp. 1-5, 2023.); the other is feature selection, also known as band selection [7](X. Zhang and Z. Wang, "Spatial Proximity Feature Selection With Residual Spatial–SpectralAttention Network for Hyperspectral Image Classification," in IEEE Access, vol. 11, pp. 23268-23281, 2023.). Traditional band selection methods include three main categories: filters [8][9]([8]D. Chutia, DKBhattacharyya, J. Sarma and PNLRaju, "An effective ensemble classification framework using random forests and a correlation-based feature selection technique", Trans.GIS, vol.21, no.6, pp.1165-1178, 2017.[9]K.Koonsanit, C. Jaruskulchai and A. Eiumnoh, "Band selection for dimension reduction in hyper spectral image using integrated information gain and principal components analysis technique", Int.J.Mach. Learn. Comput., vol. 2, no. 3, pp. 248-251, Jun.2012.), and wrappers

[10] -

[13] (

[10] A. Jović, K. Brkić and N.Bogunović, "A review of feature selection methods with applications", Proc.38th Int. Conv. Inf. Commun. Technol. Electron. Microelectron. (MIPRO), pp.1200-1205, May 2015.

[11] A. Rady, N. Ekramirad, A. A. Adedeji, M. Li and R.Alimardani, "Hyperspectral imaging for detection of codling moth infestationin GoldRush apples", Postharvest Biol. Technol., vol. 129, pp. 37-44, Jul.2017.

[12] K. Nagasubramanian, S. Jones, S. Sarkar, A. K. Singh, A. Singh andB. Ganapathysubramanian, Hyperspectral band selection using genetic algorithmand support vector machines for early identification of charcoal rot diseasein soybean, 2017.

[13] K. R. Thorp, G. Wang, K. F. Bronson, M. Badaruddin andJ. Mon, "Hyperspectral data mining to identify relevant canopy spectralfeatures for estimating durum wheat growth nitrogen status and grain yield",Comput. Electron. Agricult., vol. 136, pp. 1-12, Apr. 2017.) and embedded

[14]

[15] (

[14] Z. M. Hira and D. F.Gillies, "A review of feature selection and feature extraction methods applied on microarray data", Adv. Bioinf., vol. 2015, May2015.

[15] Y. Qian, M. Ye and J. Zhou, "Hyperspectral image classification based on structured sparse logistic regression and three-dimensional wavelettexture features", IEEE Trans. Geosci. Remote Sens., vol. 51, no. 4, pp.2276-2291, Apr. 2013.). Although filter-based methods can remove noise, they may introduce phase distortion and are not completely frequency selective, while also having high implementation complexity. Wrapper-based methods rely on specific classifiers, have high computational load, and lack flexibility, making it difficult to adapt to changes in classifiers. Although embedded methods combine band selection and classifier design, they consume high resources, face real-time challenges, and have strong hardware dependence. With the development of technology, deep learning has made great progress in various fields

[16] (C. Shi, T. Wang and L. Wang, "Branch FeatureFusion Convolution Network for Remote Sensing Scene Classification," in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 13, pp. 5194-5210, 2020.). Convolutional neural networks, graph neural networks and autoencoders are quite excellent in band selection methods. Cai et al.

[17] (Y. Cai, X. Liu and Z. Cai, "BS-Nets: An End-to-End Framework for Band Selection of Hyperspectral Image," in IEEE Transactions on Geoscience and Remote Sensing, vol. 58, no. 3, pp.In 1969-1984, March 2020, Feng et al. proposed a BS-Nets (BS-Net-Conv) based on convolutional neural networks, which uses 2D convolutional layers and global pooling layers to extract spectral and spatial information simultaneously. Feng et al.

[18] (J. Feng et al., "Convolutional Neural Network Based on Bandwise-Independent Convolution and Hard Thresholding for Hyperspectral Band Selection," in IEEE Transactions on Cybernetics, vol. 51, no. 9, pp. 4414-4428, Sept.2021.) proposed a method called BandSelection Convolutional Neural Network (BHCNN) for hyperspectral image band selection. In order to solve the accuracy problem of traditional methods in the detection of changes in dual-temporal hyperspectral images, a method called Slow-fast band selection and feature fusion grouping (SFBS-FFGNET) was proposed by Ou et al.

[19] (X. Ou, L. Liu, B. Tu, G. Zhang and Z. Xu, "A CNN Framework With Slow-Fast Band Selection and Feature Fusion Grouping for Hyperspectral Image Change Detection," in IEEE Transactions on Geoscience and Remote Sensing, vol. 60, pp. 1-16, 2022). Feng et al.

[20] (J. Feng, G. Bai, D. Li, X. Zhang, R. Shangand L. Jiao, "MR-Selection: A Meta-Reinforcement Learning Approach for Zero-Shot Hyperspectral Band Selection," in IEEE Transactions on Geoscience and Remote Sensing, vol. 61, pp. 1-20, 2023).This paper proposes a novel zero-shot hyperspectral image band selection method (MR-Selection), which transforms the band selection problem into a meta-learning problem, thus solving the problem of zero-shot hyperspectral image band selection.

[0004] According to whether labeled samples are used, existing band selection methods can be further divided into supervised

[22] (X. Cao, T. Xiong and L. Jiao, "Supervised Band Selection Using Local Spatial Information for Hyperspectral Image," in IEEE Geoscience and Remote Sensing Letters, vol.13, no.3, pp.329-333, March 2016.), semi-supervised

[23] -

[25] (

[23] J. Feng, L. Jiao, F. Liu, T. Sun and X. Zhang, "Mutual-Information-Based Semi-Supervised Hyperspectral Band Selection With High Discrimination, High Information, and Low Redundancy," in IEEE Transactions on Geoscience and Remote Sensing, vol.53, no.5, pp.2956-2969, May 2015.

[24] X. Cao, C. Wei, Y. Ge, J. Feng, J. Zhao and L. Jiao, "Semi-Supervised Hyperspectral Band Selection Based on Dynamic Classifier Selection," in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol.12, no.4, pp.1289-1298, April 2019.

[25] L. Jiao, J. Feng, F. Liu, T. Sun and X. Zhang, "Semisupervised Affinity Propagation Based on Normalized Trivariable Mutual Information for Hyperspectral Band Selection," in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol.8, no. 6, pp. 2760-2773, June 2015. There are three types: supervised, semi-supervised, and unsupervised. Compared to supervised and semi-supervised methods, unsupervised band selection methods are suitable for more downstream tasks because they do not require pre-labeling and are not constrained by labeled data. Unsupervised band selection methods do not rely on any known category information and can autonomously explore the intrinsic structure of sample data to achieve efficient band selection, making them a current hot topic in band selection research. Meanwhile, unsupervised methods can be further divided into the following four types: ranking

[26]

[27] (

[26] A. Datta, S. Ghosh and A. Ghosh, "Combination of Clustering and Ranking Techniques for Unsupervised Band Selection of HyperspectralImages," in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 8, no. 6, pp. 2814-2823, June 2015.

[27] B. Xu, X. Li, W.Hou, Y. Wang and Y. Wei, "A Similarity-Based Ranking Method for HyperspectralBand Selection," in IEEE Transactions on Geoscience and Remote Sensing, vol.59, no. 11, pp. 9585-9599, Nov. 2021.), search

[28]

[29] (

[28] X. Ou, M. Wu, B. Tu,G. Zhang and W. Li, "Multi-Objective Unsupervised Band Selection Method for Hyperspectral Images Classification," in IEEE Transactions on ImageProcessing, vol. 32, pp. 1952-1965, 2023.

[29] SN Ibrahim, AS Eyigulu,and EEAttah, “The fast changing phase of land titling and registration using GIS application and its benefits in Nasarawa State, Nigeria,” Int. J. Innov. Res. Develop., vol. 5, no. 14, pp. 84–91, 2016.), clustering

[30] -

[32] (

[30] Y. Yuan, J. Lin and Q. Wang, "Dual-Clustering-Based Hyperspectral Band Selection by Contextual Analysis," in IEEE Transactions on Geoscience and Remote Sensing, vol. 54, no. 3, pp. 1431-1445, March 2016.

[31] M. Zeng, Y. Cai, Z. Cai, X. Liu, P. Hu and J. Ku, "Unsupervised Hyperspectral Image Band Selection Based on Deep Subspace Clustering," in IEEE Geoscience and Remote Sensing Letters, vol. 16, no. 12, pp. 1889-1893, Dec. 2019.

[32] Z. Tian, K. Gao, X. Zhang, J. Wang and Y. Feng, "Successive Clustering-Based Outlier Resistant Band Selection Method for Hyperspectral Images With Spatial Information Difference Metrics," in IEEE Geoscience and Remote Sensing Letters, vol. 20, pp. 1-5, 2023.) and sparse representation methods

[33]

[34] (

[33] P. Hu, X. Liu, Y. Cai and Z.Cai, "Band Selection of Hyperspectral Images Using MultiobjectiveOptimization-Based Sparse Self-Representation," in IEEE Geoscience and RemoteSensing Letters, vol. 16, no. 3, pp. 452-456, March 2019.

[34] F. Li, P. Zhangand L. Huchuan, "Unsupervised Band Selection of Hyperspectral Images viaMulti-Dictionary Sparse Representation," in IEEE Access, vol. 6, pp. 71632-71643, 2018.). In order to solve the problem that traditional single-objective optimization methods cannot consider multiple optimization objectives at the same time, Xu et al.

[35] (X. Xu, Z. Shi and B. Pan, "A New Unsupervised Hyperspectral Band Selection Method Based on Multiobjective Optimization", IEEE Geoscience and Remote Sensing Letters, vol. 14, no. 11, pp. 2112-2116, Nov. 2017.) proposed an infrared spectral data band selection method (IRMoBS) based on multi-objective optimization. Sun et al.

[36] (H. Sun, L. Zhang, L. Wang, and H. Huang, “Stochastic gate-based autoencoder for unsupervised hyperspectral band selection,” Pattern Recognit., vol. 132, p.108969, 2022.) introduced a deep learning-based unsupervised hyperspectral image band selection method—random gate autoencoder (SGAE), which improves the performance and robustness of the algorithm through autoencoder. To address the problem that traditional band selection methods struggle to effectively capture the dependencies between nodes in high-dimensional data, Yu et al.

[37] (C. Yu, S. Zhou, M. Song, B. Gong, E. Zhao and C. -I.Chang, in "Unsupervised Hyperspectral Band Selection via Hybrid Graph Convolutional Network," in IEEE Transactions on Geoscience and Remote Sensing, vol. 60, pp. 1-15, 2022, proposed a method for hyperspectral image feature extraction and band selection based on Convolutional Neural Networks (CNNs) and Graph Convolutional Networks (GCNs), which accurately selects representative bands. You et al.

[38] (M. You, A. Yuan, M. Zou and K. Konno, "Robust Unsupervised Hyperspectral Band Selection via Global Affinity Matrix Reconstruction," in IEEE Journal of Selected Topics in Applied Earth Observations and RemoteSensing, vol. 16, pp. 7374-7384, 2023.) proposed a new pseudo-label construction method, namely Global Similarity Reconstruction (GAMR), to solve the problem of hyperspectral image band selection and classification, and better preserve the spatial structure information of the original data. Ma et al.

[39] (Z. Ma and B. Yang, "Spatial–Spectral Hypergraph-BasedUnsupervised Band Selection for Hyperspectral Remote Sensing Images," in IEEE Sensors Journal, vol. 24, no. 17, pp. 27870-27882, 1 Sept.1, 2024.) proposed a novel unsupervised band selection method based on a spatial-spectral hypergraph (SSHUBS), which constructs a spatial-spectral hypergraph to reduce redundant information and noise in high-dimensional remote sensing images.

[0005] At present, an important application scenario of hyperspectral images is pixel classification. Based on feature merging, existing band techniques can be divided into spatial, spectral, or hybrid spectral and spatial information. Cai et al.

[40] (Y.Cai,Z.Zhang,X.Liu and Z.Cai, "Efficient Graph Convolutional Self-Representation for Band Selection of Hyperspectral Image", in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol.13, pp.4869-4880, 2020.) proposed a spectral clustering algorithm called Efficient Graph Convolutional Self-Representation (EGCSR), which uses spectral graph embedding to capture structural information between spectral data. By segmenting the original image into multiple homogeneous regions and fusing the features of these regions into a low-dimensional latent space, Feng et al.

[41] (W.Feng et al., "Hyperspectral band selection via region-wise latent feature fusion and graph filter embedded subspace clustering", Eng.Appl.Artif.Intell., vol.132,2024.) proposed a band selection algorithm called region-wise latent feature fusion and graph filter embedded subspace clustering (RFGEC), which effectively captures spatial information.Roy et al.

[42] (SKRoy,S.Das,T.Song and B.Chanda, "DARecNet-BS: Unsupervised Dual-Attention Reconstruction Network for Hyperspectral Band Selection," in IEEE Geoscience and Remote Sensing Letters, vol.18, no.12, pp.2152-2156, Dec.2021.) proposed a new hyperspectral image band selection method (DARecNet-BS). Compared with traditional convolutional neural networks, the DARECNet-BS network introduces pixel attention modules and channel attention modules, which can automatically learn and extract useful spectral and spatial features. Wang et al.

[43] (J. Wang, C. Tang, X. Zheng, X. Liu, W. Zhang and E. Zhu, "Graphregularized spatial–spectral subspace clustering for hyperspectral band selection", Neural Netw., vol. 153, pp. 292-302, 2022.) proposed a hyperspectral image sub-band selection method (GRSC) based on low-dimensional feature representation and graph regularization of regional segmentation, which explores the spectral correlation between all bands and preserves the spatial structure information in different sub-regions. An unsupervised band selection method called Dual Global–Local Attention Network (DGLAnet) was proposed

[44] (K. He et al., "A Dual Global–Local Attention Network for Hyperspectral Band Selection", in IEEE Transactions on Geoscience and Remote Sensing, vol. 60, pp. 1-13, 2022.), which makes full use of the features of spatial and spectral information in both global and local aspects.By utilizing the interaction between pixels and superpixels, spatial and structural information is embedded into the model. A new hyperspectral band selection method based on global local graph autoencoder (Tensorial Global-Local Graph Self-Representation, TGSR) has been proposed

[45] (

[45] Y.Zhang,J.Qi,X.Wang,Z.Cai,J.Peng and Y.Zhou, "Tensorial Global-Local GraphSelf-Representation for Hyperspectral Band Selection,"in IEEE Transactions on Circuits and Systems for Video Technology,doi: 10.1109 / TCSVT.2024.), which improves the accuracy of band selection. Ma et al.

[46] (M.Ma, F.Li, Y.Hu, Z.Wang and S.Mei, "Joint Spatial and Spectral Graph-Based Consistent Self-Representation for Unsupervised Hyperspectral Band Selection," in IEEE Transactions on Geoscience and RemoteSensing, vol.62, pp.1-16, 2024.) proposed a method based on joint spatial and spectral graph consistent self-representation (JSSGCSR), which effectively combines spatial and spectral information while utilizing graph convolutional networks to capture structured information in the data.

[0006] From the above analysis, it can be found that

[41] only considered spatial information and ignored spectral information.

[40] considered spectral relationships but ignored spatial structure information. While methods

[42] -

[46] all considered spatial structure information and spectral relationships, they all ignored detailed features in the frequency domain. Summary of the Invention

[0007] The purpose of this invention is to address the problems of low accuracy in band selection in existing methods, which leads to data redundancy; and the inability of traditional spatial attention modules to effectively capture long-distance or cross-regional dependencies when capturing local context information, resulting in the inability to effectively capture global and local spatial dependencies. Therefore, this invention proposes a greedy band selection algorithm based on dual-frequency collaborative feature fusion for unsupervised band selection of hyperspectral images.

[0008] The specific process of the unsupervised hyperspectral greedy band selection strategy based on dual-frequency collaborative feature fusion is as follows:

[0009] Step 1: Obtain the training set; Step 2: Construct a dual-frequency collaborative feature fusion and greedy band selection model; the dual-frequency collaborative feature fusion and greedy band selection model includes a dual-frequency collaborative feature fusion module and a band selection module based on a greedy strategy; Step 3: Train the dual-frequency collaborative feature fusion and greedy band selection model based on the training set to obtain a trained dual-frequency collaborative feature fusion and greedy band selection model; Step 4: Collect the hyperspectral image data to be tested, input the hyperspectral image data to be tested into the trained dual-frequency collaborative feature fusion and greedy band selection model, and the trained dual-frequency collaborative feature fusion and greedy band selection model outputs the selected bands.

[0010] The beneficial effects of this invention are as follows:

[0011] This invention proposes a greedy strategy based on dual-frequency collaborative feature fusion for hyperspectral image band selection. It aims to capture local and global features of spatial structure and spectral information in HSI by learning frequency domain features, and then use a subsequent greedy band selection strategy to select the optimal subset of bands. Considering that traditional Spatial Attention Modules (SAMs) cannot effectively capture long-distance or cross-regional dependencies when capturing local contextual information, a Grouping Spatial Attention Module (GSAM) is introduced to effectively capture global and local spatial dependencies. To overcome the shortcomings of using a single metric, a greedy selection strategy is proposed that effectively combines Pearson correlation coefficient and mutual information, thereby achieving better selection results. In summary, the specific contributions of this invention are as follows:

[0012] This invention proposes a greedy band selection algorithm based on dual-frequency collaborative feature fusion for unsupervised band selection in hyperspectral images. It combines the frequency feature learning capability of a dual-frequency collaborative feature fusion module with the local optimal selection characteristic of a greedy selection strategy, effectively improving the accuracy of band selection and reducing data redundancy. The invention proposes a dual-frequency collaborative feature fusion module (DFCFFM) that effectively captures local and global features of spatial and spectral information in HSI by learning depth and detail features under different frequency domain components. The invention also proposes a greedy selection strategy that effectively combines MI and PCC, simultaneously capturing nonlinear relationships and linear correlations between bands; effectively overcoming the shortcomings of a single metric, ensuring that the selected bands are both highly informative and relatively independent. Band selection plays a crucial role as a preprocessing step in hyperspectral image processing. However, most current band selection methods, while effectively fusing spatial spectral features, often lack the learning of frequency domain features. This invention proposes a greedy band selection algorithm based on dual-frequency collaborative feature fusion for unsupervised band selection in hyperspectral images. First, a dual-frequency cooperative feature fusion module (DFCFFM) is used to extract and fuse spatial spectral features from the two frequency domain components. Second, the correlation and nonlinear relationship between bands are learned by calculating the mutual information (MI) of the fused features and the Pearson correlation coefficient (PCC) of the original data. Finally, a greedy selection strategy is adopted to combine MI and PCC to select the optimal subset of bands. Classification is used to validate the extracted band selection method. Extensive experiments demonstrate that, compared with other state-of-the-art methods, the proposed method achieves the best classification performance on three public datasets: Indian Pines, Pavia University, and Houston 2013, fully demonstrating the effectiveness of the proposed band selection method. Attached Figure Description

[0013] Figure 1 The overall structural block diagram of the method of the present invention includes two main stages: dual-frequency collaborative feature fusion and greedy band selection; Figure 2 Here is a block diagram of the Grouped Channel Attention Module (GSAM). Figure 3 This is a flowchart of the GBS processing method proposed in this invention; Figure 4 Visualization of classification results for different methods on the Indian Pines dataset; Figure 5 Visualizations of classification results using different methods on the Pavia University dataset; Figure 6This is a visualization of the classification results of different methods on the Houston 2013 dataset. Detailed Implementation

[0014] Specific Implementation Method 1: The specific process of this implementation method based on the unsupervised hyperspectral greedy band selection strategy of dual-frequency collaborative feature fusion is as follows:

[0015] Step 1: Obtain the training set; Step 2: Construct a dual-frequency collaborative feature fusion and greedy band selection model; the dual-frequency collaborative feature fusion and greedy band selection model includes a dual-frequency collaborative feature fusion module (DFCFFM) and a greedy band selection module (Greedy Band Selection, GBS); Step 3: Train the dual-frequency collaborative feature fusion and greedy band selection model based on the training set to obtain a trained dual-frequency collaborative feature fusion and greedy band selection model; Step 4: Collect the hyperspectral image data to be tested, input the hyperspectral image data to be tested into the trained dual-frequency collaborative feature fusion and greedy band selection model, and the trained dual-frequency collaborative feature fusion and greedy band selection model outputs the selected bands.

[0016] Specific Implementation Method Two: This implementation method differs from Specific Implementation Method One in that: in step one, the training set is obtained; the specific process is as follows:

[0017] Obtain raw HSI data HSI data is hyperspectral image data;

[0018] in, and These are the height and width of the original HSI data, respectively. Indicates the original number of bands; Represent real numbers;

[0019] The hyperspectral image is randomly divided into blocks, with each hyperspectral image block being... , , express The other steps and parameters are the same as in Specific Implementation Method 1.

[0020] Specific Implementation Method Three: This implementation method differs from Specific Implementation Method One or Two in that: in step two, a dual-frequency collaborative feature fusion and greedy band selection model is constructed; the dual-frequency collaborative feature fusion and greedy band selection model includes a dual-frequency collaborative feature fusion module (DFCFFM) and a greedy band selection module (Greedy Band Selection, GBS); the specific process is as follows:

[0021] The dual-frequency collaborative feature fusion module (DFCFFM) includes DWT downsampling, GSAM module, spectral spatial attention module SSAM, high-frequency feature fusion module HFF, and feature fusion; DWT represents discrete wavelet transform;

[0022] The working process of the dual-frequency collaborative feature fusion module (DFCFFM) is as follows:

[0023] 1) Hyperspectral image patches Downsampling was performed using 2D DWT and 3D DWT respectively, such as Figure 1 As shown;

[0024] Hyperspectral image patch After 2D DWT downsampling, a low-frequency component containing the main structural information is obtained. and three high-frequency components containing detailed features , and ;

[0025] Hyperspectral image patch A low-frequency component was obtained after 3D DWT downsampling. and seven high-frequency components , , , , , and ;

[0026] Leveraging the powerful local feature learning capabilities of convolutional kernels, low-frequency components can be processed. Feature extraction is performed using two layers of 2D convolution to obtain features. ; indicates as:

[0027]

[0028] Leveraging the powerful local feature learning capabilities of convolutional kernels, low-frequency components can be processed. Feature extraction is performed using two layers of 3D convolution to obtain features. ; indicates as:

[0029]

[0030] in, This indicates that the number of channels is 32 and the kernel size is [value missing]. 2D convolution; This indicates that the number of channels is 64 and the kernel size is [value missing]. 2D convolution; This indicates that the number of channels is 8 and the kernel size is [value missing]. 3D convolution; This indicates that the number of channels is 16 and the kernel size is [value missing]. 3D convolution;

[0031] The proposed dual-frequency collaborative feature fusion module (DFCFFM) aims to fully utilize the deep features in the low-frequency components and the detailed features in the high-frequency components to further achieve feature fusion of spatial spectral information in HSI data. For example... Figure 1 As shown, the DFCFFM module mainly consists of DWT downsampling, spatial spectral feature extraction, and feature fusion.

[0032] 2) Features The GSAM module outputs features. ;

[0033] 3) Features The Spectral Spatial Attention Module (SSAM) outputs features. ;

[0034] 4) Combine the three high-frequency components containing detailed features , and The input is processed by the High-Frequency Feature Fusion (HFF) module, and the HFF module outputs features. ;

[0035] 5) Seven high-frequency components , , , , , and The input is processed by the High-Frequency Feature Fusion (HFF) module, and the HFF module outputs features. ;

[0036] 6) Features and The pieces are stitched together along the channel dimension, and then passed through sequentially. The convolutional layer, batch normalization layer, and ReLU activation function layer, with the ReLU activation function layer outputting features. ;

[0037] Output features of the spectral-spatial attention module and characteristics The pieces are stitched together along the channel dimension, and then passed through sequentially. The convolutional layer, batch normalization layer, and ReLU activation function layer, with the ReLU activation function layer outputting features. ;

[0038] Feature extraction and Mapped to a unified channel dimension;

[0039] 7) Features and The fusion is performed to obtain the fusion features. ;

[0040] 8) Fusion characteristics Flatten the surface, and then input the flattened features sequentially. Convolutional layer, Batch Normalization (BN) layer, ReLU activation function layer; ReLU activation function layer outputs reconstructed features. ;

[0041] The specific working process of the greedy band selection (GBS) module is as follows:

[0042] 1) Calculate hyperspectral image patches Reconstruction features Mutual information (MI) between tags;

[0043] 2) Transfer the original HSI data Deformation to obtain new HSI data ;

[0044] III) Calculate the new HSI data The Pearson correlation coefficients among all bands are used to construct a symmetric Pearson correlation coefficient matrix. ;

[0045] Pearson correlation coefficient matrix The Middle The first band and the first Pearson correlation coefficient between bands Represented as:

[0046]

[0047] In the formula, Represents hyperspectral image patches Reconstruction features In the Values ​​for each band, Represents hyperspectral image patches Reconstruction features In the Values ​​for each band; Indicates the number of hyperspectral image patches;

[0048] Represents the reconstructed features of all hyperspectral image patches In the The average value of each band, Represents the reconstructed features of all hyperspectral image patches In the The average value of each band;

[0049] (iv) Initialize two sets: an empty set `selected_bands` to store the selected band indices, and a set `available_bands` to store the reconstructed features. All selectable band indexes;

[0050] 5) Select reconstruction features from the set available_bands The reconstructed feature corresponding to the highest mutual information (MI) value between the label and the target label Band Index Populate selected_bands and retrieve the selected band index from available_bands. Remove;

[0051] 6) Extract the Pearson correlation coefficients between the bands selected in step 5 and all selected bands in the set selected_bands, and take the average of all Pearson correlation coefficients. The mutual information corresponding to the selected bands (v) Subtract the mean As the gain of the selected band (v). ;

[0052] VII) If the gain of the band selected in V) Less than or equal to the optimal gain set in the original configuration When, then the optimal gain constant;

[0053] If the gain of the selected band in (v) Greater than the optimal gain set by the original parameters When the optimal gain value is updated, the optimal gain will be adjusted accordingly. Corresponding band Select and enter the values ​​into selected_bands;

[0054] 8) Repeat steps 5) through 7) until the number of selected bands is greater than or equal to the set number of selected bands. At that time, the final selected band is obtained.

[0055] A greedy strategy is an algorithm design strategy that makes locally optimal choices at each step, hoping that these locally optimal choices will lead to a globally optimal solution. The proposed Greedy Band Selection (GBS) method aims to select the most representative bands from hyperspectral image data. By combining the Pearson correlation coefficient (PCC) between bands and the mutual information (MI) between bands and labels, the optimal band combination is selected using a greedy strategy.

[0056] This invention proposes a greedy band selection algorithm based on dual-frequency collaborative feature fusion for unsupervised band selection in hyperspectral images. First, DFCFFM is used to learn features from the high-frequency and low-frequency components of the patch data separately. The learned features from different frequency domains are extracted and fused to capture the global and local features of HSI. Second, the PCC and MI of the data are calculated separately to learn the inter-band correlation and nonlinear relationships. Finally, by combining PCC and MI with a greedy selection strategy, the optimal band subset is selected from all bands, and the classification performance of the selected band subset is verified using an SSRN classifier. Other steps and parameters are the same as in specific implementation methods one or two.

[0057] Specific Implementation Method Four: This implementation method differs from one of Specific Implementation Methods One to Three in that: in step 2), the feature... The GSAM module outputs features. The specific process is as follows:

[0058] To adaptively learn important features in low-frequency components and enhance the effectiveness of feature extraction, GSAM ( Figure 2 Without performing channel dimensionality reduction, it effectively captures global and local spatial dependencies through grouping and multi-scale parallel subnetwork structures.

[0059] Features Divided into channels Sub-characteristics, the first Each sub-feature is represented as , ; express Much smaller than ;

[0060] The data passes through a 1D global average pooling layer (along the horizontal direction) sequentially, and the global average pooling layer outputs an attention map. ; The data passes through a 1D global average pooling layer (along the vertical direction) sequentially, and the global average pooling layer outputs an attention map. ; Passing through in sequence The convolutional layer, the softmax activation function, and the attention map output by the softmax activation function. ; Passing through in sequence Convolutional layers, average pooling, and average pooling output attention map. ;

[0061] attention map Attention map Concatenate along the channel dimension to generate features ;feature go through Convolutional layers generate features ;feature Output features after sigmoid activation function (along the horizontal direction) ;feature Output features after sigmoid activation function (along the vertical direction) ;

[0062] Features ,feature , Perform dot product to generate features ;feature Features are generated after group normalization using GroupNorm. ;feature The output features are obtained after the Softmax activation function. ;feature Output features after average pooling ;

[0063] attention map With features Perform matrix multiplication to obtain the features Attention map With features Perform matrix multiplication to obtain the features ;

[0064] Features With features By adding elements one by one, we can obtain the features. ;feature The output features are obtained after the sigmoid activation function. ;feature With features Perform dot product to generate features ;feature As an output feature of the GSAM module .

[0065] The other steps and parameters are the same as those in one of the specific implementation methods one to three.

[0066] Specific Implementation Method Five: This implementation method differs from one of Specific Implementation Methods One to Four in that: the feature in 3) The Spectral Spatial Attention Module (SSAM) outputs features. The specific process is as follows:

[0067] 31) Features Input to a max pooling layer, output features from the max pooling layer ;

[0068] 32) Features Input to average pooling layer, output features from average pooling layer ;

[0069] 33) Features With features Concatenate along the channel dimension to generate features ;feature The input is sequentially processed by a 3D convolutional layer, a layer normalization layer (BN), and a ReLU activation function layer. The ReLU activation function layer outputs features, which are then input into a sigmoid activation function, which outputs features. ; Features With features Perform element-wise multiplication to generate features ; indicates as:

[0070]

[0071] in, Indicates the max pooling layer. Indicates the average pooling layer. This represents a 3D convolutional layer. This represents the sigmoid activation function. This indicates element-wise multiplication.

[0072] For 3D low-frequency components, a Spectral-Spatial Attention Module (SSAM) is used, which can operate simultaneously in the spectral and spatial dimensions, comprehensively capturing the correlation between bands and in pixel space, and reducing redundant features. Other steps and parameters are the same as in specific implementation methods one to four to one.

[0073] Specific Implementation Method Six: This implementation method differs from Specific Implementation Methods One to Five in that: in step 4), three high-frequency components containing detailed features are included. , and The input is processed by the High-Frequency Feature Fusion (HFF) module, and the HFF module outputs features. The specific process is as follows:

[0074] Three high-frequency components containing detailed features , and The pieces are stitched together along the channel dimension, and then passed through... Convolution yields features ; indicates as:

[0075] in, This indicates splicing along the channel dimension. express The convolution.

[0076] The other steps and parameters are the same as those in specific implementation methods one through five.

[0077] Specific Implementation Method Seven: This implementation method differs from Specific Implementation Methods One through Six in that: in step 5), seven high-frequency components are... , , , , , and The input is processed by the High Frequency Feature Fusion (HFF) module, which outputs the features. The specific process is as follows:

[0078] Seven high-frequency components , , , , , and The pieces are stitched together along the channel dimension, and then passed through... Convolution yields features ; indicates as:

[0079] .

[0080] The other steps and parameters are the same as those in one of the specific implementation methods one to six.

[0081] Specific Implementation Method Eight: This implementation method differs from one of Specific Implementation Methods One to Seven in that: in 7), the feature... and The fusion is performed to obtain the fusion features. ; indicates as:

[0082] In the formula, This indicates summation pixel by pixel. This indicates the weighting coefficients set.

[0083] The other steps and parameters are the same as those in any of the specific implementation methods one to seven.

[0084] Specific Implementation Method Nine: This implementation method differs from Specific Implementation Methods One through Eight in that: in step six), the Pearson correlation coefficients between the bands selected in step five and all selected bands in the set selected_bands are extracted, and the average of all Pearson correlation coefficients is taken. The mutual information corresponding to the selected bands (v) Subtract the mean As the gain of the selected band (v). ; indicates as: .

[0085] The other steps and parameters are the same as those in one of the specific implementation methods one to eight.

[0086] Specific Implementation Method Ten: This implementation method differs from Specific Implementation Methods One through Nine in that: in step seven), if the gain of the band selected in step five is... Less than or equal to the optimal gain set in the original configuration When, then the optimal gain constant;

[0087] If the gain of the selected band in (v) Greater than the optimal gain set by the original parameters When the optimal gain value is updated, the optimal gain will be adjusted accordingly. Corresponding band Select and enter the information into selected_bands; specifically:

[0088] .

[0089] The other steps and parameters are the same as those in any of the specific implementation methods one to nine.

[0090] Mutual Information: Assuming the input data is , tag as The mutual information (MI) between data and labels is defined as:

[0091]

[0092] In the formula, It's a pixel value. These are the maximum pixel values ​​of the input data and the label, respectively. It is the joint probability distribution function. and It is the marginal probability distribution function. As a method to measure the degree of interdependence between two variables, mutual information can measure the linear or nonlinear correlation between them. When the two variables are completely independent... Due to its ability to capture nonlinear relationships, MI has been widely used in feature selection. Features of classification targets with higher MI values ​​are more suitable for classification

[47] (

[47] J. Feng, L. Jiao, F. Liu, T. Sun and X. Zhang, "Mutual-Information-Based Semi-Supervised Hyperspectral Band SelectionWith High Discrimination, High Information, and Low Redundancy," in IEEE Transactions on Geoscience and Remote Sensing, vol. 53, no. 5, pp. 2956-2969, May 2015.). MI can also be calculated using entropy values.

[0093] Discrete Wavelet Transform (DWT) is a method that can decompose data into components of different frequency ranges and is widely used in image processing

[48] (

[48] J. Xu, J. Zhao and Y. Fu, "An Efficient Hyperspectral Image Classification Method Using Deep Fusion of 3-D Discrete Wavelet Transform and CNN," in IEEE Geoscience and RemoteSensing Letters, vol. 20, pp. 1-5, 2023.). In deep learning, max pooling is often used for downsampling operations, but it is easy to destroy the basic object structure

[49] (

[49] Q. Li, L. Shen, S. Guo, and Z. Lai, "Wavelet integrated CNNs for noise-robust image classification," in Proc.IEEE / CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), Jun. 2020, pp. 7245–7254.). DWT decomposes the original image into low-frequency and high-frequency components. The low-frequency components represent basic object structure, while the high-frequency components contain a wealth of image details. Compared to max pooling, DWT can preserve high-frequency details during downsampling, better capturing subtle changes and features in the data. This is achieved by combining a series of 1D, 2D, and 3D DWTs. For 2D signals... Perform 1D DWT on its rows and columns, i.e.

[0094]

[0095] In the formula, L and L represent the high-pass filter and low-pass filter, respectively. Therefore, a low-frequency component can be obtained through 2D DWT. and three high-frequency components , and Similarly, using 3D DWT for 3D signals can yield a low-frequency component. And seven corresponding high-frequency components.

[0096] The beneficial effects of the present invention are verified using the following embodiments:

[0097] Example 1:

[0098] This invention compares the proposed method with eight other band selection methods on three publicly available datasets. The three publicly available datasets used for the experiments and the experimental setup are also described.

[0099] A. Dataset Introduction: The experiment uses three commonly used public datasets: Indian Pines, Pavia University, and Houston 2013. Indian Pines: The Indian Pines dataset is a classic agricultural hyperspectral remote sensing dataset, collected by NASA's AVIRIS sensor in 1992, covering an area in northern Indiana, USA. Pavia University: The Pavia University dataset reflects the urban environment of Pavia, Italy, acquired by the ROSIS sensor in 2004. Houston 2013: The Houston 2013 dataset is complex urban environmental data collected from Houston, Texas, USA, including information from multiple sensors. Furthermore, this dataset incorporates radar data, increasing the diversity of the data.

[0100] B. Experimental Setup: The proposed method was implemented in the PyTorch 2.0 framework and experimentally verified using an NVIDIA RTX 4090D GPU with 24GB of RAM. The training parameters were set to 20 for training epochs and 20 for learning rate. In the DFCFFM parameter configuration, the patch size is set to 8, and the feature fusion weight coefficient is... The initial value is set to 0.5, which is the channel packet size for GSAM. Set it to 8.

[0101] C. Performance Analysis: This experiment selected eight state-of-the-art band selection methods and compared their performance with the proposed method. These included one traditional method, PCA

[50] (Uddin, Md. P., Mamun, Md. A. and Hossain, Md. A. (2020) 'PCA-based Feature Reduction for Hyperspectral Remote Sensing ImageClassification', IETE Technical Review, 38(4), pp. 377–396.), two attention-based methods, namely BS-Net-Conv

[17] and DARecNet-BS

[42] , and four graph network-based methods, namely EGCSR

[40] , RFGEC

[41] , TGSR

[45] and SPCA_AMGL

[51] (W. Zhang, A. Yuan, J. Tang and X. Li, "Sparse Principal Component Analysis and Adaptive Multigraph Learning for Hyperspectral Band Selection," in IEEE Journal of Selected Topics in AppliedEarth Observations and Remote Sensing, vol. 17, pp. 1419-1433, 2024.), and an encoder-based method, DropoutCAE

[21] (

[21] Xu L, Ahishali M, Gabbouj M. DropoutConcrete Autoencoder for Band Selection on HSI Scenes[J]. arXiv preprintarXiv:2401.16522, 2024.).To better compare the classification performance of different methods, the spectral–spatial residual network (SSRN)

[52] (Z. Zhong, J. Li, Z. Luo and M. Chapman, "Spectral–Spatial Residual Network for Hyperspectral ImageClassification: A 3-D Deep Learning Framework," in IEEE Transactions on Geoscience and Remote Sensing, vol. 56, no. 2, pp. 847-858, Feb. 2018.) was used to evaluate the selected band subsets of all methods in an end-to-end training manner. In addition, the classification metrics used included overall accuracy (OA), average accuracy (AA), and [missing information]. .

[0102] 1) Quantitative analysis under different methods: Table 1-3 shows the OA, AA, and other parameters obtained by the proposed method and all comparative methods on three datasets with different numbers of bands selected. The accuracy for each category was measured. The Indian Pines, Pavia University, and Houston 2013 datasets used 25, 15, and 20 bands respectively, with 5% of the samples used for training. The results show that the proposed method achieves the highest classification results on all three datasets. As shown in Table 1, the accuracy for each category is as follows: OA, AA, and... The best results in this area are shown in bold. For the Indian Pines dataset, although the spectral clustering algorithm EGCSR using graph convolutional neural networks outperforms the proposed method in terms of accuracy for a single class, the proposed method achieves a 1.6% improvement in OA and a 1.4% improvement in AA compared to the proposed method. This represents a 1.9% improvement. Compared to other methods, the proposed method achieves a maximum improvement of 3.1% in OA and a maximum improvement of 6.5% in AA. The proposed method achieves a maximum performance improvement of 3.6%. As shown in Table 2, BS-Net-Conv, RFGEC, and SPCA_AMGL all achieve high classification performance on the Pavia University dataset. In comparison, the proposed method outperforms the other methods in all aspects. Compared to RFGEC, the best-performing method among the comparisons, the proposed method improves performance by at least 0.7% on OA and at least 1.6% on AA. It improved by at least 0.9%. As shown in Table 3, for the Houston 2013 dataset, among all methods, the proposed method improved in OA, AA, and The proposed method achieves the best classification performance in all aspects. In particular, it achieves 100% classification results for the four categories: Synthetic Grass, Residential, Tennis court, and Running track.

[0103]

[0104]

[0105]

[0106] 2) Visualization results analysis: Figure 4-6 The original ground truth maps of the three datasets and the visualized classification results of all methods on the three datasets are shown. It is evident that the visual classification results of the proposed method on all three datasets are closer to the original ground truth maps, demonstrating better classification performance. On the Indian Pines dataset (… Figure 4 Most methods misclassify Alfalfa and Grass pasture. While the proposed method also misclassifies, it misclassifies the fewest regions compared to other methods. On the Pavia University dataset (…), Figure 5 BS-Net-Conv and TGSR exhibit large-area misclassification in the Gravel class. Compared to other methods, the proposed method maintains higher regional consistency and provides visualizations closer to the ground truth map. Unlike the previous two datasets, on the Houston 2013 dataset ( Figure 6 Most methods achieve high region consistency, but compared to other methods, the proposed method not only maintains high region consistency but also achieves better edge information localization. In summary, the proposed method demonstrates strong classification performance on all three datasets, and its effectiveness is fully demonstrated through quantitative analysis and visualization results.

[0107] D. Quantitative Analysis of Band Selection Quantity: This experiment compares the classification performance of all methods by varying the number of selected bands. Specifically, the band selection range is [10:50]. Two comparison results are presented for each dataset.

[0108] On three datasets, different band selection methods yielded average OA, AA, and [other values] within the selected band range. .

[0109] Average classification performance of different methods within a selected band range

[0110] Table 4 leads to the conclusion that the proposed method achieves good results in terms of average OA, AA, and In all aspects, the proposed method outperforms other comparative methods. Specifically, on the three datasets, the proposed method achieves at least OA improvements of 1.42%, 1.26%, and 2.54%, respectively. Furthermore, when the number of bands selected for the three datasets is 25, 15, and 20, respectively, the classification results all exceed the average classification accuracy.

[0111]

[0112] Therefore, the number of bands selected was determined to be 25, 15, and 20 for the Indian Pines, Pavia University, and Houston 2013 datasets, respectively.

[0113] E. Analysis of Selected Bands: The selected bands are analyzed based on their location and entropy value. Tables 5, 6, and 7 show the index values ​​of the optimal bands obtained using different methods on the three datasets. As can be seen, on the Indian Pines dataset, RFGEC selects most of its bands in the [15-30] and [35-45] intervals, but ignores bands with higher entropy values ​​in the [145-200] interval. TGSR selects more dispersed bands, but it can be seen that the selected bands have lower entropy values ​​in certain intervals. Similarly, in the [75-105] interval, BSNet-Conv and DARecNet-BS select a large number of low-entropy bands. SPCA_AMGL selects most of its bands in the [35-200] interval, but ignores bands with higher entropy values ​​in the [0-30] interval. On the Pavia University dataset, while BSNet-Conv, DARecNet-BS, EGCSR, TGSR, and DropoutCAE all selected bands that were relatively dispersed, they all chose bands with low entropy values ​​in the [65-75] interval, thus reducing classification performance to some extent. SPCA_AMGL selected bands with high entropy values ​​and a uniform distribution, but ignored a large number of bands with high entropy values ​​in the [90-103] interval. On the Houston 2013 dataset, RFGEC selected bands concentrated in the [25-110] interval, ignoring bands with high entropy values ​​in the [130-144] interval.

[0114] On the one hand, due to the high correlation between adjacent bands, selecting bands in a denser arrangement would lead to greater information redundancy. On the other hand, bands with higher entropy values ​​contain more information, and selecting bands with lower entropy values ​​would limit classification performance. The proposed method combines MI and PCC, considering both the correlation and nonlinear relationships between bands. Therefore, the bands selected on the first two datasets are relatively discrete and have relatively high entropy values. While the bands selected on the Houston 2013 dataset are denser, most of the selected bands have high entropy values. Thus, compared to other methods, the bands selected by the proposed method significantly outperform other state-of-the-art comparative methods in terms of classification performance.

[0115]

[0116]

[0117]

[0118] F. Ablation Experiments: To comprehensively evaluate the effectiveness of DFCFFM and GBS in the proposed band selection method, detailed ablation experiments were conducted. Specifically, classification performance was tested on the Indian Pines, Pavia University, and Houston2013 datasets using only DFCFFM, only GBS, and both DFCFFM and GBS. The best results are highlighted in bold. The experimental results are listed in Table 8. As shown in Table 8, on the Indian Pines dataset, when using both DFCFFM and GBS, the proposed method achieved an OA of 95.18% and an AA of 96.73%. The accuracy rate was 94.51%, representing improvements of 2.17% and 4.1% in OA, and 1.76% and 4.96% in AA, respectively, compared to using DFCFFM or GBS alone. The improvements were 2.5% and 4.7% respectively. On the Pavia University dataset, when using both DFCFFM and GBS, the proposed method achieved an OA of 98.90% and an AA of 98.73%. The accuracy rate was 98.54%, representing improvements of 1.22% and 0.41% respectively in OA compared to using DFCFFM or GBS alone, and improvements of 2.22% and 0.31% respectively in AA. The improvements were 1.61% and 0.55% respectively. When using both DFCFFM and GBS on the Houston 2013 dataset, the proposed method achieved an OA of 96.59% and an AA of 97.17%. The accuracy rate was 96.31%, which was 1.31% and 0.71% higher in OA and AA, respectively, compared to using DFCFFM or GBS alone. The performance was improved by 1.41% and 0.76% respectively. In summary, the experimental results show that both DFCFFM and GBS can achieve good classification results when used alone, but the classification performance is significantly improved when the two are used in combination. This proves the importance of detailed information in different frequency domains for band selection and classification results, and further verifies the rationality and necessity of the proposed method.

[0119]

[0120] This invention may have other embodiments. Without departing from the spirit and essence of this invention, those skilled in the art can make various corresponding changes and modifications according to this invention, but these corresponding changes and modifications should all fall within the protection scope of the appended claims.

Claims

1. An unsupervised hyperspectral greedy band selection strategy based on dual-frequency collaborative feature fusion, characterized in that: The strategy The specific process is as follows: Step 1: Obtain the training set; Step 2: Construct a dual-frequency collaborative feature fusion and greedy band selection model; The dual-frequency collaborative feature fusion and greedy band selection model includes a dual-frequency collaborative feature fusion module and a band selection module based on a greedy strategy; Step 3: Train the dual-frequency collaborative feature fusion and greedy band selection model based on the training set to obtain the trained dual-frequency collaborative feature fusion and greedy band selection model; Step 4: Collect the hyperspectral image data to be tested, input the hyperspectral image data to be tested into the trained dual-frequency collaborative feature fusion and greedy band selection model, and output the selected bands.

2. The unsupervised hyperspectral greedy band selection strategy based on dual-frequency collaborative feature fusion according to claim 1, characterized in that: The training set is obtained in step one; the specific process is as follows: Obtain raw HSI data HSI data are hyperspectral image data; in, and These are the height and width of the original HSI data, respectively. Indicates the original number of bands; Represent real numbers; The hyperspectral image is randomly divided into blocks, with each hyperspectral image block being... , , express .

3. The unsupervised hyperspectral greedy band selection strategy based on dual-frequency collaborative feature fusion according to claim 2, characterized in that: Step two involves constructing a dual-frequency collaborative feature fusion and greedy band selection model. This model includes a dual-frequency collaborative feature fusion module and a band selection module based on a greedy strategy. The specific process is as follows: The dual-frequency collaborative feature fusion module includes DWT downsampling, GSAM module, spectral spatial attention module SSAM, high-frequency feature fusion module HFF, and feature fusion; DWT represents discrete wavelet transform; The working process of the dual-frequency collaborative feature fusion module is as follows: 1) Hyperspectral image blocks A low-frequency component was obtained after 2D DWT downsampling. and three high-frequency components , and ; Hyperspectral image patch A low-frequency component was obtained after 3D DWT downsampling. and seven high-frequency components , , , , , and ; For low-frequency components Feature extraction is performed using two layers of 2D convolution to obtain features. ; indicates as: For low-frequency components Feature extraction is performed using two layers of 3D convolution to obtain features. ; indicates as: in, This indicates that the number of channels is 32 and the kernel size is [value missing]. 2D convolution; This indicates that the number of channels is 64 and the kernel size is [value missing]. 2D convolution; This indicates that the number of channels is 8 and the kernel size is [value missing]. 3D convolution; This indicates that the number of channels is 16 and the kernel size is [value missing]. 3D convolution; 2) Features The GSAM module outputs features. ; 3) Features The Spectral Spatial Attention Module (SSAM) outputs features. ; 4) Combine the three high-frequency components , and The input is processed by the High Frequency Feature Fusion (HFF) module, which outputs the features. ; 5) Seven high-frequency components , , , , , and The input is processed by the High Frequency Feature Fusion (HFF) module, which outputs the features. ; 6) Features and The pieces are stitched together along the channel dimension, and then passed through sequentially. The convolutional layer, batch normalization layer, and ReLU activation function layer, with the ReLU activation function layer outputting features. ; Output features of the spectral-spatial attention module and characteristics The pieces are stitched together along the channel dimension, and then passed through sequentially. The convolutional layer, batch normalization layer, and ReLU activation function layer, with the ReLU activation function layer outputting features. ; 7) Features and The fusion is performed to obtain the fusion features. ; 8) Fusion characteristics Flatten the surface, and then input the flattened features sequentially. Convolutional layer, Batch Normalization (BN) layer, ReLU activation function layer; ReLU activation function layer outputs reconstructed features. ; The specific working process of the band selection module based on the greedy strategy is as follows: 1) Calculate hyperspectral image patches Reconstruction features Mutual information (MI) between tags; 2) Transfer the original HSI data Deformation to obtain new HSI data ; III) Calculate the new HSI data The Pearson correlation coefficients among all bands are used to construct a symmetric Pearson correlation coefficient matrix. ; Pearson correlation coefficient matrix The Middle The first band and the first Pearson correlation coefficient between bands Represented as: In the formula, Represents hyperspectral image patches Reconstruction features In the Values ​​for each band, Represents hyperspectral image patches Reconstruction features In the Values ​​for each band; Indicates the number of hyperspectral image patches; Represents the reconstructed features of all hyperspectral image patches In the The average value of each band, Represents the reconstructed features of all hyperspectral image patches In the The average value of each band; (iv) Initialize two sets: an empty set `selected_bands` to store the selected band indices, and a set `available_bands` to store the reconstructed features. All selectable band indexes; 5) Select reconstruction features from the set available_bands The reconstructed feature corresponding to the highest mutual information (MI) value between the label and the target label Band Index Populate selected_bands and retrieve the selected band index from available_bands. Remove; 6) Extract the Pearson correlation coefficients between the bands selected in step 5 and all selected bands in the set selected_bands, and take the average of all Pearson correlation coefficients. The mutual information corresponding to the selected bands (v) Subtract the mean As the gain of the selected band (v). ; VII) If the gain of the band selected in V) Less than or equal to the optimal gain set by the original parameters When, then the optimal gain constant; If the gain of the selected band in (v) Greater than the optimal gain set by the original parameters When the optimal gain value is updated, the optimal gain will be adjusted accordingly. Corresponding band Select and enter the values ​​into selected_bands; 8) Repeat steps 5) through 7) until the number of selected bands is greater than or equal to the set number of selected bands. At that time, the final selected band is obtained.

4. The unsupervised hyperspectral greedy band selection strategy based on dual-frequency collaborative feature fusion according to claim 3, characterized in that: The features mentioned in 2) The GSAM module outputs features. ; The specific process is as follows: Features Divided into channels Sub-characteristics, the first Each sub-feature is represented as , ; express Much smaller than ; The data passes through a 1D global average pooling layer, which outputs an attention map. ; The data passes through a 1D global average pooling layer, which outputs an attention map. ; Passing through in sequence The convolutional layer, the softmax activation function, and the attention map output by the softmax activation function. ; Passing through in sequence Convolutional layers, average pooling, and average pooling output attention map. ; attention map Attention map Concatenate along the channel dimension to generate features ; feature go through Convolutional layers generate features ; feature The output features are obtained after the sigmoid activation function. ; feature The output features are obtained after the sigmoid activation function. ; Features ,feature , Perform dot product to generate features ; feature Features are generated after group normalization (GroupNorm). ; feature The output features are obtained after the Softmax activation function. ; feature Output features after average pooling ; attention map With features Perform matrix multiplication to obtain the features ; attention map With features Perform matrix multiplication to obtain the features ; Features With features By adding elements one by one, we can obtain the features. ; feature The output features are obtained after the sigmoid activation function. ; feature With features Perform dot product to generate features ;feature As an output feature of the GSAM module .

5. The unsupervised hyperspectral greedy band selection strategy based on dual-frequency collaborative feature fusion according to claim 4, characterized in that: The features in 3) The Spectral Spatial Attention Module (SSAM) outputs features. The specific process is as follows: 31) Features Input to a max pooling layer, output features from the max pooling layer ; 32) Features Input to average pooling layer, output features from average pooling layer ; 33) Features With features Concatenate along the channel dimension to generate features ;feature The input is sequentially processed by a 3D convolutional layer, a layer normalization layer (BN), and a ReLU activation function layer. The ReLU activation function layer outputs features, which are then input into a sigmoid activation function, which outputs features. ; Features With features Perform element-wise multiplication to generate features ; indicates as: in, Indicates the max pooling layer. Indicates the average pooling layer. This represents a 3D convolutional layer. This represents the sigmoid activation function. This indicates element-wise multiplication.

6. The unsupervised hyperspectral greedy band selection strategy based on dual-frequency collaborative feature fusion according to claim 5, characterized in that: In section 4), three high-frequency components are included. , and The input is processed by the High Frequency Feature Fusion (HFF) module, which outputs the features. The specific process is as follows: Three high-frequency components , and The pieces are stitched together along the channel dimension, and then passed through... Convolution yields features ; indicates as: in, This indicates splicing along the channel dimension. express The convolution.

7. The unsupervised hyperspectral greedy band selection strategy based on dual-frequency collaborative feature fusion according to claim 6, characterized in that: The seven high-frequency components in step 5) , , , , , and The input is processed by the High Frequency Feature Fusion (HFF) module, which outputs the features. ; The specific process is as follows: Seven high-frequency components , , , , , and The pieces are stitched together along the channel dimension, and then passed through... Convolution yields features ; indicates as: 。 8. The unsupervised hyperspectral greedy band selection strategy based on dual-frequency collaborative feature fusion according to claim 7, characterized in that: The features in 7) and The fusion is performed to obtain the fusion features. ; indicates as: In the formula, This indicates summation pixel by pixel. This indicates the weighting coefficients set.

9. The unsupervised hyperspectral greedy band selection strategy based on dual-frequency collaborative feature fusion according to claim 8, characterized in that: In step six), extract the Pearson correlation coefficients between the bands selected in step five and all selected bands in the set `selected_bands`, and then take the average of all Pearson correlation coefficients. The mutual information corresponding to the selected bands (v) Subtract the mean As the gain of the selected band (v). ; indicates as: 。 10. The unsupervised hyperspectral greedy band selection strategy based on dual-frequency collaborative feature fusion according to claim 9, characterized in that: If the gain of the band selected in (v) is as described in (vii), then... Less than or equal to the optimal gain set in the original configuration When, then the optimal gain constant; If the gain of the selected band in (v) Greater than the optimal gain set by the original parameters When the optimal gain value is updated, the optimal gain will be adjusted accordingly. Corresponding band Select and enter the values ​​into selected_bands; specifically: 。

Images