Hyperspectral band selection method based on multi-feature enhancement

Abstract

This invention relates to the field of remote sensing image processing technology, and particularly to a hyperspectral band selection method based on multi-feature enhancement. The method involves spatial max pooling of the original hyperspectral data to obtain information representation data for the corresponding bands; filtering the original hyperspectral data using a mask constructed from the information representation data set to obtain a preliminary hyperspectral data set; dividing the preliminary hyperspectral data set and extracting multi-dimensional features from the preliminary hyperspectral data of each band; calculating the relative proximity of the corresponding bands based on the multi-dimensional features of each preliminary hyperspectral data set, and constructing an initial candidate band set by selecting the bands with the highest relative proximity; filtering the bands in the initial candidate band set; and finding the optimal band set from the resulting simplified band set. This invention significantly improves the feature representation capability of band subsets, achieving a significant improvement in computational efficiency while maintaining high classification accuracy.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing image processing technology, and in particular relates to a hyperspectral band selection method based on multi-feature enhancement. Background Technology

[0002] Hyperspectral remote sensing imaging technology boasts extremely high spectral resolution, capable of capturing hundreds or even thousands of consecutive narrow spectral bands, thus playing a crucial role in fields such as material identification, precision agriculture, and environmental monitoring. However, with the continuous improvement of spatial and spectral resolution of spaceborne and airborne sensors, the volume of hyperspectral data is experiencing explosive growth.

[0003] To address the "curse of dimensionality" in high-dimensional data, band selection (BS) has been widely studied as an effective dimensionality reduction technique, aiming to select the most representative subset of bands from the original bands. Existing band selection methods mainly include those based on ranking, clustering, evolutionary algorithms, and hybrid strategies. Among these, hybrid strategies based on subspace partitioning (such as curve fitting subspace partitioning (CFSP)) have attracted considerable attention due to their ability to better balance computational efficiency and feature representation. However, existing CFSP-like methods face two main limitations when deployed on edge hardware: limited algorithmic features and high computational complexity.

[0004] To meet the real-time processing demands of edge computing, FPGA (Field-Programmable Gate Array)-based hardware acceleration solutions have become the preferred choice over general-purpose CPUs and GPUs due to their customizable pipeline architecture and high energy efficiency. However, the design of FPGA accelerators for high-performance hybrid band selection algorithms still faces three major technical bottlenecks: 1) Hybrid algorithms typically contain a large number of branching structures and iterative logic. This irregularity in control flow conflicts with the deterministic nature of FPGA synchronous pipelines, hindering efficient parallel mapping. 2) Hyperspectral data throughput is enormous, and general computing architectures lack dedicated data flow scheduling mechanisms to overcome irregular memory access patterns, making it difficult to meet strict timing constraints.

[0005] 3) On resource-constrained edge devices, quantization is an essential optimization technique. However, existing uniform quantization strategies ignore the differences in nonlinear sensitivity to numerical precision among different algorithm operators (such as variance calculation and texture calculation). Using a uniform bit width often leads to the introduction of noise or loss of information in some features, making it difficult to achieve the optimal balance between hardware resource consumption and algorithm feature preservation.

[0006] Therefore, there is an urgent need to develop a hyperspectral band selection method and device that can fully integrate multidimensional spatial-spectral features to ensure algorithm accuracy, and overcome computational and storage bottlenecks through hardware and software co-design to achieve efficient FPGA deployment. Summary of the Invention

[0007] In view of this, the present invention aims to provide a hyperspectral band selection method based on multi-feature enhancement, which solves the problem of balancing accuracy, speed and resource efficiency in hyperspectral edge processing, and provides a complete solution for real-time intelligent processing of hyperspectral remote sensing.

[0008] To achieve the above objectives, the technical solution created by this invention is implemented as follows: A hyperspectral band selection method based on multi-feature enhancement includes: S1: Spatial max pooling is performed on the hyperspectral data of each band in the original hyperspectral data to obtain the information characterization data of the corresponding band. S2: Construct a mask based on the set of information characterization data obtained in step S1, and filter the original hyperspectral data in step S1 based on the constructed mask to obtain a preliminary set of hyperspectral data. S3: Divide the preliminary hyperspectral data set obtained in step S2 into sets according to bands, and extract the multidimensional features of the preliminary hyperspectral data of each band in each set. S4: Based on the multidimensional features of each preliminary hyperspectral data obtained in step S3, calculate the relative proximity of the corresponding bands, and select the bands with the highest relative proximity to construct an initial band candidate set. S5: Filter the bands in the initial band candidate set in step S4 to obtain a simplified band set; find the optimal band set from the simplified band set, and the optimal band set has the largest determinant of the covariance matrix.

[0009] Furthermore, step S1 includes: The raw hyperspectral data is divided into blocks, and the gray-level histogram of the raw hyperspectral data for each band is calculated for each block. Based on the gray level, the hyperspectral data of the corresponding band is spatially max-pooled to obtain the information characterization data of the corresponding band.

[0010] Furthermore, based on the gray level, spatial max pooling is performed on the hyperspectral data of the corresponding band to obtain the information representation data of the corresponding band. The process includes: In the raw hyperspectral data for each band, the frequency of gray levels appearing in each raw hyperspectral data block is calculated using the following formula: ; in, This represents the frequency of the k-th gray level appearing in the original hyperspectral data block at position (u,v). Let X represent the original hyperspectral data block at position (u,v) in the original hyperspectral data of the i-th band, where δ represents the Dirac function, and X represents the original hyperspectral data of the i-th band. i (x,y) represents the hyperspectral data at position (x,y) in the original hyperspectral data of the i-th band; The Shannon entropy for each raw hyperspectral data block is calculated using the following formula: ; Among them, H (u,v) represents the Shannon entropy of the original hyperspectral data block at position (u,v), K represents the maximum gray level, and S represents the block size; By taking the upper bound of the Shannon entropy, we obtain the information representation data of the original hyperspectral data for the current band.

[0011] Furthermore, step S2 includes: The distribution threshold is obtained by taking the mean and standard deviation of the set of information representation data and using the following formula: T=μ ε +γσ ε ; Where T represents the distribution threshold, μ ε and σ ε represents the mean and standard deviation of the set of information representation data, respectively, and γ represents the screening degree parameter; Each information representation data is compared with a distribution threshold. The bands corresponding to information representation data that are greater than the distribution threshold are retained, while the hyperspectral data corresponding to the bands of information representation data that are less than or equal to the distribution threshold are set to 0, thus obtaining a preliminary hyperspectral data set.

[0012] Furthermore, the set partitioning process in step S3 includes: Stream sampling was performed on the preliminary hyperspectral dataset to obtain the average spectrum for each band. The average spectrum of each band is calculated using the central finite difference method to obtain the first and second derivatives of the average spectrum of each band. The curvature of the average spectrum of each band on the spectral curve is calculated based on the first and second derivatives. The bands corresponding to the maximum curvature values ​​are used as the dividing boundaries to divide the preliminary hyperspectral dataset by band.

[0013] Furthermore, the multidimensional features in step S3 include spectral features, structural features, and texture features.

[0014] Furthermore, the process of calculating the relative proximity in step S4 includes: After normalizing the multidimensional features, the information entropy of each feature dimension is calculated to obtain the objective weight of each feature dimension. The weighted distance from the i-th band to the positive and negative ideal solutions is calculated using the following formula: ; ; in, and w represents the weighted distance from the i-th band to the positive and negative ideal solutions. j Represents the j-th dimension feature; The relative similarity is obtained using the following formula: ; Among them, C i This indicates the relative proximity of the i-th band.

[0015] Furthermore, the process of obtaining the simplified band set in step S5 includes: The initial band candidate set is sorted in descending order of TOPSIS score, and the first band in the initial band candidate set is taken as the baseline band; Calculate the cosine similarity between the reference band and the hyperspectral data of other bands in the initial band candidate set; The cosine similarity is compared with a preset similarity threshold. If the cosine similarity is greater than the similarity threshold, the hyperspectral data of those with a cosine similarity greater than the similarity threshold but a lower TOPSIS score are set to 0. By iterating through all bands in the simplified band set, a simplified band set is obtained.

[0016] Furthermore, step S5, which involves finding the optimal band set from the simplified band set, includes: Select hyperspectral data from a specific band in the simplified band set to form a subspace; Project the other bands in the simplified band set, excluding the selected band, into the subspace, and calculate the projection distance of the other bands into the subspace; According to the preset number of bands to select, select the bands with the largest projection distance to form the optimal band set.

[0017] Furthermore, the process of calculating the projected distance includes: The projection coefficients of the bands other than the selected bands in the simplified band set onto the subspace basis vectors are obtained by the following formula: ; Where Y represents the projection coefficient. Φ represents the hyperspectral data of the bands other than the selected bands in the simplified band set, Φ represents the subspace, and L represents the lower triangular Cholesky factor of the covariance matrix of the subspace. The projection distance is obtained by the following formula: .

[0018] Compared with the prior art, the present invention can achieve the following beneficial effects: This invention presents a hyperspectral band selection method based on multi-feature enhancement. It introduces a curve-fitting subspace partitioning strategy with multi-feature enhancement, extracts multi-dimensional features from spectral data, and employs an entropy-weighted L1-TOPSIS model for adaptive fusion, significantly improving the feature representation capability of the band subset. Simultaneously, it designs a two-stage global optimization mechanism based on recursive Cholesky updates, reducing the algorithm's search complexity from O(K) to O(K). 4 ) decreased to O(K 2 This method achieves a significant improvement in computational efficiency while maintaining high classification accuracy. Attached Figure Description

[0019] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments and descriptions of the invention are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 A schematic flowchart of the hyperspectral band selection method based on multi-feature enhancement described in the embodiments of the present invention; Figure 2 The flowchart illustrates the hyperspectral band selection method based on multi-feature enhancement as described in the embodiments of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and do not constitute a limitation thereof. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other.

[0021] The invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0022] like Figures 1 to 2 As shown in the embodiments of the present invention, the hyperspectral band selection method based on multi-feature enhancement includes: S1: Spatial max pooling is performed on the hyperspectral data of each band in the original hyperspectral data to obtain the information characterization data of the corresponding band.

[0023] Hyperspectral images typically exhibit significant spatial heterogeneity, with targets of interest (such as vehicles and buildings) often occupying only a very small spatial area. Traditional global entropy algorithms tend to reflect the average information content of an image, easily leading to the "averaging" masking of bands containing key local targets but with a monotonous background. To overcome this deficiency and achieve early data reduction, this invention proposes an adaptive gating mechanism based on spatial block partitioning. Specifically, in some embodiments, step S1 includes: S11: Divide the raw hyperspectral data into blocks, and statistically analyze the gray-level histogram of the raw hyperspectral data for each band in each block. In this embodiment of the invention, the raw hyperspectral data {X1,X2,...,X...} is divided into blocks. i ,....,X I The data is divided into U×V non-overlapping local sub-blocks according to the block size S. For the i-th band, the original hyperspectral data block at position (u,v) is represented as: ; Within each raw hyperspectral data block, the probability distribution is estimated by statistically analyzing gray-level histograms. In this embodiment of the invention, the maximum gray level is set to 256.

[0024] S12: Based on the gray level, perform spatial max pooling on the hyperspectral data of the corresponding band to obtain the information representation data of the corresponding band. In some embodiments, this process specifically includes: In the raw hyperspectral data for each band, the frequency of gray levels appearing in each raw hyperspectral data block is calculated using the following formula: ; in, X represents the frequency of the k-th gray level at position (u,v) in the original hyperspectral data block, δ represents the Dirac function, and X represents the frequency of the k-th gray level at position (u,v). i (x,y) represents the hyperspectral data at position (x,y) in the original hyperspectral data of the i-th band; To capture the most significant information features in the bands and suppress background noise interference, the Shannon entropy of each raw hyperspectral data block is calculated using the following formula: ; Among them, H (u,v) The Shannon entropy of the original hyperspectral data block at position (u,v) is represented by K, where K represents the maximum gray level. Taking the upper bound of the Shannon entropy, we obtain the information representation data of the original hyperspectral data for the current band, as shown in the following formula: ; Among them, E i Information characterization data representing the raw hyperspectral data of the i-th band.

[0025] This invention ensures that even if a band contains only one high-information sub-block (e.g., containing textured features), it can still achieve a high score by determining information representation data, thus avoiding being overwhelmed by the global smoothing effect.

[0026] S2: Construct a mask based on the set of information characterization data obtained in step S1, and filter the original hyperspectral data in step S1 based on the constructed mask to obtain a preliminary hyperspectral data set.

[0027] To achieve automated processing without human intervention, this invention constructs a dynamic threshold based on the full-spectrum entropy distribution. Specifically, in some embodiments, step S2 includes: The distribution threshold is obtained by taking the mean and standard deviation of the set of information representation data and using the following formula: T=μ ε +γσ ε ; Where T represents the distribution threshold, γ represents the screening severity parameter, used to control the strictness of the screening, and μ ε and σ ε Let represent the mean and standard deviation of the set of information representation data, respectively. The distribution threshold can also be obtained from the following formula: ; Each information representation data is compared with a distribution threshold. Bands corresponding to information representation data greater than the distribution threshold are retained, while bands corresponding to information representation data less than or equal to the distribution threshold are marked as redundant bands. The hyperspectral data corresponding to these bands are set to 0 and removed from the hardware pipeline, and do not participate in subsequent feature extraction and subspace partitioning calculations, thus obtaining a preliminary hyperspectral data set.

[0028] S3: Divide the preliminary hyperspectral data set obtained in step S2 into sets according to bands, and extract the multidimensional features of the preliminary hyperspectral data of each band in each set.

[0029] Traditional partitioning methods rely on cubic spline interpolation to fit continuous curves, resulting in high computational complexity. To reduce redundancy between bands and maintain the physical continuity of the spectrum, this invention improves the CFSP partitioning strategy. To adapt to streaming cytometry, this invention employs discrete curvature estimation. Specifically, in some embodiments, the set partitioning process in step S3 includes: Streaming sampling is performed on the preliminary hyperspectral dataset to obtain the average spectrum for each band. In this embodiment of the invention, the set of average spectra formed by the preliminary hyperspectral dataset can be represented as {λ1,λ2,...,λ,....,λ IThe streaming sampling in this embodiment of the invention is an existing method, specifically derived from Chinese Patent Publication No. CN116737349A, published on September 12, 2023, entitled "Streaming Data Processing Method, System and Storage Medium"; By performing a central finite difference calculation on the average spectrum of each band, the first and second derivatives of the average spectrum of each band can be obtained, which can be expressed by the following formula: ; ; in, This represents the first derivative of the average spectrum of the i-th band. Let represent the second derivative of the average spectrum of the i-th band; The curvature of the average spectrum on the spectral curve for each band is calculated using the first and second derivatives, as follows: ; Among them, Q i This represents the curvature of the average spectrum of the i-th band on the spectral curve; Using the bands corresponding to the curvature maxima as the dividing boundaries, the preliminary hyperspectral dataset is divided into sets according to bands.

[0030] To address the problem of insufficient perception of spatial structural information (such as texture and edges) caused by traditional methods relying solely on variance, this invention constructs a complementary multidimensional feature description. Specifically, in some embodiments, the multidimensional features include spectral features, structural features, and texture features. In this invention, the global variance of hyperspectral data for each band is used to characterize the spectral features and quantify the overall information abundance of the band. The global variance is obtained by the following formula: ; Among them, F var The global variance is represented by H and W, which represent the width and height of the hyperspectral data, respectively, and μ. i This represents the mean of the hyperspectral data in the i-th band; The edge intensity of the hyperspectral data for each band is used to characterize structural features. Specifically, the Sobel operator is used to extract gradients, and bands containing clear ground feature outlines are given higher weights, as shown in the following equation: ; Among them, F edge Indicates structural features, Represents the Sobel operator; Texture features are characterized by the local contrast of hyperspectral data in each band. Specifically, neighborhood absolute difference and micro-texture information are employed to enhance the discriminative power for smooth regions, as shown in the following equation: ; Among them, F text The texture features are represented by m and n, where m and n represent the pixels in the m-th row and n-th column of the current i-th band.

[0031] S4: Based on the multidimensional features of each preliminary hyperspectral data obtained in step S3, calculate the relative proximity of the corresponding bands, and select the bands with the highest relative proximity to construct an initial band candidate set. To identify the most representative bands from the constructed multidimensional feature space, this invention employs a decision-making strategy combining the Entropy Weight Method (EWM) and the TOPSIS model. Addressing the issue of high hardware overhead (such as square root and multiplier arrays) in traditional TOPSIS Euclidean distance calculation, we introduce a Manhattan distance metric suitable for streaming processing, optimizing computational efficiency while ensuring decision robustness.

[0032] In some embodiments, the process of calculating relative proximity includes: After normalizing the multidimensional features, the information entropy of each dimension feature is calculated to obtain the objective weight of each dimension feature. This step implicitly fixes the positive ideal solution as an all-1 vector and the negative ideal solution as an all-0 vector, avoiding the overhead of dynamically searching for the maximum value at runtime. Using the L1 norm (Manhattan distance) instead of the L2 norm, the weighted distance from the i-th band to the positive and negative ideal solutions is calculated using the following formula: ; ; in, and w represents the weighted distance from the i-th band to the positive and negative ideal solutions. j Representing the j-th dimension feature, this measurement method simplifies the operation to addition, subtraction and shift operations, which significantly reduces the DSP utilization rate of the FPGA; The relative similarity is obtained using the following formula: ; Among them, C i This indicates the relative proximity of the i-th band.

[0033] The relative proximity scores of the corresponding bands are sorted. The higher the relative proximity score, the more significant the comprehensive characteristics of the band. The bands with the highest relative proximity scores are selected. This completes the sorting based on Manhattan distance and constructs the initial band candidate set.

[0034] S5: Filter the bands in the initial candidate band set from step S4 to obtain a simplified band set; find the optimal band set from the simplified band set, the optimal band set having the largest determinant of its covariance matrix. Although local optimization selects high-quality bands, due to spectral continuity, there may still be high redundancy between candidate bands in adjacent subspaces. Directly performing a global search would result in a huge computational burden, becoming a bottleneck for real-time processing. Therefore, this invention designs a two-stage filtering and optimization strategy, which significantly compresses the search space while maximizing the information content of the selected subset.

[0035] The bands in the initial candidate band set are filtered to obtain a simplified band set, which aims to quickly eliminate candidate bands with highly similar spectra. Specifically, in some embodiments, the process of obtaining the simplified band set includes: The initial band candidate set is sorted in descending order of TOPSIS score, and the first band in the initial band candidate set is taken as the baseline band; The cosine similarity between the reference band and the hyperspectral data of other bands in the initial candidate band set is calculated as follows: ; Where sim represents the cosine similarity, X a and X b These represent the hyperspectral data of the reference band (i.e., the a-th band in the initial band candidate set) and the b-th band in the initial band candidate set, respectively. The cosine similarity is compared with a preset similarity threshold, and the hyperspectral data of those with a cosine similarity greater than the similarity threshold but a low TOPSIS score are set to 0; By iterating through all bands in the simplified band set, a simplified band set is obtained.

[0036] The screening method provided by this invention can reduce the candidate size by 40%-60%, thus offloading the computational burden for subsequent high-precision searches.

[0037] Traditional MEV algorithms typically employ a sequential forward search (SFS) strategy, requiring a complete determinant calculation for an increasingly dimensional matrix in each iteration, resulting in a computational cost of O(k). 4 This constitutes a computational bottleneck for real-time processing. To address this issue, this invention utilizes the geometric properties of the hyperellipsoidal volume to propose a recursive update algorithm without redecomposition. Specifically, in some embodiments, the process of finding the optimal band set from the simplified band set includes: According to matrix geometry theory, the determinant of the covariance matrix is ​​equivalent to the product of the projection lengths of the band vectors in the orthogonal subspace. Therefore, we first select hyperspectral data of a certain band from the simplified band set to form a subspace Φ. The lower triangular Cholesky factor L corresponding to the subspace Φ satisfies: ; in, Let represent the covariance matrix of the subspace Φ; For a candidate band in the simplified band set, its marginal contribution to the total volume is equivalent to the squared projection distance of that vector onto the orthogonal complement of the current subspace. Therefore, the other bands in the simplified band set, excluding the selected band, are projected onto the subspace, and the projection distances of the other bands onto the subspace are calculated. The process of calculating the projection distance includes: The projection coefficients of the bands other than the selected bands in the simplified band set onto the subspace basis vectors are obtained by the following formula: ; Where Y represents the projection coefficient. This represents the hyperspectral data of all bands in the simplified band set except for the selected band; The projection distance is obtained by the following formula: ; According to the preset number of bands to select, select the bands with the largest projection distance to form the optimal band set.

[0038] At this point, the band selection problem is transformed into finding the band that maximizes the squared projected distance d. 2 Maximize the frequency band: ; in, Let S represent the optimal band and S represent the simplified band set.

[0039] Once the optimal band is selected There is no need to re-decompose the new covariance matrix generated by the iteration. The new covariance matrix refers to the covariance matrix corresponding to the optimal band selection. The new Cholesky factor can be achieved by directly expanding one row and one column based on the previous Cholesky factor. ; in, Represents the new Cholesky factor, y opt d represents the projection coefficient corresponding to the optimal band. 2 max Representing the squared maximum projected distance, this update mechanism leverages the structural sparsity of the matrix, reducing the computational complexity of a single iteration from O(k...3 ) reduced to O(k 3 ).

[0040] It should be understood that the various forms of processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this invention disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this invention can be achieved, and this is not limited herein.

[0041] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A hyperspectral band selection method based on multi-feature enhancement, characterized in that, include: S1: Spatial max pooling is performed on the hyperspectral data of each band in the original hyperspectral data to obtain the information characterization data of the corresponding band. S2: Construct a mask based on the set of information characterization data obtained in step S1, and filter the original hyperspectral data in step S1 based on the constructed mask to obtain a preliminary set of hyperspectral data. S3: Divide the preliminary hyperspectral data set obtained in step S2 into sets according to bands, and extract the multidimensional features of the preliminary hyperspectral data of each band in each set. S4: Based on the multidimensional features of each preliminary hyperspectral data obtained in step S3, calculate the relative proximity of the corresponding bands, and select the bands with the highest relative proximity to construct an initial band candidate set. S5: Filter the bands in the initial band candidate set in step S4 to obtain a simplified band set; find the optimal band set from the simplified band set, and the optimal band set has the largest determinant of the covariance matrix.

2. The hyperspectral band selection method based on multi-feature enhancement according to claim 1, characterized in that, Step S1 includes: The raw hyperspectral data is divided into blocks, and the gray-level histogram of the raw hyperspectral data for each band is calculated for each block. Based on the gray level, the hyperspectral data of the corresponding band is spatially max-pooled to obtain the information characterization data of the corresponding band.

3. The hyperspectral band selection method based on multi-feature enhancement according to claim 2, characterized in that, Based on the gray level, spatial max pooling is performed on the hyperspectral data of the corresponding band to obtain the information representation data of the corresponding band. The process includes: In the raw hyperspectral data for each band, the frequency of gray levels appearing in each raw hyperspectral data block is calculated using the following formula: ； in, This represents the frequency of the k-th gray level appearing in the original hyperspectral data block at position (u,v). Let X represent the original hyperspectral data block at position (u,v) in the original hyperspectral data of the i-th band, where δ represents the Dirac function, and X represents the original hyperspectral data of the i-th band. i (x,y) represents the hyperspectral data at position (x,y) in the original hyperspectral data of the i-th band; The Shannon entropy for each raw hyperspectral data block is calculated using the following formula: ； Among them, H (u,v) represents the Shannon entropy of the original hyperspectral data block at position (u,v), K represents the maximum gray level, and S represents the block size; By taking the upper bound of the Shannon entropy, we obtain the information representation data of the original hyperspectral data for the current band.

4. The hyperspectral band selection method based on multi-feature enhancement according to claim 1, characterized in that, Step S2 includes: The distribution threshold is obtained by taking the mean and standard deviation of the set of information representation data and using the following formula: T=μ ε +gs ε ; Where T represents the distribution threshold, μ ε and σ ε represents the mean and standard deviation of the set of information representation data, respectively, and γ represents the screening degree parameter; Each information representation data is compared with a distribution threshold. The bands corresponding to information representation data that are greater than the distribution threshold are retained, while the hyperspectral data corresponding to the bands of information representation data that are less than or equal to the distribution threshold are set to 0, thus obtaining a preliminary hyperspectral data set.

5. The hyperspectral band selection method based on multi-feature enhancement according to claim 1, characterized in that, The process of partitioning the set in step S3 includes: Stream sampling was performed on the preliminary hyperspectral dataset to obtain the average spectrum for each band. The average spectrum of each band is calculated using the central finite difference method to obtain the first and second derivatives of the average spectrum of each band. The curvature of the average spectrum of each band on the spectral curve is calculated based on the first and second derivatives. The bands corresponding to the maximum curvature values ​​are used as the dividing boundaries to divide the preliminary hyperspectral dataset by band.

6. The hyperspectral band selection method based on multi-feature enhancement according to claim 1, characterized in that, The multidimensional features in step S3 include spectral features, structural features, and texture features.

7. The hyperspectral band selection method based on multi-feature enhancement according to claim 1, characterized in that, The process of calculating the relative proximity in step S4 includes: After normalizing the multidimensional features, the information entropy of each feature dimension is calculated to obtain the objective weight of each feature dimension. The weighted distance from the i-th band to the positive and negative ideal solutions is calculated using the following formula: ； ； in, and w represents the weighted distance from the i-th band to the positive and negative ideal solutions. j Represents the j-th dimension feature; The relative similarity is obtained using the following formula: ； Among them, C i This represents the relative proximity of the i-th band.

8. The hyperspectral band selection method based on multi-feature enhancement according to claim 1, characterized in that, The process of obtaining the simplified band set in step S5 includes: The initial band candidate set is sorted in descending order of TOPSIS score, and the first band in the initial band candidate set is taken as the baseline band; Calculate the cosine similarity between the reference band and the hyperspectral data of other bands in the initial band candidate set; The cosine similarity is compared with a preset similarity threshold. If the cosine similarity is greater than the similarity threshold, the hyperspectral data of those with a cosine similarity greater than the similarity threshold but a lower TOPSIS score are set to 0. By iterating through all bands in the simplified band set, a simplified band set is obtained.

9. The hyperspectral band selection method based on multi-feature enhancement according to claim 1, characterized in that, Step S5, which involves finding the optimal band set from the simplified band set, includes: Select hyperspectral data from a specific band in the simplified band set to form a subspace; Project the other bands in the simplified band set, excluding the selected band, into the subspace, and calculate the projection distance of the other bands into the subspace; According to the preset number of bands to select, select the bands with the largest projection distance to form the optimal band set.

10. The hyperspectral band selection method based on multi-feature enhancement according to claim 9, characterized in that, The process of calculating the projected distance includes: The projection coefficients of the bands other than the selected bands in the simplified band set onto the subspace basis vectors are obtained by the following formula: ； Where Y represents the projection coefficient. Φ represents the hyperspectral data of the bands other than the selected bands in the simplified band set, Φ represents the subspace, and L represents the lower triangular Cholesky factor of the covariance matrix of the subspace. The projection distance is obtained by the following formula: 。

Images