Hyperspectral classification methods and related equipment

Abstract

This invention discloses a hyperspectral classification method and related equipment. The method includes: acquiring a hyperspectral image to be classified, and performing region segmentation processing on the hyperspectral image to obtain multiple superpixel regions and their adjacency relationships; constructing a multi-scale neighborhood based on the adjacency relationships, and extracting relational features from the superpixel regions based on the multi-scale neighborhood to obtain superpixel-level features; performing multi-directional feature extraction processing on the hyperspectral image using a convolution method with orientation adjustment capabilities to obtain pixel-level features; performing multi-level feature fusion processing on the superpixel-level features and pixel-level features to obtain fused features; and determining the target category of each pixel in the hyperspectral image based on the fused features and a preset classifier. By performing multi-level fusion of superpixel-level features representing regional relationships and pixel-level features, both regional contextual information and local details can be utilized simultaneously, improving the classification accuracy of hyperspectral images.

Description

Technical Field

[0001] This invention relates to the field of image processing, and more particularly to a hyperspectral classification method and related equipment. Background Technology

[0002] Hyperspectral images contain a large number of continuous spectral channels, playing a crucial role in fine-grained scene differentiation. To extract effective spatial and spectral information from them, existing techniques typically employ methods based on local feature extraction or region relation modeling for hyperspectral image classification. These methods can, to some extent, reflect the local structural features of the image and leverage the correlations between pixels to improve classification accuracy.

[0003] However, hyperspectral images often exhibit strong spatial heterogeneity, with different land cover regions displaying complex morphologies and significant boundary variations. This makes processing methods based on regular structures or fixed neighborhood relationships difficult to adapt to real-world scenarios. In some regions, local structural orientations vary, while in others, semantic connections exist across distances between different regions. Fixed local feature extraction strategies or fixed methods for characterizing regional relationships often cannot simultaneously address these issues.

[0004] Therefore, how to provide an effective hyperspectral classification method to improve classification accuracy has become a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0005] Therefore, it is necessary to provide a hyperspectral classification method and related equipment to address the above-mentioned technical problems, which has the advantage of improving classification accuracy.

[0006] A hyperspectral classification method, the method comprising:

[0007] A hyperspectral image to be classified is acquired, and a region segmentation process is performed based on the hyperspectral image to obtain multiple superpixel regions and the adjacency relationship between the superpixel regions.

[0008] Based on the aforementioned adjacency relationship, a multi-scale neighborhood is constructed, and relational features are extracted from the superpixel region based on the multi-scale neighborhood to obtain superpixel-level features.

[0009] The hyperspectral image is processed by multi-directional feature extraction based on a convolution method with orientation adjustment capability to obtain pixel-level features;

[0010] Multi-level feature fusion processing is performed based on the superpixel-level features and the pixel-level features to obtain fused features;

[0011] Based on the fusion features and a preset classifier, the target category of each pixel in the hyperspectral image is determined.

[0012] Optionally, the region segmentation process based on the hyperspectral image to obtain multiple superpixel regions includes:

[0013] Based on the spectral features of each pixel in the hyperspectral image, the spectral similarity between each pixel is calculated.

[0014] Based on the spatial location of each pixel in the hyperspectral image, connectivity constraint processing is performed to obtain the spatial connection relationship between pixels;

[0015] Based on the spectral similarity and spatial connectivity, pixels are locally aggregated to obtain multiple superpixel regions.

[0016] Optionally, constructing a multi-scale neighborhood based on the adjacency relationship includes:

[0017] Based on the aforementioned adjacency relationship, hop-count-based adjacency expansion processing is performed to determine the degree of association of each superpixel region within multiple hop count ranges;

[0018] Based on multiple different preset thresholds, the correlation degree is subjected to hierarchical screening to determine superpixel pairs with a correlation degree not lower than the corresponding preset threshold under each preset threshold.

[0019] Each superpixel pair determined under the preset threshold is used as the neighborhood relationship at the scale corresponding to that threshold, so as to form multiple multi-scale neighborhoods with different scale characteristics.

[0020] Optionally, the adjacency expansion processing based on the adjacency relationship to determine the degree of association of each superpixel region within multiple hop count ranges includes:

[0021] The strength of the connection relationship between the multiple superpixel regions is determined based on their adjacency relationships.

[0022] Within each hop count range, multiple candidate paths are determined based on the connection strength.

[0023] The minimum value of the connection strength among the candidate paths is taken as the bottleneck strength of the corresponding candidate path.

[0024] Based on the bottleneck strength of each candidate path, a target candidate path is selected, and the bottleneck strength of the target candidate path is determined as the degree of association within the corresponding hop count range.

[0025] Optionally, the step of extracting relational features from the superpixel region based on the multi-scale neighborhood to obtain superpixel-level features includes:

[0026] Based on each of the multi-scale neighborhoods, neighborhood information aggregation processing is performed on the target superpixel region and the corresponding neighborhood superpixel regions to obtain the relation representation at each scale;

[0027] Based on the discriminative power of each relation representation in the classification task, the scale weight of the relation representation at the corresponding scale is determined.

[0028] The superpixel-level features are obtained by weighting and fusing the relational representations based on the scale weights.

[0029] Optionally, the multi-directional feature extraction processing of the hyperspectral image based on a convolution method with orientation adjustment capability to obtain pixel-level features includes:

[0030] Based on a preset asymmetric filling strategy, the features of the hyperspectral image are filled to form multiple directionally related local receptive regions;

[0031] Multi-directional feature extraction is performed based on local receptive regions with multiple directional correlations to obtain the pixel-level features.

[0032] Optionally, the multi-directional feature extraction process based on multiple directionally related local receptive regions to obtain the pixel-level features includes:

[0033] Based on learnable orientation parameters, the convolution kernel is subjected to orientation transformation to obtain convolution kernels with multiple orientations;

[0034] Based on the convolution kernels in multiple directions, convolution operations are performed on the local receptive regions related to multiple directions to obtain convolution response results in multiple directions;

[0035] The pixel-level features are obtained by concatenating the convolution response results from multiple directions.

[0036] A hyperspectral classification device, the device comprising:

[0037] The acquisition module is used to acquire the hyperspectral image to be classified, and perform region segmentation processing based on the hyperspectral image to obtain multiple superpixel regions and the adjacency relationship between the superpixel regions.

[0038] The first extraction module is used to construct a multi-scale neighborhood based on the adjacency relationship, and extract relation features from the superpixel region based on the multi-scale neighborhood to obtain superpixel-level features.

[0039] The second extraction module is used to perform multi-directional feature extraction processing on the hyperspectral image based on a convolution method with orientation adjustment capability to obtain pixel-level features;

[0040] The fusion module is used to perform multi-level feature fusion processing based on the superpixel-level features and the pixel-level features to obtain fused features;

[0041] The classification module is used to determine the target category of each pixel in the hyperspectral image based on the fusion features and a preset classifier.

[0042] A computer device includes a memory, a processor, and computer-readable instructions stored in the memory and executable on the processor, wherein the processor implements the hyperspectral classification method described above when executing the computer-readable instructions.

[0043] A readable storage medium having computer-readable instructions stored thereon, which, when executed by a processor, implement the above-described hyperspectral classification method.

[0044] The aforementioned hyperspectral classification method and related equipment acquire a hyperspectral image to be classified, and perform region segmentation processing on the hyperspectral image to obtain multiple superpixel regions and their adjacency relationships. Based on these adjacency relationships, a multi-scale neighborhood is constructed, and relational features are extracted from the superpixel regions based on the multi-scale neighborhood to obtain superpixel-level features. Multi-directional feature extraction processing is performed on the hyperspectral image using a convolution method with directional adjustment capabilities to obtain pixel-level features. Multi-level feature fusion processing is performed based on the superpixel-level features and the pixel-level features to obtain fused features. Based on the fused features and a preset classifier, the target category of each pixel in the hyperspectral image is determined. By constructing multi-scale neighborhoods at the superpixel level to obtain regional relational features, and using a convolution method with directional adjustment capabilities to extract multi-directional spatial features at the pixel level, and then fusing the two at multiple levels, local detail information and overall regional context information can be utilized simultaneously, effectively improving the feature representation ability and classification accuracy of hyperspectral images in complex scenes. Attached Figure Description

[0045] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 This is a flowchart illustrating a hyperspectral classification method according to an embodiment of the present invention;

[0047] Figure 2 This is a schematic diagram of a threshold-guided adaptive multi-hop graph construction module in one embodiment of the present invention;

[0048] Figure 3 This is a schematic diagram of the structure of a hyperspectral classification model in one embodiment of the present invention;

[0049] Figure 4 This is a schematic diagram of a convolution module with orientation adjustment capability in one embodiment of the present invention;

[0050] Figure 5 This is a schematic diagram of a hyperspectral classification device in one embodiment of the present invention;

[0051] Figure 6 This is a schematic diagram of a computer device according to an embodiment of the present invention. Detailed Implementation

[0052] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0053] In one embodiment, such as Figure 1 As shown, a hyperspectral classification method is provided, including the following steps:

[0054] 101. Obtain the hyperspectral image to be classified, and perform region segmentation processing based on the hyperspectral image to obtain multiple superpixel regions and the adjacency relationship between superpixel regions.

[0055] In this embodiment of the invention, the hyperspectral image to be classified typically consists of multiple continuous spectral channels, with each pixel carrying rich spectral features. Based on this hyperspectral image, the pixels are divided into regions to aggregate pixels with similar spectral characteristics and spatial locations into several superpixel regions.

[0056] The superpixel region can be understood as a local region composed of a group of pixels that are similar in spectral features and connected in spatial location. It is used to establish an intermediate structure between the pixel level and the region level, thereby reducing the complexity of subsequent processing and enhancing the region's expressive power.

[0057] After the region division is completed, the adjacency relationship between each superpixel region can be determined based on the relative positional relationship of pixels on the two-dimensional plane of the image, which can be used to describe the direct adjacency of different superpixel regions in space.

[0058] In practical implementation, adjacency relationships can be represented as shared boundaries, shared pixels, or pairs of superpixel regions whose spatial distance meets certain conditions within a preset range. It should be noted that region partitioning can be implemented using various methods such as those based on spectral features, spatial distance, edge information, or clustering strategies. The specific implementation method is not limited to one; any method that can obtain superpixel regions and their corresponding adjacency relationships can be applied to this invention.

[0059] 102. Based on the adjacency relationship, construct a multi-scale neighborhood, and extract relation features from the superpixel region based on the multi-scale neighborhood to obtain superpixel-level features.

[0060] In this embodiment of the invention, based on the aforementioned adjacency relationships between superpixel regions, a multi-scale neighborhood can be further constructed to characterize the correlation between superpixel regions in different spatial ranges. A multi-scale neighborhood can be understood as a hierarchical set of region relationships established within different hop counts, such as direct adjacency, crossing one hop, or crossing multiple hops, to reflect the connections between superpixel regions in local and larger-scale spatial structures.

[0061] Based on the constructed multi-scale neighborhood, relational features are extracted for each superpixel region. This involves aggregating spatial and spectral information from neighboring superpixel regions in neighborhoods at different scales, enabling each superpixel region to obtain a comprehensive representation of its surrounding structure at multiple scales. The resulting superpixel-level features can be understood as region-level feature representations that reflect the structural attributes and contextual information of the region under spatial relationships at different scales.

[0062] Specifically, the construction of multi-scale neighborhoods can adopt an adjacency expansion method based on hop count. For example, the degree of association can be determined within different hop count ranges and multiple thresholds can be used to form a set of neighborhoods with different scale characteristics. Then, aggregation processing can be performed on superpixel regions based on these neighborhoods to obtain a hierarchical relation representation.

[0063] As an alternative approach, multi-scale neighborhoods can also be constructed using methods such as distance radius expansion, graph structure hierarchies, or adaptive local region partitioning, as long as they can form a neighborhood set that describes the multi-range relationships of superpixel regions. There is no limitation to any specific construction method. Through the aforementioned feature extraction based on multi-scale neighborhoods, subsequent classification models can possess richer contextual representation capabilities at the region level.

[0064] 103. Multi-directional feature extraction processing of hyperspectral images is performed based on convolution with orientation adjustment capability to obtain pixel-level features.

[0065] In this embodiment of the invention, in order to more fully explore the spatial structure information of hyperspectral images at the pixel level, multi-directional feature extraction processing can be performed on hyperspectral images based on a convolution method with orientation adjustment capability.

[0066] Convolution methods with orientation adjustment capabilities can refer to convolution operations that can adjust the sampling direction, response direction, or receptive region shape of the convolution kernel according to preset or learnable orientation parameters, so that the same location can obtain local responses in multiple different directions, thereby adapting to possible directional changes, edge contour differences, or non-uniform texture structures in the image.

[0067] The pixel-level features obtained through the above convolution method can be understood as fine-grained spatial features extracted based on the pixel location itself, which can reflect detailed information such as the directional structure, boundary shape or texture difference of the image in the local area.

[0068] Specifically, convolution methods with orientation adjustment capabilities can employ orientation-variable convolution strategies. For example, multiple orientation-related local receptive regions can be formed based on asymmetric padding, and the convolution kernel can be oriented by combining learnable orientation parameters. This allows for convolution processing of the image in multiple directions, ultimately fusing the responses in each direction into pixel-level features.

[0069] As an alternative, the above convolution method can also be implemented by rotating the convolution kernel, dynamically selecting sampling points, constructing dilated convolution receptive regions in different directions, or using directional sensitive operators such as guided filtering, as long as the feature responses of the local pixel structure can be obtained in multiple directions. Therefore, this invention is not limited to using a specific direction adjustment mechanism; any method that can achieve multi-directional feature extraction and obtain effective pixel-level features can be applied to the method flow of this invention.

[0070] 104. Multi-level feature fusion processing is performed based on superpixel-level features and pixel-level features to obtain fused features.

[0071] In this embodiment of the invention, after obtaining superpixel-level features and pixel-level features, multi-level feature fusion processing can be performed on the two to fully combine global relationship information at the region level and local detail information at the pixel level. Superpixel-level features typically have the characteristics of large spatial range and strong contextual expression, enabling the establishment of more stable structural relationships between regions; while pixel-level features focus more on depicting fine-grained information such as edges, textures, and shapes. When the two types of features are used alone, problems such as insufficient spatial relationships or insufficient detail expression may easily occur. Therefore, it is necessary to fuse the two types of features in a unified representation space to obtain fused features that combine overall structure and local accuracy.

[0072] In practical implementation, superpixel-level features can first be spatially mapped, expanding or weighting them within the pixel domain according to superpixel regions, thus maintaining a correspondence with pixel-level features at the pixel scale. Alternatively, alignment of regional features to the pixel domain can be achieved through upsampling, feature broadcasting, or regional mean expansion. Subsequently, the aligned superpixel-level features and pixel-level features are concatenated, weighted, or adaptively fused based on attention mechanisms along the channel or feature dimensions. This allows the fusion process to automatically adjust the contribution ratio of regional and local features according to the image content. For example, in regions with complex boundaries and significant variations, the fusion process can enhance the weight of pixel-level features; while in scenarios with similar spectra and important regional relationships, the fusion process can increase the influence of superpixel-level features.

[0073] Furthermore, normalization operations, linear transformations, or small feature transformation networks can be introduced during the fusion process to enhance the compatibility between different feature sources and avoid fusion failures caused by excessive differences in feature scale or distribution. After completing the above fusion, a fused feature is obtained, which simultaneously contains structural relationship information at the region level and local detail information at the pixel level. This provides subsequent classifiers with more comprehensive and discriminative input features, thereby improving the accuracy and stability of the final hyperspectral image classification.

[0074] 105. Based on the fusion features and the preset classifier, determine the target category of each pixel in the hyperspectral image.

[0075] In this embodiment of the invention, after obtaining the aforementioned fusion features, the category of each pixel in the hyperspectral image can be determined based on a preset classifier. The fusion features integrate regional relationship information from superpixel level and directional local features from pixel level, giving them a more complete and discriminative expressive ability. Therefore, using these fusion features as input to the classifier can effectively improve the stability and accuracy of the classification results.

[0076] Specifically, the preset classifier can be any discrimination model suitable for hyperspectral data, such as fully connected neural networks, support vector machines, lightweight convolutional networks, attention-based classification modules, or other models with classification capabilities selected according to application requirements. The classifier typically performs feature mapping or nonlinear transformations on the fused features to learn the discrimination boundaries between different categories. In actual processing, the fused features can be input to the classifier pixel by pixel, and the classifier outputs the classification score or probability of each pixel in each candidate category, determining the target category of each pixel based on preset decision rules.

[0077] In some application scenarios, to further improve classification robustness, the classifier can also incorporate regularization strategies, class weight adjustments, post-processing smoothing mechanisms, or confidence filtering mechanisms to reduce the adverse effects of noisy pixels or boundary regions on the classification results. Ultimately, using the above classification process, corresponding classification labels can be accurately generated for each pixel in a hyperspectral image, enabling automated recognition and classification analysis of image content. This is applicable to various hyperspectral remote sensing applications such as agricultural monitoring, ecological identification, and ground feature analysis.

[0078] In this embodiment of the invention, a hyperspectral image to be classified is acquired, and region segmentation is performed based on the hyperspectral image to obtain multiple superpixel regions and their adjacency relationships. Based on these adjacency relationships, a multi-scale neighborhood is constructed, and relational features are extracted from the superpixel regions based on the multi-scale neighborhood to obtain superpixel-level features. Multi-directional feature extraction is performed on the hyperspectral image using a convolution method with directional adjustment capabilities to obtain pixel-level features. Multi-level feature fusion is performed based on the superpixel-level features and the pixel-level features to obtain fused features. Based on the fused features and a preset classifier, the target category of each pixel in the hyperspectral image is determined. By constructing multi-scale neighborhoods at the superpixel level to obtain regional relational features, and using a convolution method with directional adjustment capabilities to extract multi-directional spatial features at the pixel level, and then fusing the two at multiple levels, the model can simultaneously utilize local detail information and overall regional context information, effectively improving the feature representation ability and classification accuracy of hyperspectral images in complex scenes.

[0079] It is understood that in the specific implementation of this application, data such as hyperspectral images are involved. When the embodiments in this application are applied to specific products or technologies, user permission or consent is required. Furthermore, the collection, use and processing of related data, as well as the construction, training and use of convolution and classifiers, must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0080] Optionally, in the step of performing region segmentation processing based on hyperspectral images to obtain multiple superpixel regions, the spectral similarity between pixels can be calculated based on the spectral features between pixels in the hyperspectral image; connectivity constraint processing can be performed based on the spatial location of pixels in the hyperspectral image to obtain the spatial connection relationship between pixels; and local aggregation processing can be performed on pixels based on spectral similarity and spatial connection relationship to obtain multiple superpixel regions.

[0081] In this embodiment of the invention, the purpose of constructing superpixel regions is to reasonably aggregate a large number of original pixels according to their spectral characteristics and spatial structure, thereby forming regional units with similar spectral properties and spatially adjacent, making subsequent regional relationship modeling more stable, more resistant to noise, and effectively reducing computational complexity.

[0082] Specifically, the spectral similarity between pixels can be calculated based on their spectral features in a hyperspectral image. These spectral features can be the original spectral curves or spectral vectors processed by principal component compression, spectral smoothing, or noise suppression. Spectral similarity can be calculated using methods such as Euclidean distance, spectral angular distance, or correlation coefficient, and is used to measure the closeness of spectral attributes between any two pixels.

[0083] Subsequently, based on the spatial location of each pixel in the hyperspectral image, connectivity constraints are applied to the pixels to obtain the spatial connectivity relationships between them. Connectivity constraints are typically based on the neighborhood relationships of a two-dimensional image, such as determining whether pixels are directly adjacent based on 4-neighborhood or 8-neighborhood. This constraint allows for the removal of spectrally similar but spatially isolated pixels, avoiding erroneous cross-regional clustering and ensuring the spatial continuity of the generated regions.

[0084] Based on this, pixels are locally clustered according to spectral similarity and spatial connectivity to obtain multiple superpixel regions. Local clustering can employ Simple Linear Iterative Clustering (SLIC), region growing methods, energy function-based partitioning methods, or other clustering strategies that can simultaneously utilize information from spectral similarity and spatial connectivity. The goal of this clustering process is to ensure that pixels within each superpixel region maintain high spectral consistency and spatial coherence, while minimizing cross-class boundary blurring.

[0085] Through the above processing steps, multiple superpixel regions with good spectral consistency and spatial structure can be obtained to support subsequent multi-scale neighborhood construction and relation feature modeling, making the entire hyperspectral classification process more robust and precise.

[0086] Optionally, in the step of constructing a multi-scale neighborhood based on adjacency relationships, adjacency expansion processing based on hop count can also be performed based on adjacency relationships to determine the degree of association of each superpixel region within multiple hop count ranges; based on multiple different preset thresholds, the degree of association is hierarchically filtered to determine superpixel pairs with an association degree not lower than the corresponding preset threshold under each preset threshold; the superpixel pairs determined under each preset threshold are used as the neighborhood relationship under the scale corresponding to that threshold to form multiple multi-scale neighborhoods with different scale characteristics.

[0087] In this embodiment of the invention, during the construction of a multi-scale neighborhood based on adjacency relationships, hop-count-based adjacency expansion processing can be performed first to obtain the degree of association of each superpixel region within different hop count ranges. The degree of association can be obtained by the propagation, accumulation, or compression of the aforementioned connection strength within the multi-hop range, reflecting the strength of the indirect connection between any two superpixel regions in the graph structure. In specific implementation, it can be determined by combining information such as the number of paths, path strength, or graph diffusion results.

[0088] To comprehensively explore and characterize the multi-level graph structure information, after obtaining the correlation between each superpixel region, m thresholds can be set. By applying multi-threshold hierarchical filtering to assess the degree of association, a set of graph structures with differentiated topological features is constructed. Specifically, for any given threshold... The corresponding adjacency matrix can be constructed based on the correlation matrix. When the correlation between a pair of superpixel regions is not lower than a threshold At that time, in the matrix The connectivity between the superpixel regions is preserved; otherwise, it is set to zero. By changing the threshold, various graph structures, ranging from dense to sparse and from local to global, can be obtained, enabling the model to perceive spatial topological information at different scales when modeling subsequent relationships.

[0089] After obtaining the adjacency matrix set corresponding to multiple thresholds, a threshold-guided graph attention module can be introduced to focus on superpixel node features. Relationship feature extraction is performed. Specifically, the adjacency matrices at each scale and superpixel features can be jointly input into the adaptive k-hop graph attention module TKGA. The graph attention mechanism weights and aggregates neighborhood information across various graph structures, and adaptively determines the importance of different threshold graph structures in the final relation representation through multi-head attention or scale weight allocation strategies. This process can be represented as:

[0090]

[0091] in, To comprehensively consider the superpixel-level relationship feature representation after considering multiple thresholds and multi-scale topology, For different thresholds The corresponding adjacency matrix, TKGA, represents the threshold-guided adaptive k-hop graph attention module.

[0092] In another alternative implementation, multi-scale neighborhoods can also function without relying on a fixed set of discrete thresholds. Instead, they can segment or softly classify the degree of association using a continuous threshold function or a weight function based on learnable parameters. For example, a single adjacency matrix combined with a scale weight channel can be used to weighted model the association relationships in different intensity ranges, thereby achieving a "soft multi-scale" neighborhood characterization. Alternatively, the aggregation results of the same graph structure within different k-hop ranges can be jointly modeled to implicitly represent multi-scale spatial dependencies without explicitly constructing adjacency matrices corresponding to multiple thresholds. Any of the above methods can be used to construct multi-scale neighborhoods, and this invention is not limited to specific threshold setting strategies or specific implementation forms.

[0093] Optionally, in the step of performing hop-count-based adjacency expansion processing based on adjacency relationships to determine the degree of association of each superpixel region within multiple hop count ranges, the connection strength between superpixel regions can also be determined based on the adjacency relationships between multiple superpixel regions; within each hop count range, multiple candidate paths are determined based on the connection strength; the minimum value of the connection strength among each candidate path is taken as the bottleneck strength of the corresponding candidate path; based on the bottleneck strength of each candidate path, a target candidate path is selected, and the bottleneck strength of the target candidate path is determined as the degree of association within the corresponding hop count range.

[0094] In this embodiment of the invention, the connection strength can be understood as the degree of similarity between two superpixels when they are directly connected in the graph structure, for example, calculated based on spectral similarity, boundary continuity, spatial adhesion or other indicators that can reflect regional correlation.

[0095] Within a specified hop count range, several possible candidate paths can be constructed based on the strength of the connectivity relationships to represent the indirect connections between two superpixel regions established through different intermediate nodes. For each candidate path, the connectivity strength of all its edges can be obtained, and the minimum of these strengths is taken as the bottleneck strength of the path, representing the degree of connection of the "weakest link" in the path. The greater the bottleneck strength, the more reliable the indirect connections of the entire path.

[0096] Subsequently, based on the bottleneck strength of multiple candidate paths, the target candidate path with the highest score can be selected as the optimal indirect association path between the two superpixel regions within that hop count range. The bottleneck strength of this target path can then be used as the degree of association within the corresponding hop count range to construct the subsequent multi-scale neighborhood structure.

[0097] Specifically, a graph structure can be further constructed based on the adjacency relationships between multiple superpixel regions.

[0098] Where V represents the set of superpixel nodes, E represents the set of edges, and A is the edge similarity weight matrix, which describes the strength of the connection between any two directly adjacent superpixel regions.

[0099] Within a specified hop count range, for any two superpixel region nodes It is possible to determine the origin based on the adjacency relationship. arrive All candidate path sets For any path, the strength of the connectivity between its edges can be represented by A(e), where e is any edge in the path. To characterize the weakest link in the path, the minimum strength of the connection relationships among all edges in the path can be taken as the bottleneck strength of the path, which is defined as: .

[0100] Among all candidate paths, the path with the highest bottleneck strength can be further selected as the optimal path for the two nodes within this hop count range. Its maximum bottleneck strength (Maximin correlation) can be defined as:

[0101]

[0102] The Maximin value mentioned above can be used as a superpixel region. and The degree of correlation within this hop count range. When this value is greater than or equal to a preset threshold. If a node has a reliable indirect connection at that scale, then the connection between the two nodes can be considered insufficient; otherwise, the connection is considered weak. Based on this, a new adjacency matrix can be constructed. Its elements are defined as:

[0103]

[0104] The above matrix This represents the neighborhood structure obtained under given hop count and threshold conditions.

[0105] In practical applications, in order to characterize different strengths of correlations over a larger spatial range, multiple different thresholds can be set. And construct the corresponding adjacency matrix set respectively. ,in Different thresholds correspond to different scales of neighborhood structure. A lower threshold yields a dense local neighborhood, while a higher threshold yields a sparse long-range neighborhood, enabling subsequent relation feature extraction to cover multiple spatial scales from local to global.

[0106] For example, such as Figure 2 As shown, Figure 2This is a schematic diagram of a threshold-guided adaptive multi-hop graph construction module in one embodiment of the present invention. Figure 2 It can be seen that when the given threshold is 0.6, for a certain pair of superpixel regions, the bottleneck intensities obtained through multiple possible indirect paths are 0.45 and 0.70, respectively, where 0.70 is greater than 0.6. This path can be considered an effective path, and its bottleneck strength can be used as the degree of correlation at this scale; while 0.45 is less than If a path does not meet the threshold requirement, it does not constitute a neighborhood connection. This bottleneck strength-based filtering method can effectively eliminate weak connection paths and enhance the robustness of the neighborhood construction process.

[0107] Optionally, in the step of extracting relational features from superpixel regions based on multi-scale neighborhoods to obtain superpixel-level features, it is also possible to perform neighborhood information aggregation processing on the target superpixel region and the corresponding neighboring superpixel regions based on each multi-scale neighborhood to obtain relational representations at each scale; determine the scale weight of the relational representations at the corresponding scale based on the discriminative power of each relational representation in the classification task; and perform weighted fusion processing on each relational representation based on the scale weights to obtain superpixel-level features.

[0108] In this embodiment of the invention, graph relationship modeling can be performed on each multi-scale neighborhood to obtain regional relationship representations at different scales. For a target superpixel region, graph structure features can be extracted from the target superpixel region and its neighboring superpixel regions based on the neighborhood structure corresponding to that scale. Then, using neighborhood information aggregation, the features of the neighboring superpixel regions are aggregated onto the target superpixel region according to their connectivity in the graph, thereby obtaining the relationship representation at that scale. Neighborhood information aggregation can be implemented based on graph convolution, graph attention, similarity-based weighted aggregation, or other aggregation methods that can reflect local topological dependencies.

[0109] Under multi-scale conditions, the neighborhood structures constructed by different thresholds typically have different topological densities and connectivity ranges, resulting in differences in spatial dependency modeling at different scales. To fully utilize these differences, the discriminative power of relation representations at each scale for classification tasks can be further evaluated. For example, based on information such as the gradient contribution, classification confidence distribution, feature variance, or responsiveness of learnable parameters during training, corresponding scale weights can be determined, enabling the model to automatically identify the more informative scale type in different scenarios.

[0110] After obtaining the relation representations and their scale weights corresponding to multiple scales, these representations can be weighted and fused according to their scale weights to construct the final superpixel-level features. The weighted fusion method can be linear weighting, channel-wise weighting, attention-based dynamic fusion, or other feature fusion strategies that can integrate multi-scale relation information. The superpixel-level features obtained through these methods can simultaneously encode local neighborhood information and long-range dependency information across regions, giving them stronger expressive power and providing more stable and discriminative structured feature support for subsequent pixel-level feature fusion and classification prediction.

[0111] In an alternative implementation, relation feature extraction does not necessarily require building separate graph learning modules for each scale. Instead, neighborhood structures at multiple scales can be stacked and input into a single graph network, automatically learning scale differences through shared parameters. Alternatively, a hypergraph structure can be used to simultaneously model neighborhoods at different scales, allowing relation fusion to be implicitly completed during graph computation. This invention does not limit specific feature aggregation and weighting strategies; suitable implementation methods can be selected based on actual application requirements.

[0112] Optionally, in the step of performing multi-directional feature extraction processing on hyperspectral images based on convolution with orientation adjustment capability to obtain pixel-level features, the features of hyperspectral images can also be filled based on a preset asymmetric filling strategy to form multiple orientation-related local receptive regions; multi-directional feature extraction processing is then performed based on the multiple orientation-related local receptive regions to obtain pixel-level features.

[0113] In this embodiment of the invention, a direction-sensitive filling operation can first be performed on the features of the hyperspectral image based on a preset asymmetric filling strategy to form multiple direction-related local receptive regions. This asymmetric filling can ensure that the convolution kernel completely covers the local region in a specific direction, while avoiding the introduction of too much redundant information in other directions, so that subsequent convolutions can focus on structural features in different directions.

[0114] For example, in one exemplary implementation, the features denoised by 1×1 convolution can be denoted as... Where H and W represent the height and width of the feature map, respectively, and C represents the number of channels. For different directions, [the following can be done separately]. Different asymmetric filling methods can be used to obtain local receptive regions with multiple directions of correlation, such as:

[0115]

[0116] The four filling methods mentioned above correspond to four local interest directions. Different filling biases can make the convolution kernel form different effective receptive regions in space, thereby capturing local structural differences in the horizontal, vertical and tilt directions.

[0117] After obtaining local receptive regions with multiple directions, multi-directional feature extraction can be further performed based on each local region. For example, learnable directional parameters can be configured for each direction, and the convolution kernel can be spatially transformed so that it samples on the feature map according to different rotation angles or directional biases. The transformed convolution kernel is then applied to the corresponding local receptive region, resulting in convolutional responses in multiple directions. These convolutional results in multiple directions can be further processed through feature concatenation, channel-wise fusion, or attention weighting to obtain the final pixel-level features.

[0118] The aforementioned orientation adjustment capability can be achieved not only through convolution kernel rotation, but also through orientation-selective filters, deformable convolutions, orientation-sensitive attention modules, etc.; the asymmetric padding method can also be adjusted to more directions or different biases according to actual application needs to enhance the modeling capability of local orientation information.

[0119] Optionally, in the step of performing multi-directional feature extraction based on local receptive regions with multiple directions to obtain pixel-level features, the convolution kernel can be oriented by performing directional transformation based on learnable directional parameters to obtain convolution kernels with multiple directions; convolution operations can be performed on local receptive regions with multiple directions based on the convolution kernels with multiple directions to obtain convolution response results with multiple directions; and feature concatenation can be performed based on the convolution response results with multiple directions to obtain pixel-level features.

[0120] In this embodiment of the invention, the convolution kernel can be further oriented based on a learnable orientation parameter to generate multiple oriented convolution kernels for feature extraction in different directions. Specifically, a learnable angle parameter can be set for each direction. And convert it to the corresponding radian value:

[0121]

[0122] So that it can be used in space transformation operations.

[0123] For the original convolution kernel It can be done through rotation transformation function Or a rotation operator based on bilinear interpolation Rotate the convolution kernel counterclockwise by an angle α to obtain the directionally transformed convolution kernel:

[0124]

[0125] in, This represents a convolution kernel rotation operation implemented using interpolation, where α is a learnable direction angle parameter. By applying rotation transformations of different angles to the convolution kernels, a set of convolution kernels for multi-directional sensing can be constructed.

[0126] After obtaining convolution kernels in multiple directions, each convolution kernel can be applied to the corresponding local receptive regions related to those directions. This allows us to obtain convolutional response results in multiple directions. The output can be represented as:

[0127]

[0128] Where * represents convolution operation, BN( ) represents the batch normalization operation, SiLU( ) is a non-linear activation function, and b is a bias term. By performing the above convolution operation in multiple directions, convolutional activation features in different directions can be obtained.

[0129] After obtaining the convolutional response results in multiple directions, these directional features can be integrated according to the channel dimension or other dimensions based on feature concatenation methods to obtain the final pixel-level features. For example, the convolutional outputs in four directions can be combined. By concatenating along the channel dimension, combined features are obtained. This enhances directional sensitivity and local structural expression capabilities.

[0130] It should be noted that this orientation transformation operation is not limited to kernel rotation. It can also be achieved by deformable convolution, orientation attention, and orientation-encoded dynamic convolution to extract information in different directions. Similarly, the fusion of multi-directional features can also be achieved by weighted fusion, channel-by-channel selection, attention mechanism fusion, etc.

[0131] Furthermore, such as Figure 3 As shown in the diagram, this embodiment of the invention also provides a structural schematic diagram of a hyperspectral classification model. Figure 3 It can be seen that the model adopts a dual-branch structure that combines "superpixel-level relationship modeling" and "pixel-level directional feature extraction", and achieves hierarchical fusion based on multi-scale neighborhood and multi-directional convolution, thereby enhancing the feature expression capability at both the regional structure and local detail levels.

[0132] Specifically, the model first preprocesses the input hyperspectral image in terms of spectral and spatial dimensions, obtains multiple superpixel regions through a superpixel partitioning module, and constructs a multi-scale neighborhood structure based on adjacency relationships. In the superpixel-level branch, graph attention network (GAT) or other graph-based relationship modeling methods can be executed on the graph structures constructed under different threshold conditions to extract superpixel relationship features at different scales and realize the modeling of complex spatial topological relationships.

[0133] Meanwhile, in the pixel-level branch, the model performs orientation-sensitive convolutional feature extraction on the image through normalization processing and multiple DRPC (Dynamic Rotating Windmill Convolution) modules. It captures local texture features at multiple directions and scales using asymmetric padding and orientation-learnable convolutional kernels. Furthermore, a channel attention module (SEBlock) can be cascaded after the pixel-level branch, enabling the model to adaptively adjust the importance of different convolutional channels. This module integrates and weights features from different convolutional kernel sizes (e.g., 3×3, 5×5, 7×7), thereby enhancing the model's ability to perceive multi-scale and multi-directional information and improving the richness and discriminative power of the final pixel-level features.

[0134] Subsequently, the multi-scale relationship features obtained from the superpixel-level branch and the orientation-sensitive features obtained from the pixel-level branch will be integrated in the feature fusion module. The fusion method can include feature concatenation, attention-guided weighted fusion, or other fusion strategies. Through fusion, a high-dimensional feature representation that simultaneously contains regional structural information and local detail information can be obtained.

[0135] After feature fusion, the model can process the fused features through fully connected layers and a Softmax classifier to output the final class label for each pixel. To ensure that the superpixel-level branch and the pixel-level branch can optimize collaboratively, this invention can also introduce a joint training loss function, unifying the classification tasks of the two branches into a single optimization objective, achieving end-to-end joint training, so that region relationship modeling and local feature extraction promote each other during training, thereby further improving the overall classification performance.

[0136] It should be noted that, Figure 3 The functional modules, combination methods, and signal routes in the examples are a feasible implementation of the present invention. In actual deployment, the number, arrangement, or internal operation details of each module can be adjusted according to the specific data scale, target scenario, or calculation requirements. The present invention does not impose any restrictions on this.

[0137] Furthermore, such as Figure 4 As shown, this embodiment of the invention also provides a structural schematic diagram of a convolution module with orientation adjustment capability. Through Figure 4As can be seen, this module aims to enhance the model's ability to perceive directional textures and spatial structure changes by enabling the convolutional kernel to respond to features in multiple different directions through asymmetric padding and learnable orientation rotation operations.

[0138] Specifically, in Figure 4 In the pixel-level branches shown, the input feature map is first used to construct multiple orientation-dependent local receptive regions based on different asymmetric filling strategies, such as different filling methods in the vertical, horizontal, or diagonal directions, thus forming several feature blocks with orientation bias. Subsequently, the same set of basic convolutional kernels is applied to each orientation branch, and a learnable rotation angle parameter is used to perform a dynamic rotation operation on the convolutional kernels, enabling the convolutional kernels to produce an orientation-dependent rotation filtering effect in space. The rotation angle can be obtained from the learnable orientation parameter θ in the form of θ×π / 180, and the rotation of the convolutional kernel can be achieved based on spatial transformation methods such as bilinear interpolation.

[0139] After the rotated convolutional kernel performs convolution operations on the local receptive regions in each direction, the module concatenates or fuses the convolutional response results from different directions to form a feature representation with multi-directional semantics. Furthermore, an attention mechanism can be introduced to adaptively weight the response intensity in different directions, thereby highlighting salient directions and suppressing invalid directional information. The fused directional features then pass through a feature interaction module to enhance the complementarity between directional features, ultimately outputting pixel-level features with enhanced directions.

[0140] It should be noted that, Figure 4 The padding method, convolution kernel rotation method, attention structure, etc. in the examples are all feasible implementations of the present invention. In actual deployment, the number of directions, rotation angle sampling method or fusion strategy can be adjusted according to computing resources, image size or application scenario. The present invention does not limit these aspects.

[0141] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0142] In one embodiment, a hyperspectral classification device is provided, which corresponds one-to-one with the hyperspectral classification methods described in the above embodiments. For example... Figure 5 As shown, the hyperspectral classification device includes an acquisition module 501, a first extraction module 502, a second extraction module 503, a fusion module 504, and a classification module 505. Detailed descriptions of each functional module are as follows:

[0143] The acquisition module 501 is used to acquire the hyperspectral image to be classified, and perform region division processing based on the hyperspectral image to obtain multiple superpixel regions and the adjacency relationship between the superpixel regions.

[0144] The first extraction module 502 is used to construct a multi-scale neighborhood based on the adjacency relationship, and extract relation features from the superpixel region based on the multi-scale neighborhood to obtain superpixel-level features.

[0145] The second extraction module 503 is used to perform multi-directional feature extraction processing on the hyperspectral image based on a convolution method with orientation adjustment capability to obtain pixel-level features;

[0146] The fusion module 504 is used to perform multi-level feature fusion processing based on the superpixel-level features and the pixel-level features to obtain fused features;

[0147] The classification module 505 is used to determine the target category of each pixel in the hyperspectral image based on the fusion features and a preset classifier.

[0148] Optionally, the acquisition module 501 is further configured to:

[0149] Based on the spectral features of each pixel in the hyperspectral image, the spectral similarity between each pixel is calculated.

[0150] Based on the spatial location of each pixel in the hyperspectral image, connectivity constraint processing is performed to obtain the spatial connection relationship between pixels;

[0151] Based on the spectral similarity and spatial connectivity, pixels are locally aggregated to obtain multiple superpixel regions.

[0152] Optionally, the first extraction module 502 is further configured to:

[0153] Based on the aforementioned adjacency relationship, hop-count-based adjacency expansion processing is performed to determine the degree of association of each superpixel region within multiple hop count ranges;

[0154] Based on multiple different preset thresholds, the correlation degree is subjected to hierarchical screening to determine superpixel pairs with a correlation degree not lower than the corresponding preset threshold under each preset threshold.

[0155] Each superpixel pair determined under the preset threshold is used as the neighborhood relationship at the scale corresponding to that threshold, so as to form multiple multi-scale neighborhoods with different scale characteristics.

[0156] Optionally, the first extraction module 502 is further configured to:

[0157] The strength of the connection relationship between the multiple superpixel regions is determined based on their adjacency relationships.

[0158] Within each hop count range, multiple candidate paths are determined based on the connection strength.

[0159] The minimum value of the connection strength among the candidate paths is taken as the bottleneck strength of the corresponding candidate path.

[0160] Based on the bottleneck strength of each candidate path, a target candidate path is selected, and the bottleneck strength of the target candidate path is determined as the degree of association within the corresponding hop count range.

[0161] Optionally, the first extraction module 502 is further configured to:

[0162] Based on each of the multi-scale neighborhoods, neighborhood information aggregation processing is performed on the target superpixel region and the corresponding neighborhood superpixel regions to obtain the relation representation at each scale;

[0163] Based on the discriminative power of each relation representation in the classification task, the scale weight of the relation representation at the corresponding scale is determined.

[0164] The superpixel-level features are obtained by weighting and fusing the relational representations based on the scale weights.

[0165] Optionally, the second extraction module 503 is further configured to:

[0166] Based on a preset asymmetric filling strategy, the features of the hyperspectral image are filled to form multiple directionally related local receptive regions;

[0167] Multi-directional feature extraction is performed based on local receptive regions with multiple directional correlations to obtain the pixel-level features.

[0168] Optionally, the second extraction module 503 is further configured to:

[0169] Based on learnable orientation parameters, the convolution kernel is subjected to orientation transformation to obtain convolution kernels with multiple orientations;

[0170] Based on the convolution kernels in multiple directions, convolution operations are performed on the local receptive regions related to multiple directions to obtain convolution response results in multiple directions;

[0171] The pixel-level features are obtained by concatenating the convolution response results from multiple directions.

[0172] Each module in the aforementioned hyperspectral classification device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0173] In one embodiment, a computer device is provided, which may be a terminal device, and its internal structure diagram may be as follows: Figure 6 As shown, the computer device includes a processor, memory, and a network interface connected via a system bus. The processor provides computational and control capabilities. The memory includes a readable storage medium storing computer-readable instructions. The network interface communicates with external terminals via a network connection. When executed by the processor, the computer-readable instructions implement a hyperspectral classification method. The readable storage medium provided in this embodiment includes both non-volatile and volatile readable storage media.

[0174] In this application embodiment, a computer device is provided, including a memory, a processor, and computer-readable instructions stored in the memory and executable on the processor. When the processor executes the computer-readable instructions, it implements the steps of the hyperspectral classification method described above.

[0175] In one embodiment of the application, a readable storage medium is provided, which stores computer-readable instructions. When the computer-readable instructions are executed by a processor, they implement the steps of the hyperspectral classification method described above.

[0176] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by instructing related hardware with computer-readable instructions. These computer-readable instructions can be stored in a non-volatile readable storage medium or a volatile readable storage medium. When executed, these computer-readable instructions can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0177] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0178] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A hyperspectral classification method, characterized in that, The method includes: A hyperspectral image to be classified is acquired, and a region segmentation process is performed based on the hyperspectral image to obtain multiple superpixel regions and the adjacency relationship between the superpixel regions. Based on the aforementioned adjacency relationship, a multi-scale neighborhood is constructed, and relational features are extracted from the superpixel region based on the multi-scale neighborhood to obtain superpixel-level features. The hyperspectral image is processed by multi-directional feature extraction based on a convolution method with orientation adjustment capability to obtain pixel-level features; Multi-level feature fusion processing is performed based on the superpixel-level features and the pixel-level features to obtain fused features; Based on the fusion features and a preset classifier, the target category of each pixel in the hyperspectral image is determined; The construction of a multi-scale neighborhood based on the adjacency relationship includes: Based on the aforementioned adjacency relationship, hop-count-based adjacency expansion processing is performed to determine the degree of association of each superpixel region within multiple hop count ranges; Based on multiple different preset thresholds, the correlation degree is subjected to hierarchical screening to determine superpixel pairs with a correlation degree not lower than the corresponding preset threshold under each preset threshold. Each superpixel pair determined under the preset threshold is used as the neighborhood relationship at the scale corresponding to that threshold, so as to form multiple multi-scale neighborhoods with different scale characteristics. The adjacency expansion process based on the adjacency relationship to determine the degree of association of each superpixel region within multiple hop count ranges includes: The strength of the connection relationship between the multiple superpixel regions is determined based on their adjacency relationships. Within each hop count range, multiple candidate paths are determined based on the connection strength. The minimum value of the connection strength among the candidate paths is taken as the bottleneck strength of the corresponding candidate path. Based on the bottleneck strength of each candidate path, a target candidate path is selected, and the bottleneck strength of the target candidate path is determined as the degree of association within the corresponding hop count range.

2. The hyperspectral classification method as described in claim 1, characterized in that, The hyperspectral image is used for region segmentation to obtain multiple superpixel regions, including: Based on the spectral features of each pixel in the hyperspectral image, the spectral similarity between each pixel is calculated. Based on the spatial location of each pixel in the hyperspectral image, connectivity constraint processing is performed to obtain the spatial connection relationship between pixels; Based on the spectral similarity and spatial connectivity, pixels are locally aggregated to obtain multiple superpixel regions.

3. The hyperspectral classification method as described in claim 1, characterized in that, The step of extracting relational features from the superpixel region based on the multi-scale neighborhood to obtain superpixel-level features includes: Based on each of the multi-scale neighborhoods, neighborhood information aggregation processing is performed on the target superpixel region and the corresponding neighborhood superpixel regions to obtain the relation representation at each scale; Based on the discriminative power of each relation representation in the classification task, the scale weight of the relation representation at the corresponding scale is determined. The superpixel-level features are obtained by weighting and fusing the relational representations based on the scale weights.

4. The hyperspectral classification method as described in claim 1, characterized in that, The hyperspectral image is processed using a convolution method with orientation adjustment capabilities to extract multi-directional features, resulting in pixel-level features, including: Based on a preset asymmetric filling strategy, the features of the hyperspectral image are filled to form multiple directionally related local receptive regions; Multi-directional feature extraction is performed based on local receptive regions with multiple directional correlations to obtain the pixel-level features.

5. The hyperspectral classification method as described in claim 4, characterized in that, The multi-directional feature extraction process based on local receptive regions with multiple directional correlations to obtain the pixel-level features includes: Based on learnable orientation parameters, the convolution kernel is subjected to orientation transformation to obtain convolution kernels with multiple orientations; Based on the convolution kernels in multiple directions, convolution operations are performed on the local receptive regions related to multiple directions to obtain convolution response results in multiple directions; The pixel-level features are obtained by concatenating the convolution response results from multiple directions.

6. A hyperspectral classification device, characterized in that, The device includes: The acquisition module is used to acquire the hyperspectral image to be classified, and perform region segmentation processing based on the hyperspectral image to obtain multiple superpixel regions and the adjacency relationship between the superpixel regions. The first extraction module is used to construct a multi-scale neighborhood based on the adjacency relationship, and extract relation features from the superpixel region based on the multi-scale neighborhood to obtain superpixel-level features. The second extraction module is used to perform multi-directional feature extraction processing on the hyperspectral image based on a convolution method with orientation adjustment capability to obtain pixel-level features; The fusion module is used to perform multi-level feature fusion processing based on the superpixel-level features and the pixel-level features to obtain fused features; A classification module is used to determine the target category of each pixel in the hyperspectral image based on the fusion features and a preset classifier. The first extraction module is further configured to: Based on the aforementioned adjacency relationship, hop-count-based adjacency expansion processing is performed to determine the degree of association of each superpixel region within multiple hop count ranges; Based on multiple different preset thresholds, the correlation degree is subjected to hierarchical screening to determine superpixel pairs with a correlation degree not lower than the corresponding preset threshold under each preset threshold. Each superpixel pair determined under the preset threshold is used as the neighborhood relationship at the scale corresponding to that threshold, so as to form multiple multi-scale neighborhoods with different scale characteristics. The first extraction module is further configured to: The strength of the connection relationship between the multiple superpixel regions is determined based on their adjacency relationships. Within each hop count range, multiple candidate paths are determined based on the connection strength. The minimum value of the connection strength among the candidate paths is taken as the bottleneck strength of the corresponding candidate path. Based on the bottleneck strength of each candidate path, a target candidate path is selected, and the bottleneck strength of the target candidate path is determined as the degree of association within the corresponding hop count range.

7. A computer device comprising a memory, a processor, and computer-readable instructions stored in the memory and running on the processor, characterized in that, When the processor executes the computer-readable instructions, it implements the hyperspectral classification method as described in any one of claims 1 to 5.

8. A readable storage medium having computer-readable instructions stored thereon, characterized in that, When the computer-readable instructions are executed by a processor, they implement the hyperspectral classification method as described in any one of claims 1 to 5.

Images