Method for identifying spectral and spatial features of hyperspectral image

Abstract

The invention belongs to the technical field of hyperspectral image recognition, and relates to a method for recognizing spectral and spatial features of a hyperspectral image, which comprises the following steps of: dividing the hyperspectral image into N image blocks with spectral and spatial features; the method comprises the following steps: extracting and activating image blocks from a hyperspectral image, performing batch standardization processing on the image blocks to generate first-order statistical features, and performing element-by-element multiplication on the first-order statistical features and transposed outer products thereof to obtain second-order statistical features; converting the second-order statistical features into a local feature map, inputting the local feature map into a cross-orthogonal attention mechanism, calculating a horizontal attention matrix and a vertical attention matrix through an affinity matrix used for measuring correlation between feature mappings, and calculating a Chebyshev distance according to the horizontal attention matrix and the vertical attention matrix to adjust weighted features of the local feature map; and carrying out two-dimensional convolution processing on the weighted features to generate a weighted feature map, carrying out vectorization processing through a matrix, and completing identification by using a full connection layer.

Description

Technical Field

[0001] This invention belongs to the field of hyperspectral image recognition technology, specifically relating to a method for recognizing the spectral and spatial features of hyperspectral images. Background Technology

[0002] Hyperspectral remote sensing technology, due to its high-resolution capability in capturing spectral and spatial information, has been widely applied in fields such as agricultural monitoring, environmental protection, mineral exploration, urban planning, and traffic management. Hyperspectral images can provide detailed spectral information about materials through the continuous spectral curve of each pixel, giving them significant advantages in object classification, target recognition, and feature extraction. In recent years, with the rapid development of hyperspectral imaging technology and deep learning, classification research based on hyperspectral images has become an important direction in the field of remote sensing.

[0003] Classifying hyperspectral images faces several key challenges: First, hyperspectral images are highly dimensional, with spectral dimensions far exceeding those of ordinary images, significantly increasing data processing and computational complexity. Second, hyperspectral images often contain substantial amounts of redundant information and noise, which can significantly degrade the performance of classification models. Furthermore, the high cost and difficulty in obtaining labeled samples often result in small training sets, making it difficult for models to fully learn the complex features of hyperspectral data. Especially under complex environmental conditions (such as changes in illumination, shadow interference, or occlusion), the classification performance of hyperspectral images is frequently reduced by external disturbances, further limiting its application scope.

[0004] To address the aforementioned issues, numerous methods have been proposed in recent years. For instance, classification methods based on Support Vector Machines (SVM) and deep learning have achieved some success in improving the accuracy of hyperspectral image classification. Deep Support Vector Machines (DSVM) enhance classification performance through multi-kernel functions, while transfer learning methods based on Convolutional Neural Networks (CNN) utilize pre-trained models to improve adaptability for small-sample classification. Furthermore, techniques such as multi-scale feature extraction, morphological filtering, and attention mechanisms have also been widely applied in hyperspectral classification research. However, these methods often suffer from high computational complexity, strong dependence on large-scale datasets and computational resources, and insufficient stability and robustness when handling high-dimensional, high-noise data. In particular, traditional deep learning methods typically rely on large-scale pre-trained models, requiring additional optimization and parameter tuning when applied to specific scenarios or small-sample datasets, making it difficult to directly generalize to different data types and environments.

[0005] To address these issues, recent research has introduced more efficient feature extraction and classification techniques, such as attention-based classification methods and subspace learning techniques. Attention mechanisms improve the accuracy and efficiency of classification models by focusing on important spectral and spatial features, but they require significant computational resources when processing high-dimensional data. Subspace learning methods, such as Chebyshev Subspace Similarity Learning (CSSL), reduce redundant information and noise through dimensionality reduction and feature optimization, improving the model's classification ability on high-dimensional data. However, these methods still have certain limitations in practical applications, such as insufficient generalization ability to complex datasets and weak adaptability to dynamic scenes.

[0006] Therefore, to address the challenges of high dimensionality, noise interference, and adaptability to complex environments in hyperspectral image classification, a novel method combining Chebyshev subspace similarity learning and cross-orthogonal attention is proposed. This method can more efficiently extract spectral and spatial features from hyperspectral images, significantly improving classification accuracy and robustness. This approach effectively solves the bottleneck problems of traditional methods in hyperspectral image classification, providing strong technical support for remote sensing image analysis, target recognition, and scene understanding, and promoting the application and development of hyperspectral remote sensing technology in more fields. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for identifying the spectral and spatial features of hyperspectral images.

[0008] To achieve the objectives of this invention, the present invention will be implemented using the technical solutions described below.

[0009] A method for identifying spectral and spatial features of hyperspectral images: Principal component analysis (PCA) is used to divide the hyperspectral image into N image blocks. Image blocks are extracted row by row and column by column using a sliding window and activated by the ReLU function. The activated image blocks are batch normalized and the input of each layer is standardized to generate first-order statistical features. Second-order statistical features are obtained by multiplying the outer product of the first-order statistical features and their transposes element by element. The cross-orthogonal attention mechanism transforms the input second-order statistical features into local feature maps through 2D convolution. These local feature maps are then used to calculate horizontal and vertical attention matrices. Specifically, the horizontal attention matrix is ​​generated by passing the input local feature maps through two convolutional layers to create feature maps. and Through calculation The dot product between K and K is used to measure the feature mapping. and The affinity matrix is ​​calculated based on the correlation between the input and output features; the vertical attention matrix is ​​generated by convolving the input local feature maps through a convolutional layer to produce feature maps. Obtained through calculation; Chebyshev distance is calculated based on the horizontal and vertical attention matrices, and Chebyshev similarity is defined to adjust the weighted features of the local feature maps. The weighted features are processed by two-dimensional convolution to generate a weighted feature map. Positive and negative feature maps are generated through positive and negative feature processing branches respectively, and then merged into a merged feature map, generating a sum matrix. The sum matrix is ​​vectorized using a Flatten layer, and recognition is completed using a fully connected layer incorporating softmax loss. Wherein: The positive feature processing branch maps the weighted feature map to the (0,1) interval through the Sigmoid function, and optimizes the output of the Sigmoid function using a positive threshold strategy to output a positive feature map. The negative feature processing branch processes the weighted feature map through the ELU function and optimizes the output of the ELU function using a negative threshold strategy to output a negative feature map.

[0010] As a preferred embodiment of the present invention, the expression for the principal component analysis (PCA) method is: (1); In the formula, The matrix represents the original hyperspectral image data, containing multiple continuous spectral bands from the visible to the near-infrared range; the matrix By calculating the data matrix covariance matrix ,in: Let be the number of pixels in the image, and let be the eigenvector matrix obtained by performing eigenvalue decomposition on the covariance matrix, representing the principal components of the data. This represents the feature space after dimensionality reduction, where each column corresponds to a principal component.

[0011] As a preferred embodiment of the present invention, the first-order statistical feature is: (4); in: To extract the equation, , Representing an image The position index of the sliding window. This represents the extraction of the 1st dimension from the dimensionality-reduced image. Image blocks, Specify the size of the image patch. Indicates the number of channels; Represents the image patch activated by the ReLU function; This indicates batch normalization.

[0012] As a preferred embodiment of the present invention, the second-order statistical feature is: (5).

[0013] In the formula: This represents element-wise multiplication.

[0014] As a preferred embodiment of the present invention, the specific process of obtaining the horizontal attention matrix is ​​as follows: The input local feature map is passed through two convolutional layers to generate feature maps. And K, construct a set of feature maps ,in: Less than ; through calculation The dot product between K and K is used to measure the feature mapping. Affinity matrix of correlation between K and K Then, the horizontal attention matrix is ​​calculated using the horizontal attention mechanism; where: the affinity matrix The expression is: (6); in, Indicates the transpose of K; The expression for the lateral attention mechanism is: (7); (8); In the formula, Representing the affinity matrix The number of rows in the middle, This represents the attention matrix in the horizontal direction.

[0015] As a preferred embodiment of the present invention, the expression for the horizontal attention mechanism is: (9); (10); in, express the number of rows, This represents the vertical attention matrix.

[0016] As a preferred embodiment of the present invention, the weighted feature map Represented as: (15).

[0017] As a preferred embodiment of the present invention, the weighted feature map The formula for mapping to the (0,1) interval using the Sigmoid function is as follows: (16).

[0018] As a preferred embodiment of the present invention, the positive threshold strategy is specifically as follows: The formula for the positive threshold strategy is as follows: (17); when As it approaches 0, its impact on the model decreases; when Its importance increases as it approaches 1; when At that time, the corresponding features Generated by magnification of 5 times ; when At that time, the feature retains its original value without being amplified.

[0019] As a further aspect of the present invention, the threshold parameter Set it to 0.5.

[0020] As a preferred embodiment of the present invention, the negative threshold strategy is specifically as follows: The formula for the negative threshold strategy is as follows: (19); when hour, The value is multiplied by 0.5; when hour, The value remains unchanged.

[0021] As a preferred embodiment of the present invention, the process of merging the positive value feature map and the negative value feature map is described as follows: (20).

[0022] Beneficial effects: 1. Chebyshev Subspace Similarity Learning (CSSL): This method innovatively introduces Chebyshev distance to measure the maximum directional difference between different subspaces, improving the classification accuracy and computational efficiency of the model in high-dimensional hyperspectral data. This approach focuses on significant feature differences, effectively solving the problems of noise interference and redundant information in hyperspectral classification.

[0023] 2. Cross-orthogonal attention mechanism: An innovative orthogonal attention mechanism is designed, which enables horizontal and vertical attention to operate independently, effectively capturing multi-dimensional spatial information, reducing feature redundancy, and enhancing the robustness and generalization ability of the model.

[0024] 3. Feature Filtering and Enhancement (FFE) Module: A dynamic feature weight adjustment method is proposed. By using the Sigmoid function and threshold strategy, reliable features are strengthened and low-quality features are suppressed, ensuring that the model focuses on the most discriminative features and effectively improving classification performance.

[0025] 4. Innovative application of ELU activation function: To effectively utilize negative features, a processing method based on ELU function was designed to preserve the semantic information in negative values, avoid information loss, and enhance feature representation ability and network training stability.

[0026] 5. End-to-end trainable CCFFE model: Innovatively integrates CSSL, cross-orthogonal attention mechanism and FFE module into an end-to-end classification framework, ensuring the differentiability of each component and improving the efficiency and accuracy of hyperspectral image classification tasks.

[0027] 6. High practicality and validation results: Experimental validation on multiple public hyperspectral datasets shows that the CCFFE model outperforms existing hyperspectral classification methods in terms of classification accuracy, robustness, and computational efficiency, and has significant practical application value and theoretical research significance.

[0028] In summary, this invention, through the collaborative design of Chebyshev subspace similarity learning and cross-orthogonal attention mechanism, successfully solves the problems of noise interference, feature redundancy, and insufficient environmental adaptability in hyperspectral data classification, and proposes an innovative and efficient hyperspectral classification solution. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of the network architecture of the method described in this invention; Figure 2 This is a schematic diagram of the cross-orthogonal attention mechanism described in this invention; Figure 3 This is an explanatory diagram of the Chebyshev distance and Euclidean distance described in this invention; wherein: the red path is the Chebyshev distance between two points; the blue dashed line is the Euclidean distance between two points; the Chebyshev distance represented by the red path moves first along the horizontal axis and then along the vertical axis; the blue dashed line directly connects two points, representing the Euclidean distance between them; Figure 4 This is a flowchart illustrating step S8 of the present invention; Figure 5 This is a diagram showing the car classification results of different models on the QUH-Qingyun dataset according to this invention; Figure 6 This is a diagram showing the road classification results of different models on the QUH-Qingyun dataset according to this invention; Figure 7 This is a graph showing the car classification results of different models on the QUH-Pingan dataset according to the present invention; Figure 8 This is a diagram showing the road classification results of different models on the QUH-Pingan dataset according to this invention; Figure 9 This image shows the road classification results of different models on the Houston 2013 dataset according to the present invention. Detailed Implementation

[0030] The present invention will be further described in conjunction with the accompanying drawings and embodiments.

[0031] As an embodiment of the present invention, such as Figures 1 to 4 As shown, a method for identifying the spectral and spatial features of a hyperspectral image includes the following specific details: In the study of hyperspectral image classification, effectively extracting the spectral-spatial features of hyperspectral data and improving classification accuracy and efficiency has always been a pressing issue. The high dimensionality, redundancy, and complex environmental interference of hyperspectral images pose significant challenges to classification models, a problem widely present in various practical applications such as remote sensing image analysis, environmental monitoring, and traffic management. Therefore, this study proposes a feature filtering and enhancement model (CCFFE) guided by Chebyshev subspace similarity learning and cross-orthogonal attention mechanism for hyperspectral image classification tasks. The overall architecture of this model is as follows: Figure 1 As shown.

[0032] Specifically, the CCFFE model first preprocesses the hyperspectral image data using a first-order spectral operator to extract local spectral and spatial features to form data blocks, providing input support for subsequent processing. In the cross-orthogonal attention module, the model aggregates contextual information and calculates the weight distribution among multi-dimensional features, thereby generating an attention weight map for feature enhancement, and constructing a Chebyshev subspace based on this. Through Chebyshev subspace similarity learning, the model maximizes the spectral differences between pixels during dimensionality reduction, improving the model's sensitivity to key features while effectively reducing redundant information and noise interference.

[0033] (1) First-order and second-order statistical operators In the analysis and processing of hyperspectral images, the standard workflow typically includes principal component analysis (PCA) to reduce the dimensionality of the image data. This step helps extract key features from the image, reducing the complexity of subsequent processing.

[0034] PCA reduces data dimensionality by extracting the most salient features from hyperspectral images, thus simplifying data representation. This reduction in complexity not only accelerates the subsequent classification process but also improves overall efficiency. The mathematical representation of PCA is: (1); in, This represents the original hyperspectral image data matrix, containing multiple spectral bands. It is the eigenvector matrix obtained by eigenvalue decomposition of the covariance matrix, representing the principal components of the data. This represents the feature space after dimensionality reduction, where each column corresponds to a principal component.

[0035] After dimensionality reduction, the application size is A sliding window captures image patches sequentially from the image. These patches are crucial for generating features that simultaneously contain spatial and spectral information, which plays a key role in accurately understanding the image content.

[0036] Sliding window size The window size is a crucial parameter that directly impacts feature extraction performance. A larger window can capture a wider range of spatial context, which is a significant advantage for hyperspectral image classification because it enables the model to better understand the relationships between different regions in the image.

[0037] Each image patch extracted through the sliding window is treated as an independent data unit, which helps in learning localized features crucial for distinguishing categorical targets (such as different land cover types or materials). The extraction process for each image patch can be represented by the following equation: (2); in, Representing an image The position index of the sliding window. This represents the extraction of the 1st dimension from the dimensionality-reduced image. Image blocks, Specify the size of the image patch. Indicates the number of channels.

[0038] After image patch extraction, the image is activated using the ReLU function to enhance the model's nonlinear representation capability. The specific definition is as follows: (3); The activated data undergoes batch normalization to improve the model's convergence speed and generalization ability, and the input of each layer is standardized. This process generates first-order statistical features, expressed as: (4); First-order statistical features are fundamental elements for capturing basic spatial-spectral information in hyperspectral images. These features are crucial for distinguishing different categories and provide a robust data foundation for subsequent advanced processing techniques, thereby ensuring more accurate and efficient classification results.

[0039] To further enhance feature representation, second-order statistical features are computed. Second-order operators are used to capture higher-order dependencies between features that cannot be directly obtained from first-order statistical features. Specifically, second-order features... It is obtained by element-wise multiplication of the outer product of the first-order statistical feature and its transpose, as defined below: (5); In the formula: This represents element-wise multiplication.

[0040] This operation, often referred to as feature fusion or interaction, allows models to integrate and utilize the relationships between different features. In hyperspectral images, by capturing these interrelationships, models are able to better represent complex patterns and structures, thereby improving classification performance. Therefore, second-order features provide richer, more expressive representations, complementing the information obtained from first-order statistical features.

[0041] (2) Cross-orthogonal attention mechanism Attention mechanisms in remote sensing hyperspectral classification are a key technology for improving classification accuracy and efficiency. By focusing on the most informative features, attention mechanisms effectively reduce redundant data, lower the computational burden on the model, and reduce the risk of overfitting. This selective attention also suppresses noise and irrelevant information, showing significant advantages, especially in tasks involving complex backgrounds or distinguishing spectrally similar categories. Remote sensing data processing often faces various complex scenarios, such as image data from different seasons and weather conditions. Attention mechanisms, by dynamically adjusting the focus, enable the model to better adapt to these changes, maintain stable classification performance in variable environments, and enhance the generalization ability of deep learning models.

[0042] Inspired by the cross-attention model, Criss Orthogonal Attention further enhances feature representation capabilities by simultaneously considering contextual information in both the horizontal and vertical directions of the feature map. Its core innovation lies in introducing orthogonal constraints to decouple attention mechanisms in different directions. This design ensures that horizontal and vertical attention operate independently, effectively reducing redundancy and avoiding potential mutual interference. By capturing diverse and complementary features across more dimensions, the Criss Orthogonal Attention mechanism significantly improves the model's ability to distinguish between different categories, strengthening the classification performance of hyperspectral images.

[0043] The overall structure of the Criss Orthogonal Attention mechanism, such as Figure 2 As shown, the mechanism is accomplished through a series of steps: first, attention maps in the horizontal and vertical directions are generated through 1×1 convolutional layers; then, affinity matrices are calculated, and orthogonal constraints are applied to ensure that the attention mechanisms in the horizontal and vertical directions operate independently.

[0044] Unlike traditional attention mechanisms (such as spatial attention or channel attention), which primarily focus on enhancing specific regions or channels of the feature map, cross-orthogonal attention processes feature variations orthogonally in both spatial and channel dimensions. This approach is able to capture more detailed and comprehensive feature representations when processing high-dimensional data (such as hyperspectral images).

[0045] By introducing orthogonal constraints, the cross-orthogonal attention mechanism has the following key advantages: Reduced redundancy: By operating spatial and channel attention independently, feature redundancy between different dimensions is reduced, resulting in more efficient feature representation.

[0046] Enhanced generalization ability: Independently captures diverse features across different dimensions, enabling the model to adapt to complex scenarios, especially suitable for high-dimensional data applications such as hyperspectral image classification.

[0047] Enhanced robustness: Through independent feature learning, the model exhibits stronger stability and robustness when facing data changes.

[0048] In summary, the cross-orthogonal attention mechanism effectively improves the model's classification performance and generalization ability in hyperspectral data through orthogonal constraints and multidimensional feature learning, making it suitable for complex and ever-changing remote sensing scenarios.

[0049] a. Horizontal attention First-order features As fundamental elements for capturing important spatial-spectral information in hyperspectral images, second-order features are computed to further enrich these features. By combining convolution operations The outer product of the transpose and its transpose is used to effectively capture high-order dependencies between features.

[0050] Subsequently, these second-order features are transformed into local feature maps through 2D convolution operations. The model uses two... Convolutional layer pairs The process generates two feature maps. and .in, ,in Less than This effectively reduces the dimensionality.

[0051] Affinity Matrix Through calculation The dot product between K and is used to measure the feature mapping. and The correlation between them. Affinity matrix. Normalization is performed using the softmax function, and its calculation formula is as follows: (6); in, express The transpose of .

[0052] Let the affinity matrix be denoted as In hyperspectral image classification tasks, in order to classify targets, it is necessary to base... The specific formula for calculating horizontal attention is as follows: (7); (8); in, Representing the affinity matrix The number of rows in the middle, This represents the attention matrix in the horizontal direction.

[0053] b. Vertical attention one Convolutional layers are used for Perform convolution operations to generate To highlight the characteristics of different ground features in hyperspectral images, based on The formula for calculating vertical attention is as follows: (9); (10); in, express the number of rows, This represents the vertical attention matrix.

[0054] This mechanism improves classification performance by calculating attention weights in the vertical direction, highlighting important features in hyperspectral images that are relevant to land cover classification.

[0055] (2) Chebyshev subspace similarity learning In hyperspectral images, each pixel corresponds to a high-dimensional spectral vector, representing reflectance at different wavelengths. Chebyshev distance effectively captures these differences by considering the maximum discrepancies across each dimension (wavelength) of the spectral vector, making it particularly suitable for pixels exhibiting significant differences within a specific wavelength range. Hyperspectral data typically contains a large number of wavelengths, some of which may contain noise or irrelevant information. Compared to Euclidean distance, Chebyshev distance is more sensitive to outliers because it focuses on the maximum deviation between dimensions, rather than a comprehensive measure of differences across all dimensions. This gives Chebyshev distance a significant advantage in hyperspectral data, particularly for detecting significant differences at key wavelengths, while also demonstrating excellent computational efficiency and the ability to identify key differences. The computation of Chebyshev distance is relatively simple, especially in high-dimensional space, requiring only the determination of the maximum difference between each wavelength. This makes its computation speed far faster than Euclidean distance and other more complex metrics, making it particularly important for real-time or large-scale hyperspectral data processing tasks. The differences between Chebyshev distance and Euclidean distance, such as... Figure 3 As shown.

[0056] Our method introduces the concept of Chebyshev subspace similarity, which plays a crucial role in capturing the maximum directional differences between different subspaces in hyperspectral data. Chebyshev Similarity Based on the definition of Chebyshev distance, this distance measures the maximum difference between corresponding elements of two matrices.

[0057] Given two attention matrices—the horizontal attention matrix and the horizontal attention matrix. and vertical attention matrix The formula for calculating the Chebyshev distance is as follows: (11); (12); in, This represents the position of an element in the matrix. This distance captures... and The maximum absolute difference between corresponding elements provides a robust measure of significant changes in both directions.

[0058] To quantify the similarity between these subspaces, we define Chebyshev similarity. as follows: (13); Due to the symmetry of the Chebyshev distance, this similarity can ultimately be simplified to... The value of similarity. It reflects the degree of correlation between subspaces, mainly measuring their proximity in terms of maximum directional difference.

[0059] Chebyshev similarity This approach is crucial for enhancing the robustness and accuracy of hyperspectral classification. By focusing on the maximum differences, it ensures the model highlights the most critical features, thus effectively distinguishing classes. This method facilitates feature selection and weighting, prioritizing features with significant directional differences. Furthermore, it improves the model's robustness to noise and data variations because it focuses on the most discriminative features.

[0060] Furthermore, Chebyshev similarity provides a deeper understanding of the consistency of feature orientations within a subspace. Higher similarity values ​​indicate consistency of features across different directions, which is crucial for ensuring the generalization and stability of models in hyperspectral data analysis.

[0061] Euclidean distance measures overall difference across all dimensions, which can dull significant differences in key features due to averaging. In contrast, Chebyshev distance focuses on the maximum difference, making it more effective at capturing crucial variations in high-dimensional data such as hyperspectral images. While Mahalanobis distance considers correlations between features, it is often computationally complex and sensitive to the quality of the covariance matrix, especially in high-dimensional spaces. Chebyshev distance is simpler and more robust to compute because it focuses only on the maximum difference between features, making it well-suited for scenarios with limited data.

[0062] Deep learning methods typically require large datasets and significant computational resources, and their decision-making processes are difficult to interpret. In contrast, Chebyshev similarity offers a more interpretable and computationally efficient alternative, particularly suitable for applications requiring high transparency and efficiency.

[0063] Given the significant advantages of Chebyshev similarity in terms of computational efficiency, robustness, and interpretability, it has been integrated as a core component of our hyperspectral classification framework. After calculating the Chebyshev similarity between two attention matrices, this similarity is used to adjust the values ​​derived from second-order features. Weighting of derived local feature maps ensures that the most critical features are emphasized.

[0064] Subsequently, local feature maps Weighted feature adjustment is performed using the following formula: (14); This adjustment process enables the output matrix to better capture the relationships and similarities between the input matrices, ultimately yielding the second-order features most relevant to the classification task.

[0065] In summary, Chebyshev subspace similarity learning provides a robust and efficient method for hyperspectral data analysis. By focusing on the most significant differences within the subspace, this method improves classification accuracy while maintaining computational efficiency. This makes it particularly suitable for real-time or large-scale hyperspectral data processing tasks where both speed and accuracy are critical.

[0066] (3) Feature filtering and enhancement Convolutional Neural Networks (CNNs) effectively capture feature dependencies and patterns within local regions through their convolutional layers, which is crucial for understanding the spatial distribution and contextual relationships of objects in an image. To ensure that the model can acquire features in the spectral dimension and learn their spatial distribution and interactions, thereby enhancing the model's generalization ability, this invention focuses on second-order features. Two-dimensional convolution was employed. This process generates a feature map, represented as follows: : (15); While the initial processing steps offer certain advantages, the features obtained at this stage still contain a significant amount of redundant information, which can negatively impact the classification performance of roads and vehicles in hyperspectral images. Therefore, these features must be further optimized to filter out interfering signals and enhance the model's discriminative ability. To address this issue, we designed a conditional filtering mechanism that improves model performance by rewarding salient features and penalizing redundant features. Ultimately, this mechanism significantly enhances the classifier's performance. A flowchart of the method is shown below. Figure 4 As shown.

[0067] The core concept of the Feature Filtering and Enhancement (FFE) module is to map the feature matrix output by Chebyshev Subspace Similarity Learning (CSSL) to the (0,1) interval using the Sigmoid function. The calculation formula is as follows: (16); like Figure 4 As shown, the FFE module processes the output of the Sigmoid function and applies a thresholding strategy to further optimize the features. Specifically, when At that time, the corresponding features Magnified 5 times, thus generating This amplification aims to enhance the impact of these more reliable features on the classification results. Conversely, when In this case, the features retain their original values ​​without being amplified, thus preventing unreliable features from excessively influencing the classification results. The formula is as follows: (17); when As it approaches 0, its impact on the model decreases; when Its importance increases as it approaches 1. In this study, the threshold parameter... Setting it to 0.5 ensures that only positive features are significantly enhanced, thereby increasing their impact on classification results. For negative features, their original values ​​are retained to avoid their negative impact on model performance. This setting allows the model to prioritize the most informative features, thus comprehensively improving classification accuracy and robustness. Furthermore, It is considered a hyperparameter that can be adjusted according to specific datasets and application scenarios. This adaptability makes the model more general and effective in different hyperspectral scenarios and operating environments.

[0068] Considering that even negative values ​​in the feature points of the convolved hyperspectral image (HSI) may contain semantic information, this invention plans to process the features output from CSSL using the ELU function instead of the ReLU function to remove negative values ​​from the feature map. The calculation formula is as follows: (18); in, This is an adjustable parameter used to control the saturation level of the negative portion of the ELU function. The final threshold strategy calculation formula is as follows: (19); when hour, The value is multiplied by 0.5. This method reduces the model's sensitivity to negative inputs and reduces interference caused by noise or other redundant information. Furthermore, it reduces the gradient magnitude in the affected region, which helps control the range of gradients during training, thereby enhancing the stability of network training. When hour, The value remains unchanged.

[0069] Next, the feature maps output from the two branches are merged to generate a feature map sum matrix. The merging process is described below: (20); Finally, the output features are vectorized through a Flatten layer. Classification is then performed using a fully connected (dense) layer incorporating softmax loss. Within this framework, a novel CCFFE model for hyperspectral image classification was developed. It is worth noting that because each component of the CCFFE model is differentiable, the entire network can be trained end-to-end.

[0070] Evaluation indicators To quantify the performance of the CCFFE model in road and vehicle classification, this invention employs three evaluation metrics: Kappa coefficient, average accuracy (AA), and overall accuracy (OA).

[0071] Overall Accuracy (OA): In hyperspectral image classification research, Objective Allocation (OA) is a commonly used metric to measure the overall performance of a classifier. OA represents the proportion of correctly classified samples out of the total number of samples. This metric provides a quick benchmark to reflect the classifier's performance across the entire dataset. However, a potential drawback of this metric is that it can be affected by imbalanced datasets. If the number of samples in one class is significantly greater than those in other classes, the classifier may tend to favor this majority class, thereby improving overall accuracy, but this does not necessarily mean that the classification performance for all classes is satisfactory. The formula for calculating OA is as follows: (twenty one); in, Indicates the total number of categories. Indicates the first The number of samples that are correctly classified in the class. Indicates the first Total number of samples in the class.

[0072] This metric can be used to quickly evaluate the model's performance in overall classification tasks, but it should be combined with other metrics for comprehensive evaluation when dealing with imbalanced data.

[0073] Average Accuracy (AA): In hyperspectral image classification, AA (Average Accuracy) is used to measure the average performance of a classifier across all classes. AA is the arithmetic mean of the single-class accuracy for each class. Single-class accuracy represents the proportion of correctly classified samples in a given class out of the total number of samples in that class. By calculating the average of the single-class accuracy across all classes, AA provides a comprehensive metric that reflects the overall performance of the classifier across all classes. The formula for calculating AA is as follows: (twenty two); in, Indicates the total number of categories. Indicates the first The number of correctly classified samples in a class. Indicates the first Total number of samples in the class.

[0074] Kappa coefficient: The Kappa coefficient is a statistical metric used to evaluate classification accuracy, particularly reliable in imbalanced data. The Kappa statistic considers the random consistency present in the data, thus providing a more comprehensive performance evaluation than accuracy alone (the proportion of correctly classified samples). In hyperspectral image processing, due to the high dimensionality of the samples and the complexity of land cover categories, accuracy alone may not fully reflect the classifier's performance. The Kappa coefficient, by considering the influence of random classification, provides a more accurate and robust performance evaluation method. The formula for calculating the Kappa coefficient is as follows: (twenty three); in, Indicates overall accuracy (OA). The probability of random classification is expressed by the following formula: (twenty four); in, Indicates the first The sum of the elements in the column, This represents the total number of samples.

[0075] The Kappa coefficient, by evaluating the deviation between the classification result and the random classification, can more accurately measure the performance of the classifier in imbalanced data and is one of the important indicators in hyperspectral image classification research.

[0076] Experimental environment The CCFFE model proposed in this invention is built using Python and the Keras deep learning framework. The experimental model runs in the following environment: Table 1 Model Experimental Environment To minimize bias caused by different training samples, this study conducted 20 experiments under the same conditions and took the average result to ensure the reliability of the experimental results.

[0077] Table 2 Model Experiment Parameter Settings Through the above-mentioned environmental configuration and parameter adjustment, this study ensured the stability and efficiency of model training, providing reliable support for the performance verification of the CCFFE model in hyperspectral data classification.

[0078] Experimental results To evaluate the performance of the CCFFE model, this study compared it with several classic models and state-of-the-art deep learning models. Classic models included 2DSSA, F-SSA, DenseNet, and ResNet, while state-of-the-art deep learning models covered Tri-CNN, MCNN-CP, SAGP, SSUN, and RIAN. These models were selected as baselines for performance comparison with the CCFFE model, comprehensively validating the accuracy, robustness, and superiority of CCFFE in hyperspectral image classification tasks.

[0079] (1) Comparison experiment of QUH-Qingyun dataset Experimental results on the QUH-Qingyun dataset demonstrate that the CCFFE model performs exceptionally well in handling imbalanced sample classes and complex background interference (such as shadow effects). The "vehicle" category has the lowest sample proportion and is most severely affected by shadows in this dataset, increasing classification difficulty. However, the CCFFE model still achieves a significant advantage, surpassing other comparative models with an overall accuracy (OA) of 95.45%, an average accuracy (AA) of 93.22%, and a Kappa coefficient of 93.97%. Machine learning models (such as 2DSSA and F-SSA) perform poorly in classification due to their inability to effectively handle the high-dimensionality of hyperspectral data. Deep learning models (such as ResNet and DenseNet), while offering improved accuracy, are limited in utilizing the rich spectral information of hyperspectral data because they were originally designed for RGB images. In contrast, the CCFFE model effectively integrates spectral and spatial information by introducing Chebyshev similarity learning and cross-orthogonal attention mechanisms, achieving the best performance, particularly in the "vehicle" (79.52%) and "road" (94.06%) categories. This demonstrates that the CCFFE model exhibits superior robustness and generalization ability in hyperspectral data classification tasks.

[0080] Table 3. Classification results of different models on the QUH-Qingyun dataset.

[0081] Note: The colors for QT3:Car and QT6:Road correspond to... Figure 5 and Figure 6 The categories shown are: Figure 5 Each image in the table represents the model representation in the Car (QT3) class as listed in Table 3; Figure 6 Each image in the table represents the classification output of the Road (QT6) class, as listed in Table 3.

[0082] like Figure 5 and Figure 6As shown, in road and vehicle classification tasks, deep learning-based F-SSA and 2DSSA models are significantly affected by noise, resulting in a large number of misclassified pixels in the final classification results. The QUH-Qingyun dataset contains complex shadows cast by roads, buildings, and vegetation, which pose a particular challenge to vehicle recognition and cause models such as DenseNet and ResNet, which cannot extract deep features, to misclassify road targets as buildings or other categories. In vehicle recognition, the classification accuracy of all models is not high, further illustrating the impact of data imbalance in the hyperspectral dataset. This imbalance also makes MCNN-CP and Tri-CNN perform poorly in extracting spatial features, but they still outperform SSUN, RIAN, and MSPF-NFAT models in vehicle classification accuracy.

[0083] In comparison, the CCFFE model proposed in this invention demonstrates significant advantages in vehicle classification and road classification tasks, achieving a vehicle classification accuracy of 79.52% and a road classification accuracy of 94.06%, respectively, far exceeding other comparative models. This indicates that the CCFFE model exhibits superior robustness and accuracy when dealing with complex backgrounds and data imbalance issues in hyperspectral datasets.

[0084] (2) Comparative experiment of QUH-Pingan dataset Experimental results on the QUH-Pingan dataset demonstrate that the CCFFE model significantly outperforms other comparative models in classification performance (as shown in Table 4). Despite the highly dispersed distribution of different land cover types in the QUH-Pingan dataset and the extreme imbalance of vehicle samples compared to other samples, posing a significant challenge to the classification task, the CCFFE model still achieves a classification accuracy of 91.94% in the vehicle target classification task, showcasing its powerful ability to extract detailed land cover features. Simultaneously, the model also achieves a high accuracy of 99.38% in the road target classification task, fully demonstrating its superiority in large-scale land cover feature extraction. In contrast, CNN-based models (such as MCNN-CP and Tri-CNN) perform poorly when handling vehicle targets with a small sample size, reflecting the limitations of CNN networks in dealing with data imbalance. Although attention-based models have improved vehicle target classification accuracy, they are still affected by the dispersed sample distribution and limited sample quantity, making it difficult to further improve classification accuracy.

[0085] Table 4. Classification results of different models on the QUH-Pingan dataset.

[0086] Note: The colors for PT9: Car and PT10: Road correspond to... Figure 7 and Figure 8 The categories shown; where: Figure 7 Each image in the table represents the model representation in the Car (vehicle) class (PT9) as listed in Table 4; Figure 8 Each image in the table represents the classification output of the Road (PT10) class, as listed in Table 4.

[0087] like Figure 7 and Figure 8 As shown, the superior performance of the CCFFE model is clearly demonstrated visually, with clear and accurate vehicle and road classification results and fewer misclassified noise points. In contrast, machine learning-based models (such as 2DSSA and F-SSA) are unable to effectively handle high-dimensional data and have insufficient feature extraction capabilities, resulting in significant noise or incomplete target classification results in vehicle and road categories. Deep learning-based models, with their deeper network architecture, possess stronger feature extraction capabilities and therefore perform better in classification results. However, in vehicle classification tasks, due to the small size and dispersed distribution of target samples, some models still suffer from misclassification. Overall, the CCFFE model significantly improves classification accuracy through effective high-dimensional feature extraction and noise suppression, especially excelling in vehicle and road target recognition.

[0088] (2) Comparative experiment using the Houston 2013 dataset The Houston 2013 dataset is a widely used benchmark dataset containing a wide variety of road sample types. In this study, only 1% of the samples were used as the training set to evaluate the classification ability of the CCFFE model. (See Table 5 and...) Figure 9 As shown, the CCFFE model excels in capturing detailed features over large regions and accurately classifying various road types. While other models also perform well on the Houston 2013 dataset, they often underperform in fine-grained classification tasks. The CCFFE model effectively assigns attention weights to multidimensional features through a Criss-Cross Orthogonal Attention mechanism, fully utilizing the rich spectral information in hyperspectral images. Simultaneously, Chebyshev Subspace Similarity Learning further enhances the ability to distinguish similar pixel features, thereby improving the model's overall classification accuracy.

[0089] Table 5. Classification results of different models on the Houston 2013 dataset.

[0090] Note: The different colors for HT9 (Road), HT10 (Highway), and HT11 (Railway) correspond to... Figure 9 The categories shown are: Figure 9 Each image in the table represents the classification output of the Road, Highway, and Railway classes (HT9, HT10, and HT11) as listed in Table 5.

[0091] like Figure 9 As shown, the CCFFE model proposed in this invention performs excellently in capturing dataset details, and can comprehensively represent the features of large-scale hyperspectral remote sensing images. A key challenge of the Houston 2013 dataset is that its sample annotation ratio is relatively small, and a large number of samples lack accurate labels. In this case, if the model expands the prediction boundary to improve the classification accuracy of labeled regions during training, it may lead to overfitting of unlabeled regions, resulting in the loss of image details.

[0092] The CCFFE model effectively addresses this issue through Chebyshev Subspace Similarity Learning, enabling it to distinguish different types of samples at the pixel level, thus achieving more refined classification results. This mechanism not only enhances the model's ability to capture detailed features in hyperspectral images but also significantly improves the robustness and accuracy of classification, fully demonstrating its advantages in dealing with the problem of scarce labeled samples.

[0093] Other materials This experiment evaluated the classification performance of the CCFFE model for vehicles and roads in hyperspectral images. Details of the four public datasets used are shown in Tables 6 to 8.

[0094] (1) QUH-Qingyun dataset The data was acquired by the drone while flying at an altitude of 300 meters. The image size is 880×1360 pixels, containing 270 bands with a wavelength range of 400-1000 nanometers and a spatial resolution of approximately 0.15 meters.

[0095] Table 6 Land Cover Types and Number of Training Samples in the QUH-Qingyun Dataset

[0096] (2) QUH-Pingan dataset The data was acquired by the drone while flying at an altitude of 200 meters. The image size is 1230×1000 pixels, containing 176 bands with a wavelength range of 400-1000 nanometers and a spatial resolution of approximately 0.10 meters.

[0097] Table 7 Land cover types and number of training samples in the QUH-Pingan dataset

[0098] (3) Houston 2013 dataset The image size is 349×1905 pixels, contains 144 bands, has a wavelength range of 380-1050 nanometers, and a spatial resolution of approximately 2.5 meters.

[0099] Table 8 Land cover types and number of training samples in the Houston 2013 dataset

[0100] During model training, appropriate training sample ratios were adopted based on the characteristics of each dataset, and model parameters were optimized using 10-fold cross-validation. Experiments involved multiple training runs to determine the optimal hyperparameters, avoiding overfitting and thus improving the model's generalization ability and robustness. These datasets provide a reliable benchmark for a comprehensive evaluation of hyperspectral remote sensing image classification performance.

Claims

1. A method for identifying spectral and spatial features of a hyperspectral image, characterized in that: The hyperspectral image is divided into N image blocks using principal component analysis (PCA). The image blocks are extracted row by row and column by column using a sliding window and activated by the ReLU function. The activated image blocks are then batch normalized, and the input of each layer is standardized to generate first-order statistical features. Second-order statistical features are obtained by multiplying the outer product of the first-order statistical features and their transposes element by element. The cross-orthogonal attention mechanism transforms the input second-order statistical features into local feature maps through 2D convolution. These local feature maps are then used to calculate horizontal and vertical attention matrices. Specifically, the horizontal attention matrix is ​​generated by passing the input local feature maps through two convolutional layers to create feature maps. and Through calculation The dot product between K and K is used to measure the feature mapping. and The affinity matrix is ​​calculated based on the correlation between the input and output features; the vertical attention matrix is ​​generated by convolving the input local feature maps through a convolutional layer to produce feature maps. Obtained through calculation; Chebyshev distance is calculated based on the horizontal and vertical attention matrices, and Chebyshev similarity is defined to adjust the weighted features of the local feature maps. The weighted features are processed by two-dimensional convolution to generate a weighted feature map. Positive and negative feature maps are generated through positive and negative feature processing branches respectively, and then merged into a merged feature map, generating a sum matrix. The sum matrix is ​​vectorized using a Flatten layer, and recognition is completed using a fully connected layer incorporating softmax loss. Wherein: The positive feature processing branch maps the weighted feature map to the (0,1) interval through the Sigmoid function, and optimizes the output of the Sigmoid function using a positive threshold strategy to output a positive feature map. The negative feature processing branch processes the weighted feature map through the ELU function and optimizes the output of the ELU function using a negative threshold strategy to output a negative feature map.

2. The method for identifying spectral and spatial features of a hyperspectral image according to claim 1, characterized in that: The expression for the principal component analysis (PCA) method is: ； In the formula, The matrix represents the original hyperspectral image data, containing multiple continuous spectral bands from the visible to the near-infrared range; the matrix By calculating the data matrix covariance matrix ,in: Let be the number of pixels in the image, and let be the eigenvector matrix obtained by performing eigenvalue decomposition on the covariance matrix, representing the principal components of the data. This represents the feature space after dimensionality reduction, where each column corresponds to a principal component.

3. The method for identifying spectral and spatial features of a hyperspectral image according to claim 2, characterized in that: The first-order statistical feature is: ； in: To extract the equation, , Representing an image The position index of the sliding window. This represents the extraction of the 1st dimension from the dimensionality-reduced image. Image blocks, Specify the size of the image patch. Indicates the number of channels; Represents the image patch activated by the ReLU function; This indicates batch normalization.

4. The method for identifying spectral and spatial features of a hyperspectral image according to claim 3, characterized in that: The second-order statistical feature is: In the formula: This represents element-wise multiplication.

5. The method for identifying spectral and spatial features of a hyperspectral image according to claim 4, characterized in that: The specific process of obtaining the horizontal attention matrix is ​​as follows: The input local feature map is passed through two convolutional layers to generate feature maps. and Construct a set of feature maps ,in: Less than ; through calculation The dot product between K and K is used to measure the feature mapping. and Affinity matrix of correlation between Then, the horizontal attention matrix is ​​calculated using the horizontal attention mechanism; where: the affinity matrix The expression is: ； in, express Transpose of; The expression for the lateral attention mechanism is: ； ； In the formula, Representing the affinity matrix The number of rows in the middle, This represents the attention matrix in the horizontal direction.

6. The method for identifying spectral and spatial features of a hyperspectral image according to claim 5, characterized in that: The expression for the horizontal attention mechanism is: ； ； in, express the number of rows, This represents the vertical attention matrix.

7. The method for identifying spectral and spatial features of a hyperspectral image according to claim 6, characterized in that: The weighted feature map Represented as: 。 8. The method for identifying spectral and spatial features of a hyperspectral image according to claim 7, characterized in that: The positive threshold strategy is as follows: The formula for the positive threshold strategy is as follows: ； when As it approaches 0, its impact on the model decreases; when Its importance increases as it approaches 1; when At that time, the corresponding features Generated by magnification of 5 times ; when At that time, the feature retains its original value without being amplified.

9. The method for identifying spectral and spatial features of a hyperspectral image according to claim 8, characterized in that: The negative threshold strategy is as follows: The formula for the negative threshold strategy is as follows: ； when hour, The value is multiplied by 0.5; when hour, The value remains unchanged.

10. The method for identifying spectral and spatial features of a hyperspectral image according to claim 9, characterized in that: The process of merging the positive and negative feature maps is described as follows: 。

Images