A wavelet guided network ground object classification method for on-site security scene

Abstract

This invention discloses a wavelet-guided network-based land cover classification method for near-site security scenarios. It proposes dynamic wavelet convolution to dynamically aggregate high-frequency texture information into the convolution kernel, improving the perception of local features. To capture long-range dependencies, a multi-head wavelet spatial self-attention module is constructed, introducing wavelet transformations and spectral features to reduce information loss caused by downsampling average pooling downsampling, while simultaneously improving the perception of spatial texture and spectral information. Based on this, a wavelet-guided hybrid network is designed to extract local and global features from hyperspectral remote sensing images. Results on three public datasets demonstrate that this method outperforms other representative hyperspectral remote sensing land cover classification methods.

Description

Technical Field

[0001] This invention belongs to the field of pattern recognition technology, specifically relating to a wavelet-guided network method for ground feature classification in near-site security scenarios. Background Technology

[0002] Temporary security refers to a diversified, cross-domain, three-dimensional, collaborative, and intelligent technical system addressing the needs of defense, protection, production, safety, and rescue within space. Specific application scenarios include low-altitude security, underwater security, and cross-domain security. The temporary security technology system encompasses six major research directions: related light detection, stable detection, multimodal cognitive computing, swarm intelligence decision-making, waterborne optics, and cross-domain remote sensing. Obtaining accurate information on ground features is crucial for the effective implementation of typical temporary security tasks such as low-altitude UAV detection, extreme disaster early warning and rescue, and target reconnaissance. Hyperspectral images contain fine-grained spectral information and rich spatial information, serving as an important information carrier in the cross-domain remote sensing research direction within the temporary security technology system. Analyzing hyperspectral remote sensing images for ground feature classification is key to the successful implementation of temporary security applications; therefore, hyperspectral remote sensing ground feature classification has received widespread attention.

[0003] Due to the complex and varied nature of ground features, their irregular regional distribution, the large number of ground feature categories, and the imbalance in sample size, coupled with the fact that the acquisition of hyperspectral remote sensing images is affected by meteorological conditions, sensor type, and shooting angle, and is accompanied by issues of "different spectra for the same object" and "different objects for the same spectrum," accurately identifying the ground feature categories contained in hyperspectral remote sensing images is a challenging task. Constructing a discriminative feature extraction method for hyperspectral remote sensing images is a key step in overcoming these difficulties and achieving accurate identification of ground feature categories in hyperspectral remote sensing.

[0004] Currently, utilizing deep networks to mine high-level semantic features embedded in hyperspectral remote sensing images has become the mainstream method for remote sensing ground object identification. Among them, convolutional neural networks (CNNs) have attracted much attention due to their excellent local feature extraction capabilities. In recent years, the Transformer architecture has been proposed to compensate for the shortcomings of CNNs in modeling long-range dependencies in images. Therefore, hybrid networks combining CNNs and Transformer architectures can fully leverage the advantages of both models and have received widespread attention. However, these hybrid network methods only capture joint spatial and spectral features from the spatial domain perspective, ignoring the texture information contained in the frequency domain. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, this invention provides a wavelet-guided network-based land cover classification method for near-site security scenarios. It proposes wavelet dynamic convolution (WDConv) to dynamically aggregate high-frequency texture information into the convolution kernel, improving the perception of local features. To capture long-range dependencies, a multi-head wavelet spectral-spatial self-attention (MHWS2A) module is constructed, incorporating wavelet transformations and spectral features to reduce information loss caused by downsampling average pooling, while simultaneously improving the perception of spatial texture and spectral information. Based on this, a wavelet-guided hybrid network is designed to extract local and global features from hyperspectral remote sensing images. Results on three publicly available datasets demonstrate that this method outperforms other representative hyperspectral remote sensing land cover classification methods.

[0006] The technical solution adopted by this invention to solve its technical problem is as follows:

[0007] Step 1: After spectral dimensionality reduction of the hyperspectral remote sensing image, image patch sampling is performed as the input to the model;

[0008] Step 2: Construct the wavelet dynamic convolution operation WDConv;

[0009] Step 3: Construct a wavelet dynamic convolutional spatial-spectral joint feature extraction module based on WDConv;

[0010] The wavelet dynamic convolutional spatial-spectral joint feature module contains three WDConv-Resblock residual connection blocks; each WDConv-Resblock residual connection block is a residual connection block based on WDConv; spatial attention and spectral attention are embedded after each WDConv-Resblock residual connection block.

[0011] Step 4: Construct the multi-head wavelet spatial spectrum self-attention module MHWS2A;

[0012] Step 5: Construct a wavelet-space-spectral joint Transformer module based on MHWS2A;

[0013] The wavelet-space-spectral joint Transformer module includes a feature shrinkage module, an expansion module, and an MHWS2A module. The feature shrinkage module includes a 3×3 convolutional layer and a Batch Normalization layer to shrink the channel dimension of the feature map. Subsequently, the features are fed into the MHWS2A module to extract long-range dependencies. The output features are expanded through the feature expansion module to extend the channel dimension of the feature map.

[0014] Step 6: Construct the wavelet-guided hybrid network WHNet;

[0015] Image patches after PCA dimensionality reduction and neighborhood sampling First, a 3×3 convolutional layer is used for initial feature extraction. Then, a wavelet dynamic convolutional spatial-spectral joint feature extraction module is used for local feature extraction. After that, the feature is fed into a wavelet spatial-spectral joint Transformer module to capture global features. Finally, the spatial-spectral joint feature F used for classification is obtained. spe-spa ;

[0016] Step 7: Combine the captured spatial and spectral features F spe-spa The data is input into the Softmax classifier to predict the category of the pixel to be classified; the loss function of WHNet consists of the cross-entropy function and the center loss function, which are used to train the weights of the network to obtain the predicted land cover category.

[0017] Further, step 1 specifically includes:

[0018] Step 1-1: For a given hyperspectral image Where r×c is the spatial dimension of the hyperspectral image X, and b represents the number of spectra in image X; after principal component analysis (PCA) dimensionality reduction, only the first p principal components are retained; the dimensionality-reduced image is represented as... p represents the image X after dimensionality reduction. p The number of spectra contained; subsequently, neighborhood sampling centered on the pixel to be classified is used as the input to the dual heterogeneous network, and the neighborhood sampled image patch is represented as... Where s is the spatial size of the image patch neighborhood;

[0019] Steps 1-2: For a given hyperspectral remote sensing image Where R×C represents the spatial dimension of the hyperspectral image H, and B represents the spectral dimension of image H; Principal Component Analysis (PCA) is used to reduce the spectral dimension of the hyperspectral remote sensing image H, retaining only the first p principal components; after dimensionality reduction, sampling is performed centered on the pixel to be classified as the input to the wavelet-guided hybrid network, and the neighborhood sampled image patch is represented as... The spatial size of the neighboring image patch is s×s.

[0020] Furthermore, step 2 specifically includes:

[0021] For input features The Discrete Wavelet Transform (DWT) decomposes it into four subbands, including three high-frequency subbands: LH, HL, and HH. Subsequently, two fully connected layers are used to generate the weights ω for each high-frequency subband. LH ω HL and ω HLSum the weights of the three high-frequency subbands to obtain the normalized weights ω of the N convolutional kernels. n The calculation process is expressed as follows:

[0022]

[0023] ω n =softmax(ω LH +ω HL +ω HH (4)

[0024] Where FC represents a fully connected layer, ReLU is the modified linear unit activation function ReLU, and softmax represents the normalization function softmax; after obtaining the weights ω n Then, the N convolutional kernels are aggregated to obtain WDConv, and its calculation process is as follows:

[0025] Y = σ(W) T (X)X+b(X)) (5)

[0026]

[0027] in, Let W(X) represent the input feature map of WDConv, and W(X) and b(X) represent the aggregated weights and biases in WDConv, respectively. σ represents the activation function, and W... n and b n This represents the weights and biases of the nth convolutional kernel.

[0028] Furthermore, step 4 specifically includes:

[0029] Step 4-1: For input features Where H×W is the spatial dimension of the feature, and d is the channel dimension of the feature; the query Q is obtained after linear transformation of the input feature F; DWT decomposes the input feature into four sub-bands, each a low-frequency sub-band F. LL and three high-frequency subbands F LH F HL F HH The four wavelet subbands are concatenated along the channel dimension and then fed into a 3×3 convolution to further extract feature maps with local context. Then, a linear transformation is performed to obtain the key K and value V. A global pooling layer is used to map the feature F to the channel dimension to obtain the spectral feature S. pe Spectral characteristics S pe It is represented as follows:

[0030]

[0031] Introducing spectral features S peSubsequently, the wavelet-space-spectral joint self-attention mechanism is defined as follows:

[0032]

[0033] Among them, Q, K, V, S pe These are Query, Key, Value, and spectral features; These are features obtained by reconstructing wavelet subbands using the inverse wavelet transform IDWT;

[0034] Step 4-2: For the input feature F, the output feature is obtained by passing it through MHWS2A with H heads. The calculation process is as follows:

[0035]

[0036] Among them, W 0 This represents the learnable parameters, and `concat` indicates the concatenation operation. Each attention head... h The calculation process can be represented as follows:

[0037]

[0038] Where h∈[1,2,…,H] represents the h-th head. and For the learnable parameters in the h-th head, in MHWS2A, different heads learn and capture different attention through a parallel method. Concatenating all attention yields the output features of MHWS2A.

[0039] The beneficial effects of this invention are as follows:

[0040] The hyperspectral remote sensing land cover classification method based on a wavelet-guided hybrid network proposed in this invention has achieved certain results in the theoretical research of hyperspectral land cover classification. This invention seamlessly embeds wavelet transform, a frequency domain analysis method, into a deep learning architecture in the spatial domain. This method can consider both frequency and spatial domain features at the feature domain level, and capture both local and global features at the feature level. It overcomes the influence of climate conditions, sensor type, and shooting angle on the accuracy of hyperspectral land cover classification in near-site security tasks, as well as the uneven distribution of land cover in hyperspectral remote sensing images, "different objects with the same spectrum" and "different spectra for the same object," etc. It provides effective and accurate data support for typical near-site security tasks requiring precise land cover category information, such as low-altitude UAV security, disaster early warning and rescue, and target reconnaissance. The method is compared with 10 other representative hyperspectral land cover classification methods on three publicly available hyperspectral remote sensing image datasets containing different land cover categories (Indian Pines, Pavia University, Salinas), and the experimental results show that the method of this invention is effective. Attached Figure Description

[0041] Figure 1 This is a schematic diagram of the structure of the wavelet-guided hybrid network WHNet of the present invention;

[0042] Figure 2 This is a schematic diagram of the wavelet dynamic convolution WDConv structure of the present invention;

[0043] Figure 3 This is a schematic diagram of the structure of the multi-head wavelet spatial spectrum self-attention module of the present invention;

[0044] Figure 4 The classification results are shown for different training samples on three datasets in the embodiments of the present invention: (a) IndiaPines, (b) Pavia University, (c) Salines.

[0045] Figure 5 The following are experimental results on three datasets with different neighborhood image patch sizes and preserved spectral dimensions in this embodiment of the invention: (a) India Pines, (2) Pavia University, (3) Salines. Detailed Implementation

[0046] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0047] This invention proposes a land cover classification method based on a wavelet-guided hybrid network (WHNet) for near-site security. Specifically, a wavelet dynamic convolution (WDConv) is proposed to dynamically aggregate high-frequency texture information into the convolution kernel, thereby improving the perception of local features. To capture long-range dependencies, a multi-head wavelet spectral-spatial self-attention (MHWS2A) mechanism is constructed, introducing wavelet transformations and spectral features to reduce information loss caused by downsampling average pooling, while improving the perception of spatial texture and spectral information. Based on this, a wavelet-guided hybrid network is designed to extract local and global features from hyperspectral remote sensing images. Results on three public datasets (Indian Pines, Pavia University, Salinas) show that this method outperforms other representative hyperspectral remote sensing land cover classification methods. The method of this invention can overcome the influence of climate conditions, sensor type, shooting angle, as well as the uneven distribution of ground objects in hyperspectral remote sensing images, "different objects with the same spectrum" and "different spectra with the same object", on the accuracy of hyperspectral remote sensing ground object classification. It provides effective and accurate data support for typical on-site security tasks that require precise ground object category information, such as low-altitude UAV security, disaster early warning and rescue, and target reconnaissance.

[0048] The contributions of this invention are as follows: First, it proposes WDConv to dynamically aggregate high-frequency texture information into the convolution kernel, achieving a seamless combination of frequency domain feature extraction and spatial domain convolution, thus improving feature representation capabilities. Second, it designs MHWS2A to fully describe the long-range dependencies of hyperspectral remote sensing images by introducing frequency domain texture information and spectral information. Third, the WHNet constructed based on WDConv and MHWS2A can capture both local and global features, and cleverly combine frequency and spatial domain features, thereby improving feature discriminative power.

[0049] The hyperspectral remote sensing ground cover classification method based on a wavelet-guided hybrid network proposed in this invention has achieved certain results in the theoretical research of hyperspectral ground cover classification. This invention seamlessly embeds wavelet transform, a frequency domain analysis method, into a deep learning architecture in the spatial domain. This method can consider both frequency and spatial domain features at the feature domain level, and capture both local and global features at the feature level. It overcomes the susceptibility of hyperspectral remote sensing ground cover classification accuracy to factors such as climate conditions, sensor type, shooting angle, uneven distribution of ground covers in hyperspectral remote sensing images, and the presence of "different objects with the same spectrum" and "different spectra for the same object" in near-site security tasks. It provides effective and accurate data support for typical near-site security tasks requiring precise ground cover category information, such as low-altitude UAV security, disaster early warning and rescue, and target reconnaissance.

[0050] Step 1: After performing spectral dimensionality reduction on the hyperspectral remote sensing image, image patch sampling is performed as the input to the model;

[0051] To reduce the computation time of feature extraction, principal component analysis (PCA) is used to reduce the dimensionality of the original hyperspectral image. For a given hyperspectral image... Where r×c is the spatial dimension of the hyperspectral image X, and b represents the number of spectra in image X. After PCA dimensionality reduction, only the first p principal components are retained. The image after dimensionality reduction is represented as follows: Where r×c is the spatial dimension, and p represents the dimension-reduced image X. p The number of spectra contained. Subsequently, neighborhood sampling centered on the pixel to be classified is used as input to the dual heterogeneous network; the neighborhood sampled image patch is represented as... Where s is the spatial size of the image patch neighborhood, and p represents the number of spectra contained after dimensionality reduction.

[0052] For a given hyperspectral remote sensing image Where R is the spatial dimension of the R×C hyperspectral image H, and B is the spectral dimension of image H. To reduce the computation time of feature extraction, principal component analysis (PCA) is used to reduce the spectral dimension of the hyperspectral remote sensing image H, retaining only the first p principal components, meaning the spectral dimension after reduction is p. After dimensionality reduction, sampling is performed centered on the pixel to be classified as the input to the wavelet-guided hybrid network. The neighborhood sampled image patch is represented as... The spatial size of the neighboring image patch is s×s.

[0053] Step 2: Construct the wavelet dynamic convolution operation (WDConv), such as... Figure 2 As shown;

[0054] To better utilize texture information, a dynamic convolution operation WDConv is constructed to dynamically aggregate high-frequency texture details captured by Discrete Wavelet Transform (DWT) into the convolution kernel, thereby improving the network's feature extraction capability. For the input features... The Discrete Wavelet Transform (DWT) decomposes the wavelet into four subbands, including three high-frequency subbands: LH, HL, and HH. Subsequently, two fully connected layers are used to generate the weights ω for each high-frequency subband. LH ω HL and ω HL The weights of the three high-frequency subbands are summed to obtain the normalized weights ω of the N convolutional kernels. n The calculation process can be represented as follows:

[0055]

[0056] ω n =softmax(ω LH +ω HL +ω HH (4)

[0057] Where FC represents a fully connected layer, ReLU is the modified linear unit activation function, and softmax represents the normalization function. After obtaining the weights ω... n Then, the N convolutional kernels are aggregated to obtain WDConv, and its calculation process is as follows:

[0058] Y = σ(W) T (X)X+b(X)) (5)

[0059]

[0060] in, Let W(X) represent the input feature map of WDConv, and W(X) and b(X) represent the aggregated weights and biases in WDConv, respectively. σ represents the activation function, such as ReLU. n and b n This represents the weights and biases of the nth convolutional kernel.

[0061] Step 3: Construct a wavelet dynamic convolutional spatial-spectral joint feature extraction module based on WDConv;

[0062] Based on the designed WDConv, a wavelet dynamic convolutional spatial-spectral joint feature extraction module was constructed for local feature extraction of hyperspectral remote sensing images. Similar to residual networks, residual connecting blocks based on WDConv, called WDConv-Resblock, were designed. Spatial attention and spectral attention are embedded after each WDConv-Resblock to refine spatial and spectral features. The wavelet dynamic convolutional spatial-spectral joint feature extraction module contains a total of three WDConv-Resblock residual connecting blocks.

[0063] Step 4: Construct a multi-head wavelet spectral-spatial self-attention module (MHWS2A);

[0064] DWT is introduced into multi-head attention, incorporating rich texture structure information into the encoding process. Simultaneously, spectral features are introduced to construct a multi-head wavelet spatial-spectral self-attention module, such as... Figure 3 As shown. For input features Where H×W is the spatial dimension of the feature, and d is the channel dimension of the feature. The query Q is obtained after a linear transformation of the input feature F. DWT decomposes the input feature into four sub-bands, each a low-frequency sub-band F. LL and three high-frequency subbands F LH F HL F HH The four wavelet subbands are concatenated along the channel dimension and then fed into a 3×3 convolution to further extract feature maps with local context. Then, a linear transformation is performed to obtain the key K and value V. To incorporate spectral information, a global pooling layer is used to map the feature F to the channel dimension to obtain the spectral feature S. pe Spectral characteristics S pe It is represented as follows:

[0065]

[0066] Where H×W is the spatial dimension of the feature, and d is the channel dimension of the feature. When introducing spectral feature S... pe Subsequently, the wavelet-space-spectral joint self-attention mechanism is defined as follows:

[0067]

[0068] Where Q, K, V, S pe These are the Query, Key, Value, and Spectral Features. Features obtained by reconstructing wavelet subbands using inverse wavelet transform IDWT.

[0069] For input features F, the output features are obtained by passing them through MHWS2A with H heads. The calculation process is as follows:

[0070]

[0071] Among them, W 0 This represents the learnable parameters, and `concat` indicates the concatenation operation. Each attention head... h The calculation process can be represented as follows:

[0072]

[0073] Where h∈[1,2,…,H] represents the h-th head. and For the learnable parameters in the h-th head, in MHWS2A, different heads learn and capture different attention through a parallel method. Concatenating all attention yields the output features of MHWS2A.

[0074] Step 5: Construct a wavelet-space-spectral joint Transformer module based on MHWS2A;

[0075] To extract long-range dependencies from the joint spatial-spectral features, a wavelet-space-spectral joint Transformer module was designed, which includes a shrinking module, an expanding module, and the core MHWS2A module. The feature shrinking module consists of a 3×3 convolutional layer and a Batch Normalization layer to shrink the channel dimension of the feature map. Subsequently, the features are fed into MHWS2A to extract long-range dependencies. The output features are then expanded through the feature expanding module to extend the channel dimension of the feature map.

[0076] Step Six: Construct the wavelet-guided hybrid network WHNet, such as... Figure 1 As shown;

[0077] Image patches after PCA dimensionality reduction and neighborhood sampling First, a 3×3 convolutional layer is used for initial feature extraction. Then, a wavelet dynamic convolutional spatial-spectral joint feature extraction module is used for local feature extraction. After that, the feature is fed into a wavelet spatial-spectral joint Transformer module to capture global features. Finally, the spatial-spectral joint feature F used for classification is obtained. spe-spa .

[0078] Step 7: Combine the captured spatial and spectral features F spe-spaThe data is fed into a Softmax classifier to predict the category of the pixel to be classified. The loss function of WHNet consists of a cross-entropy function and a center loss function, which are used to train the network weights to obtain the predicted land cover category.

[0079] Example:

[0080] like Figure 4 and Figure 5 As shown, hyperspectral remote sensing land cover classification based on the wavelet-guided hybrid network WHNet was conducted on three hyperspectral remote sensing datasets containing different land cover scenes: Indian Pines, Pavia University, and Salina. In the experiments, spectral dimensionality reduction was performed on the three datasets. The PCA algorithm was used to reduce the spectral dimension of the hyperspectral images, retaining the first 16 principal components, resulting in a spectral dimension of 16 after dimensionality reduction. Subsequently, neighborhood sampling based on image patches was performed, with the category of the center pixel as the category of the image patch. The neighborhood size for image patch sampling was 19x19 on the Indian Pines and Pavia University datasets, and 23x23 on the Salinas dataset. Specific experimental details are as follows:

[0081] (1) Hyperspectral image dataset: Indian Pines: The Indian Pines dataset was collected by the airborne visible / infrared imaging spectrometer AVIRIS in a farmland area in northern Indiana. The dataset has a dimension of 145x145x200, with a spatial dimension of 145x145 and 200 spectral bands. It contains 16 different land cover categories, including various materials in real-world scenes such as coniferous forests and farmland, such as corn, soybeans, and pine trees. 5% of the samples in each category were randomly selected as the training set, and the remaining 95% were used as the test set. The Pavia University dataset consists of hyperspectral urban remote sensing images acquired near the University of Pavia, Italy, using the ROSIS-3HS imaging spectrometer sensor. The dataset has dimensions of 610x340x103 (spatial dimension 610x340) and 103 spectral bands, encompassing 9 land cover categories, primarily including urban land cover types such as asphalt, grassland, and gravel. 30 samples were randomly selected from each category as the training set, and the remaining samples were used as the test set. The Salinas dataset, on the other hand, is a rural hyperspectral remote sensing dataset acquired in the Salinas Valley, California, using the AVIRIS airborne visible / infrared imaging spectrometer sensor. The dataset has dimensions of 512x217x204 (spatial dimension 512x217) and 204 spectral bands, encompassing 16 land cover categories, including agricultural land cover types such as corn, soybeans, cucumbers, and tomatoes. 30 samples were randomly selected from each category as the training set, and the remaining samples were used as the test set.

[0082] (2) Experimental setup: In the experiment, the training optimizer was set to the adaptive Nesterov momentum algorithm (Adan). The parameter β in Adan was set to (0.02, 0.08, 0.01), the weight decay was 0.02, the learning rate was 1e-3, the batch size was set to 32, and a total of 200 epochs were trained.

[0083] (3) Evaluation metrics: Overall Accuracy (OA) is the ratio of correctly classified pixels to the total number of pixels in the sample; Average Accuracy (AA) is the average classification accuracy across different categories; Kappa coefficient (KC) is used to evaluate the consistency between the classification results and the ground truth. To ensure fairness, each experiment was repeated 10 times, and the mean and standard deviation of the results were recorded.

[0084] (4) Comparative Experiments: Comparative experiments were conducted with 10 representative hyperspectral land cover classification methods on three publicly available hyperspectral image datasets (Indian Pines, Pavia University, and Salinas). The experimental results on the Indian Pines dataset are shown in Table 1, the results on the Pavia University dataset are shown in Table 2, and the results on the Salinas dataset are shown in Table 3. In summary, the experimental results on the three publicly available hyperspectral image datasets demonstrate that the method of this invention has higher classification accuracy.

[0085] Table 1. Comparison experiments with 10 representative methods on the Indian Pines dataset.

[0086] SSRN 96.20±0.52 78.24±1.45 95.67±0.59 RSSAN 95.67±0.59 85.69±1.01 92.36±1.14 HybridSN 93.89±0.55 83.02±2.05 93.02±0.63 A2S2KNet 96.28±0.49 90.51±3.94 95.75±0.56 CVSSN 94.77±1.36 93.57±1.84 94.03±1.55 ENL-FCN 96.01±1.11 89.37±1.52 95.46±1.26 FContNet-PSP 94.96±1.04 89.77±3.77 94.26±1.19 FContNet-ASPP 95.66±0.43 91.28±1.91 95.05±0.49 SpectralFormer 74.71±2.50 63.80±3.48 71.16±2.65 MorphFormer 94.49±2.03 92.06±2.52 93.73±2.30 WHNet 96.76±0.25 93.83±2.11 96.31±0.29

[0087] Table 2 Comparison experiments with 10 representative methods on the Pavia University dataset.

[0088]

[0089]

[0090] Table 3. Comparison experiments with 10 representative methods on the Salinas dataset.

[0091] SSRN 94.76±0.35 97.62±0.11 94.17±0.38 RSSAN 95.98±0.56 97.68±0.28 95.53±0.62 HybridSN 97.15±0.50 98.80±0.15 96.83±0.55 A2S2KNet 96.42±0.52 98.47±0.20 96.01±0.58 CVSSN 93.06±0.79 95.79±0.61 92.29±0.86 ENL-FCN 95.76±0.79 98.25±0.36 95.28±0.88 FContNet-PSP 97.21±0.92 98.65±0.27 96.90±1.02 FContNet-ASPP 97.49±0.50 98.80±0.24 97.30±0.50 SpectralFormer 83.55±3.36 90.05±3.10 81.77±3.67 MorphFormer 95.09±1.81 97.77±0.54 94.54±2.00 WHNet 98.81±0.22 99.32±0.16 98.68±0.08

Claims

1. A wavelet-guided network-based ground object classification method for a security scene facing a ground, characterized in that, Includes the following steps: Step 1: After spectral dimensionality reduction of the hyperspectral remote sensing image, image patch sampling is performed as the input to the model; Step 2: Construct the wavelet dynamic convolution operation WDConv; specifically: For the input feature , the discrete wavelet transform (DWT) decomposes it into four subbands, including three high-frequency subbands, respectively , and ; Subsequently, two fully connected layers were used to generate the weights for each high-frequency subband. , and The aggregate is obtained by summing the weights of the three high-frequency subbands. Normalized weights of each convolutional kernel The calculation process is expressed as follows: (1) (2) (3) (4) in, Indicates a fully connected layer. Modify the ReLU activation function of the linear unit. This represents the normalization function softmax; after obtaining the weights... back, The convolutional kernels are aggregated to obtain WDConv, and the calculation process is as follows: (5) (6) (7) in, This represents the input feature map of WDConv. and These are the aggregated weights and biases in WDConv. This represents the activation function. and Indicates the first The weights and biases of each convolutional kernel; Step 3: Construct a wavelet dynamic convolutional spatial-spectral joint feature extraction module based on WDConv; The wavelet dynamic convolutional spatial-spectral joint feature module contains three WDConv-Resblock residual connection blocks; each WDConv-Resblock residual connection block is a residual connection block based on WDConv; spatial attention and spectral attention are embedded after each WDConv-Resblock residual connection block. Step 4: Construct the multi-head wavelet spatial spectral self-attention module MHWS2A; specifically: Step 4-1: For input features ,in The spatial dimensions that are characterized The channel dimension of the feature; for the input feature After linear transformation, the query Q is obtained; DWT decomposes the input features into four sub-bands, each a low-frequency sub-band. and three high-frequency subbands , , The four wavelet subbands are cascaded in the channel dimension and then fed into... The convolution is used to further extract feature maps with local context; then, a linear transformation is performed to obtain the key K and value V; and a global pooling layer is used to transfer the features. Mapping to the channel dimension yields spectral features spectral characteristics It is represented as follows: (8) Introducing spectral features Subsequently, the wavelet-space-spectral joint self-attention mechanism is defined as follows: (9) in, , , , These are Query, Key, Value, and spectral features; These are features obtained by reconstructing wavelet subbands using the inverse wavelet transform IDWT; Step 4-2: For input features After having The output characteristics of the MHWS2A are obtained from the size. The calculation process is as follows: （10） in, Represents the learnable parameters. This represents a chain operation; where each attention head... The calculation process can be represented as follows: (11) in, Indicates the first A head. , , , and For the first The learnable parameters are found in each head; in MHWS2A, different heads learn to capture different attention through a parallel method; all attentions are concatenated to obtain the output features of MHWS2A. ; Step 5: Construct a wavelet-space-spectral joint Transformer module based on MHWS2A; The wavelet-space-spectral joint Transformer module includes a feature shrinkage module, an expansion module, and an MHWS2A module; the feature shrinkage module includes a The convolutional layers and batch normalization layers are used to shrink the channel dimension of the feature map; then, the features are fed into the MHWS2A module to extract long-range dependencies; the output features are expanded to extend the channel dimension of the feature map through the feature expansion module. Step 6: Construct the wavelet-guided hybrid network WHNet; Image patches after PCA dimensionality reduction and neighborhood sampling First, go through a The convolutional layers perform initial feature extraction, followed by local feature extraction using a wavelet dynamic convolutional spatial-spectral joint feature extraction module. These local features are then fed into a wavelet spatial-spectral joint Transformer module to capture global features, ultimately yielding the spatial-spectral joint features used for classification. ; Step 7: Combine the captured spatial and spectral features The data is input into the Softmax classifier to predict the category of the pixel to be classified; the loss function of WHNet consists of the cross-entropy function and the center loss function, which are used to train the weights of the network to obtain the predicted land cover category.

2. The wavelet-guided network-based feature classification method for near-site security scenarios according to claim 1, characterized in that, Step 1 specifically involves: Step 1-1: For a given hyperspectral image ,in For hyperspectral images Space dimensions, Representing an image The number of spectra; after dimensionality reduction by principal component analysis (PCA), only the first... The principal components are preserved; the image after dimensionality reduction is represented as follows: , Represents the image after dimensionality reduction The number of spectra contained; subsequently, neighborhood sampling centered on the pixel to be classified is used as the input to the dual heterogeneous network, and the neighborhood sampled image patch is represented as... ,in The spatial size of the image patch neighborhood; Steps 1-2: For a given hyperspectral remote sensing image ,in For hyperspectral images Space dimensions, For image The spectral dimension; Principal component analysis (PCA) was used to analyze hyperspectral remote sensing images. Perform dimensionality reduction of the spectral dimension, retaining only the first dimensionality. One principal component; after dimensionality reduction, sampling is performed centered on the pixel to be classified as the input to the wavelet-guided hybrid network, and the neighborhood sampled image patch is represented as... The spatial size of the neighboring image patch is .

Images