A hyperspectral land cover classification method based on three-dimensional dense connection network
By introducing efficient channel and group channel attention mechanisms into a three-dimensional densely connected network, the feature extraction capability of hyperspectral data is improved, the problems of insufficient accuracy and robustness in hyperspectral classification are solved, and more efficient land cover classification is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- FUYANG NORMAL UNIVERSITY
- Filing Date
- 2026-03-18
- Publication Date
- 2026-06-05
AI Technical Summary
Existing technologies are insufficient in extracting discriminative features and spectral-spatial joint features in hyperspectral classification, resulting in low classification accuracy and robustness.
A hyperspectral land cover classification method based on a three-dimensional densely connected network is adopted, which combines an efficient channel attention mechanism and a grouped channel attention mechanism with a parallel learnable attention module. The feature extraction capability is improved by fusing dynamic grouping strategies and learnable weights.
It significantly improves the ability to extract discriminative features and spectral-spatial joint features, enhances classification accuracy and robustness, solves the problem of overfitting to small sample classes, and reduces the number of model parameters and computational cost.
Smart Images
Figure CN122156778A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the fields of deep learning and hyperspectral imaging, and specifically relates to a hyperspectral land cover classification method based on a three-dimensional densely connected network. Background Technology
[0002] Hyperspectral remote sensing images contain hundreds of continuous, narrow spectral bands, providing rich spectral information on land cover and serving as a key data source for achieving refined land cover classification. However, the inherent high dimensionality, high correlation between spectral bands, and phenomena such as "different spectra for the same object" and "different objects with the same spectrum" in hyperspectral data present significant challenges to classification tasks.
[0003] Traditional machine learning methods, such as Support Vector Machines (SVM), while widely used, are shallow models and struggle to automatically learn deep, nonlinear features from high-dimensional data. In recent years, deep learning, especially Convolutional Neural Networks (CNNs), has become the mainstream method for hyperspectral image classification due to its powerful feature learning capabilities. From the initial one-dimensional CNNs (1D-CNNs) utilizing only spectral information, to two-dimensional CNNs (2D-CNNs) utilizing spatial neighborhood information, and then to three-dimensional CNNs (3D-CNNs) capable of simultaneously processing spatial and spectral dimensions, model performance has continuously improved. 3D-CNNs have a natural advantage because they can directly use hyperspectral data cubes as input and simultaneously extract joint spatial-spectral features using three-dimensional convolutional kernels.
[0004] The DenseNet network greatly promotes feature reuse and gradient flow by establishing dense connections between each layer and all previous layers, alleviating the gradient vanishing problem in deep networks and performing well in image classification tasks. Extending the DenseNet idea to three dimensions to form DenseNet3D provides a new approach for hyperspectral classification. However, directly applying the standard DenseNet3D to hyperspectral classification still has the following problems: (1) Insufficient feature selection ability: Different bands and different spatial locations of hyperspectral data contribute differently to classification. Standard convolution operations treat all features equally and lack an adaptive focusing mechanism for discriminative information, ultimately resulting in insufficient ability to extract discriminative information. (2) Uneven utilization of spectral-spatial information: Hyperspectral data has hundreds of bands, resulting in extremely high feature spatial dimensions. The strong correlation between adjacent bands introduces significant redundancy, which easily leads to the "curse of dimensionality," ultimately making it difficult to fully extract the spectral-spatial joint features of hyperspectral images. Therefore, the technical problem to be solved is how to simultaneously improve the ability to extract discriminative features and the ability to extract spectral-spatial joint features, so as to achieve more accurate and robust land cover classification. Summary of the Invention
[0005] The purpose of this invention is to provide a method for preparing a hyperspectral land cover classification method based on a three-dimensional densely connected network, in order to solve the technical problem that the existing technology is insufficient in extracting discriminative features and spectral-spatial joint features in hyperspectral classification, which reduces classification accuracy and robustness.
[0006] The hyperspectral land cover classification method based on a three-dimensional densely connected network includes the following steps: Step 1) Obtain the dataset; Step 2) Standardize the spectral dimensions; Step 3) Construct a classification model based on hyperspectral imaging technology and deep learning; Step 4) Train and optimize the model; Step 5) Save the model and apply it; Based on hyperspectral imaging technology and deep learning, a neural network model integrates parallel learnable attention modules into the dense blocks of the DenseNet3D model to form an improved dense block. The parallel learnable attention module includes an efficient channel attention mechanism module and a grouped channel attention mechanism module. The efficient channel attention and spectral grouped attention are computed in parallel and adaptively weighted fusion is performed through a learnable weight fusion subnetwork. Finally, the output is a feature map that has been enhanced by the synergistic enhancement of spectral and spatial dimensions.
[0007] Preferably, the efficient channel attention mechanism module first performs three-dimensional global average pooling on the input features to compress the spatial-spectral features of each channel to obtain the channel description vector; then, it reshapes the pooled one-dimensional channel features into a 3D shape and uses adaptive one-dimensional convolution to capture local cross-channel interactions; finally, it maps the convolution output to the [0,1] interval through Sigmoid activation to obtain the attention weight of each channel, and then multiplies the weights with the original input features channel by channel; the output is the enhanced feature of the efficient channel attention mechanism module.
[0008] Preferably, the kernel size of the adaptive one-dimensional convolution is k, and the formula for calculating k is as follows:
[0009] In the formula: and For hyperparameters, This represents the number of channels.
[0010] Preferably, the grouped channel attention mechanism module dynamically determines the number of groups based on the depth of the current dense block and the number of input channels, adapting to the feature granularity of different stages. Indices 0, 1, and 2 are established according to the network stage, representing shallow, medium, and deep layers respectively. Then, a basic grouping is performed along the channel dimension; followed by a dynamic grouping. After grouping, global average pooling is used to compress the 3D features into one-dimensional features along the channel dimension, and the resulting features are then decomposed into... Each group has its own channel feature vector.
[0011] Preferably, the grouped channel attention mechanism module includes two fully connected layers. The first fully connected layer takes the feature vector of the grouped channels as input, reduces the dimensionality of the grouped channel vector, and introduces nonlinearity through ReLU activation to learn the implicit dependencies of the channels within the group. The second fully connected layer takes the output of the first fully connected layer as input, restores the dimensionality-reduced feature vector to the original number of channels in the group, and obtains the attention weights within the group through Sigmoid activation, realizing the weight allocation of each channel within the group. The weight allocated to each channel is multiplied by a full-channel attention weight based on global average pooling to obtain a spectral grouped attention weight that integrates local and global information. The spectral grouped attention weight is reshaped into a 3D shape and finally multiplied with the original input features channel by channel to obtain the enhanced features of the grouped channel attention mechanism module.
[0012] Preferably, the dynamic grouping process is as follows:
[0013] In the formula: It is the basic number of groups. It is the number of channels for the input feature.
[0014] Preferably, the learnable weight fusion subnetwork adds the enhancement features corresponding to the two attention mechanism modules element-wise to obtain the fused features; the fused features are then subjected to three-dimensional global average pooling to obtain a global feature vector; the global feature vector is passed through a fully connected layer with two output nodes and activated with Softmax to obtain the adaptive fusion weights corresponding to the two attention mechanism modules, and then the corresponding adaptive fusion weights are reshaped into 3D shapes and the corresponding efficient channel attention mechanism weights and group channel attention mechanism weights are decomposed; the efficient channel attention mechanism weights and group channel attention mechanism weights are multiplied element-wise with the enhancement features of the corresponding attention mechanism modules, and the results are added to obtain the final fused features.
[0015] Preferably, in step 4), the deep neural network optimization algorithm used is the RMSprop optimization algorithm, and the formula is shown below.
[0016] In the formula: This represents the moving average of the squared gradient at iteration t. The attenuation coefficient represents the squared gradient. Represents the parameters at iteration t. gradient, Represents the learning rate. This represents the model parameters at iteration t. To prevent very small numerical constants with a denominator of zero.
[0017] Preferably, in step 4), the selected loss function is the class cross-entropy, and the formula for calculating the single-sample loss of the class cross-entropy loss function is as follows:
[0018] In the formula: Representing the Single-sample class cross-entropy loss for each sample. Represents the number of categories. Representing the The true label of the sample is the first The values that a class can take. Representing the After the sample is activated by the model Softmax, it is predicted to be the th sample. The probability of a class.
[0019] Preferably, the formula for calculating the batch average loss of the CCE loss function is as follows:
[0020] In the formula: This represents the average loss of a batch. Indicates the batch size.
[0021] The technical advantages of this invention are as follows: This invention adapts the feature granularity of different network depths through a dynamic grouping strategy, combining a parallel attention mechanism consisting of an Efficient Channel Attention (ECA) module and a Grouped Channel Attention (GCA) module. This simultaneously captures the channel interaction features and spectral grouping features of hyperspectral images, improving the extraction and representation capabilities of discriminative features. Furthermore, this invention adaptively adjusts the weights of different attention branches through learnable weight fusion, adaptively combining global spectral dependence with local spectral correlation. This avoids the limitations of fixed fusion strategies, achieves feature complementarity, and thus enhances the extraction capability of spectral-spatial joint features.
[0022] This invention also includes a sampling strategy and network architecture optimized for hyperspectral data, effectively addressing the problem of overfitting to small sample classes, with significant improvements in OA, AA, and Kappa coefficients. Furthermore, dynamic grouping strategies and redundant connection pruning reduce the number of model parameters and computational cost, balancing classification accuracy and inference speed. Attached Figure Description
[0023] Figure 1 This is a basic flowchart of a hyperspectral land cover classification method based on a three-dimensional densely connected network according to the present invention.
[0024] Figure 2 This is a network structure diagram of the improved dense blocks, transition layers, and classification head within the PAD-Net3D model of this invention.
[0025] Figure 3 This is a structural diagram of the parallel learnable attention module in this invention.
[0026] Figure 4 This is the normalized confusion matrix in an embodiment of the present invention.
[0027] Figure 5 The image shown is a diagram of the actual results in an embodiment of the present invention.
[0028] Figure 6 This is a classification result diagram from an embodiment of the present invention. Detailed Implementation
[0029] The following detailed description of the embodiments, with reference to the accompanying drawings, will further illustrate the specific implementation of the present invention, in order to help those skilled in the art to have a more complete, accurate, and in-depth understanding of the inventive concept and technical solution of the present invention.
[0030] like Figures 1-6 As shown, this invention provides a hyperspectral land cover classification method based on a three-dimensional densely connected network, comprising the following steps: Step 1) Obtain the dataset.
[0031] This step involves obtaining the publicly available Indian Pines hyperspectral dataset. This dataset contains an image data cube with a spatial resolution of 145 rows × 145 columns and 200 spectral bands, along with a corresponding land cover label map annotating 16 different land cover types, including corn, soybeans, and forests. To reduce noise, water-absorbing bands are typically removed, ultimately using 176 effective spectral bands. The original hyperspectral data has dimensions (145, 145, 176), and the label data has dimensions (145, 145), where unlabeled areas have values of 0, and labeled areas have values from 1 to 16.
[0032] Step 2) Standardize the spectral dimensions.
[0033] To avoid the impact of differences in the dimensions and numerical ranges of different spectral bands on model training, this step involves standardizing the spectral dimensions before training the neural network model. First, the three-dimensional hyperspectral data is reshaped into a two-dimensional matrix with a shape of (21025, 176). Then, each spectral band is standardized so that the mean of each band is 0 and the standard deviation is 1. The standardized data is then reshaped back into a three-dimensional structure, restoring the shape to (145, 145, 176). This step helps accelerate model convergence and improve training stability.
[0034] Step 3) Construct a neural network model based on hyperspectral imaging technology and deep learning.
[0035] In this step, the hyperspectral imaging technology and deep learning neural network (PAD-Net3D) model can be implemented using the Keras framework, with the main structure as follows: Input layer: The input to the classification model is defined as a four-dimensional tensor with shape (9,9,176,1). The last dimension is added during training via a reshape operation, representing a single channel, to accommodate the model's requirements for the input shape (batch_size, height, width, depth,1).
[0036] Initial feature extraction layer: A three-dimensional convolutional layer is used. This layer first uses a 3×3×3 convolutional kernel with a stride of (1,1,1) to perform preliminary feature extraction, and the number of output channels is 24. This operation initially fuses local spatial-spectral information.
[0037] Batch normalization and activation layer: Batch normalization and parameterized linear rectified unit (PReLU) activation are then performed to accelerate convergence and provide nonlinearity.
[0038] Spectral Dimension Pooling Layer: A three-dimensional max-pooling layer with a pooling size of (2,1,1) and a stride of (2,1,1) is used to downsample the spectral dimension by a factor of 2, reducing the spectral depth from 176 to 88. This aims to reduce computational cost and enhance feature robustness.
[0039] Spatial dimension pooling layer: A three-dimensional max pooling layer with a pooling size of (1,2,2) and a stride of (1,2,2) is used to downsample the spatial dimension by 2 times, reducing the spatial size from 9×9 to 5×5.
[0040] Improved dense blocks and transition layers: The main body of the network contains three improved dense blocks, which are connected by transition layers.
[0041] Each improved dense block sequentially comprises a bottleneck layer, a main convolutional layer, and a parallel learnable attention module. Specifically: Bottleneck layer: After batch normalization of the input feature map, 1×1×1 convolution is used to compress the number of channels. The compression factor is usually 4 to reduce the amount of computation.
[0042] Main convolutional layer: After batch normalization again, a new feature map is generated using a 3×3×3 convolutional kernel.
[0043] Parallel Learnable Attention Module: This module performs feature calibration on the output feature map of the main convolutional layer. Utilizing the core mechanism of DenseNet feature reuse, the calibrated new feature map is concatenated with the outputs of all previous layers of the current dense block along the channel dimension.
[0044] The parallel learnable attention module includes an efficient channel attention mechanism module and a grouped channel attention mechanism module. The outputs of the two attention mechanism modules are fused using learnable weights to obtain the result. The structure of the parallel learnable attention module is as follows: Figure 3 As shown. Specifically: The efficient channel attention mechanism module first performs 3D global average pooling on the input features, compressing the spatial-spectral features of each channel to obtain a channel description vector. Then, it reshapes the pooled one-dimensional channel features into a 3D shape and uses an adaptive one-dimensional convolution with a kernel size of k to capture local cross-channel interactions. The formula for calculating k is as follows:
[0045] In the formula: and These are hyperparameters, which are set to 1 and 2 respectively in this invention. The number of channels is denoted as . The efficient channel attention branch performs convolutions only on the channel dimension to capture local dependencies between channels; finally, the convolution output is mapped to the [0,1] interval through Sigmoid activation to obtain the attention weight of each channel. The weights are then multiplied with the original input features channel by channel to enhance the features of important channels, suppress redundant channels, and output the enhanced features of the ECA module.
[0046] Grouped Channel Attention Mechanism Module: This branch dynamically determines the number of groups based on the depth of the current dense block and the number of input channels, adapting to the feature granularity of different stages. Indices 0, 1, and 2 are established according to the network stage, representing shallow, medium, and deep layers respectively. Then, a basic grouping is performed along the channel dimension, specifically 16 groups for shallow layers, 12 groups for medium layers, and 8 groups for deep layers. A further dynamic grouping is then performed, as shown in the following equation:
[0047] In the formula: The basic number of groups is 16, 12, and 8 in the three dense block stages, respectively. This refers to the number of channels in the input features. After grouping, global average pooling is used to compress the 3D features into one-dimensional features with different channel dimensions. The resulting features are then split into... The system generates a set of in-group channel feature vectors. Two fully connected layers are then connected. The first fully connected layer takes the grouped in-group channel feature vectors as input, reduces the dimensionality of the grouped in-group channel vectors, and introduces non-linearity through ReLU activation to learn the implicit dependencies between in-group channels. The second fully connected layer takes the output of the first fully connected layer as input, restores the dimensionality-reduced feature vectors to the original number of channels in the group, and uses Sigmoid activation to obtain in-group attention weights, thus allocating weights to each channel within the group. Each channel's assigned weight is multiplied by a global average pooling-based full-channel attention weight to obtain spectral grouping attention weights that integrate local and global information. These spectral grouping attention weights are reshaped into a 3D shape and finally multiplied channel-by-channel with the original input features to obtain the enhanced features of the GCA module.
[0048] The learnable weighted fusion subnetwork: For the same input feature, corresponding augmented features are generated through the ECA and SGA modules respectively. The two augmented features (the augmented features from the ECA and SGA modules) are added element-wise to obtain the fused feature. The fused feature is then subjected to 3D global average pooling to obtain a global feature vector. This vector is passed through a fully connected layer with two output nodes and activated using Softmax to obtain adaptive fusion weights corresponding to the two attention mechanism modules. The corresponding adaptive fusion weights are then reshaped into a 3D shape, and the corresponding efficient channel attention mechanism weights (ECA weights) and group channel attention mechanism weights (SGA weights) are extracted. Each weight is multiplied element-wise with the augmented feature of its corresponding branch, and the results are then added to obtain the final fused feature.
[0049] Each transition layer is located between two dense blocks and includes a batch normalization layer, a 1×1×1 convolutional block, and a 3D average pooling layer. The pooling size of the 3D average pooling layer is (1,2,2), and it is downsampled by a factor of 2 only in the spatial dimension, while the spectral dimension remains unchanged. This design aims to progressively reduce the spatial size and control the model complexity while preserving discriminative spectral information to the maximum extent.
[0050] Classification Head: The output of the last improved dense block enters the classification head, where batch normalization is performed first, followed by 3D global average pooling to aggregate the spatial and spectral dimensions of the feature map into a single vector. This vector contains comprehensive information from all locations and bands. A Dropout layer is added with a dropout rate of 0.5 to randomly deactivate some neurons and prevent overfitting. Finally, a fully connected layer with the number of neurons equal to the number of land cover categories (16) is connected, and the Softmax activation function is used to output the predicted probability for each category.
[0051] Step 4) Train and optimize the model.
[0052] Before training optimization, the samples of each category in the public dataset are randomly shuffled, and n% are randomly selected as the training sample set. The remaining 100%-n% of the samples in each category are used as the test sample set, and 50% of the test set is used as the validation set.
[0053] This step involves selecting the loss function and evaluation metrics: the choice of loss function affects the quality of the model and can reflect the difference between the predicted value and the true value.
[0054] The loss function chosen is Categorical Cross-Entropy (CCE). The formula for calculating the single-sample loss using the CCE loss function is shown below:
[0055] In the formula: Representing the Single-sample class cross-entropy loss for each sample. Represents the number of categories. Representing the The true label of the sample is the first The values that a class can take. Representing the After the sample is activated by the model Softmax, it is predicted to be the th sample. The probability of a class.
[0056] The formula for calculating the batch average loss of the CCE loss function is as follows:
[0057] In the formula: This represents the average loss of a batch. Indicates the batch size.
[0058] The evaluation metrics selected include overall accuracy, average accuracy, and Kappa coefficient, which can effectively assess the accuracy of classification results.
[0059] The deep neural network optimization algorithm used in this step is the RMSprop optimization algorithm. The formula for the RMSprop optimization algorithm is shown below.
[0060] In the formula: This represents the moving average of the squared gradient at iteration t. The attenuation coefficient represents the squared gradient. Represents the parameters at iteration t. gradient, Represents the learning rate. This represents the model parameters at iteration t.
[0061] Hyperspectral data has an extremely high number of bands, resulting in significant gradient fluctuations in each round of calculation. Directly adjusting parameters using a single gradient can lead to training oscillations and difficulty in convergence. The RMSProp algorithm can eliminate gradient noise in high-dimensional features, making parameter updates more stable. While the learning rate in ordinary gradient descent is globally fixed, the gradient fluctuations differ for different parameters in hyperspectral models. In the RMSProp algorithm, the learning rate is adaptively reduced for parameters with large gradient fluctuations to prevent excessively large parameter updates that could lead to non-convergence, and increased for parameters with small gradient fluctuations to accelerate convergence to the optimal value. To prevent the denominator from reaching zero, a very small value is used. To smooth the data, a value of 1e-7 is typically used, which does not affect the calculation while ensuring numerical stability.
[0062] Step 5) Save the model and apply it.
[0063] This step saves the optimal model structure and parameters, and inputs the hyperspectral images from the test set into the network for classification. The normalized confusion matrix is obtained, such as... Figure 4 As shown. A comparison between the actual results and the classification results, as shown below. Figure 5 and Figure 6 As shown.
[0064] The classification results map and normalized confusion matrix show that the present invention can accurately identify small features and boundary areas, effectively reduce category confusion, and has certain reference value for judging the classification of hyperspectral land cover.
[0065] The present invention has been described above by way of example with reference to the accompanying drawings. Obviously, the specific implementation of the present invention is not limited to the above-described manner. Any non-substantial improvements made using the inventive concept and technical solution of the present invention, or the direct application of the inventive concept and technical solution of the present invention to other occasions without modification, are all within the protection scope of the present invention.
Claims
1. A hyperspectral land cover classification method based on a three-dimensional densely connected network, characterized in that, Includes the following steps: Step 1) Obtain the dataset; Step 2) Standardize the spectral dimensions; Step 3) Construct a classification model based on hyperspectral imaging technology and deep learning; Step 4) Train and optimize the model; Step 5) Save the model and apply it; Based on hyperspectral imaging technology and deep learning, a neural network model integrates parallel learnable attention modules into the dense blocks of the DenseNet3D model to form an improved dense block. The parallel learnable attention module includes an efficient channel attention mechanism module and a grouped channel attention mechanism module. It computes efficient channel attention and spectral grouped attention in parallel, and performs adaptive weighted fusion through a learnable weight fusion sub-network, finally outputting a feature map that has been enhanced by spectral and spatial dimensions.
2. The hyperspectral land cover classification method based on a three-dimensional densely connected network according to claim 1, characterized in that, The efficient channel attention mechanism module first performs three-dimensional global average pooling on the input features to compress the spatial-spectral features of each channel to obtain the channel description vector; then, it reshapes the pooled one-dimensional channel features into a 3D shape and uses adaptive one-dimensional convolution to capture local cross-channel interactions; finally, it uses Sigmoid activation to map the convolution output to the [0,1] interval to obtain the attention weight of each channel, and then multiplies the weights with the original input features channel by channel. Enhanced features of the output efficient channel attention mechanism module.
3. The hyperspectral land cover classification method based on a three-dimensional densely connected network according to claim 2, characterized in that, The kernel size of adaptive one-dimensional convolution is k, and the formula for calculating k is as follows: In the formula: and For hyperparameters, This represents the number of channels.
4. The hyperspectral land cover classification method based on a three-dimensional densely connected network according to claim 1, characterized in that, The grouped channel attention mechanism module dynamically determines the number of groups based on the depth of the current dense block and the number of input channels, adapting to the feature granularity at different stages. Indices 0, 1, and 2 are established according to the network stage, representing shallow, medium, and deep layers respectively. Then, a basic grouping is performed along the channel dimension; followed by a dynamic grouping. After grouping, global average pooling is used to compress the 3D features into one-dimensional features along the channel dimension, and the resulting features are then split into... Each group has its own channel feature vector.
5. The hyperspectral land cover classification method based on a three-dimensional densely connected network according to claim 4, characterized in that, The grouped channel attention mechanism module consists of two fully connected layers. The first fully connected layer takes the feature vector of the grouped channels as input, reduces the dimensionality of the grouped channel vector, and introduces non-linearity through ReLU activation to learn the implicit dependencies of the channels within the group. The second fully connected layer takes the output of the first fully connected layer as input, restores the dimensionality-reduced feature vector to the original number of channels in the group, and obtains the attention weights within the group through Sigmoid activation, realizing the weight allocation of each channel within the group. The weight allocated to each channel is multiplied by a full-channel attention weight based on global average pooling to obtain a spectral grouped attention weight that integrates local and global information. The spectral grouped attention weight is reshaped into a 3D shape and finally multiplied with the original input features channel by channel to obtain the enhanced features of the grouped channel attention mechanism module.
6. The hyperspectral land cover classification method based on a three-dimensional densely connected network according to claim 5, characterized in that, The dynamic grouping process is shown in the following formula: In the formula: It is the basic number of groups. It is the number of channels for the input feature.
7. The hyperspectral land cover classification method based on a three-dimensional densely connected network according to claim 1, characterized in that, The learnable weighted fusion subnetwork adds the enhanced features corresponding to the two attention mechanism modules element by element to obtain the fused features; The fused features are subjected to three-dimensional global average pooling to obtain a global feature vector. The global feature vector is passed through a fully connected layer with two output nodes and activated with Softmax to obtain the adaptive fusion weights corresponding to the two attention mechanism modules. Then, the corresponding adaptive fusion weights are reshaped into 3D shapes and the corresponding efficient channel attention mechanism weights and grouped channel attention mechanism weights are separated. The weights of the efficient channel attention mechanism and the group channel attention mechanism are multiplied element-wise with the enhanced features of the corresponding attention mechanism modules, and the results are added together to obtain the final fused feature.
8. The hyperspectral land cover classification method based on a three-dimensional densely connected network according to claim 1, characterized in that, In step 4), the deep neural network optimization algorithm used is the RMSprop optimization algorithm, and the formula is shown below. In the formula: This represents the moving average of the squared gradient at iteration t. The attenuation coefficient represents the squared gradient. Represents the parameters at iteration t. gradient, Represents the learning rate. This represents the model parameters at iteration t. To prevent very small numerical constants with a denominator of zero.
9. The hyperspectral land cover classification method based on a three-dimensional densely connected network according to claim 1, characterized in that, In step 4), the selected loss function is the class cross-entropy. The formula for calculating the single-sample loss of the class cross-entropy loss function is shown below: In the formula: Representing the Single-sample class cross-entropy loss for each sample. Represents the number of categories. Representing the The true label of the sample is the first The values that a class can take. Representing the After the sample is activated by the model Softmax, it is predicted to be the th sample. The probability of a class.
10. The hyperspectral land cover classification method based on a three-dimensional densely connected network according to claim 9, characterized in that, The formula for calculating the batch average loss of the CCE loss function is as follows: In the formula: This represents the average loss of a batch. Indicates the batch size.