Dynamic resampling partial label hyperspectral image ground object identification method and system
By employing a dynamic resampling biased label learning method, combined with dynamic prototype resampling and a multi-directional frequency domain attention network, the problems of land cover type imbalance and label ambiguity in hyperspectral images are solved, achieving efficient land cover recognition and improving the robustness and accuracy of the model.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-19
- Publication Date
- 2026-04-07
AI Technical Summary
Existing hyperspectral image land cover recognition methods struggle to simultaneously achieve efficient feature extraction and accurate land cover recognition when dealing with label ambiguity and class imbalance, especially given the performance degradation caused by land cover type imbalance and label ambiguity in hyperspectral images.
We employ a dynamic resampling biased label learning method, which combines a dynamic prototype resampling network and a multi-directional frequency domain attention network with smooth biased label loss and cross-entropy loss to optimize model parameters, achieve class balance and label disambiguation, and improve the robustness and generalization ability of the model.
It effectively solves the problems of land cover type imbalance and label ambiguity in hyperspectral images, significantly improves the accuracy of land cover identification and the robustness of the network, and enhances the applicability of the model in various land cover identification scenarios.
Smart Images
Figure CN121811240A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of image recognition and relates to a method and system for identifying ground features in dynamic resampling biased label hyperspectral images. Background Technology
[0002] With the advancement of remote sensing technology, hyperspectral imaging technology can acquire spatial information and continuous spectral characteristics of ground objects, providing an important data foundation for Earth observation. Hyperspectral image-based ground object identification, as one of the most commonly used and important tasks in Earth observation, is widely applied in key national development areas such as military reconnaissance, urban planning, and precision agriculture. Hyperspectral image-based ground object identification focuses on locating and identifying the pixel types of ground objects in hyperspectral images, aiming to distinguish ground objects of interest from a variety of ground features. Traditional remote sensing ground object identification methods face bottlenecks in distinguishing between "same object, different spectral types" or "different objects, same spectral type" ground objects. Hyperspectral image-based ground object identification, by analyzing the unique spectral characteristics of ground objects, achieves direct identification of the intrinsic properties of materials, driving a leapfrog development in remote sensing capabilities.
[0003] In existing research, leveraging the advantages of deep learning technology in feature extraction, hyperspectral image-based land cover recognition based on recurrent neural networks, convolutional neural networks, attention mechanisms, and their variants has been proposed and achieved satisfactory land cover detection performance. However, these deep learning-based methods generally require a large number of accurately labeled prior samples, which cannot adapt to existing labeling limitations. Due to the presence of pixel mixing regions and uneven pixel distribution in hyperspectral images, label ambiguity and class imbalance pose significant challenges to land cover recognition. Specifically, label ambiguity manifests as unclear boundaries between spectral classes within areas of land cover confusion, which hinders annotators from accurately determining the true ground surface. Simultaneously, class imbalance stems from significant differences in the number of samples between classes in the land cover distribution, ultimately leading to insufficient learning of feature representations for a few classes during model optimization. To address these issues, researchers have proposed a biased label learning strategy for label ambiguity and a class balancing strategy to address class imbalance. Partial labeling learning originates from the fuzzy labeling / partial labeling paradigm, where each land cover sample's label set includes true labels and confusing candidate labels. However, the ideal assumption behind this approach is that the number of land cover samples of each class is approximately uniformly distributed in the collected data. This condition is difficult to guarantee during hyperspectral imaging data acquisition. Class balance strategies aim to systematically alleviate the problem of class imbalance in classification tasks, thereby maintaining the integrity of the decision boundary. However, these methods require accurately labeled data to precisely statistically analyze class distribution, which cannot solve the class imbalance problem in some labeled data. In general, existing methods only focus on alleviating label ambiguity or class imbalance, and no existing framework can comprehensively solve both problems simultaneously. Therefore, in the field of hyperspectral image land cover recognition, there is an urgent need to develop robust algorithms that can simultaneously address class imbalance and label ambiguity, thereby expanding the applicable scenarios for land cover recognition tasks. Summary of the Invention
[0004] In view of this, the purpose of this invention is to provide a method and system for land cover recognition based on dynamic resampling and partial labeling in hyperspectral images. The core idea is to iteratively update land cover categories based on prediction results during training and perform dynamic category-balanced resampling. This invention consists of a dynamic prototype resampling network and a partial labeling learning network. In the dynamic prototype resampling stage, land cover samples are extracted from the normal distribution of the category prototypes to ensure that the number of samples from the minority categories in each training batch reaches their mean. Subsequently, in the partial labeling learning stage, a multi-directional frequency domain attention network is introduced to improve the quality of features used for label disambiguation, and the model parameters are optimized through smooth partial labeling loss and cross-entropy loss. To achieve the above objectives, this invention provides the following technical solution: A method and system for identifying ground features in a dynamic resampling biased label hyperspectral image, the method comprising the following steps: S1: acquiring hyperspectral remote sensing images S1: The image dimensions are h and w, and b is the spectral dimension. S2: The number of samples of interest in the hyperspectral remote sensing image is counted, and the land cover with a lower than average number of samples is dynamically resampled. S3: The sampled samples are input into the partial label learning network for training, and label disambiguation is performed. S4: After training, the test samples are used to identify land cover and obtain the results.
[0005] Furthermore, in step S1, the initial land cover identification data for the hyperspectral image is prepared, specifically including: firstly, cutting all samples according to patch size and dividing them into training samples according to a certain ratio. N is the number of samples; then, by randomly converting the non-true labels of pixels into candidate labels according to a certain threshold R, the candidate label set for each sample is obtained. ,in Includes samples The set of true labels is used; then, the candidate label set of the samples is normalized to obtain the initial sample candidate label confidence vector matrix. ],in Let represent the label confidence vector of the i-th sample. This represents the probability that label j is its true land cover category.
[0006] Furthermore, in step S2, during the dynamic prototype resampling stage, a certain number of samples are synthesized for land features with small sample sizes. First, using the initial category labeling information of the data, the initial prototypes of each category in the land feature category space are obtained. Where c is the number of land cover categories of interest:
[0007] in It is the first in the entire tag space i The class's category prototype, These are sample data belonging to category i. Then, training sample batches are loaded from the training sample set according to batch size n, and the number of samples for each land cover category is counted in each training batch. And set the sampling threshold for This determines which type of land cover requires additional sampling. If the number of samples for the i-th land cover is less than... If the sample size is large enough, it is considered a minority land cover class that requires additional samples to enhance feature extraction; conversely, if the sample size is large enough, it is considered a minority land cover class. Taking minority class i as an example, assume that all samples of this class follow a Gaussian distribution. The corresponding distribution parameters are derived from the prototype of this type of land cover. The calculation yielded the result. Then, based on this assumption, additional samples were drawn from the corresponding Gaussian distribution in the following manner. :
[0008] in, and These are the expectation and variance of the normal distribution of the prototype features of the category, respectively, which are used in the subsequent sampling process of the corresponding category samples.
[0009] Then, to ensure that the categorical features of the additional samples are representative, they are compared with the corresponding categorical prototypes in the following manner. The synthesized sample is obtained by weighted summation:
[0010] Here, w represents the weights for values between 0 and 1. Finally, due to label ambiguity, the true label of a sample is hidden within the candidate label set, making it difficult to obtain a highly accurate category prototype under initial conditions. Therefore, based on the subsequent label disambiguation process, the category prototype is dynamically updated during model training according to the label confidence matrix Q, serving as an important basis for subsequent sampling.
[0011] in, For the sample The pseudo-label is of category i. The weighting coefficient has a value between 0 and 1.
[0012] Further, in step S3, feature extraction and label disambiguation are performed on the training samples. First, a multi-directional frequency domain attention network is designed, consisting of three feature extraction modules. Each layer uses the same structure, with the input of each layer being the output of the previous layer and the output becoming the input of the next layer. In each layer, spatial information is first extracted using a convolutional network, as shown in the following formula:
[0013] in , , These represent the dropout function, the GELU activation function, and the BatchNormalization function, respectively. For sample i, the output result at layer m. Let i be the feature representation of sample i at the (m-1)th layer, i.e., the output of the i-th layer module, where the input of the first layer is the initial representation of the sample. ,Right now Then, a multi-directional frequency domain attention module is designed for further feature extraction. Specifically, it will... Convolutions with kernels of 1 and 5 as inputs yield the results used for attention computation. , Then, , Fourier transforms are performed along the channel direction, vertical direction, and horizontal direction, and the characteristic correlations in the frequency domain are estimated using the following formula:
[0014] in This represents the Fourier transform function along the channel direction. This represents the Fourier transform function along the vertical direction. This represents the Fourier transform function along the horizontal direction. , , These are the corresponding inverse transform functions. , , Attention scores are calculated along the channel direction, vertical direction, and horizontal direction, respectively; then, the relevance features are obtained through matrix multiplication, as shown in the following formula:
[0015] To prevent gradient vanishing, input features The output is obtained by adding the relevant features, as shown in the following formula:
[0016] Finally, the output features of each feature extraction module are fused to obtain a sample feature representation for ground feature identification. :
[0017] in This is the MaxPooling function. Similarly, the synthetic samples are input into a multi-directional frequency domain attention network in the same way for feature extraction, yielding the feature representation. Furthermore, to enhance the model's representation learning ability, Gaussian noise was added to the original samples for data augmentation, and an encoding process sharing weights with the aforementioned feature extraction process was implemented to obtain the feature representation. .
[0018] Furthermore, for the original sample Feature representation and A linear classifier with shared weights is used to predict the category, thereby obtaining the predicted label probability. as well as Then use as well as The sum of these values serves as the basis for label disambiguation, and the formula is as follows:
[0019] in , All are about samples The predicted probability on label j Based on , Add and normalize the samples Label confidence on label j.
[0020] Furthermore, to prevent overfitting during the biased label learning process, a smoothed biased label loss is designed to improve the model's generalization ability, i.e., smoothing the probability distribution of the label confidence matrix, as shown in the following formula:
[0021] in It is a sample The initial label confidence vector, Confidence is established for the smoothed labels. Then, based on... Calculate classification loss The formula is as follows:
[0022] Further calculation and Mutual information entropy loss To further constrain the model training process, the formula is as follows:
[0023] in This refers to the Softmax function.
[0024] Furthermore, regarding the feature representation of synthetic samples The classifier outputs its corresponding prediction result. Calculate the cross-entropy loss directly as follows:
[0025] in One-hot encoding of the synthetic sample labels. The final overall loss used for model optimization is given by the following formula:
[0026] After training, a trained hyperspectral image ground feature recognition framework is obtained; the input sample is detected by the trained hyperspectral image ground feature detection framework, and the hyperspectral image ground feature recognition result map is output.
[0027] The beneficial effects of this invention are as follows: This invention proposes a dynamic resampling biased label hyperspectral image land cover recognition algorithm, which consists of two modules: a dynamic prototype dynamic resampling stage and biased labeling. It achieves accurate identification of land cover types, can simultaneously identify multiple land covers in hyperspectral images, enhances the model's ability to capture land cover features in hyperspectral images, solves the problems of land cover type imbalance and label ambiguity in hyperspectral images, enhances the network's robustness and generalization ability, and significantly improves land cover recognition performance. Experimental results on hyperspectral image datasets show that the proposed model outperforms current state-of-the-art hyperspectral image land cover recognition methods.
[0028] Other advantages, features, and characteristics of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The features and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description
[0029] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein: Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a diagram illustrating the overall framework of a dynamic resampling biased label hyperspectral image land cover identification method and system according to the present invention. Figure 3 This is a flowchart of the dynamic prototype resampling process of the present invention; Figure 4 This is a structural diagram of the multi-directional frequency domain attention network of the present invention; Figure 5 The images show the visualization results of different recognition methods on the hyperspectral image dataset of the present invention, where (a) is a pseudo-color image, (b) is a ground truth image, (c) is CrCD, (d) is MAGS, (e) is CEL, (f) is Records+SSFTT, (g) is Records+MSSTT, (h) is Records+MDFF, (i) is HTC-LTPLL+SSFTT, (g) is HTC-LTPLL+MSSTT, (k) is HTC-LTPLL+MDFF, (l) is DrPLL+SSFTT, (m) is DrPLL+MSSTT, and (n) is DrPLL+MDFF. Detailed Implementation
[0030] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings.
[0031] Figure 1 This is a flowchart of the method of the present invention. The present invention provides a method and system for land cover recognition in dynamically resampled, partially labeled hyperspectral images. As shown in the figure, in the data preparation stage, sample patches are divided with the labeled sample as the center based on a specific patch size, and a candidate label set for the sample is obtained according to a certain threshold. The dynamic resampled, partially labeled hyperspectral image land cover recognition is as follows: Figure 2 As shown, this method can simultaneously solve the problems of class imbalance and label ambiguity in hyperspectral image land cover recognition tasks. The method consists of a dynamic prototype resampling network and a partial label learning network. First, utilizing the idea of dynamic change, new samples are synthesized for the minority classes with smaller sample sizes based on the class prototypes of the minority classes, thereby balancing the distribution of samples from different classes in the data. Furthermore, a multi-directional frequency domain attention network is designed to extract discriminative features from the samples. Finally, a smoothed partial label loss is designed to prevent model overfitting and improve its generalization ability. Specifically, the technical solution of this invention includes the following: 1. Acquire multi-feature hyperspectral remote sensing images: Cut all ground feature samples according to patch size and divide them into training samples according to a certain ratio. N is the number of samples; then, by randomly converting the non-true labels of pixels into candidate labels according to a certain threshold R, the candidate label set for each sample is obtained. ,in Includes samples The set of true labels is used; then, the candidate label set of the samples is normalized to obtain the initial sample candidate label confidence vector matrix. ],in Let represent the label confidence vector of the i-th sample. This represents the probability that label j is its true land cover category.
[0032] 2. Dynamic resampling of minority class samples: such as Figure 2 , Figure 3 As shown, in the dynamic prototype resampling stage, a certain number of samples are synthesized for land features with small sample sizes. First, the initial category labeling information of the land feature data is used to obtain the initial prototypes of each category in the land feature category space. Where c is the number of land cover categories of interest:
[0033] in It is the first in the entire tag space i The prototype of each land cover category is then determined. Next, training sample batches are loaded from the training sample set according to batch size n, and the number of samples for each land cover category is counted within each training batch. And set the sampling threshold for This determines which type of land cover requires additional sampling. If the number of samples for the i-th land cover is less than... If the sample size is large enough, it is considered a minority land cover class that requires additional samples to enhance feature extraction; conversely, if the sample size is large enough, it is considered a minority land cover class. Taking minority class i as an example, assume that all samples of this class follow a Gaussian distribution. The corresponding distribution parameters are derived from the prototype of this type of land cover. The calculation yielded the result. Then, based on this assumption, additional samples were drawn from the corresponding Gaussian distribution in the following manner. :
[0034] Then, to ensure that the categorical features of the additional samples are representative, they are compared with the corresponding categorical prototypes in the following manner. The synthesized sample is obtained by weighted summation:
[0035] Here, w represents the weights for values between 0 and 1. Finally, due to label ambiguity, the true label of a sample is hidden within the candidate label set, making it difficult to obtain a highly accurate category prototype under initial conditions. Therefore, based on the subsequent label disambiguation process, the category prototype is dynamically updated during model training according to the label confidence matrix Q, serving as an important basis for subsequent sampling.
[0036] in, For the sample The pseudo-label is of category i.
[0037] 3. Hyperspectral image biased label learning: such as Figure 2 As shown, in the partially labeled learning network, feature extraction is first performed based on a multi-directional frequency domain attention mechanism, followed by land cover recognition, label disambiguation, loss calculation, and model updating. Figure 4 As shown, the multi-directional frequency domain attention network consists of three feature extraction modules, each with the same structure. The input of each layer is the output of the previous layer, and the output is the input of the next layer. In each layer, spatial information is first extracted using a convolutional network, as shown in the following formula:
[0038] in , , These represent the dropout function, the GELU activation function, and the BatchNormalization function, respectively. Let i be the feature representation of sample i at the (m-1)th layer, i.e., the output of the i-th layer module, where the input of the first layer is the initial representation of the sample. ,Right now Then, a multi-directional frequency domain attention module is designed for further feature extraction. Specifically, it will... Convolutions with kernels of 1 and 5 as inputs yield the results used for attention computation. , Then, , Fourier transforms are performed along the channel direction, vertical direction, and horizontal direction, and the characteristic correlations in the frequency domain are estimated using the following formula:
[0039] in This represents the Fourier transform function along the channel direction. This represents the Fourier transform function along the vertical direction. This represents the Fourier transform function along the horizontal direction. , , These are the corresponding inverse transformation functions; then, the correlation features are obtained through matrix dot product, as shown in the following formula:
[0040] To prevent gradient vanishing, input features The output is obtained by adding the relevant features, as shown in the following formula:
[0041] Finally, the output features of each feature extraction module are fused to obtain the sample feature representation for multi-class ground cover detection:
[0042] in This is the MaxPooling function. Similarly, the synthetic samples are input into a multi-directional frequency domain attention network in the same way for feature extraction, yielding the feature representation. Furthermore, to enhance the model's representation learning ability, Gaussian noise was added to the original samples for data augmentation, and an encoding process sharing weights with the aforementioned feature extraction process was implemented to obtain the feature representation. .
[0043] 4. After feature extraction, the original sample is obtained. Feature representation and A linear classifier with shared weights is used to predict the category, thereby obtaining the predicted label probability. as well as Then use as well as The sum of these values serves as the basis for label disambiguation, and the formula is as follows:
[0044] in , All are about samples Predicted probability on label j.
[0045] 5. To prevent overfitting during the biased label learning process, a smoothed biased label loss is designed to improve the model's generalization ability, i.e., smoothing the probability distribution of the label confidence matrix, as shown in the following formula:
[0046] in It is a sample The initial label confidence vector. Then based on... The classification loss is calculated using the following formula:
[0047] Then, calculate and The mutual information entropy loss between them further constrains the model training process, as shown in the following formula:
[0048] in This refers to the Softmax function.
[0049] Secondly, regarding the feature representation of synthetic samples The classifier outputs its corresponding prediction result. Calculate the cross-entropy loss directly as follows:
[0050] in One-hot encoding of the synthetic sample labels. The final overall loss used for model optimization is given by the following formula:
[0051] Finally, the network is trained based on the overall loss, enabling it to identify the true types of ground features from the candidate label set.
[0052] The trained ground feature recognition framework is used to identify ground feature types from input samples, and the resulting ground feature recognition map is output. For example... Figure 4This paper presents a comparison of the DrPLL hyperspectral image dynamic balanced biased labeling land cover recognition framework and the proposed multi-directional attention network MDFF within this framework with existing imbalanced biased labeling frameworks Records, HTC-LTPLL, and existing hyperspectral image feature extraction networks SSFTT and MSSTT, using different combinations. It also compares existing biased labeling models: CrCD, MAGS, and CEL. The comparison results show that land cover types are well identified. Table 1 presents the performance indicators of different methods on OA, KA, and AA. Table 1. Comparison of AGMS with various methods on the Indian Pines hyperspectral image dataset (mean values)
[0053] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications should be covered within the scope of the claims of the present invention.
Claims
1. A method for identifying ground features in dynamically resampled, biased-label hyperspectral images, characterized in that: The method includes the following steps: S1: Establish partial label samples for hyperspectral remote sensing images; S2: Dynamic resampling of minority class samples; S3: Hyperspectral image biased label learning; S4: After training, perform ground feature recognition on the complete image to obtain prediction results; In step S1, a hyperspectral remote sensing image is acquired. Where h and w are the image dimensions, and b is the spectral dimension, all samples are cut according to patch size and divided into training samples according to a certain ratio. N is the number of samples; then, by randomly converting the non-true labels of image pixels into candidate labels according to a certain threshold R, the candidate label set for each sample is obtained. ,in Includes samples The set of true labels is used; then, the candidate label set of the samples is normalized to obtain the initial sample candidate label confidence vector matrix. ],in Let represent the label confidence vector of the i-th sample. This represents the probability that its label j is its true land cover category; In step S2, the dynamic prototype resampling includes category distribution information statistics, resampling, and category prototype initialization and updating. For initializing the category prototype... It is obtained by averaging the features of all samples in each class: ,in It is the first in the entire tag space i The class category prototype, where m represents the class to which the class belongs. i Total number of samples of land-like features It belongs to the category i Sample data; In dynamic prototype resampling, based on the subsequent label disambiguation process, the category prototypes are dynamically updated during model training according to the label confidence moments Q, serving as an important basis for subsequent resampling. ,in, For the sample The pseudo-label is of category i, and M is the total number of samples with pseudo-labels of category i. The weighting coefficients are values between 0 and 1. The position function that takes the maximum value; During dynamic resampling, n training samples are loaded into the network in each batch, and the number of samples for each land cover class is counted in each training batch. And set the sampling threshold = To determine which type of land cover requires additional sampling; if the number of samples for the i-th type of land cover is less than If the sample size is small, it is considered a minority land cover class that needs additional samples to enhance feature extraction; conversely, if the sample size is small, it is considered a minority land cover class. i For example, assume that all samples of this type of land cover follow a Gaussian distribution. The corresponding distribution parameters are derived from the prototype of this type of land cover. The calculation yields the result; then, based on this assumption, additional samples are drawn from the corresponding Gaussian distribution in the following manner. : ,in, and These are the expectation and variance of the normal distribution of the class prototype features, respectively, which are used in the subsequent sampling process of the corresponding class samples; Then, to ensure that the categorical features of the additional samples are representative, they are compared with the corresponding categorical prototypes in the following manner. The synthesized sample is obtained by weighted summation: Where w is the weight of values between 0 and 1. These are synthetic samples used for class balancing.
2. The method for identifying ground features in a dynamic resampling biased label hyperspectral image according to claim 1, characterized in that: In step S3, the partial label learning process includes feature extraction, land cover recognition, and label disambiguation. During feature extraction, land cover features are acquired through a multi-directional frequency domain attention network composed of three feature extraction modules. In each layer, spatial information is first extracted using a convolutional network, as shown in the following formula: ,in, , , These represent the dropout function, the GELU activation function, and the Batch Normalization function, respectively. For sample i, the output result at layer m. The feature representation of sample i at layer m-1 is obtained, which is the output of the module at layer i. Then, deep feature extraction is performed through a multi-directional frequency domain attention module. Convolutions with kernels of 1 and 5 as inputs yield the results used for attention computation. , Then, , Fourier transforms are performed along the channel direction, vertical direction, and horizontal direction, and the characteristic correlations in the frequency domain are estimated using the following formula: ,in, This represents the Fourier transform function along the channel direction. This represents the Fourier transform function along the vertical direction. This represents the Fourier transform function along the horizontal direction. , , These are the inverse transform functions for the corresponding directions. This is a conjugate operation. , , Attention scores are calculated along the channel direction, vertical direction, and horizontal direction, respectively; then, the relevance features are obtained through matrix multiplication, as shown in the following formula: ,in, The relevance features obtained in the m-th layer feature extraction network will be used as input features. Correlation characteristics The two numbers are added together to obtain the output, as shown in the following formula: Finally, the output features of each feature extraction module are combined to obtain the sample feature representation for ground feature identification. : ,in, Using the MaxPooling function, the synthesized samples are input into a multi-directional frequency domain attention network in the same way for feature extraction, resulting in feature representations. Gaussian noise is added to the original samples for data augmentation, and an encoding process that shares weights with the feature extraction process described above is implemented to obtain the feature representation. .
3. The method for identifying ground features in a dynamic resampling biased label hyperspectral image according to claim 2, characterized in that: In the label disambiguation process, the original samples are used Feature representation and The predicted label probability is obtained by predicting the category using a linear classifier with shared weights. as well as ,use as well as The sum of these values serves as the basis for label disambiguation, and the formula is as follows: ,in, , All are about samples The predicted probability on label j Based on , Add and normalize the samples To prevent overfitting during the biased label learning process, the label confidence level for label j is calculated by smoothing the biased label loss, which is essentially smoothing the probability distribution of the label confidence matrix. The formula is as follows: ,in, It is a sample The initial label confidence vector, To establish confidence in the smoothed labels, based on Calculate classification loss The formula is as follows: Further calculation and Mutual information entropy loss To further constrain the model training process, the formula is as follows: ,in, For the Softmax function, further, for the feature representation of synthetic samples The classifier outputs its corresponding prediction result. Calculate the cross-entropy loss directly as follows: ,in, The one-hot encoding of the synthetic sample labels is used to obtain the overall loss for model optimization, as shown in the following formula: The test set is input into the trained network to predict land cover types, and the prediction results are output.
4. A dynamic resampling biased label hyperspectral image feature recognition system, characterized in that: The system employs the method described in any one of claims 1 to 3.