Noise remote sensing image classification method based on consistency regularization

Abstract

The invention relates to a noise remote sensing image classification method based on consistency regularization, and belongs to the field of image classification. According to the method, pseudo labels from a weak data set are used, refining is carried out through confidence scores, consistency regularization is adopted to deal with noise in a training data set, and data quality is improved by identifying and modifying images containing related noise, so that classification accuracy and robustness of the model are enhanced. According to the method, the self-adaptive threshold value is creatively applied to correct the label, and the pseudo label generation process in the SSL is optimized. In this way, the number of iterations required by label correction is reduced, and the remote sensing image classification effect in a noise label scene is remarkably improved.

Description

Technical Field

[0001] The invention belongs to the field of image classification, and in particular relates to a noisy remote sensing image classification method based on consistency regularization. Background Art

[0002] In order to break away from the impact of inaccurate annotations on datasets, several classification algorithms for noisy labels have recently emerged. For example, Pakrashi et al. [1] introduced the Kalman Tune framework, which uses the Kalman filter (KF) to optimize an ensemble model in a noisy environment. This method uses the fusion ability of the KF to reduce label noise, thereby improving the performance of the boosting-based ensemble model. Yuan et al. [2] demonstrated this through their iterative cross-learning (ICL) method, which is designed for deep learning datasets with noisy labels. Iterative cross-learning divides the dataset into multiple subsets to train separate networks and adjusts the labels based on the prediction consistency. Labels that lack consistency are replaced with random values, thereby diluting the established noise pattern and distributing the noise more evenly among the labels. Wei et al. [3] designed a dual-network framework for parallel training. One network predicts the labels of the training dataset, while the other network calculates a co-regularized loss. Their joint loss function update enhances the collective resistance of the networks to noisy labels. A common limitation of these methods is that they are less effective when the labels are heavily contaminated, which is also confirmed by our experiments.

[0003] Existing remote sensing image classification techniques face multiple challenges. First, many traditional methods such as support vector machines and random forests usually assume that the labels of the training data are completely accurate, so their performance will drop significantly when encountering noisy labels, mainly because they lack effective mechanisms for handling incorrect labels. In addition, even when using advanced deep learning techniques such as convolutional neural networks, these models often show insufficient generalization ability when dealing with large-scale, noisy data sets. Incorrect labels lead to incorrect feature learning, which in turn affects the overall performance of the model. At the same time, label cleaning usually requires complex preprocessing steps, which is not only computationally expensive, but also relies on specific algorithms or manual intervention, limiting its practicality in large-scale applications. Finally, existing technologies often find it difficult to effectively deal with the extremely complex geographical and environmental features in remote sensing images, especially under the influence of label noise, and these technologies often suffer from a significant reduction in classification accuracy and reliability. These problems highlight the shortcomings of existing technologies in dealing with noisy remote sensing data, and more effective solutions are urgently needed to improve classification performance and adaptability. Summary of the invention

[0004] The purpose of the present invention is to address the defects of the prior art and provide a noisy remote sensing image classification method based on consistency regularization. The method uses pseudo labels from weak data sets and refines them through confidence scores. Consistency regularization is used to deal with the noise in the training data set. By identifying and modifying images containing correlated noise, the data quality is improved, thereby enhancing the classification accuracy and robustness of the model. The method of the present invention innovatively applies an adaptive threshold to correct labels and optimizes the pseudo label generation process in SSL. This reduces the number of iterations required for label correction and significantly improves the remote sensing image classification effect in noisy label scenarios.

[0005] To achieve the above object, the technical solution of the present invention is: a noisy remote sensing image classification method based on consistency regularization, comprising:

[0006] Train a DCNN model that integrates basic noise learning technology to roughly distinguish the labels of images in the original dataset to divide the initial dataset into a strong dataset and a weak dataset;

[0007] A remote sensing image classification model is trained based on strong data sets and weak data sets. An adaptive threshold mechanism is introduced in the process of training the remote sensing image classification model with weak data sets. The trained remote sensing image classification model can classify noisy remote sensing images.

[0008] In one embodiment of the present invention, the DCNN model integrating the basic noise learning technology is trained to roughly distinguish the labels of the images in the original data set to divide the initial data set into a strong data set and a weak data set. The specific implementation process is as follows:

[0009] Using an initial dataset that may contain incorrect classification labels Training the DCNN model integrated with the basic noise learning technology to obtain a trained DCNN model integrated with the basic noise learning technology;

[0010] Predict the initial dataset using the trained DCNN model with the base noise learning technique The label of the image and the label confidence corresponding to the label of the image, N is the total number of data in the initial data set, x i ,y i Represents the image and its corresponding label respectively. If the predicted label of the image The true label y of the corresponding image in the initial dataset i If they are consistent, the corresponding images are assigned to the strong dataset Where M represents the number of correctly labeled images; if the predicted label of the image The true label y of the corresponding image in the initial dataset iInconsistent, for images whose predicted label confidence is higher than the confidence threshold, the predicted label confidence of the corresponding image is re-labeled as the highest predicted label confidence among all images, and the corresponding image is placed in the weak dataset In the example, images whose predicted label confidence is lower than the confidence threshold are directly placed in the weak dataset. middle.

[0011] In one embodiment of the present invention, before training the DCNN model integrating the basic noise learning technology, a Mixup enhancement operation needs to be performed on the initial data set.

[0012] In one embodiment of the present invention, the DCNN model is a PENCIL model.

[0013] In one embodiment of the present invention, the adaptive threshold mechanism is specifically implemented as follows:

[0014] First, define the following formula:

[0015]

[0016] Here, T represents the learning duration, E represents the total number of training cycles, and e represents the current training count;

[0017] Second, using the weak dataset D w The number of instances exceeding the threshold in is used to determine the degree of learning of the model, that is, the following formula is used to evaluate the degree of learning of the model:

[0018]

[0019] Here, M(c) represents the degree of learning of category c in the image, and c(p>τ) represents the weak dataset D w The number of instances in which the predicted label confidence p of images belonging to category c exceeds the threshold τ, O represents the weak dataset D w Total number of instances in

[0020] Set different thresholds for the model's learning level based on different time intervals:

[0021] A(c,e)=H(c)×(1-M(c)-T(e))

[0022] Among them, A(c,e) represents the threshold of category c in the image at the current training count e, and H(c) represents the initial threshold of category c in the image.

[0023] In one embodiment of the present invention, after each training iteration, the weak data set D is reduced by a factor α. w The number of instances in the corresponding formula is expressed as:

[0024] C(t)=α·C(t-1)

[0025] Where C(t) represents the weak data set D at the current training count e w The number of instances in the dataset, C(t-1) represents the weak dataset D in the previous iteration. w The number of instances in .

[0026] In one embodiment of the present invention, before training the remote sensing image classification model, a common enhancement operation is performed on the strong data set, and a weak enhancement operation and a strong enhancement operation are performed on the weak data set.

[0027] In one embodiment of the present invention, after the weak data set is subjected to the weak enhancement operation, the following operations need to be performed on the image in the weak data set subjected to the weak enhancement operation:

[0028] (1) If the label confidence of the image after the weak enhancement operation is greater than the confidence threshold, the label of the image after the weak enhancement operation is matched with the label of the image after the strong enhancement operation, and the count is performed. When the count exceeds the count threshold, the label of the image after the weak enhancement operation is updated to the label of the image after the strong enhancement operation, i.e., the pseudo label;

[0029] (2) If the label confidence of the image after the weak enhancement operation does not exceed the confidence threshold, the label of the image after the weak enhancement operation is not updated;

[0030] The confidence thresholds in steps (1) and (2) are adaptively adjusted through the introduced adaptive threshold mechanism.

[0031] In one embodiment of the present invention, the overall loss function L of the method is expressed as:

[0032] L=L s +b*L p =L x +L m +b*(L w +L r )

[0033] Among them, b is a fixed threshold used to dynamically adjust the trust level of unlabeled data, L s represents the strong dataset training loss function, L p represents the weak dataset training loss function; L s =L x +L m , L m represents the loss function obtained by Mixup enhancement, L x Represents the loss function obtained without Mixup enhancement. If the initial data set is not enhanced with Mixup, L m =0;L p =L w +L r , Lw represents the loss function in step (1), L r represents the loss function in step (2).

[0034] In one embodiment of the present invention, the specific calculation formulas of each loss function are as follows:

[0035]

[0036] in, y i Respectively represent the labels of the i-th image after Mixup enhancement and without Mixup enhancement, x i They represent the i-th image enhanced by Mixup and the i-th image not enhanced by Mixup, respectively. is the output probability of category c in the image enhanced by Mixup, p(x i =c) is the output probability of category c in the image without Mixup enhancement; represents the label of the i-th image in the weak dataset replaced with a pseudo label, represents the label of the i-th image in the weak dataset, represents the i-th image in the weak dataset, Represents the output probability of category c in the image for the strong augmentation operation of the weak dataset.

[0037] Compared with the prior art, the present invention has the following beneficial effects:

[0038] The present invention effectively improves the performance and accuracy of remote sensing image classification in noisy label processing by integrating adaptive threshold mechanism and hybrid data set application. In addition, the integrated Mixup data enhancement method and the special training strategy for weak labels further enhance the robustness and generalization ability of the model. These innovations make this technology have significant advantages and application potential in the field of remote sensing image classification.

[0039] This invention belongs to the Fujian Province Key Technology Innovation and Industrialization Project (School-Enterprise Joint Project) (2024X0002). BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 It is a schematic diagram of the process of the present invention. DETAILED DESCRIPTION

[0041] The technical solution of the present invention is described in detail below in conjunction with the accompanying drawings.

[0042] The present invention provides a noisy remote sensing image classification method based on consistency regularization, comprising:

[0043] Train a DCNN model that integrates basic noise learning technology to roughly distinguish the labels of images in the original dataset to divide the initial dataset into a strong dataset and a weak dataset;

[0044] A remote sensing image classification model is trained based on strong data sets and weak data sets. An adaptive threshold mechanism is introduced in the process of training the remote sensing image classification model with weak data sets. The trained remote sensing image classification model can classify noisy remote sensing images.

[0045] The following is the specific implementation process of the present invention.

[0046] like Figure 1 As shown, a noisy remote sensing image classification method based on consistency regularization in an example of the present invention mainly includes:

[0047] 1. Choose reliable data

[0048] The initial dataset contained noisy labels, resulting in a low confidence level for the model trained using a standard DCNN. However, implementing a DCNN model that incorporates basic noise learning techniques can achieve better results. Therefore, in the label selection stage, the present invention uses this enhanced DCNN model to train a more robust model to help roughly distinguish labels. Unlike the old methods that rely on high threshold predictions or multiple model combinations to select the optimal image, the method of the present invention is effective in label acquisition while also reducing the reduction of relevant labels.

[0049] At the beginning of training, an initial dataset is used that may contain misclassified labels. The DCNN model designed for noisy labels is trained. The DCNN model used in the present invention is PENCIL.

[0050] Next, the trained DCNN model is used to predict the labels of all remote sensing images and their corresponding confidence scores. Compared with the true label y of the image in the original dataset i If the image is consistent, it is assigned to the strong dataset. Where M represents the number of correctly labeled images. On the other hand, if the predicted label of an image is and the true label y of the image in the initial dataset i Inconsistent, for images with prediction confidence higher than the threshold, re-label them as the prediction value with the highest confidence and put them in the weak dataset , and other images are directly put into the weak dataset. It is worth noting that D w and D s The combination of constitutes the entire original data set D, so D w ∪Ds =D.

[0051] 2. Adaptive Threshold

[0052] In the label correction step, a higher threshold was initially used to evaluate the reliability of the labels. However, this approach resulted in an uneven distribution of the corrected labels in the minority classes, possibly due to differences in the degree of learning of different classes. To address this imbalance, the present invention introduces an adaptive threshold mechanism. This mechanism is dynamically adjusted according to the specific learning level of each class, aiming to achieve more fair and accurate label correction in each class.

[0053]

[0054] Here, T represents the learning duration, E represents the total number of training cycles, and e represents the current training count.

[0055] Second, use D w The learning intensity of the model is determined by the number of instances exceeding the threshold in . Therefore, the present invention uses the following formula to evaluate the learning degree of the model:

[0056]

[0057] Here, M(c) represents the degree of learning of category c in the image, and c(p>τ) represents the weak dataset D w The number of instances in which the predicted label confidence p of images belonging to category c exceeds the confidence threshold τ, O represents the weak dataset D w Total number of instances in

[0058] Set different thresholds for the learning intensity of the model according to different time intervals:

[0059] A(c,e)=H(c)×(1-M(c)-T(e))

[0060] Among them, A(c,e) indicates that time e is the threshold of category c, and H(c) indicates the initial threshold of category c.

[0061] As the training progresses, after each training iteration, the present invention reduces the number of instances by a factor α. Therefore, the formula can be expressed as:

[0062] C(t)=α·C(t-1)

[0063] Here, C(t) represents the number of instances at time e, and C(t-1) represents the number of instances in the previous iteration.

[0064] 3. Label correction during training

[0065] After completing the data preprocessing stage, the data is systematically divided into two subsets: the robust dataset D s and the less reliable dataset D w This division improves the efficiency of the subsequent analysis phase. In the machine learning process, a s A deep convolutional neural network model trained on the dataset. This focused approach exploits the D s reliability and integrity, aiming to improve the stability and accuracy of the model training phase.

[0066] However, in the careful investigation of the present invention, it was found that s The class imbalance tendency is as high as 60% between the most and least populated classes. This observation highlights the importance of mitigation strategies to address the imbalanced class problem.

[0067] In addition, if the Mixup method is not used, it may quickly overfit in the second stage, when the confidence of all labels can be raised to a credible level in a short period of time. When the label prediction is incorrect, it will cause the entire model to predict the wrong direction without adjusting the relevant labels as the learning progresses.

[0068] Therefore, for the new data obtained by mixup Can be written as:

[0069]

[0070] Among them, x1 and x2 represent two pictures respectively, and y1 and y2 represent related labels.

[0071] For the loss function L obtained by using mixup enhancement m And the loss function L obtained by common x , which can be expressed as:

[0072]

[0073] in, The output probability of the cth class in the Mixup enhanced image calculated by DCNN (i.e., the DCNN used in the remote sensing image classification model), p(x i =c) is the output probability of the cth class in the image calculated by DCNN (i.e., the DCNN adopted by the remote sensing image classification model).

[0074] Since the labels of the segmentation results are not fully trusted, the present invention adopts the cross entropy loss function for the weak enhancement model and performs the following two operations on the image obtained by weak enhancement:

[0075] 1. If the value of the weakly enhanced label is greater than the set threshold, the present invention matches it with the strongly enhanced label and counts it based on the new label. When the count exceeds the threshold, the weakly enhanced label is updated to the new label.

[0076] 2. If the weakly enhanced label does not exceed the threshold, the present invention directly uses the label after segmenting the data set for prediction.

[0077] The confidence thresholds in 1 and 2 are adaptively adjusted through the introduced adaptive threshold mechanism.

[0078] Therefore, the formula can be expressed as:

[0079]

[0080] in, represents the label of the i-th image in the weak dataset replaced with a pseudo label, represents the label of the i-th image in the weak dataset, represents the i-th image in the weak dataset, Represents the output probability of category c in the image for the strong augmentation operation of the weak dataset.

[0081] Therefore, the overall loss function L of the present invention can be expressed as:

[0082] L=L s +b*L p =L x +L m +b*(L w +L r )

[0083] Where b is a fixed threshold. If Mixup enhancement is not used, L m =0.

[0084] In order to evaluate the effectiveness of the present invention in the classification of noisy remote sensing images, the present invention conducts experiments on the UCM dataset and compares it with several recently published algorithms. In addition, the effects of various modules are explored on the UCM dataset. The present invention compares Co-Teaching[4], Co-Teaching+[5], JoCoR[3], Decoupling[6], RS-MVVER[7], RS-ETDL[8], PENCIL

[10] and their benchmark model ResNet50[9]. The algorithm of the present invention uses mAP as the main evaluation indicator.

[0085] (1) Experimental results

[0086] Table 1 Ablation experiment of adaptive threshold label correction module

[0087]

[0088] The present invention first examines the impact of different initial thresholds of adaptive threshold label correction (ATRL) on model performance, covering different noise ratios. The experiment involves adjusting the initial threshold to change the proportion of modified labels in the weak dataset. As shown in Table 1, which shows the results on the WHU-RS19 dataset, the choice of the initial threshold significantly affects the results. For example, at noise ratios of 20% and 50%, an initial threshold of 5 produced the best results. In contrast, at noise ratios of 20% and 30%, an initial threshold of 15 was more effective. At noise ratios of 10% and 40%, it was determined that the initial threshold of 20 was optimal. Given that the initial threshold needs to be set at the beginning of training, 15 was selected as the initial threshold to ensure stable results in different scenarios.

[0089] In addition, to verify the effectiveness of the proposed method, the present invention conducted a comparative experiment on the UCM dataset with a noise ratio of 20%. As shown in detail in Table 2, the proposed method significantly outperforms the basic framework. Additional performance improvements were observed when the mixup technique was integrated. Even without including the adaptive threshold for semi-supervised learning (ATSL), the method implementation of the present invention surpassed the baseline performance. The addition of ATSL further enhanced this improvement. In addition, the continued use of label correction and padding techniques consistently enhanced the effect of the framework, thereby confirming the practicality of the proposed framework.

[0090] Table 2 Ablation experiments on the effectiveness of each module

[0091] Mixup Tab Filling Label correction Adaptive Threshold Accuracy 76.57 √ 77.05 √ √ 77.90 √ √ √ 78.00 √ √ √ √ 78.48

[0092] (2) Quantitative comparison

[0093] Table 3 shows the evaluation results on the UCM test set. By comparing the methods of Co-Teaching[4], Co-Teaching+[5], JoCoR[3], Decoupling[6], RS-MVVER[7], RS-ETDL[8] and their benchmark model ResNet50[9], it can be found that the method of the present invention performs very well under different noise ratios, is relatively stable, and is less affected by labels.

[0094] Table 3 Quantitative comparison on the UCM dataset

[0095]

[0096] References:

[0097] [1]A.Pakrashi and B.Mac Namee,“Kalman filter-based heuristic ensemble(kfhe):A new perspective on multi-class ensemble classification using kalmanfilters,”Information Sciences,vol.485,pp.456–485,2019.

[0098] [2]B.Yuan,J.Chen,W.Zhang,H.-S.Tai,and S.McMains,“Iterative crosslearning on noisy labels,”in 2018IEEE Winter Conference on applicationsofcomputer vision(WACV).IEEE,2018,pp.757–765.

[0099] [3]H.Wei,L.Feng,X.Chen,and B.An,“Combating noisy labels by agreement:Ajoint training method with co-regularization,”in Proceedings ofthe IEEE / CVFconference on computer vision and pattern recognition,2020,pp.13726–13735.

[0100] [4]B.Han,Q.Yao,X.Yu,G.Niu,M.Xu,W.Hu,I.Tsang,and M.Sugiyama,“Co-teaching:Robust training ofdeep neural networks with extremely noisy labels,”Advances in neural information processing systems,vol.31,2018.

[0101] [5]X.Yu,B.Han,J.Yao,G.Niu,I.Tsang,and M.Sugiyama,“How doesdisagreement help generalization against label corruption?”in InternationalConference on Machine Learning.PMLR,2019,pp.7164–717.

[0102] [6]B.Kang,S.Xie,M.Rohrbach,Z.Yan,A.Gordo,J.Feng,andY.Kalantidis,“Decoupling representation and classifier for long-tailed recognition,”inEighth.

[0103] [7]J.Wang,T.Wang,M.Gan,and G.Hadjichristofi,“Robust remote sensingscene classification with multi-view voting and entropy ranking,”inInternational Conference on Machine Learning for Cyber Security.Springer,2022,pp.87–98.

[0104] [8]Y.Li,Y.Zhang,and Z.Zhu,“Error-tolerant deep learning for remotesensing image scene classification,”IEEE transactions on cybernetics,vol.51,no.4,pp.1756–1768,202.

[0105] [9]K.He,X.Zhang,S.Ren,and J.Sun,“Deep residual learning for imagerecognition,”in Proceedings ofthe IEEE conference on computer visionandpattern recognition,2016,pp.770–778.

[0106]

[10] Y. Kun and W. Jianxin, "Probabilistic End-to-endNoise Correction for Learning with Noisy Labels," in The IEEE Conference on ComputerVisionandPattern Recognition (CVPR), 2019.

[0107] The above are preferred embodiments of the present invention. Any changes made according to the technical solution of the present invention, as long as the resulting functions do not exceed the scope of the technical solution of the present invention, belong to the protection scope of the present invention.

Claims

1. A noisy remote sensing image classification method based on consistency regularization, characterized in that: include: Train a DCNN model that integrates basic noise learning technology to roughly distinguish the labels of images in the original dataset to divide the initial dataset into a strong dataset and a weak dataset; A remote sensing image classification model is trained based on strong data sets and weak data sets. An adaptive threshold mechanism is introduced in the process of training the remote sensing image classification model with weak data sets. The trained remote sensing image classification model can classify noisy remote sensing images.

2. The noisy remote sensing image classification method based on consistency regularization according to claim 1, characterized in that: The specific implementation process of training the DCNN model integrating the basic noise learning technology to roughly distinguish the labels of the images in the original data set to divide the initial data set into strong data sets and weak data sets is as follows: Using an initial dataset that may contain incorrect classification labels Training the DCNN model integrated with the basic noise learning technology to obtain a trained DCNN model integrated with the basic noise learning technology; Predict the initial dataset using the trained DCNN model with the base noise learning technique The label of the image and the label confidence corresponding to the label of the image, N is the total number of data in the initial data set, x i ,y i Represents the image and its corresponding label respectively. If the predicted label of the image The true label y of the corresponding image in the initial dataset i If they are consistent, the corresponding images are assigned to the strong dataset Where M represents the number of correctly labeled images; if the predicted label of the image The true label y of the corresponding image in the initial dataset i Inconsistent, for images whose predicted label confidence is higher than the confidence threshold, the predicted label confidence of the corresponding image is re-labeled as the highest predicted label confidence among all images, and the corresponding image is placed in the weak dataset In the example, images whose predicted label confidence is lower than the confidence threshold are directly placed in the weak dataset. middle.

3. The noisy remote sensing image classification method based on consistency regularization according to claim 1, characterized in that: Before training the DCNN model that integrates the basic noise learning technology, the initial data set needs to be enhanced by Mixup.

4. The noisy remote sensing image classification method based on consistency regularization according to claim 1, characterized in that: The DCNN model is a PENCIL model.

5. The noisy remote sensing image classification method based on consistency regularization according to claim 1, characterized in that: The adaptive threshold mechanism is specifically implemented as follows: First, define the following formula: Here, T represents the learning duration, E represents the total number of training cycles, and e represents the current training count; Second, using the weak dataset D w The number of instances that exceed the confidence threshold in determines the degree of learning of the model, that is, the following formula is used to evaluate the degree of learning of the model: Here, M(c) represents the degree of learning of category c in the image, and c(p>τ) represents the weak dataset D w The number of instances in which the predicted label confidence p of images belonging to category c exceeds the confidence threshold τ, O represents the weak dataset D w Total number of instances in Set different thresholds for the model's learning level based on different time intervals: A(c,e)=H(c)×(1-M(c)-T(e)) Among them, A(c,e) represents the threshold of category c in the image at the current training count e, and H(c) represents the initial threshold of category c in the image.

6. The noisy remote sensing image classification method based on consistency regularization according to claim 5, characterized in that: After each training iteration, the weak dataset D is reduced by a factor α w The number of instances in the corresponding formula is expressed as: C(t)=α·C(t-1) Where C(t) represents the weak data set D at the current training count e w The number of instances in the dataset, C(t-1) represents the weak dataset D in the previous iteration. w The number of instances in .

7. The noisy remote sensing image classification method based on consistency regularization according to claim 1, characterized in that: Before training the remote sensing image classification model, ordinary enhancement operations are performed on the strong data set, and weak enhancement operations and strong enhancement operations are performed on the weak data set.

8. The noisy remote sensing image classification method based on consistency regularization according to claim 7, characterized in that: After performing the weak enhancement operation on the weak data set, the following operations need to be performed on the images in the weak data set after the weak enhancement operation: (1) If the label confidence of the image after the weak enhancement operation is greater than the confidence threshold, the label of the image after the weak enhancement operation is matched with the label of the image after the strong enhancement operation, and the count is performed. When the count exceeds the count threshold, the label of the image after the weak enhancement operation is updated to the label of the image after the strong enhancement operation, i.e., the pseudo label; (2) If the label confidence of the image after the weak enhancement operation does not exceed the confidence threshold, the label of the image after the weak enhancement operation is not updated; The confidence thresholds in steps (1) and (2) are adaptively adjusted through the introduced adaptive threshold mechanism.

9. The noisy remote sensing image classification method based on consistency regularization according to claim 8, characterized in that: The overall loss function L of the method is expressed as: L=L s +b*L p =L x +L m +b*(L w +L r ) Among them, b is a fixed threshold used to dynamically adjust the trust level of unlabeled data, L s represents the strong dataset training loss function, L p represents the weak dataset training loss function; L s =L x +L m , L m represents the loss function obtained by Mixup enhancement, L x Represents the loss function obtained without Mixup enhancement. If the initial data set is not enhanced with Mixup, L m =0;L p =L w +L r , L w represents the loss function in step (1), L r represents the loss function in step (2).

10. The noisy remote sensing image classification method based on consistency regularization according to claim 9, characterized in that: The specific calculation formulas of each loss function are as follows: in, y i Respectively represent the labels of the i-th image after Mixup enhancement and without Mixup enhancement, x i They represent the i-th image enhanced by Mixup and the i-th image not enhanced by Mixup, respectively. is the output probability of category c in the image enhanced by Mixup, p(x i =c) is the output probability of category c in the image without Mixup enhancement; represents the label of the i-th image in the weak dataset replaced with a pseudo label, represents the label of the i-th image in the weak dataset, represents the i-th image in the weak dataset, Represents the output probability of category c in the image for the strong augmentation operation of the weak dataset.

Images