Remote sensing data classification method based on transfer learning and label noise filtering

Abstract

The invention discloses a remote sensing data classification method based on transfer learning and label noise filtering, and the method comprises the steps: introducing feature weight dynamic adjustment through a dynamic adaptive transfer filtering module, calculating the applicability of a source domain sample feature in real time, and adjusting the influence of the feature in a transfer process; target domain sample labels are optimized through credibility regression driving; introducing a multi-scale space-time consistency learning framework, constructing a multi-modal feature fusion network, and dynamically adjusting the fusion weight of the multi-modal features; target domain sample label noise is corrected through interactive target domain sample label correction in combination with a prediction driving feedback mechanism and manual labeling cooperation; filtering noise target domain sample labels layer by layer and optimizing sample weight distribution by combining migration information of source domain sample features and target domain sample label distribution through a migration sensing nested noise filtering network; the scheme has the characteristic of realizing efficient and accurate classification of large-scale remote sensing data.

Description

Technical Field

[0001] The present invention relates to the field of remote sensing data processing, and in particular to a remote sensing data classification method based on transfer learning and label noise filtering. Background Art

[0002] With the rapid development of remote sensing technology, various types of remote sensing images, such as optical, SAR, and multispectral images, are widely used in fields such as land resources management, environmental protection, agricultural monitoring, and disaster warning. Remote sensing image classification, a core task in remote sensing data analysis, aims to classify pixels or objects into specific feature categories based on the spectral, spatial, and temporal characteristics of remote sensing images, thereby extracting valuable information.

[0003] Traditional remote sensing image classification methods primarily rely on feature-based machine learning methods, such as support vector machines (SVMs), random forests (RFs), and k-nearest neighbors (KNNs). These methods typically require manual extraction of spectral, texture, or shape features and rely on abundant labeled data for supervised training. However, with the continuous improvement in the resolution and complexity of remote sensing imagery, the difficulty of remote sensing data classification has also increased. On the one hand, high-resolution remote sensing images contain more detailed object information, but the differences between object classes become more subtle. On the other hand, with the increase in image dimensionality, such as multispectral and even hyperspectral data, the curse of dimensionality becomes more prominent, and the number of samples required by the model also increases significantly. Furthermore, annotating high-quality remote sensing data is costly, especially in fine-grained scenes requiring high-precision classification. Manual annotation often requires the participation of domain experts, which is not only time-consuming and labor-intensive but also susceptible to subjective factors of the annotator, resulting in a certain degree of noise in the target domain sample labels. This noise in the target domain sample labels can arise from spectral aliasing, blurred boundaries between object classes, interference from lighting and shadow effects in the image, and human error during the annotation process. As the noise ratio of target domain sample labels increases, the features learned by the model may not meet the actual classification requirements, thereby significantly reducing the model performance. Especially in deep learning models, the risk of overfitting noisy target domain sample labels is more serious.

[0004] In recent years, transfer learning has gradually become an important research direction in remote sensing data classification due to its outstanding performance in scenarios with small samples and insufficient data. By transferring knowledge from models pre-trained in large-scale general datasets (such as ImageNet) to the target remote sensing task, transfer learning methods can improve classification performance when labeled data is scarce. However, the application of transfer learning in the field of remote sensing still faces the problem of difficulty in cross-domain migration. There are often significant differences in data distribution between the source domain (such as natural images) and the target domain (such as remote sensing images). For example, the spectral distribution of remote sensing data is very different from that of ordinary RGB images, which may cause traditional transfer learning methods to introduce negative transfer effects, affecting the classification effect of the target task. In addition, in remote sensing data classification, there are the following problems that need to be solved: (1) Scarcity of labeled data: The cost of obtaining high-quality labeled data is high, and due to the complexity of remote sensing data, such as multi-spectral dimensions and fine-grained classification requirements, the coverage of labeled samples is often insufficient, which easily leads to insufficient diversity of model training data.

[0005] (2) Noise interference of target domain sample labels: Due to the spectral aliasing of remote sensing data, the similarity between ground object categories and subjective labeling errors, the labeled data may contain a high proportion of noisy target domain sample labels, which poses a serious challenge to the robustness of the classification model.

[0006] (3) Insufficient model robustness: Under the interference of target domain sample label noise, deep learning models are prone to overfitting noisy data, thereby reducing the model's generalization ability and classification performance.

[0007] (4) High computational complexity: Some current target domain sample label noise processing algorithms, such as filtering methods based on model uncertainty or two-stage training, have high computational overhead and are difficult to deploy efficiently in remote sensing big data environments.

[0008] To this end, the present invention is devoted to studying a remote sensing data classification method that combines transfer learning with target domain sample label noise filtering. By making full use of the knowledge transfer advantages in transfer learning and the robustness improvement capability of target domain sample label noise filtering, the bottleneck problems of the current remote sensing classification, namely the scarcity of labeled data and the interference of target domain sample label noise on the model, are solved, thereby achieving efficient and accurate classification of large-scale remote sensing data. Summary of the Invention

[0009] To achieve the above objectives, the inventors provide a remote sensing data classification method based on transfer learning and label noise filtering, comprising the following steps: S1, through the dynamic adaptive migration filtering module, introduces dynamic adjustment of feature weights, calculates the applicability of source domain sample features in real time according to the feature distribution of remote sensing images, and automatically adjusts their influence in the migration process; S2, after the initial alignment of the migration features, optimizes the target domain sample labels by driving credibility regression to address the target domain sample label noise in the remote sensing image data; S3, based on the optimization of target domain sample labels, introduces a multi-scale spatiotemporal consistency learning framework to build a multimodal feature fusion network, enhances the expressiveness of target category features, and dynamically adjusts the fusion weights of multimodal features to fully capture complex spatiotemporal features; S4, through interactive target domain sample label correction, combined with the prediction-driven feedback mechanism and manual annotation collaboration, further corrects the target domain sample label noise; S5, through the transfer-aware nested noise filtering network, combines the migration information of source domain sample features with the target domain sample label distribution, filters the noisy target domain sample labels layer by layer and optimizes the sample weight distribution.

[0010] As a preferred embodiment of the present invention, step S1 includes: S101 introduces a source domain sample feature adaptability calculation mechanism to adjust the adaptability of the source domain sample features in real time according to the feature distribution of the target domain samples, and calculates the adaptability weight of the source domain sample features. The expression is: ; Where exp represents the natural exponential function, and Represents the index, represents the number of sample features in the target domain, represents the hyperparameter that controls the range of influence of similarity, Indicates the Source domain sample characteristics The adaptive weight of Indicates the target domain sample features; S102, based on the adaptive calculation of source domain sample features, dynamically adjust the learning rate and feature alignment strategy in transfer learning, and dynamically adjust the learning rate The expression is: ; in, represents the initial learning rate, Represents a hyperparameter that controls the adjustment rate of the learning rate. represents the number of training steps, Indicates the number of weights; S103 introduces negative transfer risk assessment, calculates the transfer effect of source domain sample features in the target domain, and dynamically adjusts the transfer strategy according to the difference in transfer loss. If negative transfer is found in the transfer of source domain sample features in the target domain, the feature weight or transfer path is automatically adjusted to reduce the negative transfer risk. The expression is: ; in, Indicates the negative migration risk, which is used to reflect the degree of negative migration during the migration process. Indicates the source domain The loss value of the feature, Represents the loss value of the source domain sample feature in the target domain.

[0011] As a preferred embodiment of the present invention, step S2 includes: S201: Construct a credibility distribution function for target domain sample labels, analyze the source domain sample feature consistency, prediction confidence, and historical prediction trajectory, assign a credibility score to each target domain sample, and identify and mark target domain sample labels with noise. The expression is: ; in, Indicates the Source domain sample characteristics Hedi Target domain sample features The consistency measurement function between Represents the historical prediction trajectory for the Source domain sample characteristics The impact of is used to reflect the stability of its prediction, Represents the prediction confidence, which is used to measure the classifier’s confidence in the sample characteristics of the target domain. , , Represents the weight coefficient, which is used to balance the influence of different factors. Indicates the The credibility score of the target domain sample assignment; S202: Introduce a credibility regression network, train it based on the credibility score of each target domain sample, and output the corrected target domain sample label distribution, which is expressed as: ; in, represents the original target domain sample label, Represents the weight of the correction factor, which is used to control the magnitude of the correction. Indicates the set credibility threshold, represents the corrected target domain sample label, The credibility regression network is composed of multi-layer perceptrons; S203, using an iterative optimization algorithm, repeatedly optimize the target domain sample label set based on the corrected target domain sample labels. The expression is: ; in, Indicates the The corrected target domain sample labels after iterations, Indicates the The credibility score of the target domain sample after iterations, express The credibility threshold after iterations, Represents the adjustment coefficient, which is used to control the speed of correction of target domain sample labels. Indicates the Corrected target domain sample labels after iteration.

[0012] As a preferred embodiment of the present invention, step S3 includes: S301, through the spatial-spectral mapping function, the modal features of different dimensions in the remote sensing image are mapped in a unified spatial framework to achieve modal information fusion, which is expressed as: ; in, Represents the overall source domain sample characteristics and the overall target domain sample characteristics The distribution alignment between them is used to measure the adaptability of the source domain sample features in the target domain. Indicates the The characteristic representation of the modality, Indicates the Adaptive weights of modalities, represents the total number of modes, Represents the final fusion feature output; S302 uses a graph convolutional network to transform the modal features in the remote sensing image into a graph structure and process the feature consistency constraints at different scales. The expression is: ; in, Indicates the Graph convolution of modality feature maps, represents the graph convolution generated by the target domain sample label, represents the vector two-norm operation, Represents the loss function of the graph convolutional network, which is used to enhance the consistency of features between different spatial scales. The total number of graph convolutional feature maps representing the modality, represents the regularization coefficient, which is used to balance the consistency constraint and the graph convolutional network loss. represents the cross-scale consistency loss; S303 introduces an adaptive spatiotemporal feature adjustment mechanism, combines spatiotemporal feature adaptive learning with classification optimization algorithms, dynamically adjusts feature weights at different scales, and optimizes classification results. The expression is: ; in, Represents the spatiotemporal feature adaptive adjustment loss function, Indicates the The spatiotemporal characteristics of samples, represents the spatiotemporal characteristics of the prediction, Represents the weight adjustment coefficient, which is used to control the balance between spatiotemporal feature adjustment and classification error. Represents the standard classification loss function, used to calculate the difference between the predicted and true target domain sample labels The error between .

[0013] As a preferred embodiment of the present invention, the modal features include spectral features, spatial features and temporal features.

[0014] As a preferred embodiment of the present invention, step S4 includes: S401: Identify low-confidence target domain samples based on prediction confidence, introduce a quality assessment function for target domain sample features, filter out abnormal samples through quality assessment, and identify a low-confidence sample set by minimizing the objective function. The expression is: ; in, represents the minimization objective function, represents the overall target domain sample characteristics, represents the confidence of the model prediction, represents the quality assessment function, which is used to measure the consistency or abnormality of image features. represents the regularization coefficient, which is used to balance the weight between confidence and feature consistency. Represents a set of low-confidence target domain samples; S402: After identifying the low-confidence area, the automatic target domain sample label correction algorithm generates correction suggestions and optimizes the target domain sample labels in combination with human feedback. The expression is: ; in, represents an automatic target domain sample label correction algorithm, which is used to optimize the target domain sample labels by generating correction suggestions. Represents a correction function based on feedback information, which is used to combine the manually annotated feedback and the model's prediction information to generate correction suggestions for the target domain sample labels. represents the target domain sample label set for manual feedback, Represents the target domain sample label set after automatic and manual correction; S403: After the target domain sample label is corrected, the corrected target domain sample label and the current target domain sample label are updated synchronously through the online update mechanism. The difference between the corrected target domain sample label and the model target domain sample label is evaluated through the difference measurement function, and the target domain sample label weight is adjusted according to the difference. The expression is: ; in, represents the final revised target domain label set, is the online update function, which indicates the process of online correction and update. Represents the label difference measurement function, which is used to calculate the difference between the corrected label and the current target domain sample label. Represents the target domain sample labels predicted by the training dataset.

[0015] As a preferred embodiment of the present invention, step S5 includes: S501, through a multi-level nested structure, align the source domain sample features and the target domain sample features layer by layer. Each layer maps the source domain features to the target domain features based on the target domain sample label set mapping, realizing the dynamic fusion of the source domain features and the target domain features. The expression is: ; in, Represents the features after the target domain sample features are fused with the source domain sample features, Represents the fusion weight of each level, with a total of layer, Represents the overall source domain sample characteristics The mathematical expectation value between the target domain sample label set, Represents the overall target domain sample characteristics; S502 is the feature after the target domain sample feature and the source domain sample feature are fused. Calculate dynamic feature weights and dynamically adjust by minimizing the label filtering loss function Contribution to training, using regularization terms to ensure stable updates of weights, the expression is: ; in, represents the label filtering loss function, represents the dynamic feature weight, represents the dynamic feature weights learned in the previous iteration, used to maintain stability, represents the vector two-norm operation, represents the indicator function, when When the value is 1, represents the regularization coefficient of weight change, represents the cross entropy loss, which evaluates the accuracy of label prediction, represents the true target domain sample label, represents the predicted target domain sample label, represents the minimization of dynamic feature weight function; S503: Identify and filter out noise labels in the target domain using the source domain sample features as a guide, and optimize the filtering strategy using the adaptability evaluation function of the source domain sample features. The expression is: ; in, represents the minimization noise filtering strategy function, represents the noise filtering strategy, represents the source domain sample feature adaptation function, Represents the difference value of the predicted target domain sample label, represents the quality assessment function, which is used to measure the consistency or abnormality of image features. represents the label filtering loss function, Represents the adaptability evaluation function of the source domain sample features.

[0016] Different from the existing technology, the above technical solution has the following beneficial effects: (1) This method dynamically adjusts the influence weight of source domain sample features on the target domain during the transfer learning process to solve the negative transfer problem caused by the distribution difference between the source and target domains. This method not only enhances the feature transfer effect but also avoids the performance degradation of the model due to the distribution difference. In other words, it effectively solves the problem of classification performance degradation caused by the scarcity of labeled data and the noise interference of target domain sample labels in remote sensing data classification tasks. (2) This method constructs a credibility distribution function of the target domain sample noise label and scores the credibility of the target domain sample label of each sample in the dataset, thereby achieving accurate identification of the noisy target domain sample label and dynamic adjustment of the classification weight, effectively improving the robustness of this method; (3) This method comprehensively utilizes the characteristic information of spectral, spatial and temporal dimensions in the remote sensing data classification task, and strengthens the cross-scale discrimination ability of target categories through the consistency constraint method, so as to adapt to the complexity of multi-source heterogeneous remote sensing data; (4) This method effectively realizes the intelligent correction of the labels of noisy target domain samples in remote sensing image datasets through feedback updating and iterative optimization, ensuring that the classification model can learn based on high-quality data and improve classification accuracy.

[0017] (5) This method dynamically optimizes the filtering strategy of sample labels in the noisy target domain, enabling the model to more efficiently cope with large-scale remote sensing data classification tasks under high noise ratios, significantly improving the scalability and applicability of the task. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 The figure is a flow chart of the method described in the specific implementation manner. DETAILED DESCRIPTION

[0019] In order to explain the technical content, structural features, achieved objectives and effects of the technical solution in detail, the following is a detailed description in conjunction with specific embodiments and accompanying drawings.

[0020] like Figure 1 As shown, this embodiment provides a remote sensing data classification method based on transfer learning and label noise filtering, including the following steps: S1, through a dynamic adaptive transfer filtering module, introduces dynamic adjustment of feature weights. This module calculates the applicability of source domain sample features in real time based on the feature distribution of remote sensing images and automatically adjusts their influence during the transfer process. Specifically, a dynamic adaptive transfer filtering module is designed to address the distribution offset between the source domain (natural images) and the target domain (remote sensing images) through a real-time feature adaptation mechanism. The adaptive transfer filtering module introduces dynamic adjustment of feature weights. This technology calculates the applicability of source domain sample features in real time based on the feature distribution of remote sensing images and automatically adjusts their influence during the transfer process. By constructing a domain-aware optimizer, the adaptive transfer filtering module can dynamically optimize the learning rate and feature alignment strategy in the transfer path, effectively reducing the risk of negative transfer and laying the foundation for stable training of subsequent models.

[0021] S2: After the initial alignment of the migrated features, the target domain sample labels are optimized through credibility regression to address the target domain sample label noise present in remote sensing image data. Specifically: After the initial alignment of the migrated features, a credibility regression-driven target domain sample label optimization method is proposed to address the target domain sample label noise problem that is prevalent in remote sensing data. By constructing a credibility distribution function for the target domain sample labels, a multi-dimensional analysis of the sample's feature consistency, historical prediction trajectory, and confidence is performed to generate a credibility score. Combined with a credibility regression network, the credibility regression-driven target domain sample label optimization can dynamically correct the target domain sample label distribution of low-credibility samples and improve the overall quality of the dataset through an iterative optimization strategy. After this stage of optimization, the classification model's robustness to target domain sample label noise is significantly enhanced, providing a reliable foundation for high-quality classification tasks.

[0022] S3, based on the optimization of target domain sample labels, introduces a multi-scale spatiotemporal consistency learning framework, constructs a multimodal feature fusion network, strengthens the target category feature expression capability, and dynamically adjusts the fusion weights of multimodal features to achieve comprehensive capture of complex spatiotemporal features; Specifically: Based on the optimization of target domain sample labels, in order to further improve the adaptability of this method to complex remote sensing data, a multi-scale spatiotemporal consistency learning framework is introduced. This framework solves the fusion problem caused by the multi-source heterogeneous characteristics of remote sensing data by constructing a multimodal feature fusion network in spectral, spatial and temporal dimensions. By introducing a cross-scale consistency constraint method, the multi-scale spatiotemporal consistency learning framework strengthens the consistent expression capability of target category features at different time points and spatial scales, and dynamically adjusts the fusion weights of multimodal features to achieve comprehensive capture of complex spatiotemporal features by the classification model.

[0023] S4, through interactive target domain sample label correction, combined with a prediction-driven feedback mechanism and manual labeling collaboration, further corrects the target domain sample label noise; specifically: to address residual indistinguishable noisy target domain sample labels and data anomalies in specific scenarios, this method designs interactive target domain sample label correction, combines a prediction-driven feedback mechanism and manual labeling collaboration, and further corrects the target domain sample label noise through a closed-loop optimization system. Interactive target domain sample label correction captures low-confidence areas predicted by the model in real time, generates high-priority correction suggestions, and gradually optimizes the target domain sample label quality by combining automatic algorithms with manual feedback. Through online correction and synchronous model updates, interactive target domain sample label correction significantly improves the effectiveness of the dataset and provides more accurate training data for complex remote sensing classification tasks.

[0024] S5, through a transfer-aware nested noise filtering network, combines the migration information of source domain sample features with the target domain sample label distribution, filters the noisy target domain sample labels layer by layer, and optimizes the sample weight distribution. Specifically: To address the classification bottleneck problem in high-noise scenarios, this method proposes a transfer-aware nested noise filtering network. This network uses a multi-level nested structure to combine the migration information of source domain sample features with the target domain sample label distribution in the target domain, filters the noisy target domain sample labels layer by layer, and optimizes the sample weight distribution. The transfer-aware nested noise filtering network dynamically adjusts the noise filtering strategy to ensure that correct samples are not lost due to false filtering, significantly improving the model's adaptability to large-scale high-noise data.

[0025] In the above embodiment, step S1 specifically includes the following steps: S101, by introducing the source domain sample feature adaptability calculation mechanism, adjusts the adaptability of the source domain sample, i.e., the natural image feature, in real time according to the feature distribution of the target domain sample, i.e., the remote sensing image data, and calculates the adaptability weight of each feature. The core of this process is to assign a source domain sample feature adaptability weight to each source domain sample feature. This weight will dynamically adjust the influence of the source domain sample feature on the target domain sample feature during the migration process. The overall source domain sample feature is , the overall target domain sample characteristics are , the distribution of target domain sample features is , the calculation expression of the adaptive weight of the source domain sample feature is: ; Where exp represents the natural exponential function, and Represents the index, represents the number of sample features in the target domain, represents the hyperparameter that controls the range of influence of similarity, Indicates the Source domain sample characteristics The adaptability weight represents the influence of the source domain sample features when migrating to the target domain sample features. Indicates the target domain sample features.

[0026] S102, based on the adaptive calculation of the source domain sample features, dynamically adjust the learning rate and feature alignment strategy in the transfer learning to further improve the transfer effect and ensure smooth convergence during the transfer process. This step introduces a domain-aware optimizer, combined with the adaptive weights of the source domain sample features, to optimize the learning rate in the transfer process in real time, and dynamically adjust the feature alignment strategy according to the degree of alignment of the source domain and target domain sample features. This step ensures that the feature alignment of the target domain and the source domain is more accurate during the transfer process based on the adaptive weights of the source domain sample features in the above steps, thereby reducing the risk of negative transfer and dynamically adjusting the learning rate. The expression is: ; in, represents the initial learning rate, Represents a hyperparameter that controls the adjustment rate of the learning rate. represents the number of training steps, Indicates the The adaptive weight of the source domain sample features, Indicates the number of weights.

[0027] S103, by evaluating the negative transfer risk during the migration process, avoid the negative transfer effect of the source domain sample features in the target domain. This step further optimizes the source domain sample features based on the results of the dynamic adjustment of the feature adaptation weights and learning rates in the previous two steps to eliminate potential negative transfer risks. By introducing a negative transfer risk assessment mechanism, the migration effect of the source domain sample features in the target domain is calculated, and the migration strategy is dynamically adjusted according to the difference in migration loss. If it is found that the migration effect of the source domain sample features in the target domain is not ideal, that is, there is negative transfer, the feature weights or migration paths are automatically adjusted to avoid negative transfer effects and negative transfer risks. The expression is: ; in, Indicates the negative migration risk, reflecting the degree of negative migration during the migration process. Indicates the source domain The loss value of the feature, Represents the loss value of the source domain sample feature in the target domain.

[0028] In the above embodiment, step S2 specifically includes the following steps: S201: Construct a credibility distribution function for target domain sample labels. This function comprehensively considers the consistency of source domain sample features, prediction confidence, and historical prediction trajectory. Through multi-dimensional analysis of these factors, this embodiment can assign a credibility score to each target domain sample, thereby identifying and labeling target domain sample labels that may contain noise. The expression is: ; in, Indicates the Source domain sample characteristics Hedi Target domain sample features The consistency measurement function between Represents the historical prediction trajectory for the Source domain sample characteristics The impact of Represents the prediction confidence, which measures the classifier’s confidence in the sample characteristics of the target domain. , , Represents the weight coefficient, which is used to balance the influence of different factors. Indicates the The credibility scores assigned to target domain samples.

[0029] S202 introduces a credibility regression network to dynamically correct low-credibility target domain sample labels by modeling the credibility scores. The credibility regression network is trained based on the credibility scores of each target domain sample and outputs a corrected target domain sample label distribution, effectively eliminating the interference of target domain sample label noise on the model. The expression is: ; in, represents the original target domain sample label, Represents the weight of the correction factor, which controls the magnitude of the correction. Represents the set credibility threshold, below which the target domain sample labels will be corrected. represents the corrected target domain sample label, The credibility regression network is composed of multi-layer perceptrons.

[0030] In step S203, an iterative optimization algorithm is used to repeatedly optimize the target domain sample label set based on the corrected target domain sample labels, gradually improving the overall quality of the dataset. After each iteration, the credibility regression network updates the target domain sample label distribution based on the new target domain sample label correction results, further correcting the potential noisy target domain sample labels. The expression is: ; in, Indicates the The corrected target domain sample labels after iterations, Indicates the The credibility score of the target domain sample after iterations, express The credibility threshold after iterations, Represents the adjustment coefficient, which is used to control the speed of correction of target domain sample labels. Indicates the Corrected target domain sample labels after iteration.

[0031] In step S3 of the above embodiment, the steps are specifically included: S301 fuses the features of different modalities in the remote sensing image, such as spectral, spatial, and temporal dimensions. Using a spatial-spectral mapping function, the features of different dimensions are mapped into a unified spatial framework, thereby reducing the differences between modalities and achieving efficient fusion of multimodal information in remote sensing images. This resolves the inconsistency between features of different modalities and addresses the challenges posed by multi-source heterogeneous data. The expression is: ; in, Represents the overall source domain sample characteristics and the overall target domain sample characteristics The distribution alignment between them is used to measure the adaptability of the source domain sample features in the target domain. Indicates the Modalities, such as spectral, spatial, and temporal feature representations, Indicates the The adaptive weight of the modality is learned through training and represents the influence of the modality. represents the total number of modes, The final fused feature output combines weighted features from multiple modalities. This step introduces an adaptive attention mechanism to dynamically weight features from different modalities in remote sensing images, thereby optimizing the fusion of heterogeneous features from multiple sources. The multi-scale convolutional network extracts features at different scales, enabling the classification model to more comprehensively understand the spatial, spectral, and temporal information in the image.

[0032] S302 uses graph convolutional networks to model the spatiotemporal relationship of cross-scale features, converting the spatial, spectral, and temporal features in remote sensing images into graph structures and processing feature consistency constraints at different scales. By using cross-scale consistency constraints, the model's ability to express target categories at different scales is enhanced, avoiding information loss. The expression is: ; in, Indicates the Graph convolution of modality feature maps, represents the graph convolution generated by the target domain sample label, Represents the loss function of the graph convolutional network, which aims to enhance the consistency of features between different spatial scales. The total number of graph convolutional feature maps representing the modality, represents the regularization coefficient, which is used to balance the consistency constraint and the graph convolutional network loss. Represents cross-scale consistency loss. This step constructs a spatiotemporal graph structure through a graph convolutional network and, combined with cross-scale consistency constraints, enhances the consistent representation of remote sensing images across different spatial scales and time points. This embodiment helps more accurately capture feature relationships at different scales and improves the model's adaptability to remote sensing data.

[0033] S303 introduces an adaptive spatiotemporal feature adjustment mechanism, combining spatiotemporal feature adaptive learning and classification optimization algorithms to dynamically adjust feature weights at different time and spatial scales, thereby optimizing classification results. Through the adaptive gradient descent optimization model, the spatiotemporal feature fusion method is continuously adjusted during the training process. The expression is: ; in, Represents the spatiotemporal feature adaptive adjustment loss function, Indicates the The spatiotemporal characteristics of samples, represents the spatiotemporal characteristics predicted by the model, Represents the weight adjustment coefficient, which controls the balance between spatiotemporal feature adjustment and classification error. Represents the standard classification loss function, calculating the difference between the predicted and the actual target domain sample labels The error between .

[0034] In step S4 of the above embodiment, the steps are specifically included: S401 identifies low-confidence target domain samples based on the model's prediction confidence, i.e., remote sensing image samples with high prediction uncertainty. A quality assessment function for target domain sample features is introduced to further improve recognition accuracy. This process not only focuses on low-confidence samples but also enhances the screening of abnormal samples through quality assessment. Ultimately, a set of low-confidence samples is identified by minimizing the objective function, expressed as: ; in, represents the minimization objective function, represents the overall target domain sample characteristics, represents the confidence of the model prediction, represents the quality assessment function, which measures the consistency or abnormality of image features. represents the regularization coefficient, which is used to balance the weight between confidence and feature consistency. Represents a set of low-confidence target domain samples.

[0035] S402: After identifying low-confidence regions, the automatic target domain sample label correction algorithm generates correction suggestions and further optimizes the target domain sample labels based on manual feedback. By introducing manually annotated feedback information, the correction function intelligently adjusts the target domain sample labels based on the model's prediction results and manual correction information. This step continuously improves the accuracy and consistency of the target domain sample labels through an automatic and manual closed-loop feedback system, expressed as: ; in, represents the automatic target domain sample labeling algorithm, which optimizes the target domain sample labels by generating correction suggestions. It represents the correction function based on feedback information, which combines the manually annotated feedback and the model's prediction information to generate correction suggestions for the target domain sample labels. represents the target domain sample label set of manual feedback, and represents the correction feedback of manual annotators for low-confidence samples. represents the target domain sample label set after automatic and manual correction, represents the target domain sample label after correction, removing the noise effect, Represents a set of low-confidence target domain samples.

[0036] S403: After the target domain sample labels are corrected, this step synchronizes the corrected target domain sample labels with the current target domain sample labels through an online update mechanism. The difference metric function evaluates the difference between the corrected target domain sample labels and the model target domain sample labels, and adjusts the target domain sample label weights based on the difference. The online correction mechanism ensures that the model can update its weights in real time after the target domain sample labels are corrected, thereby enhancing the adaptability of the target domain sample label noise and further optimizing the classification performance. The expression is: ; in, represents the final corrected target domain label set, represents the final label set obtained after correction and optimization, Represents the online update function, which indicates the process of online correction and update of the model. represents the label difference measurement function, which calculates the difference between the corrected label and the current target domain sample label. represents the target domain sample label predicted by the training dataset, Represents the set of target domain sample labels after automatic and manual correction.

[0037] In the above embodiment, step S5 specifically includes the following steps: S501, through a multi-level nested structure, align the source domain sample features and the target domain sample features layer by layer. Each layer maps the source domain features to the target domain features based on the target domain sample label set mapping, ultimately achieving dynamic fusion of the source domain features and the target domain features, expressed as: ; in, Represents the features after the target domain sample features are fused with the source domain sample features, Represents the fusion weight of each level, with a total of layer, Represents the overall source domain sample characteristics The mathematical expectation value between the target domain sample label set, Represents the overall target domain sample characteristics.

[0038] S502 is the feature after the target domain sample feature and the source domain sample feature are fused. Calculate dynamic feature weights to reduce the impact of noisy labels. By minimizing label filtering loss, it is possible to dynamically adjust The contribution to training, especially the impact on low-confidence target domain sample label samples. In addition, the regularization term is used to ensure the smooth update of the weights to avoid overfitting the model to some target domain sample label noise. The expression is: ; in, represents the label filtering loss function, represents the dynamic feature weight, Represents the dynamic feature weights learned by the model in the previous iteration before the current iterative training step, which is used to maintain stability. represents the vector two-norm operation, represents the indicator function, when When the value is 1, represents the regularization coefficient of weight change, represents the cross entropy loss, which evaluates the accuracy of label prediction, represents the true target domain sample label, represents the predicted target domain sample label, represents minimizing the dynamic feature weight function.

[0039] S503: Adaptive noise filtering further optimizes the quality of the target domain sample labels by dynamically adjusting the noise filtering strategy. Guided by the source domain sample features, the noise labels in the target domain are identified and filtered out to ensure that the correct labels are not lost due to noise filtering. The adaptive evaluation function of the source domain sample features is used to optimize the filtering strategy and enhance the robustness of the target domain sample labels. The expression is: ; in, represents the minimization noise filtering strategy function, represents the noise filtering strategy, represents the source domain sample feature adaptation function, represents the overall source domain sample characteristics, represents the overall target domain sample characteristics, Represents the difference value of the predicted target domain sample label, represents the quality assessment function, which measures the consistency or abnormality of image features. represents the label filtering loss function, represents the sample characteristics of the target domain, It represents the objective function of the entire adaptive noise filtering optimization, which aims to minimize the interference of noise labels on the training process by dynamically adjusting the filtering strategy, and ultimately enhance the quality of target domain sample labels.

[0040] In order to prove the effectiveness of the above embodiment, a public dataset is used for verification. Specifically, the DeepGlobe-Land-Cover dataset is a high-resolution remote sensing image dataset focusing on land cover classification tasks. It was originally released as part of the DeepGlobe Challenge, aiming to promote research and application in the field of remote sensing. The images of this dataset come from DigitalGlobe satellites with a resolution of up to 0.5 meters per pixel, covering a variety of surface features around the world, including seven major land cover types such as cities, farmlands, forests, vegetation, water bodies, wastelands and ice and snow. The image target domain sample labels are manually annotated by professionals. Although there are certain labeling errors, its diverse geographical scenes and high-resolution characteristics provide a rich data resource for studying land cover classification in complex environments. The DeepGlobe-Land-Cover dataset can not only be used to evaluate the robustness of classification models in complex scenarios, but is also suitable for verifying the effectiveness of technologies such as target domain sample label optimization and noise filtering. It is an important benchmark dataset in the field of remote sensing classification.

[0041] Table 1: Experimental results for performance comparison of different methods The experimental results comprehensively demonstrate the superior performance of this method in remote sensing data classification tasks and provide detailed comparisons with several common methods. In terms of overall classification accuracy, traditional transfer learning methods achieved only 68.5%. Random forests and support vector machines (SVMs) achieved slight improvements, reaching 70.3% and 72.1%, respectively. However, these methods are limited by their inability to cope with complex domain shifts and noisy target domain labels. A deep learning baseline model (ResNet) improves overall classification accuracy to 74.6% by incorporating more powerful feature extraction capabilities. However, due to a lack of optimization for target domain label noise, its target domain label noise error reduction rate is only 10.2%, remaining suboptimal in high-noise scenarios. An improved method based on transfer learning combined with target domain label noise filtering significantly improves classification performance, achieving an overall classification accuracy of 78.4% and a target domain label noise error reduction rate of 18.5%. Further incorporating multimodal information increases overall classification accuracy and target domain label noise handling capabilities to 82.5% and 22.4%, respectively, demonstrating the potential advantages of multimodal fusion.

[0042] Compared with these common methods, the method of the present invention achieves comprehensive optimization of domain shift, target domain sample label noise and multimodal data fusion by integrating five innovative technical means, and has achieved remarkable experimental results. Specifically, the overall classification accuracy of this method reaches 93.1%, the average F1 score is 90.2%, the Kappa coefficient is as high as 0.89, and the target domain sample label noise error reduction rate is increased to 60.2%, which is the best performance among all the compared methods. This excellent performance is due to the dynamic adaptive migration filtering module proposed in this method, which effectively alleviates the domain shift problem through real-time feature alignment and dynamic weight adjustment; the target domain sample label credibility optimization achieves accurate filtering and optimization of noisy target domain sample labels through credibility distribution evaluation and dynamic correction; the multi-scale spatiotemporal consistency learning framework enhances the model's understanding of complex land features through the deep fusion of spectral, spatial and temporal features; the intelligent interactive target domain sample label correction combines manual feedback and automatic optimization to effectively improve the target domain sample label quality of the training data; the migration-aware nested noise filtering network significantly improves the robustness of the model in high-noise scenarios through multi-level nested optimization.

[0043] Furthermore, the overall performance improvement of this method is also reflected in its significantly enhanced ability to handle target domain sample label noise. The target domain sample label noise error reduction rate increased from 0% in traditional methods to 60.2%, far exceeding existing transfer learning methods. The improvement in the Kappa coefficient demonstrates the high reliability of this method in terms of the consistency and stability of classification results. The deep fusion of multimodal data further expands the applicable scenarios of this method, making it capable of classifying complex terrain, multi-source data, and dynamically changing scenes. In summary, this method not only excels in overall classification accuracy and robustness, but also successfully addresses the issues of domain shift and target domain sample label noise in remote sensing data classification, providing a new technical solution for high-precision remote sensing classification in complex scenarios. Its outstanding performance and wide applicability have opened up new research directions in the field of remote sensing data analysis and have important theoretical and applied value.

[0044] It should be noted that although the above embodiments have been described herein, this does not limit the scope of patent protection of the present invention. Therefore, based on the innovative concept of the present invention, changes and modifications to the embodiments described herein, or equivalent structural or equivalent process transformations made using the contents of the present invention's specification and drawings, and direct or indirect application of the above technical solutions to other related technical fields, are all included in the scope of patent protection of the present invention.

Claims

1. A remote sensing data classification method based on transfer learning and label noise filtering, characterized in that: The following steps are involved: S1, through the dynamic adaptive migration filtering module, introduces dynamic adjustment of feature weights, calculates the applicability of source domain sample features in real time according to the feature distribution of remote sensing images, and automatically adjusts their influence in the migration process; S2, after the initial alignment of the migration features, optimizes the target domain sample labels by driving credibility regression to address the target domain sample label noise in the remote sensing image data; S3, based on the optimization of target domain sample labels, introduces a multi-scale spatiotemporal consistency learning framework to build a multimodal feature fusion network, enhances the expressiveness of target category features, and dynamically adjusts the fusion weights of multimodal features to fully capture complex spatiotemporal features; S4, through interactive target domain sample label correction, combined with the prediction-driven feedback mechanism and manual annotation collaboration, further corrects the target domain sample label noise; S5, through the transfer-aware nested noise filtering network, combines the migration information of source domain sample features with the target domain sample label distribution, filters the noisy target domain sample labels layer by layer and optimizes the sample weight distribution.

2. The remote sensing data classification method based on transfer learning and label noise filtering according to claim 1 is characterized in that: Step S1 includes: S101 introduces a source domain sample feature adaptability calculation mechanism to adjust the adaptability of the source domain sample features in real time according to the feature distribution of the target domain samples, and calculates the adaptability weight of the source domain sample features. The expression is: ； Where exp represents the natural exponential function, and Represents the index, Indicates the number of sample features in the target domain, represents the hyperparameter that controls the range of similarity influence, Indicates the Source domain sample characteristics The adaptive weight of Indicates the target domain sample features; S102, based on the adaptive calculation of source domain sample features, dynamically adjust the learning rate and feature alignment strategy in transfer learning, and dynamically adjust the learning rate The expression is: ； in, represents the initial learning rate, Represents a hyperparameter that controls the adjustment rate of the learning rate. represents the number of training steps, Indicates the number of weights; S103 introduces negative transfer risk assessment, calculates the transfer effect of source domain sample features in the target domain, and dynamically adjusts the transfer strategy according to the difference in transfer loss. If negative transfer is found in the transfer of source domain sample features in the target domain, the feature weight or transfer path is automatically adjusted to reduce the negative transfer risk. The expression is: ； in, Indicates the negative migration risk, which is used to reflect the degree of negative migration during the migration process. Indicates the source domain The loss value of the feature, Represents the loss value of the source domain sample feature in the target domain.

3. The remote sensing data classification method based on transfer learning and label noise filtering according to claim 1 is characterized in that: Step S2 includes: S201: Construct a credibility distribution function for target domain sample labels, analyze the source domain sample feature consistency, prediction confidence, and historical prediction trajectory, assign a credibility score to each target domain sample, and identify and mark target domain sample labels with noise. The expression is: ； in, Indicates the Source domain sample characteristics Hedi Target domain sample features The consistency measurement function between Represents the historical prediction trajectory for the Source domain sample characteristics The impact of is used to reflect the stability of its prediction, Represents the prediction confidence, which is used to measure the classifier’s confidence in the sample characteristics of the target domain. , , Represents the weight coefficient, which is used to balance the influence of different factors. Indicates the The credibility score of the target domain sample assignment; S202: Introduce a credibility regression network, train it based on the credibility score of each target domain sample, and output the corrected target domain sample label distribution, which is expressed as: ； in, represents the original target domain sample label, Represents the weight of the correction factor, which is used to control the magnitude of the correction. Indicates the set credibility threshold, represents the corrected target domain sample label, The credibility regression network is composed of multi-layer perceptrons; S203, using an iterative optimization algorithm, repeatedly optimize the target domain sample label set based on the corrected target domain sample labels. The expression is: ； in, Indicates the The corrected target domain sample labels after iterations, Indicates the The credibility score of the target domain sample after iterations, express The credibility threshold after iterations, Represents the adjustment coefficient, which is used to control the speed of correction of target domain sample labels. Indicates the Corrected target domain sample labels after iteration.

4. The remote sensing data classification method based on transfer learning and label noise filtering according to claim 1 is characterized in that: Step S3 includes: S301, through the spatial-spectral mapping function, the modal features of different dimensions in the remote sensing image are mapped in a unified spatial framework to achieve modal information fusion, which is expressed as: ； in, Represents the overall source domain sample characteristics and the overall target domain sample characteristics The distribution alignment between them is used to measure the adaptability of the source domain sample features in the target domain. Indicates the The characteristic representation of the modality, Indicates the Adaptive weights of modalities, represents the total number of modes, Represents the final fusion feature output; S302 uses a graph convolutional network to transform the modal features in the remote sensing image into a graph structure and process the feature consistency constraints at different scales. The expression is: ； in, Indicates the Graph convolution of modality feature maps, represents the graph convolution generated by the target domain sample label, represents the vector two-norm operation, Represents the loss function of the graph convolutional network, which is used to enhance the consistency of features between different spatial scales. The total number of graph convolutional feature maps representing the modality, represents the regularization coefficient, which is used to balance the consistency constraint and the graph convolutional network loss. represents the cross-scale consistency loss; S303 introduces an adaptive spatiotemporal feature adjustment mechanism, combines spatiotemporal feature adaptive learning with classification optimization algorithms, dynamically adjusts feature weights at different scales, and optimizes classification results. The expression is: ； in, Represents the spatiotemporal feature adaptive adjustment loss function, Indicates the The spatiotemporal characteristics of samples, represents the spatiotemporal characteristics of the prediction, Represents the weight adjustment coefficient, which is used to control the balance between spatiotemporal feature adjustment and classification error. Represents the standard classification loss function, used to calculate the difference between the predicted and true target domain sample labels The error between .

5. The remote sensing data classification method based on transfer learning and label noise filtering according to claim 4 is characterized in that: The modal features include spectral features, spatial features and temporal features.

6. The remote sensing data classification method based on transfer learning and label noise filtering according to claim 1, characterized in that: Step S4 includes: S401: Identify low-confidence target domain samples based on prediction confidence, introduce a quality assessment function for target domain sample features, filter out abnormal samples through quality assessment, and identify a low-confidence sample set by minimizing the objective function. The expression is: ； in, represents the minimization objective function, represents the overall target domain sample characteristics, represents the confidence of the model prediction, represents the quality assessment function, which is used to measure the consistency or abnormality of image features. represents the regularization coefficient, which is used to balance the weight between confidence and feature consistency. Represents a set of low-confidence target domain samples; S402: After identifying the low-confidence area, the automatic target domain sample label correction algorithm generates correction suggestions and optimizes the target domain sample labels in combination with human feedback. The expression is: ； in, represents an automatic target domain sample label correction algorithm, which is used to optimize the target domain sample labels by generating correction suggestions. Represents a correction function based on feedback information, which is used to combine the manually annotated feedback and the model's prediction information to generate correction suggestions for the target domain sample labels. represents the target domain sample label set for manual feedback, Represents the target domain sample label set after automatic and manual correction; S403: After the target domain sample label is corrected, the corrected target domain sample label and the current target domain sample label are updated synchronously through the online update mechanism. The difference between the corrected target domain sample label and the model target domain sample label is evaluated through the difference measurement function, and the target domain sample label weight is adjusted according to the difference. The expression is: ； in, represents the final revised target domain label set, is the online update function, which indicates the process of online correction and update. Represents the label difference measurement function, which is used to calculate the difference between the corrected label and the current target domain sample label. Represents the target domain sample labels predicted by the training dataset.

7. The remote sensing data classification method based on transfer learning and label noise filtering according to claim 1 is characterized in that: Step S5 includes: S501, through a multi-level nested structure, align the source domain sample features and the target domain sample features layer by layer. Each layer maps the source domain features to the target domain features based on the target domain sample label set mapping, realizing the dynamic fusion of the source domain features and the target domain features. The expression is: ； in, Represents the features after the target domain sample features are fused with the source domain sample features, Represents the fusion weight of each level, with a total of layer, Represents the overall source domain sample characteristics The mathematical expectation value between the target domain sample label set, Represents the overall target domain sample characteristics; S502 is the feature after the target domain sample feature and the source domain sample feature are fused. Calculate dynamic feature weights and dynamically adjust by minimizing the label filtering loss function Contribution to training, using regularization terms to ensure stable updates of weights, the expression is: ； in, represents the label filtering loss function, represents the dynamic feature weight, represents the dynamic feature weights learned in the previous iteration, used to maintain stability, represents the vector two-norm operation, represents the indicator function, when When the value is 1, represents the regularization coefficient of weight change, represents the cross entropy loss, which evaluates the accuracy of label prediction, represents the true target domain sample label, represents the predicted target domain sample label, represents the minimization of dynamic feature weight function; S503: Identify and filter out noise labels in the target domain using the source domain sample features as a guide, and optimize the filtering strategy using the adaptability evaluation function of the source domain sample features. The expression is: ； in, represents the minimization noise filtering strategy function, represents the noise filtering strategy, represents the source domain sample feature adaptation function, Represents the difference value of the predicted target domain sample label, represents the quality assessment function, which is used to measure the consistency or abnormality of image features. represents the label filtering loss function, Represents the adaptability evaluation function of the source domain sample features.

Images