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

Abstract

The application discloses a kind of based on transfer learning and label noise filtering remote sensing data classification method, including steps: by dynamic self-adapting migration filtering module, introduce feature weight dynamic adjustment, the applicability of source domain sample feature is calculated in real time, and adjust its influence in migration process;Through the credibility regression driven target domain sample label is optimized;Introduce multiscale spatiotemporal consistency learning framework, construct multimodal feature fusion network, dynamically adjust the fusion weight of multimodal feature;Through interactive target domain sample label correction, in combination with prediction driven feedback mechanism and artificial annotation cooperation, target domain sample label noise is corrected;Through the nested noise filtering network of migration perception, in combination with the migration information of source domain sample feature and target domain sample label distribution, noise target domain sample label is filtered layer by layer and sample weight distribution is optimized;The scheme has the characteristics of realizing the efficient, accurate classification of large-scale remote sensing data.

Description

Technical Field

[0001] This 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 Technology

[0002] With the rapid development of remote sensing technology, various types of remote sensing images, such as optical images, SAR images, and multispectral images, are widely used in fields such as land resource management, environmental protection, agricultural monitoring, and disaster early warning. Remote sensing image classification, as one of the core tasks of remote sensing data analysis, aims to divide pixels or objects into specific land cover 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 (SVM), Random Forests (RF), and K-Nearest Neighbors (KNN). These methods typically require manual extraction of spectral, texture, or shape features and depend on rich labeled data for supervised training. However, with the continuous improvement of remote sensing image resolution and complexity, the difficulty of remote sensing data classification tasks has also increased. On the one hand, high-resolution remote sensing images contain more refined ground feature information, but the differences between ground feature categories become more subtle; on the other hand, with the increase in image dimensionality, such as multispectral or even hyperspectral data, the curse of dimensionality becomes more prominent, and the model's requirement for a large sample size also increases significantly. In addition, labeling high-quality remote sensing data is costly, especially in fine-grained scenarios requiring high-precision classification. Manual labeling often requires the participation of domain experts, which is not only time-consuming and labor-intensive but also easily affected by the subjective factors of the labelers, resulting in a certain degree of noise in the target domain sample label data. Target domain sample label noise may originate from spectral aliasing, blurred boundaries of ground feature categories, interference from lighting or shadow effects in the image, and human errors during the labeling process. As the proportion of noise in the target domain sample labels increases, the features learned by the model may not match the actual classification requirements, thus significantly reducing model performance. This risk is even more serious in deep learning models, where overfitting to noisy target domain sample labels is more severe.

[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 on large-scale general datasets (such as ImageNet) to the target remote sensing task, transfer learning methods can improve classification performance under conditions of scarce labeled data. However, the application of transfer learning in the field of remote sensing still faces the challenge of cross-domain transfer. The data distribution between the source domain (such as natural images) and the target domain (such as remote sensing imagery) often differs significantly. For example, the spectral distribution of remote sensing data is very different from that of ordinary RGB images, which may lead to negative transfer effects introduced by traditional transfer learning methods, affecting the classification results of the target task. In addition, the following problems urgently need to be solved in remote sensing data classification:

[0005] (1) Scarcity of labeled data: High-quality labeled data is expensive to obtain, and due to the complexity of remote sensing data, such as multispectral dimensions and fine-grained classification requirements, the coverage of labeled samples is often insufficient, which can easily lead to insufficient diversity of model training data.

[0006] (2) Noise interference in target domain sample labels: Due to spectral aliasing of remote sensing data, similarity between land cover 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.

[0007] (3) Insufficient robustness of the model: Under the interference of noise in the target domain sample labels, deep learning models are prone to overfitting noisy data, thereby reducing the model's generalization ability and classification performance.

[0008] (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 environment.

[0009] Therefore, this invention aims to study a remote sensing data classification method that combines transfer learning with target domain sample label noise filtering. By fully utilizing the knowledge transfer advantages of transfer learning and the robustness enhancement capability of target domain sample label noise filtering, this invention solves the bottleneck problems of scarce labeled data and interference of target domain sample label noise on the model in current remote sensing classification, thereby achieving efficient and accurate classification of large-scale remote sensing data. Summary of the Invention

[0010] 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:

[0011] S1 introduces a dynamic adaptive migration filtering module to dynamically adjust feature weights, calculate the applicability of source domain sample features in real time based on the feature distribution of remote sensing images, and automatically adjust their influence in the migration process.

[0012] S2, after the initial alignment of the migration features, the target domain sample labels are optimized through confidence regression to address the target domain sample label noise in the remote sensing image data;

[0013] S3, based on the optimization of target domain sample labels, introduces a multi-scale spatiotemporal consistency learning framework to construct a multimodal feature fusion network, enhances the target category feature expression capability, and dynamically adjusts the fusion weights of multimodal features to achieve comprehensive capture of complex spatiotemporal features;

[0014] S4 further corrects the noise of the target domain sample labels through interactive target domain sample label correction, combined with prediction-driven feedback mechanism and manual annotation collaboration;

[0015] S5 uses a nested noise filtering network with transfer awareness to combine the transfer information of source domain sample features with the target domain sample label distribution, filtering noisy target domain sample labels layer by layer and optimizing the sample weight distribution.

[0016] In a preferred embodiment of the present invention, step S1 includes:

[0017] S101 introduces a source domain sample feature adaptive calculation mechanism, which adjusts the applicability of source domain sample features in real time based on the feature distribution of target domain samples, and calculates the adaptive weights of source domain sample features, expressed as:

[0018] ;

[0019] Where exp represents the natural exponential function. and These represent indexes, Indicates the number of features in the target domain samples. The hyperparameter representing the range of influence of similarity. Indicates the first Features of source domain samples The adaptive weights, Indicates the first Features of each target domain sample;

[0020] S102, based on the adaptive calculation of source domain sample features, dynamically adjusts the learning rate and feature alignment strategy in transfer learning, with the learning rate being dynamically adjusted. The expression is:

[0021] ;

[0022] in, This represents the initial learning rate. This represents a hyperparameter used to control the rate at which the learning rate is adjusted. Indicates the number of training steps. Indicates the number of weights;

[0023] S103 introduces a negative migration risk assessment, calculating the migration effect of source domain sample features in the target domain and dynamically adjusting the migration strategy based on the difference in migration loss. If negative migration of source domain sample features in the target domain is detected, the feature weights or migration path are automatically adjusted to mitigate the negative migration risk. The expression is:

[0024] ;

[0025] in, This indicates the risk of negative migration and reflects the degree to which negative migration occurs during the migration process. Indicates the first in the source domain The loss value of each feature, This represents the loss value of the source domain sample features in the target domain.

[0026] In a preferred embodiment of the present invention, step S2 includes:

[0027] S201, Construct a confidence distribution function for target domain sample labels, analyze the consistency of source domain sample features, prediction confidence, and historical prediction trajectories, assign a confidence score to each target domain sample, and identify and label noisy target domain sample labels. The expression is:

[0028] ;

[0029] in, Indicates the first Features of source domain samples and the Features of target domain samples Consistency measure function between them Indicates the historical predicted trajectory for the first Features of source domain samples The impact of this is used to reflect the stability of its predictions. This represents the prediction confidence level, used to measure the degree of confidence of the classifier in the features of samples in the target domain. , , This represents the weighting coefficient, used to balance the influence of different factors. Indicates the first Credibility score of sample assignment to each target domain;

[0030] S202 introduces a credibility regression network, which is trained by combining the credibility score of each target domain sample, and outputs the corrected target domain sample label distribution, expressed as:

[0031] ;

[0032] in, Indicates the original target domain sample label. This represents the weight of the correction factor, used to control the magnitude of the correction. This indicates the set credibility threshold. This represents the corrected target domain sample label. The credibility regression network is composed of multilayer perceptrons;

[0033] S203 employs an iterative optimization algorithm, repeatedly optimizing the target domain sample label set based on the corrected target domain sample labels. The expression is:

[0034] ;

[0035] in, Indicates the first The target domain sample labels are corrected after the next iteration. Indicates the first The credibility score of the target domain samples after the next iteration. express The credibility threshold after the next iteration This represents the adjustment coefficient, used to control the speed of target domain sample label correction. Indicates the first The target domain sample labels are corrected after iteration.

[0036] In a preferred embodiment of the present invention, step S3 includes:

[0037] S301, through a space-spectral mapping function, maps modal features of different dimensions in remote sensing images into a unified spatial framework, thereby achieving modal information fusion. The expression is:

[0038] ;

[0039] in, Represents the source domain sample characteristics of the population Features of the target domain samples in the overall population The distribution alignment between them is used to measure the adaptability of source domain sample features to the target domain. Indicates the first Modal feature representation, Indicates the first Adaptive weights for modality This represents the total number of modes. This represents the final fusion feature output;

[0040] S302 utilizes a graph convolutional network to transform modal features in remote sensing images into a graph structure, handling feature consistency constraints at different scales. The expression is:

[0041] ;

[0042] in, Indicates the first Graph convolution of modal feature maps, This represents a graph convolution generated using the target domain sample labels. This represents the vector 2-norm operation. This represents the loss function of a graph convolutional network, used to enhance feature consistency across different spatial scales. The total number of graph convolutional feature maps representing modalities. This represents the regularization coefficient, used to balance consistency constraints and the loss of graph convolutional networks. This represents the cross-scale consistency loss;

[0043] S303 introduces an adaptive spatiotemporal feature adjustment mechanism, combining adaptive learning of spatiotemporal features and a classification optimization algorithm to dynamically adjust feature weights at different scales and optimize classification results. The expression is:

[0044] ;

[0045] in, This represents the loss function that adaptively adjusts to spatiotemporal features. Indicates the first The spatiotemporal characteristics of each sample Indicates the spatiotemporal characteristics of the prediction. This represents the weighting adjustment coefficient, used to control the balance between spatiotemporal feature adjustment and classification error. This represents the standard classification loss function, used to calculate the difference between the predicted and true target domain sample labels. The error between them.

[0046] In a preferred embodiment of the present invention, the modal features include spectral features, spatial features, and temporal features.

[0047] In a preferred embodiment of the present invention, step S4 includes:

[0048] S401, based on the predicted confidence level, low-confidence target domain samples are identified. A quality assessment function for the target domain sample features is introduced. Abnormal samples are screened through quality assessment. The set of low-confidence samples is identified by minimizing the objective function. The expression is:

[0049] ;

[0050] in, This represents minimizing the objective function. This represents the characteristics of the target domain samples in the population. This indicates the confidence level of the model prediction. This represents a quality assessment function used to measure the consistency or anomaly of image features. This represents the regularization coefficient, used to balance the weight between confidence and feature consistency. This represents a set of samples from the target domain with low confidence.

[0051] S402, after identifying low-confidence regions, the automatic target domain sample label correction algorithm generates correction suggestions and optimizes the target domain sample labels by combining human feedback. The expression is:

[0052] ;

[0053] in, This refers to an automatic target domain sample label correction algorithm, used to optimize target domain sample labels by generating correction suggestions. This represents a correction function based on feedback information, used to combine manually labeled feedback and model predictions to generate suggested corrections for target domain sample labels. This represents the set of target domain sample labels provided by human feedback. This represents the set of target domain sample labels after automatic and manual correction;

[0054] S403, after the target domain sample labels are corrected, the corrected target domain sample labels are updated synchronously with the current target domain sample labels through an online update mechanism. The difference between the corrected target domain sample labels and the model target domain sample labels is evaluated using a difference metric function, and the target domain sample label weights are adjusted based on the difference. The expression is:

[0055] ;

[0056] in, This represents the final corrected set of target domain labels. This is an online update function, representing the process of online correction and update. This represents the label difference metric function, used to calculate the difference between the corrected label and the label of the current target domain sample. This represents the target domain sample label predicted from the training dataset.

[0057] In a preferred embodiment of the present invention, step S5 includes:

[0058] S501, through a multi-level nested structure, aligns the source domain sample features and 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, achieving dynamic fusion of source domain features and target domain features. The expression is:

[0059] ;

[0060] in, This represents the feature obtained by fusing the features of the target domain samples and the features of the source domain samples. This represents the fusion weights for each layer, totaling [number missing]. layer, Represents the source domain sample characteristics of the population The expected value between the target domain sample label set and the target domain sample label set. This represents the characteristics of the target domain samples in the overall population.

[0061] S502 represents the feature obtained by fusing the features of the target domain samples and the features of the source domain samples. Calculate dynamic feature weights and dynamically adjust them by minimizing the label filtering loss function. The contribution to training is ensured by using a regularization term to guarantee stable weight updates; the expression is:

[0062] ;

[0063] in, This represents the label filtering loss function. Represents dynamic feature weights. This represents the dynamic feature weights learned in the previous iteration, used to maintain stability. This represents the vector 2-norm operation. Indicates an indicator function, when The value is 1 at time. The regularization coefficient represents the change in weights. Cross-entropy loss is used to evaluate the accuracy of label prediction. This represents the labels of the real target domain samples. This indicates the predicted target domain sample label. This represents minimizing the dynamic feature weight function;

[0064] S503, guided by the features of the source domain samples, identifies and filters out noise labels in the target domain, and optimizes the filtering strategy through the adaptive evaluation function of the source domain sample features, the expression of which is:

[0065] ;

[0066] in, This represents the function that minimizes the noise filtering strategy. Indicates the noise filtering strategy. Represents the source domain sample feature adaptation function. This represents the difference value of the predicted target domain sample labels. This represents a quality assessment function used to measure the consistency or anomaly of image features. This represents the label filtering loss function. An adaptive evaluation function representing the characteristics of source domain samples.

[0067] Unlike existing technologies, the above technical solution achieves the following beneficial effects:

[0068] (1) This method dynamically adjusts the influence weight of source domain sample features on target domain during the transfer learning process to solve the negative transfer problem caused by the difference in distribution between source domain and target domain. This enhances the feature transfer effect while avoiding performance degradation of the model due to distribution differences. In other words, it effectively solves the problem of classification performance degradation caused by scarce labeled data and noise interference of target domain sample labels in remote sensing data classification tasks.

[0069] (2) This method constructs a confidence distribution function for the noise label of the target domain sample and scores the confidence of the target domain sample label for each sample in the dataset, thereby achieving accurate identification of the noise target domain sample label and dynamic adjustment of the classification weight, effectively improving the robustness of this method;

[0070] (3) This method comprehensively utilizes the feature information of spectral, spatial and temporal dimensions in remote sensing data classification tasks, 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.

[0071] (4) This method effectively corrects the labels of noisy target domain samples in remote sensing image datasets through feedback updates and iterative optimization, ensuring that the classification model can learn based on high-quality data and improve classification accuracy.

[0072] (5) This method dynamically optimizes the filtering strategy of the sample labels in the noisy target domain, enabling the model to cope more efficiently with large-scale remote sensing data classification tasks with high noise ratios, and significantly improving the scalability and applicability of the task. Attached Figure Description

[0073] Figure 1 The flowchart is for a specific implementation method. Detailed Implementation

[0074] To explain in detail the technical content, structural features, objectives, and effects of the technical solution, the following description is provided in conjunction with specific embodiments and accompanying drawings.

[0075] like Figure 1As shown, this embodiment provides a remote sensing data classification method based on transfer learning and label noise filtering, including the following steps:

[0076] S1, through a dynamic adaptive transfer filtering module, introduces dynamic adjustment of feature weights. It calculates the applicability of source domain sample features in real time based on the feature distribution of the remote sensing image and automatically adjusts their influence during the transfer process. Specifically, the dynamic adaptive transfer filtering module addresses the distribution offset problem between the source domain (natural image) and the target domain (remote sensing image) through a real-time feature adaptation mechanism. The module introduces dynamic feature weight adjustment technology to calculate the applicability of source domain sample features in real time based on the feature distribution of the remote sensing image 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, thereby effectively reducing the risk of negative transfer and laying the foundation for stable training of the subsequent model.

[0077] S2, after initial alignment of transfer features, addresses the target domain sample label noise present in remote sensing image data by optimizing the target domain sample labels through credibility regression. Specifically, after initial alignment of transfer features, a credibility regression-driven target domain sample label optimization is proposed to address the common problem of target domain sample label noise in remote sensing data. By constructing a target domain sample label credibility distribution function, multidimensional analysis is performed on the feature consistency, historical prediction trajectory, and confidence of the samples to generate a credibility score. Combined with a credibility regression network, 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 robustness of the classification model to target domain sample label noise is significantly enhanced, providing a reliable foundation for high-quality classification tasks.

[0078] S3, building upon the optimization of target domain sample labels, introduces a multi-scale spatiotemporal consistency learning framework to construct a multi-modal feature fusion network. This enhances the expressive power of target category features and dynamically adjusts the fusion weights of multi-modal features to achieve comprehensive capture of complex spatiotemporal features. Specifically, to further improve the adaptability of this method to complex remote sensing data, a multi-scale spatiotemporal consistency learning framework is introduced, building upon the optimization of target domain sample labels. This framework addresses the fusion challenges arising from the multi-source heterogeneity of remote sensing data by constructing a multi-modal feature fusion network across spectral, spatial, and temporal dimensions. By introducing a cross-scale consistency constraint method, the multi-scale spatiotemporal consistency learning framework strengthens the consistent expressive power of target category features across different time points and spatial scales, while dynamically adjusting the fusion weights of multi-modal features to achieve comprehensive capture of complex spatiotemporal features by the classification model.

[0079] S4, through interactive target domain sample label correction, combined with a prediction-driven feedback mechanism and collaborative manual annotation, further corrects the noise in target domain sample labels. Specifically, for residual, difficult-to-distinguish noisy target domain sample labels and data anomalies in specific scenarios, this method designs an interactive target domain sample label correction mechanism. This mechanism, combined with a prediction-driven feedback mechanism and collaborative manual annotation, uses a closed-loop optimization system to further correct the noise in target domain sample labels. The interactive target domain sample label correction mechanism captures low-confidence regions predicted by the model in real time, generates high-priority correction suggestions, and gradually optimizes the quality of target domain sample labels using a combination of automatic algorithms and manual feedback. Through online correction and synchronous model updates, the interactive target domain sample label correction mechanism significantly improves the effectiveness of the dataset, providing more accurate training data for complex remote sensing classification tasks.

[0080] S5 proposes a transfer-aware nested noise filtering network that combines the transfer information of source domain sample features with the target domain sample label distribution to filter noisy target domain sample labels layer by layer and optimize the sample weight distribution. Specifically, to address the classification bottleneck in high-noise scenarios, this method proposes a transfer-aware nested noise filtering network. This network, through a multi-level nested structure, combines the transfer information of source domain sample features with the target domain sample label distribution to filter noisy target domain sample labels layer by layer and optimize 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.

[0081] Step S1 in the above embodiment specifically includes the following steps:

[0082] S101 introduces a source domain sample feature adaptive calculation mechanism to adjust the applicability of source domain samples (i.e., natural image features) in real time based on the feature distribution of target domain samples (i.e., remote sensing image data), and calculates the adaptive weight for each feature. The core of this process is assigning source domain sample feature adaptive weights to each source domain sample feature; these weights dynamically adjust the influence of source domain sample features on target domain sample features during the migration process. The overall source domain sample features are... The overall target domain sample characteristics are The distribution of the target domain sample features is as follows The formula for calculating the adaptive weights of source domain sample features is:

[0083] ;

[0084] Where exp represents the natural exponential function. and These represent indexes, Indicates the number of features in the target domain samples. The hyperparameter representing the range of influence of similarity. Indicates the first Features of source domain samples The adaptive weights represent the influence of source domain sample features on target domain sample features during the transfer process. Indicates the first Features of each target domain sample.

[0085] S102, based on the adaptive calculation of source domain sample features, dynamically adjusts the learning rate and feature alignment strategy in transfer learning to further improve the transfer effect and ensure smooth convergence during the transfer process. This step introduces a domain-aware optimizer, which, combined with the adaptive weights of source domain sample features, optimizes the learning rate in real time during the transfer process and dynamically adjusts the feature alignment strategy according to the alignment degree between source and target domain sample features. Based on the adaptive weights of source domain sample features from the previous steps, this step ensures more accurate feature alignment between the target and source domains during the transfer process, thereby reducing the risk of negative transfer, and dynamically adjusts the learning rate. The expression is:

[0086] ;

[0087] in, This represents the initial learning rate. This represents a hyperparameter that controls the rate at which the learning rate is adjusted. Indicates the number of training steps. Indicates the first Adaptive weights of source domain sample features Indicates the number of weights.

[0088] S103, by assessing the negative transfer risk during the transfer process, avoids negative transfer effects of source domain sample features in the target domain. This step further optimizes the source domain sample features based on the dynamic adjustment results of feature adaptation weights and learning rates from the previous two steps, eliminating potential negative transfer risks. By introducing a negative transfer risk assessment mechanism, the transfer effect of source domain sample features in the target domain is calculated, and the transfer strategy is dynamically adjusted based on the difference in transfer loss. If the transfer effect of source domain sample features in the target domain is found to be unsatisfactory, i.e., negative transfer exists, the feature weights or transfer path are automatically adjusted to avoid negative transfer effects and risks. The expression is:

[0089] ;

[0090] in, This indicates the risk of negative migration, reflecting the degree to which negative migration occurs during the migration process. Indicates the first in the source domain The loss value of each feature, This represents the loss value of the source domain sample features in the target domain.

[0091] Step S2 in the above embodiment specifically includes the following steps:

[0092] S201, Construct a confidence distribution function for target domain sample labels. This function comprehensively considers the consistency of source domain sample features, prediction confidence, and historical prediction trajectories. Through multidimensional analysis of these factors, this embodiment can assign a confidence score to each target domain sample, thereby identifying and labeling target domain sample labels that may contain noise. The expression is:

[0093] ;

[0094] in, Indicates the first Features of source domain samples and the Features of target domain samples Consistency measure function between them Indicates the historical predicted trajectory for the first Features of source domain samples The impact reflects the stability of its predictions. This represents the prediction confidence level, measuring the degree of confidence the classifier has in the features of samples in the target domain. , , This represents the weighting coefficient, used to balance the influence of different factors. Indicates the first The credibility score of the target domain sample assignment.

[0095] S202 introduces a credibility regression network, which dynamically corrects the labels of low-credibility target domain samples by modeling credibility scores. The credibility regression network is trained using the credibility scores of each target domain sample, outputting a corrected target domain sample label distribution, effectively eliminating the interference of target domain sample label noise on the model. The expression is:

[0096] ;

[0097] in, Indicates the original target domain sample label. This indicates the weight of the correction factor, controlling the magnitude of the correction. This indicates the set credibility threshold; target domain sample labels below this value will be corrected. This represents the corrected target domain sample label. The credibility regression network is composed of a multilayer perceptron.

[0098] S203 employs an iterative optimization algorithm, repeatedly optimizing the target domain sample label set based on the corrected target domain sample labels to gradually improve the overall quality of the dataset. After each iteration, the reliability regression network updates the target domain sample label distribution based on the new target domain sample label correction results, further correcting the potentially noisy target domain sample labels. The expression is:

[0099] ;

[0100] in, Indicates the first The target domain sample labels are corrected after the next iteration. Indicates the first The credibility score of the target domain samples after the next iteration. express The credibility threshold after the next iteration This represents the adjustment coefficient, used to control the speed of target domain sample label correction. Indicates the first The target domain sample labels are corrected after iteration.

[0101] Step S3 in the above embodiment specifically includes the following steps:

[0102] S301 fuses features from different modalities in remote sensing images, such as spectral, spatial, and temporal dimensions. A spatial-spectral mapping function maps features from different dimensions within a unified spatial framework, thereby reducing intermodal differences and achieving efficient fusion of multimodal information in remote sensing images. This addresses the inconsistency problem between different modal features and handles the challenges posed by multi-source heterogeneous data. The expression is:

[0103] ;

[0104] in, Represents the source domain sample characteristics of the population Features of the target domain samples in the overall population The distribution alignment between them is used to measure the adaptability of source domain sample features to the target domain. Indicates the first Modal representations, such as spectral, spatial, and temporal features. Indicates the first The adaptive weights of a modality, learned through training, represent the influence of that modality. This represents the total number of modes. This represents the final fused feature output, combining weighted features from multiple modalities. This step optimizes the fusion effect of multi-source heterogeneous features by introducing an adaptive attention mechanism to dynamically weight features from different modalities in the remote sensing image. Multi-scale convolutional networks extract features at different scales, enabling the classification model to more comprehensively understand the spatial, spectral, and temporal information in the image.

[0105] S302 utilizes graph convolutional networks to model the spatiotemporal relationships of features across scales, transforming spatial, spectral, and temporal features in remote sensing images into graph structures and handling feature consistency constraints at different scales. By applying cross-scale consistency constraints, the model's ability to represent target categories at different scales is enhanced, avoiding information loss. The expression is:

[0106] ;

[0107] in, Indicates the first Graph convolution of modal feature maps, This represents a graph convolution generated using the target domain sample labels. This represents the loss function of a graph convolutional network, designed to enhance feature consistency across different spatial scales. The total number of graph convolutional feature maps representing modalities. This represents the regularization coefficient, used to balance consistency constraints and the loss of graph convolutional networks. This represents the cross-scale consistency loss. This step constructs a spatiotemporal graph structure through a graph convolutional network, combined with cross-scale consistency constraints, enhancing the consistency representation ability of remote sensing images across different spatial scales and time points. This embodiment helps to more accurately capture feature relationships at different scales and improves the model's adaptability to remote sensing data.

[0108] S303 introduces an adaptive spatiotemporal feature adjustment mechanism, combining adaptive learning of spatiotemporal features with a classification optimization algorithm to dynamically adjust feature weights at different time and spatial scales, thereby optimizing classification results. Through an adaptive gradient descent optimization model, the fusion method of spatiotemporal features is continuously adjusted during training, expressed as:

[0109] ;

[0110] in, This represents the loss function that adaptively adjusts to spatiotemporal features. Indicates the first The spatiotemporal characteristics of each sample This represents the spatiotemporal characteristics predicted by the model. This represents the weighting adjustment coefficient, which controls the balance between spatiotemporal feature adjustment and classification error. This represents the standard classification loss function, used to calculate the predicted and true target domain sample labels. The error between them.

[0111] Step S4 in the above embodiment specifically includes the following steps:

[0112] S401, based on the model's prediction confidence, identifies low-confidence target domain samples, i.e., remote sensing image samples with high prediction uncertainty, and introduces a quality assessment function for the target domain sample features to further improve identification accuracy. This process not only focuses on samples with low confidence but also strengthens the screening of abnormal samples through quality assessment. Finally, it identifies the set of low-confidence samples by minimizing the objective function, expressed as:

[0113] ;

[0114] in, This represents minimizing the objective function. This represents the characteristics of the target domain samples in the population. This indicates the confidence level of the model prediction. This represents a quality assessment function that measures the consistency or anomaly of image features. This represents the regularization coefficient, used to balance the weight between confidence and feature consistency. This represents the set of samples from the target domain with low confidence.

[0115] 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 by incorporating human feedback. By introducing manually labeled feedback, the correction function intelligently adjusts the target domain sample labels based on the model's predictions and human correction information. This step continuously improves the accuracy and consistency of the target domain sample labels through a closed-loop feedback system of automatic and human intervention. The expression is:

[0116] ;

[0117] in, This refers to an automatic target domain sample labeling algorithm that optimizes target domain sample labels by generating correction suggestions. This represents a correction function based on feedback information. It combines manually labeled feedback and model predictions to generate suggested corrections for target domain sample labels. This represents the set of target domain sample labels provided by human annotation, indicating the correction feedback from human annotators for low-confidence samples. This represents the set of target domain sample labels after automatic and manual correction, while the second represents the corrected target domain sample labels after noise removal. This represents the set of samples from the target domain with low confidence.

[0118] S403, after the target domain sample labels are corrected, this step uses an online update mechanism to synchronously update the corrected target domain sample labels with the current target domain sample labels. A difference metric function evaluates the difference between the corrected target domain sample labels and the model's target domain sample labels, and adjusts the target domain sample label weights based on this 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 model's ability to adapt to target domain sample label noise and further optimizing classification performance. The expression is:

[0119] ;

[0120] in, This represents the final set of target domain labels after correction, and the final set of labels obtained after correction and optimization. This represents the online update function, which describes the process of making online corrections and updates to the model. This represents the label difference metric function, which calculates the difference between the corrected label and the label of the current target domain samples. This represents the target domain sample labels predicted from the training dataset. This represents the set of target domain sample labels after automatic and manual correction.

[0121] Step S5 in the above embodiment specifically includes the following steps:

[0122] S501 aligns source domain sample features and target domain sample features layer by layer through a multi-level nested structure. Each layer maps source domain features to target domain features based on the target domain sample label set, ultimately achieving dynamic fusion of source and target domain features. The expression is:

[0123] ;

[0124] in, This represents the feature obtained by fusing the features of the target domain samples and the features of the source domain samples. This represents the fusion weights for each layer, totaling [number missing]. layer, Represents the source domain sample characteristics of the population The expected value between the target domain sample label set and the target domain sample label set. This represents the characteristics of the target domain samples in the overall population.

[0125] S502 represents the feature obtained by fusing the features of the target domain samples and the features of the source domain samples. Calculate dynamic feature weights to reduce the impact of noisy labels. By minimizing the label filtering loss, the weights can be dynamically adjusted. The contribution to training, especially the impact on low-confidence target domain sample labels. Furthermore, a regularization term is used to ensure smooth weight updates and avoid overfitting the model to noise in certain target domain sample labels; the expression is:

[0126] ;

[0127] in, This represents the label filtering loss function. Represents dynamic feature weights. This represents the dynamic feature weights learned by the model in the previous iteration before the current training iteration, used to maintain stability. This represents the vector 2-norm operation. Indicates an indicator function, when The value is 1 at time. The regularization coefficient represents the change in weights. Cross-entropy loss is used to evaluate the accuracy of label prediction. This represents the labels of the real target domain samples. This indicates the predicted target domain sample label. This represents minimizing the dynamic feature weight function.

[0128] S503, Adaptive noise filtering, further optimizes the quality of predicted target domain sample labels by dynamically adjusting the noise filtering strategy. Guided by source domain sample features, it identifies and filters out noisy labels in the target domain, ensuring that 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 predicted target domain sample labels. The expression is:

[0129] ;

[0130] in, This represents the function that minimizes the noise filtering strategy. Indicates the noise filtering strategy. Represents the source domain sample feature adaptation function. Represents the characteristics of the source domain samples in the overall population. This represents the characteristics of the target domain samples in the population. This represents the difference value of the predicted target domain sample labels. This represents a quality assessment function that measures the consistency or anomaly of image features. This represents the label filtering loss function. Represents the features of the target domain samples. The objective function for the entire adaptive noise filtering optimization aims to minimize the interference of noise labels on the training process by dynamically adjusting the filtering strategy, thereby ultimately enhancing the quality of the target domain sample labels.

[0131] To demonstrate the effectiveness of the above embodiments, a public dataset is used for verification. Specifically, the DeepGlobe-Land-Cover dataset is a high-resolution remote sensing image dataset focused on land cover classification. Originally released as part of the DeepGlobe Challenge to promote research and applications in remote sensing, the dataset's images come from the DigitalGlobe satellite, with a resolution of up to 0.5 meters per pixel. It covers diverse global land surface features, including seven major land cover types: urban areas, farmland, forests, vegetation, water bodies, wasteland, and snow. The target domain sample labels are manually annotated by professionals. Although some annotation errors exist, its diverse geographical scenes and high resolution provide rich data resources 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 scenes but is also suitable for verifying the effectiveness of techniques such as target domain sample label optimization and noise filtering, making it an important benchmark dataset in the field of remote sensing classification.

[0132] Table 1: Experimental results comparing the performance of different methods

[0133]

[0134] The experimental results comprehensively demonstrate the superior performance of our proposed 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%, while random forests and support vector machines (SVMs) showed slight improvements, reaching 70.3% and 72.1% respectively. However, these methods are limited in their ability to handle complex neighborhood shifts and high-noise target domain sample labels. The deep learning baseline model (ResNet) improved overall classification accuracy to 74.6% by introducing more powerful feature extraction capabilities. However, due to the lack of optimization for target domain sample label noise, its target domain sample label noise error reduction rate was only 10.2%, indicating unsatisfactory performance in high-noise scenarios. The improved method based on transfer learning combined with target domain sample label noise filtering significantly improved classification performance, achieving an overall classification accuracy of 78.4% and a target domain sample label noise error reduction rate of 18.5%. Further fusion of multimodal information further improved the overall classification accuracy and target domain sample label noise processing capabilities to 82.5% and 22.4% respectively, demonstrating the potential advantages of multimodal fusion.

[0135] Compared to these common methods, the method of this invention integrates five innovative technologies to achieve comprehensive optimization of domain offset, target domain sample label noise, and multimodal data fusion, yielding significant 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 improved to 60.2%, which is the best among all compared methods. This superior performance is attributed to the proposed dynamic adaptive transfer filtering module, which effectively alleviates the domain offset 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 ground features through deep fusion of spectral, spatial, and temporal features; the intelligent interactive target domain sample label correction, combined with human feedback and automatic optimization, effectively improves the quality of target domain sample labels in the training data; and the transfer-aware nested noise filtering network significantly improves the robustness of the model in high-noise scenarios through multi-level nested optimization.

[0136] Furthermore, the overall performance improvement of this method is also reflected in its significantly enhanced ability to handle target domain sample label noise. The error reduction rate of target domain sample label noise increases from 0% in traditional methods to 60.2%, far exceeding existing transfer learning methods. The improved Kappa coefficient indicates that this method has extremely high reliability in terms of consistency and stability of classification results. The deep fusion of multimodal data further expands the applicable scenarios of this method, enabling it to meet the classification needs of complex terrain, multi-source data, and dynamically changing scenarios. In summary, this method not only performs excellently in overall classification accuracy and robustness, but also successfully solves the problems of neighborhood bias and target domain sample label noise in remote sensing data classification, providing a brand-new technical solution for high-precision remote sensing classification in complex scenarios. Its superior performance and wide applicability open up new research directions in the field of remote sensing data analysis and have important theoretical and applied value.

[0137] 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, any changes and modifications made to the embodiments described herein based on the innovative concept of the present invention, or equivalent structural or procedural transformations made using the content of the present invention's specification and drawings, directly or indirectly applying the above technical solutions to other related technical fields, are all included within 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, Includes the following steps: S1 introduces a dynamic adaptive migration filtering module to dynamically adjust feature weights, calculate the applicability of source domain sample features in real time based on the feature distribution of remote sensing images, and automatically adjust their influence in the migration process. S2, after the initial alignment of the migration features, the target domain sample labels are optimized through confidence 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 construct a multimodal feature fusion network, enhances the target category feature expression capability, and dynamically adjusts the fusion weights of multimodal features to achieve comprehensive capture of complex spatiotemporal features; S4 further corrects the noise of the target domain sample labels through interactive target domain sample label correction, combined with prediction-driven feedback mechanism and manual annotation collaboration; S5 uses a nested noise filtering network with transfer awareness to combine the transfer information of source domain sample features with the target domain sample label distribution, filtering noise target domain sample labels layer by layer and optimizing sample weight distribution. Step S2 includes: S201, Construct a confidence distribution function for target domain sample labels, analyze the consistency of source domain sample features, prediction confidence, and historical prediction trajectories, assign a confidence score to each target domain sample, and identify and label noisy target domain sample labels. The expression is: ； in, Indicates the first Features of source domain samples and the Target domain sample features Consistency measure function between them Indicates the historical predicted trajectory for the first Features of source domain samples The impact of this is used to reflect the stability of its predictions. This represents the prediction confidence level, used to measure how confident the classifier is about the features of samples in the target domain. , , This represents the weighting coefficient, used to balance the influence of different factors. Indicates the first Credibility score of sample assignment to each target domain; S202 introduces a credibility regression network, which is trained by combining the credibility score of each target domain sample, and outputs the corrected target domain sample label distribution, expressed as: ； in, Indicates the original target domain sample label. This represents the weight of the correction factor, used to control the magnitude of the correction. This indicates the set credibility threshold. This represents the corrected target domain sample label. The credibility regression network is composed of multilayer perceptrons; S203 employs an iterative optimization algorithm, repeatedly optimizing the target domain sample label set based on the corrected target domain sample labels. The expression is: ； in, Indicates the first The target domain sample labels are corrected after the next iteration. Indicates the first The credibility score of the target domain samples after the next iteration. express The credibility threshold after the next iteration This represents the adjustment coefficient, used to control the speed of target domain sample label correction. Indicates the first The target domain sample labels are corrected after iteration.

2. The remote sensing data classification method based on transfer learning and label noise filtering according to claim 1, characterized in that, Step S1 includes: S101 introduces a source domain sample feature adaptive calculation mechanism, which adjusts the applicability of source domain sample features in real time based on the feature distribution of target domain samples, and calculates the adaptive weights of source domain sample features, expressed as: ； Where exp represents the natural exponential function. and These represent indexes, Indicates the number of features in the target domain samples. The hyperparameter representing the range of influence of similarity. Indicates the first Features of source domain samples The adaptive weights, Indicates the first Features of each target domain sample; S102, based on the adaptive calculation of source domain sample features, dynamically adjusts the learning rate and feature alignment strategy in transfer learning, with the learning rate being dynamically adjusted. The expression is: ； in, This represents the initial learning rate. This represents a hyperparameter used to control the rate at which the learning rate is adjusted. Indicates the number of training steps. Indicates the number of weights; S103 introduces a negative migration risk assessment, calculating the migration effect of source domain sample features in the target domain and dynamically adjusting the migration strategy based on the difference in migration loss. If negative migration of source domain sample features in the target domain is detected, the feature weights or migration path are automatically adjusted to mitigate the negative migration risk. The expression is: ； in, This indicates the risk of negative migration and reflects the degree to which negative migration occurs during the migration process. Indicates the first in the source domain The loss value of each feature, This represents the loss value of the source domain sample features in the target domain.

3. The remote sensing data classification method based on transfer learning and label noise filtering according to claim 1, characterized in that, Step S3 includes: S301, through a space-spectral mapping function, maps modal features of different dimensions in remote sensing images into a unified spatial framework, thereby achieving modal information fusion. The expression is: ； in, Represents the source domain sample characteristics of the population Features of the target domain samples in the overall population The distribution alignment between them is used to measure the fitness of source domain sample features on the target domain. Indicates the first Modal feature representation, Indicates the first Modal adaptive weights, This represents the total number of modes. This represents the final fusion feature output; S302 utilizes a graph convolutional network to transform modal features in remote sensing images into a graph structure, handling feature consistency constraints at different scales. The expression is: ； in, Indicates the first Graph convolution of modal feature maps, This represents a graph convolution generated using the target domain sample labels. This represents the vector 2-norm operation. This represents the loss function of a graph convolutional network, used to enhance feature consistency across different spatial scales. The total number of graph convolutional feature maps representing modalities. This represents the regularization coefficient, used to balance consistency constraints and the loss of graph convolutional networks. This represents the cross-scale consistency loss; S303 introduces an adaptive spatiotemporal feature adjustment mechanism, combining adaptive learning of spatiotemporal features and a classification optimization algorithm to dynamically adjust feature weights at different scales and optimize classification results. The expression is: ； in, This represents the loss function that adaptively adjusts to spatiotemporal features. Indicates the first The spatiotemporal characteristics of each sample Indicates the spatiotemporal characteristics of the prediction. This represents the weighting adjustment coefficient, used to control the balance between spatiotemporal feature adjustment and classification error. This represents the standard classification loss function, used to calculate the difference between the predicted and true target domain sample labels. The error between them.

4. The remote sensing data classification method based on transfer learning and label noise filtering according to claim 3, characterized in that: The modal features include spectral features, spatial features, and temporal features.

5. 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, based on the predicted confidence level, low-confidence target domain samples are identified. A quality assessment function for the target domain sample features is introduced. Abnormal samples are screened through quality assessment. The set of low-confidence samples is identified by minimizing the objective function. The expression is: ； in, This represents minimizing the objective function. This represents the characteristics of the target domain samples in the population. This indicates the confidence level of the model prediction. This represents a quality assessment function used to measure the consistency or anomaly of image features. This represents the regularization coefficient, used to balance the weight between confidence and feature consistency. This represents a set of samples from the target domain with low confidence. S402, after identifying low-confidence regions, the automatic target domain sample label correction algorithm generates correction suggestions and optimizes the target domain sample labels by combining human feedback. The expression is: ； in, This refers to an automatic target domain sample label correction algorithm, used to optimize target domain sample labels by generating correction suggestions. This represents a correction function based on feedback information, used to combine manually labeled feedback and model predictions to generate suggested corrections for target domain sample labels. This represents the set of target domain sample labels provided by human feedback. This represents the set of target domain sample labels after automatic and manual correction; S403, after the target domain sample labels are corrected, the corrected target domain sample labels are updated synchronously with the current target domain sample labels through an online update mechanism. The difference between the corrected target domain sample labels and the model target domain sample labels is evaluated using a difference metric function, and the target domain sample label weights are adjusted based on the difference. The expression is: ； in, This represents the final corrected set of target domain labels. This is an online update function, representing the process of online correction and update. This represents the label difference metric function, used to calculate the difference between the corrected label and the label of the current target domain sample. This represents the target domain sample label predicted from the training dataset.

6. The remote sensing data classification method based on transfer learning and label noise filtering according to claim 1, characterized in that, Step S5 includes: S501, through a multi-level nested structure, aligns the source domain sample features and 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, achieving dynamic fusion of source domain features and target domain features. The expression is: ； in, This represents the feature obtained by fusing the features of the target domain samples and the features of the source domain samples. This represents the fusion weights at each level, totaling [number]. layer, Represents the source domain sample characteristics of the population The expected value between the target domain sample label set and the target domain sample label set. This represents the characteristics of the target domain samples in the overall population. S502 calculates dynamic feature weights for the features obtained by fusing the target domain sample features and the source domain sample features. This is done by dynamically adjusting the weights by minimizing the label filtering loss function. The contribution to training is ensured by using a regularization term to guarantee stable weight updates; the expression is: ； Where represents the label filtering loss function, represents the dynamic feature weights, represents the dynamic feature weights learned in the previous iteration to maintain stability, represents the vector L2 norm operation, and represents the indicator function, which takes a value of 1 when . The regularization coefficient represents the change in weights. This represents the cross-entropy loss, used to evaluate the accuracy of label prediction. The true target domain sample label represents the actual target domain sample label, and the predicted target domain sample label represents the actual target domain sample label. This represents minimizing the dynamic feature weight function; S503, guided by the features of the source domain samples, identifies and filters out noise labels in the target domain, and optimizes the filtering strategy through the adaptive evaluation function of the source domain sample features, the expression of which is: ； in, This represents the function that minimizes the noise filtering strategy, and it represents the noise filtering strategy itself. represents the source domain sample feature adaptation function, represents the difference value of the predicted target domain sample labels, and represents the quality assessment function used to measure the consistency or anomaly of image features. This represents the label filtering loss function. An adaptive evaluation function representing the features of source domain samples.

Images