Remote sensing sample training data augmentation method and system based on semi-supervised closed-loop iteration

Abstract

The application discloses a remote sensing sample training data amplification method and system based on semi-supervised closed-loop iteration, relates to the technical field of remote sensing big data processing and artificial intelligence, and comprises the following steps: S1, a small amount of remote sensing image samples manually labeled are acquired, an initial training sample set is constructed, and unlabeled remote sensing images are acquired to construct an unlabeled sample set; S2, a target detection model is trained based on the initial training sample set, and an initial target detection model is obtained; S3, the target detection model is used for target detection on the unlabeled remote sensing images, and a pseudo-label sample set containing a predicted category, a predicted bounding box and corresponding confidence is generated. The application fuses four types of indexes, namely, a confidence threshold, time sequence consistency, distribution similarity and confidence stability, jointly judges, constructs a multi-dimensional pseudo-label quality control system, can accurately filter noise and low-quality pseudo-labels, effectively suppresses model prediction drift and classification deviation, and guarantees model detection precision.

Description

Technical Field

[0001] This invention relates to the fields of remote sensing big data processing and artificial intelligence technology, specifically to a method and system for augmenting remote sensing sample training data based on semi-supervised closed-loop iteration. Background Technology

[0002] With the rapid development of aerospace remote sensing, aerial photography, and ground-based remote sensing observation technologies, massive amounts of multi-source remote sensing image data have been continuously acquired. Remote sensing intelligent interpretation technology has been applied on a large scale in many fields such as land spatial planning, natural resource surveys, geological disaster monitoring, ecological environment protection, and agricultural remote sensing yield estimation. Relying on deep learning to achieve automatic detection of remote sensing targets is the mainstream existing technology for improving image interpretation efficiency and reducing the intensity of manual work, and it has also become a core support means for the intelligent upgrading of the industry.

[0003] However, remote sensing images have inherent characteristics such as large imaging area, variable target scale, complex ground texture, and numerous occlusion interferences. Furthermore, the high professional threshold of remote sensing, the cumbersome manual annotation process, and the difficulty in unifying cross-regional sample annotation standards result in extremely high costs for obtaining high-quality labeled samples and a severe shortage of usable labeled samples. Currently, the industry commonly uses traditional semi-supervised pseudo-label iterative amplification techniques to supplement training samples, thereby alleviating the problem of small-sample training. However, this traditional approach has significant shortcomings: firstly, it relies solely on a single confidence level for pseudo-label selection, resulting in a single quality control dimension and an inability to identify fluctuations in prediction results or biases in category distribution during the iteration process. Issues such as inconsistencies in migration and temporal prediction easily introduce a large number of noisy labels, causing model localization drift and classification bias. Secondly, existing technologies do not manage the state of unlabeled samples, and images that have already been augmented will repeatedly enter the detection process, resulting in sample reuse and dataset redundancy. Thirdly, there is a lack of a mechanism to eliminate inferior samples, and low-quality samples that have been repeatedly deemed unqualified continue to participate in closed-loop iteration, constantly consuming computing resources and slowing down the model's convergence speed. Fourthly, most solutions do not pay attention to the dynamic balance of the training set's class distribution. As pseudo-labels are continuously added, the problem of class imbalance in the dataset gradually worsens, further weakening the model's generalization ability. The combination of these defects makes traditional semi-supervised sample augmentation schemes poorly applicable in remote sensing scenarios and unable to support the implementation of high-precision and high-stability remote sensing target detection services. To address this, we propose a method and system for augmenting remote sensing sample training data based on semi-supervised closed-loop iteration. Summary of the Invention

[0004] To address the aforementioned technical issues, a method and system for augmenting remote sensing sample training data based on semi-supervised closed-loop iteration is provided. This technical solution solves the problems of single quality control in traditional augmentation methods and defects in sample transfer and dataset structure.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method and system for augmenting remote sensing sample training data based on semi-supervised closed-loop iteration, including: S1. Obtain a small number of manually labeled remote sensing image samples to construct an initial training sample set, and obtain unlabeled remote sensing images to construct an unlabeled sample set; S2. Train the target detection model based on the initial training sample set to obtain the initial target detection model; S3. Use the target detection model to detect targets in unlabeled remote sensing images and generate a pseudo-label sample set containing predicted categories, predicted bounding boxes and their corresponding confidence scores. S4. Based on the preset judgment indicators, the quality of the pseudo-label sample set is judged to determine whether it meets the preset conditions for being added to the training sample set. The judgment indicators include at least one of the following: confidence threshold, temporal consistency, distribution similarity and confidence stability. S5. When the preset conditions are met, add the pseudo-labeled samples to the training sample set, and retain the corresponding unlabeled samples in the unlabeled sample set and mark them as used; when the preset conditions are not met, remove the corresponding pseudo-label information, and revert the corresponding unlabeled samples to the unlabeled sample set, keeping them in a state to be regenerated. S6. Retrain the target detection model based on the updated training sample set, and repeat steps S3 to S5 until the preset stopping strategy is triggered.

[0006] Preferably, in step S4, the prediction confidence of the pseudo-label sample includes class confidence and bounding box confidence; if the class confidence is greater than or equal to a preset class confidence threshold and the bounding box confidence is greater than or equal to a preset bounding box confidence threshold, then it is determined that the preset condition is met.

[0007] Preferably, in step S4, if the pseudo-label samples predict the same category in a series of preset number of iterations, the matching degree of the predicted bounding box is greater than or equal to a preset matching degree threshold, and the fluctuation of the bounding box confidence is less than a preset fluctuation threshold, then it is determined that the preset conditions are met, and the matching degree is measured by the intersection-union ratio or the center point offset.

[0008] Preferably, in step S4, the category distribution of the current training sample set is the proportion distribution of the number of samples in each category; The preset category distribution is a prior category distribution obtained based on the statistics of the initial training sample set. At least one of KL divergence, JS divergence or cosine similarity is used to measure the similarity between the two distributions. If the similarity is greater than or equal to the preset similarity threshold, it is determined that the preset condition is met.

[0009] Preferably, in step S5, when the preset conditions are met, the predicted category and predicted bounding box of the pseudo-label sample are used as label information and added to the training sample set along with the corresponding unlabeled remote sensing image sample, and the category distribution statistics of the training sample set are updated synchronously.

[0010] Preferably, in step S5, if the preset conditions are not met, the predicted category and predicted bounding box of the pseudo-label sample are removed, and the number of times the sample does not meet the judgment is recorded; if the number of times the judgment is not met reaches the preset upper limit, the sample is removed from the unlabeled sample set; otherwise, the corresponding unlabeled remote sensing image sample is rolled back to the unlabeled sample set and kept in a state of waiting to be regenerated, so that the pseudo-label can be regenerated by the updated target detection model in subsequent iterations.

[0011] Preferably, in step S6, the upper limit of the preset number of iterations is preset based on the size of the initial training sample set and the convergence characteristics of the target detection model; If the current iteration count reaches the preset iteration count limit, then stop executing the loop from step S3 to step S5, and output the current target detection model as the final model.

[0012] Preferably, in step S6, the mean average accuracy (mAP) is used to characterize the model performance, and the difference in mAP between two adjacent iterations is the improvement in model performance. If, in a series of preset iterations, the absolute value of the difference between two adjacent iterations of mAP is less than the preset convergence threshold, then the model is determined to have converged, the loop from step S3 to step S5 is stopped, and the current target detection model is output as the final model.

[0013] Preferably, the system includes: The training sample management module, model training module, pseudo-label generation module, pseudo-label quality judgment module, and training data control module are all included. The training sample management module, connected to the model training module and the training data control module, is used to store the initial training sample set and the unlabeled sample set, and to receive updated data from the training data control module to maintain the updated training sample set. The model training module, connected to the training sample management module and the pseudo-label generation module, is used to train the target detection model based on the training sample set and output the trained target detection model to the pseudo-label generation module. The pseudo-label generation module, connected to the model training module and the pseudo-label quality assessment module, is used to perform target detection on unlabeled remote sensing images using the target detection model, generate pseudo-label samples containing predicted categories, predicted bounding boxes and confidence scores, and output the pseudo-label samples to the pseudo-label quality assessment module. The pseudo-label quality assessment module is connected to the pseudo-label generation module and the training data control module. It is used to assess the quality of pseudo-label samples and output the assessment results and corresponding samples to the training data control module. The training data control module, connected to the pseudo-label quality judgment module and the training sample management module, is used to add pseudo-label samples that meet preset conditions to the training sample set according to the judgment results, and remove pseudo-label information from pseudo-label samples that do not meet preset conditions and revert them to the unlabeled sample set.

[0014] Preferably, the training data control module includes an add execution unit and a rollback execution unit; The execution unit is connected to the pseudo-label quality judgment module and the training sample management module. When it receives the judgment result that meets the preset conditions and the corresponding pseudo-label sample, it sends the predicted category and predicted bounding box of the pseudo-label sample as label information, along with the corresponding unlabeled remote sensing image sample, to the training sample management module. The rollback execution unit is connected to the pseudo-label quality judgment module and the training sample management module. When it receives a judgment result that does not meet the preset conditions and the corresponding pseudo-label sample, it removes the predicted category and predicted bounding box of the pseudo-label sample and rolls back the corresponding unlabeled remote sensing image sample to the training sample management module.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention integrates four indicators—confidence threshold, temporal consistency, distribution similarity, and confidence stability—to construct a multi-dimensional pseudo-label quality control system. This completely changes the traditional single-indicator screening mode, accurately filtering noise and low-quality pseudo-labels, effectively suppressing model prediction drift and classification bias, and ensuring model detection accuracy. A sample status marking mechanism is added to mark unlabeled images that have completed augmentation as used, avoiding the problem of single samples repeatedly participating in detection and augmentation, reducing dataset redundancy, and improving the utilization efficiency of all samples. A new judgment count statistics and over-limit removal strategy are added to directly remove inferior samples that fail multiple judgments, preventing invalid samples from continuously participating in iterative calculations, significantly reducing system computing power overhead, accelerating model convergence, and improving overall operating efficiency. After sample updates, the training set category distribution is maintained synchronously, and distribution similarity is used as a quality control constraint to ensure the stability of the dataset category structure before and after augmentation, improving the category imbalance problem and significantly enhancing the generalization ability and robustness of deep learning models in complex remote sensing scenarios. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the method steps of the present invention; Figure 2 This is a schematic diagram of the system modules of the present invention. Detailed Implementation

[0017] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0018] Reference Figure 1 As shown, the remote sensing sample training data augmentation method based on semi-supervised closed-loop iteration includes: S1. Obtain a small number of manually labeled remote sensing image samples to construct an initial training sample set, and obtain unlabeled remote sensing images to construct an unlabeled sample set; S2. Train the target detection model based on the initial training sample set to obtain the initial target detection model; S3. Use the target detection model to detect targets in unlabeled remote sensing images and generate a pseudo-label sample set containing predicted categories, predicted bounding boxes and their corresponding confidence scores. S4. Based on the preset judgment indicators, the quality of the pseudo-label sample set is judged to determine whether it meets the preset conditions for being added to the training sample set. The judgment indicators include at least one of the following: confidence threshold, temporal consistency, distribution similarity and confidence stability. S5. When the preset conditions are met, add the pseudo-labeled samples to the training sample set, and retain the corresponding unlabeled samples in the unlabeled sample set and mark them as used; when the preset conditions are not met, remove the corresponding pseudo-label information, and revert the corresponding unlabeled samples to the unlabeled sample set, keeping them in a state to be regenerated. S6. Retrain the target detection model based on the updated training sample set, and repeat steps S3 to S5 until the preset stopping strategy is triggered.

[0019] In step S1, the labeled remote sensing images undergo preprocessing including radiometric correction, geometric correction, atmospheric correction, image cropping, and pixel value normalization. A small number of manually labeled remote sensing image samples refer to the number of labeled samples corresponding to each type of target being between fifty and two hundred, or accounting for no more than ten percent of all available labeled data. The number of unlabeled remote sensing images is usually ten to one hundred times the number of labeled samples. The specific ratio can be adjusted according to the cost of acquiring remote sensing images and the complexity of the target detection task. The initial training sample set and the unlabeled sample set are physically independent of each other. In step S2, the object detection model can adopt the YOLOv8 network architecture. The initial training rounds can be set to one hundred rounds. After training, the model weights are saved as the initial object detection model for subsequent pseudo-label generation. In step S3, unlabeled remote sensing imagery is input into the current target detection model (the initial target detection model obtained in S2 during the first iteration). The model predicts candidate targets in the imagery and outputs the predicted category, predicted bounding box coordinates (usually expressed as the coordinates of the top left corner and width and height, or the coordinates of the center point and width and height), and the corresponding confidence score. For overlapping predicted boxes in the same imagery, non-maximum suppression (NMS) is used to remove duplicates. The intersection-over-union (IoU) threshold of NMS can be set to 0.5 for example. The prediction results after deduplication are used as pseudo-label samples and are collected to form a pseudo-label sample set. In subsequent closed-loop iterations, the target detection model used in S3 refers to the model trained in the latest iteration round, rather than always using the initial target detection model in S2.

[0020] In step S4, the prediction confidence of the pseudo-label sample includes class confidence and bounding box confidence. If the class confidence is greater than or equal to the preset class confidence threshold and the bounding box confidence is greater than or equal to the preset bounding box confidence threshold, then it is determined that the preset condition is met.

[0021] The category confidence score is obtained from the classification branch of the object detection model. For models using the YOLOv8 architecture, the category confidence score is the maximum value among the probability outputs of the model for each category to which the predicted box belongs, and its value ranges from 0 to 1. The bounding box confidence score is jointly determined by the localization branch and the target branch of the object detection model. The bounding box confidence score can be defined as the product of the target score output by the model and the quality estimate of the predicted box, or as the box confidence score directly output by the model, and its value ranges from 0 to 1. The preset category confidence threshold and the preset bounding box confidence threshold are determined based on the size of the initial training sample set, the complexity of the object detection task, and the confidence distribution of the model output. The category confidence threshold is set higher than the bounding box confidence threshold because the negative impact of category errors on model training is usually greater than that of localization bias. By increasing the category admission threshold, the probability of introducing false labels of incorrect categories can be effectively reduced. If any confidence level is lower than the corresponding threshold, it is determined that the preset conditions are not met, and the rollback process of S5 is entered.

[0022] In step S4, if the pseudo-label samples predict the same category in a series of preset iterations, the matching degree of the predicted bounding box is greater than or equal to the preset matching degree threshold, and the fluctuation of the bounding box confidence is less than the preset fluctuation threshold, then it is determined that the preset conditions are met, and the matching degree is measured by the intersection-union ratio or the center point offset.

[0023] The number of consecutive preset times can be set to three. The same prediction category means that the model predicts the same category for the sample in three iterations. The matching degree is used to measure the consistency of predicted bounding boxes in consecutive iterations. The intersection-union ratio is calculated as the intersection area of ​​the predicted bounding boxes in two iterations divided by the union area, and its value ranges from 0 to 1. The center point offset refers to the Euclidean distance or normalized distance between the center points of the predicted bounding boxes in two iterations in the image coordinate system. Confidence stability is used to measure the consistency of bounding box confidence across multiple iterations. In this embodiment, standard deviation is used as the metric. The smaller the standard deviation, the higher the stability. The bounding box confidence of the pseudo-label sample is recorded in three consecutive iterations, and its standard deviation is calculated. The preset fluctuation threshold can be set to 0.1 for example. When using the intersection-union ratio (IU) metric, the preset IU matching threshold can be set to 0.5. When using the center point offset metric, the preset center point offset threshold can be set to a normalized distance of less than 0.02. For a pseudo-label sample to be considered to meet the preset conditions for inclusion in the training sample set, the following conditions must be met simultaneously: the predicted category must be the same, the matching degree must be greater than or equal to the threshold, and the fluctuation must be less than the threshold.

[0024] In step S4, the class distribution of the current training sample set is the distribution of the proportion of samples in each class; The preset category distribution is a prior category distribution obtained based on the statistics of the initial training sample set. At least one of KL divergence, JS divergence or cosine similarity is used to measure the similarity between the two distributions. If the similarity is greater than or equal to the preset similarity threshold, it is determined that the preset condition is met.

[0025] The number of samples of each category in the current training sample set is counted, and divided by the total number of samples to obtain the proportion of each category, forming the category distribution vector of the current training sample set; after constructing the initial training sample set in S1, the proportion of the number of samples of each category in the initial training sample set is immediately counted as the prior category distribution. When using KL divergence, the KL divergence value of the current class distribution relative to the prior class distribution is calculated. The smaller the KL divergence value, the closer the two class distributions are. When using JS divergence, JS divergence is a symmetric KL divergence, which is more robust to the zero-value distribution. When using cosine similarity, the inner product operation is performed between the current class distribution vector and the prior class distribution vector, and then normalized to obtain the cosine similarity value, which ranges from 0 to 1. In step S5, when the preset conditions are met, the predicted category and predicted bounding box of the pseudo-label sample are used as label information and added to the training sample set along with the corresponding unlabeled remote sensing image sample, and the category distribution statistics of the training sample set are updated synchronously.

[0026] The unlabeled remote sensing images and their newly generated pseudo-labels are combined into labeled samples and added to the storage queue of the training sample set. The mixing ratio can be set as needed to balance the relative weight of the new samples and the real labeled samples, so as to avoid the real labeled samples being diluted due to the large number of pseudo-labeled samples. After adding new pseudo-label samples, an incremental update method is adopted, which only accumulates the category count of the newly added samples and updates the current category distribution vector, providing the latest data for the distribution similarity determination in the next round S4.

[0027] In step S5, if the preset conditions are not met, the predicted category and predicted bounding box of the pseudo-label sample are removed, and the number of times the judgment is not met is recorded. If the number of judgments is not met reaches the preset upper limit, the corresponding unlabeled remote sensing image sample is removed from the unlabeled sample set. Otherwise, the corresponding unlabeled remote sensing image sample is rolled back to the unlabeled sample set and kept in a state of waiting to be regenerated, so that the pseudo-label can be regenerated by the updated target detection model in subsequent iterations.

[0028] When rolling back, all prediction information generated in the current iteration is deleted, but the corresponding unlabeled remote sensing image sample itself is retained. A counter is maintained for the sample in the metadata of the unlabeled sample set, and the counter is incremented by one each time it rolls back. In this embodiment, the preset upper limit can be set to five times. Reaching the upper limit indicates that the unlabeled sample cannot generate high-quality pseudo-labels in multiple iterations. It may be a difficult sample or background noise. Continuing to keep it will waste computing resources. Therefore, it is removed. The removed sample can be transferred to the difficult sample set and can be kept for future reference. The model is retrained based on the expanded training sample set in each iteration, and its detection capability is gradually improved. Samples that did not meet the conditions in the past may generate pseudo-labels that meet the quality requirements in subsequent iterations.

[0029] In step S6, the upper limit of the preset number of iterations is preset based on the size of the initial training sample set and the convergence characteristics of the target detection model; If the current iteration count reaches the preset iteration count limit, then stop executing the loop from step S3 to step S5, and output the current target detection model as the final model.

[0030] The smaller the initial training sample set or the higher the complexity of the object detection model, the more iterations are usually required. This strategy acts as a hard stopping condition to ensure that closed-loop iterations do not proceed indefinitely and to control the overall training overhead.

[0031] In step S6, the mean average precision (mAP) is used to characterize the model performance, and the difference in mAP between two adjacent iterations is the improvement in model performance. If, in a series of preset iterations, the absolute value of the difference between two adjacent iterations of mAP is less than the preset convergence threshold, then the model is determined to have converged, the loop from step S3 to step S5 is stopped, and the current target detection model is output as the final model.

[0032] After each closed-loop iteration, the current target detection model is evaluated using the validation set, and the mAP value is calculated. The validation set can consist of a portion of the true label samples divided from the initial training sample set and does not participate in model training. In this embodiment, the number of consecutive preset times can be set to three, and the preset convergence threshold can be set to 0.005, that is, when the mAP improvement is less than five per thousand, the model performance is considered to have entered a plateau period, and further iteration is unlikely to bring significant improvement. The closed-loop iteration stops when either the maximum number of iterations is reached or the model performance converges.

[0033] Reference Figure 2 As shown, a remote sensing sample training data augmentation system based on semi-supervised closed-loop iteration is characterized in that the system includes: The training sample management module, model training module, pseudo-label generation module, pseudo-label quality judgment module, and training data control module are all included. The training sample management module, connected to the model training module and the training data control module, is used to store the initial training sample set and the unlabeled sample set, and to receive updated data from the training data control module to maintain the updated training sample set. The model training module, connected to the training sample management module and the pseudo-label generation module, is used to train the target detection model based on the training sample set and output the trained target detection model to the pseudo-label generation module. The pseudo-label generation module, connected to the model training module and the pseudo-label quality assessment module, is used to perform target detection on unlabeled remote sensing images using the target detection model, generate pseudo-label samples containing predicted categories and predicted bounding boxes, and output the pseudo-label samples to the pseudo-label quality assessment module. The pseudo-label quality assessment module is connected to the pseudo-label generation module and the training data control module. It is used to assess the quality of pseudo-label samples and output the assessment results and corresponding samples to the training data control module. The training data control module, connected to the pseudo-label quality judgment module and the training sample management module, is used to add pseudo-label samples that meet preset conditions to the training sample set according to the judgment results, and remove pseudo-label information from pseudo-label samples that do not meet preset conditions and revert them to the unlabeled sample set.

[0034] The training data control module includes an add execution unit and a rollback execution unit; The execution unit is connected to the pseudo-label quality judgment module and the training sample management module. When it receives the judgment result that meets the preset conditions and the corresponding pseudo-label sample, it sends the predicted category and predicted bounding box of the pseudo-label sample as label information, along with the corresponding unlabeled remote sensing image sample, to the training sample management module. The rollback execution unit is connected to the pseudo-label quality judgment module and the training sample management module. When it receives a judgment result that does not meet the preset conditions and the corresponding pseudo-label sample, it removes the predicted category and predicted bounding box of the pseudo-label sample and rolls back the corresponding unlabeled remote sensing image sample to the training sample management module.

[0035] After receiving the above data, the training sample management module adds the new sample to the training sample set and updates the status of the corresponding unlabeled sample to the used status; after receiving the rollback instruction, the training sample management module restores the status of the sample to the pending regeneration status.

[0036] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A method for augmenting remote sensing sample training data based on semi-supervised closed-loop iteration, characterized in that, include: S1. Obtain a small number of manually labeled remote sensing image samples to construct an initial training sample set, and obtain unlabeled remote sensing images to construct an unlabeled sample set; S2. Train the target detection model based on the initial training sample set to obtain the initial target detection model; S3. Use the target detection model to detect targets in unlabeled remote sensing images and generate a pseudo-label sample set containing predicted categories, predicted bounding boxes and their corresponding confidence scores. S4. Based on the preset judgment indicators, the quality of the pseudo-label sample set is judged to determine whether it meets the preset conditions for being added to the training sample set. The judgment indicators include at least one of the following: confidence threshold, temporal consistency, distribution similarity and confidence stability. S5. When the preset conditions are met, add the pseudo-labeled samples to the training sample set, and retain the corresponding unlabeled samples in the unlabeled sample set and mark them as used; when the preset conditions are not met, remove the corresponding pseudo-label information, and revert the corresponding unlabeled samples to the unlabeled sample set, keeping them in a state to be regenerated. S6. Retrain the target detection model based on the updated training sample set, and repeat steps S3 to S5 until the preset stopping strategy is triggered.

2. The remote sensing sample training data augmentation method based on semi-supervised closed-loop iteration according to claim 1, characterized in that, In step S4, the prediction confidence of the pseudo-label sample includes class confidence and bounding box confidence. If the class confidence is greater than or equal to the preset class confidence threshold and the bounding box confidence is greater than or equal to the preset bounding box confidence threshold, then it is determined that the preset condition is met.

3. The remote sensing sample training data augmentation method based on semi-supervised closed-loop iteration according to claim 1, characterized in that, In step S4, if the pseudo-label samples predict the same category in a series of preset iterations, the matching degree of the predicted bounding box is greater than or equal to the preset matching degree threshold, and the fluctuation of the bounding box confidence is less than the preset fluctuation threshold, then it is determined that the preset conditions are met, and the matching degree is measured by the intersection-union ratio or the center point offset.

4. The remote sensing sample training data augmentation method based on semi-supervised closed-loop iteration according to claim 1, characterized in that, In step S4, the class distribution of the current training sample set is the distribution of the proportion of samples in each class; The preset category distribution is a prior category distribution obtained based on the statistics of the initial training sample set. At least one of KL divergence, JS divergence or cosine similarity is used to measure the similarity between the two distributions. If the similarity is greater than or equal to the preset similarity threshold, it is determined that the preset condition is met.

5. The remote sensing sample training data augmentation method based on semi-supervised closed-loop iteration according to claim 1, characterized in that, In step S5, when the preset conditions are met, the predicted category and predicted bounding box of the pseudo-label sample are used as label information and added to the training sample set along with the corresponding unlabeled remote sensing image sample, and the category distribution statistics of the training sample set are updated synchronously.

6. The remote sensing sample training data augmentation method based on semi-supervised closed-loop iteration according to claim 1, characterized in that, In step S5, if the preset conditions are not met, the predicted category and predicted bounding box of the pseudo-label sample are removed, and the number of times the sample does not meet the judgment is recorded. If the number of times the judgment is not met reaches the preset upper limit, the sample is removed from the unlabeled sample set. Otherwise, the corresponding unlabeled remote sensing image sample is rolled back to the unlabeled sample set and kept in a state of waiting to be regenerated, so that the pseudo-label can be regenerated by the updated target detection model in subsequent iterations.

7. The remote sensing sample training data augmentation method based on semi-supervised closed-loop iteration according to claim 1, characterized in that, In step S6, the upper limit of the preset number of iterations is preset based on the size of the initial training sample set and the convergence characteristics of the target detection model; If the current iteration count reaches the preset iteration count limit, then stop executing the loop from step S3 to step S5, and output the current target detection model as the final model.

8. The remote sensing sample training data augmentation method based on semi-supervised closed-loop iteration according to claim 1, characterized in that, In step S6, the mean average precision (mAP) is used to characterize the model performance, and the difference in mAP between two adjacent iterations is the improvement in model performance. If, in a series of preset iterations, the absolute value of the difference between two adjacent iterations of mAP is less than the preset convergence threshold, then the model is determined to have converged, the loop from step S3 to step S5 is stopped, and the current target detection model is output as the final model.

9. A remote sensing sample training data augmentation system based on semi-supervised closed-loop iteration, applied to the remote sensing sample training data augmentation method based on semi-supervised closed-loop iteration as described in any one of claims 1-8, characterized in that, The system includes: The training sample management module, model training module, pseudo-label generation module, pseudo-label quality judgment module, and training data control module are all included. The training sample management module, connected to the model training module and the training data control module, is used to store the initial training sample set and the unlabeled sample set, and to receive updated data from the training data control module to maintain the updated training sample set. The model training module, connected to the training sample management module and the pseudo-label generation module, is used to train the target detection model based on the training sample set and output the trained target detection model to the pseudo-label generation module. The pseudo-label generation module, connected to the model training module and the pseudo-label quality assessment module, is used to perform target detection on unlabeled remote sensing images using the target detection model, generate pseudo-label samples containing predicted categories, predicted bounding boxes and confidence scores, and output the pseudo-label samples to the pseudo-label quality assessment module. The pseudo-label quality assessment module is connected to the pseudo-label generation module and the training data control module. It is used to assess the quality of pseudo-label samples and output the assessment results and corresponding samples to the training data control module. The training data control module, connected to the pseudo-label quality judgment module and the training sample management module, is used to add pseudo-label samples that meet preset conditions to the training sample set according to the judgment results, and remove pseudo-label information from pseudo-label samples that do not meet preset conditions and revert them to the unlabeled sample set.

10. The remote sensing sample training data augmentation system based on semi-supervised closed-loop iteration according to claim 9, characterized in that, The training data control module includes an add execution unit and a rollback execution unit; The execution unit is connected to the pseudo-label quality judgment module and the training sample management module. When it receives the judgment result that meets the preset conditions and the corresponding pseudo-label sample, it sends the predicted category and predicted bounding box of the pseudo-label sample as label information, along with the corresponding unlabeled remote sensing image sample, to the training sample management module. The rollback execution unit is connected to the pseudo-label quality judgment module and the training sample management module. When it receives a judgment result that does not meet the preset conditions and the corresponding pseudo-label sample, it removes the predicted category and predicted bounding box of the pseudo-label sample and rolls back the corresponding unlabeled remote sensing image sample to the training sample management module.

Images