Weak label expansion and screening method for remote sensing sample library construction

Abstract

The present application relates to the weak label expansion and screening field, especially to the weak label expansion and screening method for remote sensing sample library construction. The obtained multi-source remote sensing data is unified in space reference, divided into grids and constructed in features, combined with the obtained artificial annotation sample data, to form remote sensing sample units, and the remote sensing sample units are divided to obtain an annotated sample set and an unannotated sample set; based on the remote sensing sample units, a label propagation weight is obtained through a spatio-temporal coupling nonlinear decay propagation model, a category propagation intensity is fused, and a candidate weak label is generated for an unannotated sample; based on the label propagation weight corresponding to the candidate weak label, a credibility evolution model driven by multi-source consistency is constructed, the final credibility of the weak label sample is calculated, and the weak label sample meeting the requirements is screened by comparing with a set screening threshold. The technical problems of low construction efficiency of the remote sensing sample library, insufficient reliability of the weak label and difficulty in effectively screening the error samples are solved.

Description

Technical Field

[0001] This invention relates to the field of weak label expansion and screening, and more particularly to a method for weak label expansion and screening for the construction of remote sensing sample libraries. Background Technology

[0002] With the development of remote sensing technology, the scale of remote sensing data continues to expand, and remote sensing sample libraries have become an important foundation for artificial intelligence models to identify, classify, and extract information from land features. Existing remote sensing sample libraries mainly rely on manual annotation. However, remote sensing images cover a wide area and have a large data volume, making manual annotation inefficient, costly, and difficult to update, thus failing to meet the demand for rapid utilization of large-scale remote sensing data. To reduce the burden of manual annotation, existing technologies employ weakly supervised methods to expand the labels of a large amount of unlabeled remote sensing data using a small number of labeled samples. However, existing weak label expansion methods typically generate labels based solely on feature similarity or model prediction results, failing to fully consider the spatial correlation, temporal changes, and label propagation paths between remote sensing samples, easily leading to erroneous labels that accumulate during subsequent expansion. Furthermore, existing screening methods often use fixed thresholds to determine the validity of weak labels, failing to dynamically adjust screening conditions according to the sample library's status, resulting in unstable quality of expanded samples. Therefore, existing technologies suffer from low efficiency in constructing remote sensing sample libraries, insufficient reliability of weak labels, and difficulty in effectively removing erroneous samples. A method that combines spatial, temporal, and feature correlations to automatically expand and reliably screen weak labels is needed to improve the quality of remote sensing sample library construction. Summary of the Invention

[0003] This invention provides a method for weak label expansion and screening for the construction of remote sensing sample libraries, in order to solve the problems of high manual labeling costs, low utilization rate of unlabeled remote sensing data, and the accumulation of label errors due to spatial location differences, temporal changes and multiple propagations during the weak label expansion process.

[0004] The weak label expansion and screening method for remote sensing sample library construction of the present invention includes the following steps:

[0005] S1. Acquire multi-source remote sensing data and existing manually labeled sample data. Perform spatial benchmark unification, grid division, and feature construction on the remote sensing image data in the multi-source remote sensing data. Combine with the existing manually labeled sample data to form remote sensing sample units containing spatial location coordinates, temporal information, multi-dimensional feature vectors, and label status. Divide the remote sensing sample units according to the label status to obtain a set of labeled samples and a set of unlabeled samples. Based on the remote sensing sample units, calculate the label propagation weight of labeled samples to unlabeled samples using a spatiotemporal coupled nonlinear attenuation propagation model. Based on the label propagation weight, accumulate the label propagation weights of labels belonging to the same category to obtain the category propagation intensity. Based on the category propagation intensity, generate candidate weak labels for unlabeled samples and obtain the label propagation weights corresponding to the candidate weak labels.

[0006] S2. Based on the label propagation weights corresponding to the candidate weak labels, and combined with remote sensing sample units, construct a multi-source consistency-driven credibility evolution model to calculate the final credibility of the weak label samples; compare the final credibility with the set screening threshold, and add the weak label samples that meet the credibility requirements to the remote sensing sample library.

[0007] Preferably, S1 specifically includes:

[0008] In the spatiotemporal coupled nonlinear decay propagation model, based on the spatial coordinates of labeled and unlabeled samples, the normalized spatial distance between them is calculated using Euclidean distance and normalization. Based on temporal information, the difference in acquisition time between labeled and unlabeled samples is normalized to calculate the normalized time difference. Based on the multidimensional feature vectors of labeled and unlabeled samples, the angle between their multidimensional feature vectors is calculated using cosine similarity.

[0009] Preferably, S1 specifically includes:

[0010] Based on normalized spatial distance and normalized time difference, the exponential term is calculated by extending the model through Gaussian radial basis function; based on the angle between the multidimensional feature vectors of labeled and unlabeled samples, the non-negative weight factor is calculated through cosine similarity calculation and normalization; the propagation path length is introduced, and the path penalty term is calculated through a decreasing function.

[0011] Preferably, S1 specifically includes:

[0012] Multiplying the exponential term, the non-negative weight factor, and the path penalty term yields the label propagation weight of labeled samples on unlabeled samples. The calculation formula is as follows: , in, Indicates labeled samples For unlabeled samples Tag propagation weight; Indicates labeled samples and unlabeled samples Normalized spatial distance; Indicates labeled samples and unlabeled samples Normalized time difference; Indicates the spatial attenuation coefficient; Indicates the time decay coefficient; It is an exponential term; Indicates labeled samples and unlabeled samples The angle between the multidimensional feature vectors; Indicates a non-negative weighting factor; Indicates the propagation path length; Indicates the path attenuation coefficient; It is a path penalty item.

[0013] Preferably, S1 specifically includes:

[0014] Based on the category propagation strength, the category with the greatest category propagation strength is selected as the candidate weak label, and the corresponding unlabeled sample number is used as the weak label sample number; based on the label propagation weight, the label propagation weight of the labeled sample to the unlabeled sample corresponding to the candidate weak label is obtained.

[0015] Preferably, S2 specifically includes:

[0016] In the multi-source consistency-driven credibility evolution model, the spatial distance decay factor is calculated based on the normalized spatial distance between labeled and unlabeled samples corresponding to candidate weak labels plus the reciprocal of 1; based on the multi-dimensional feature vector, the feature differences between adjacent time nodes of weak label samples are accumulated to obtain the temporal change degree of weak label samples.

[0017] Preferably, S2 specifically includes:

[0018] Based on the spatial distance decay factor, the propagation weights of the candidate weak labels are weighted statistically to obtain the spatial propagation consistency factor; based on the degree of temporal change of the weak label samples, the time transformation decay factor is calculated through the exponential decay model.

[0019] Preferably, S2 specifically includes:

[0020] The number of times a weakly labeled sample has maintained the same class status in history is introduced, and a stability enhancement factor is calculated using a logarithmic function. The spatial propagation consistency factor, the time transformation decay factor, and the stability enhancement factor are multiplied together to obtain the final credibility of the weakly labeled sample.

[0021] Preferably, S2 specifically includes:

[0022] When the final confidence level is greater than or equal to the screening threshold, the weakly labeled sample and its corresponding weakly labeled category are written into the remote sensing sample library; when the final confidence level is less than the screening threshold, the weakly labeled sample is discarded or retained in the candidate sample set for subsequent re-evaluation.

[0023] The beneficial effects of the technical solution of the present invention are:

[0024] 1. The label propagation weights of unlabeled samples are calculated using labeled samples, and candidate weak labels are determined by the category propagation intensity. This allows the influence of multiple labeled samples on the same unlabeled sample to be comprehensively quantified, reducing the impact of single labeled sample errors on the expansion results. At the same time, by introducing a path decay model, the weights of weak labels obtained through multi-layer neighborhood propagation are gradually reduced, effectively suppressing the accumulation of errors in weak labels during the iterative expansion process and improving the label reliability during the expansion of the remote sensing sample library.

[0025] 2. Based on label propagation weight, spatial consistency, and temporal stability, a multi-source consistency-driven credibility evolution model is constructed. This model integrates and analyzes the propagation information, spatial distribution relationships, and multi-temporal change characteristics corresponding to candidate weak labels, enabling dynamic quantitative evaluation of the credibility of each extended sample. Simultaneously, by combining the spatial continuity and temporal stability characteristics of remote sensing ground objects, it can effectively identify unreliable weak labels caused by changes in ground objects, spatial offsets, or excessively long propagation distances, thereby reducing the probability of erroneous labels entering the remote sensing sample database.

[0026] 3. Dynamically determine the screening threshold based on the credibility distribution of samples of the same category, and automatically determine whether weakly labeled samples should be included in the remote sensing sample library by comparing the relationship between the credibility of weakly labeled samples and the screening threshold. This enables the sample expansion process to adaptively adjust according to the current quality status of the remote sensing sample library. When the number of high-credibility samples in the sample library increases, the screening requirements are raised to avoid the continuous introduction of low-quality samples. When the number of samples is insufficient, the expansion conditions are appropriately adjusted according to the credibility distribution to improve the ability to supplement effective samples. Attached Figure Description

[0027] Figure 1 This is a flowchart of the weak label expansion and screening method for constructing remote sensing sample libraries as described in this invention. Detailed Implementation

[0028] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0029] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0030] The following description, in conjunction with the accompanying drawings, details the specific scheme of the weak label expansion and screening method for constructing remote sensing sample libraries provided by this invention.

[0031] See attached document Figure 1 The diagram illustrates a flowchart of a weak label expansion and screening method for constructing a remote sensing sample library, provided by an embodiment of the present invention. The method includes the following steps:

[0032] S1. Acquire multi-source remote sensing data and existing manually labeled sample data. Perform spatial benchmark unification, grid division, and feature construction on the remote sensing image data in the multi-source remote sensing data. Combine with the existing manually labeled sample data to form remote sensing sample units containing spatial location coordinates, temporal information, multi-dimensional feature vectors, and label status. Divide the remote sensing sample units according to the label status to obtain a set of labeled samples and a set of unlabeled samples. Based on the remote sensing sample units, calculate the label propagation weight of labeled samples to unlabeled samples using a spatiotemporally coupled nonlinear attenuation propagation model. Based on the label propagation weight, accumulate the label propagation weights of labels belonging to the same category to obtain the category propagation intensity. Based on the category propagation intensity, generate candidate weak labels for unlabeled samples and obtain the label propagation weights corresponding to the candidate weak labels.

[0033] Acquire multi-source remote sensing data and existing manually labeled sample data corresponding to the area where the remote sensing sample library to be constructed is located. The multi-source remote sensing data includes remote sensing image data at different time points, corresponding spatial coordinate data, and multi-temporal change information data; the manually labeled sample data includes category labels, sample spatial location, manual confirmation status, and label source information.

[0034] After acquiring the aforementioned data, the remote sensing image data is first processed using a spatial registration method based on coordinate transformation models and geometric correction to unify the spatial reference. This maps remote sensing image data from different time points to the same geographic coordinate reference system. Then, the target area is divided into regularized grids according to a preset spatial resolution, dividing the study area into multiple continuously arranged spatial grids. Each spatial grid serves as a basic remote sensing sample unit, and a unique sample number is assigned to each grid. The spatial resolution is determined by technicians based on application requirements; for example, for building identification scenarios using meter-level remote sensing imagery, the spatial resolution can be set to 10 meters × 10 meters.

[0035] For each remote sensing sample unit Determine spatial coordinate information based on its corresponding grid range. , Time information is determined based on the time of acquisition of remote sensing images. Simultaneously, the remote sensing image data and manually labeled sample data within the grid area are associated to form a basic sample data set. For data with manual annotations, the corresponding category information, manual confirmation status, and label source information are written into the sample label field; for data without manual annotations, its label status is set to pending expansion.

[0036] Subsequently, for each remote sensing sample unit The corresponding remote sensing image data is processed to construct features and form multidimensional feature vectors. Specifically, the multi-band remote sensing image pixel data within the grid range corresponding to the remote sensing sample unit is used as the basic input data. The multi-band remote sensing image pixel data includes the pixel values ​​corresponding to each spectral band at the same acquisition time. It is assumed that the grid range corresponding to the remote sensing sample unit contains... Each pixel contains 100 pixels. The reflectance values ​​of each band are calculated as follows: First, the reflectance values ​​of each band are statistically calculated, and the average and standard deviation of that band within the grid are obtained. The average value represents the overall spectral response level of the region, and the standard deviation represents the dispersion of the reflectance values ​​of that band. The average and standard deviation are then horizontally concatenated to form the spectral feature components. Second, the gray values ​​of the pixels within the grid are quantized, and based on pixel spatial adjacency relationships, such as 8-neighborhood, a set of adjacent pixel pairs is constructed. The average variation of the gray value difference and the uniformity of the gray value distribution are statistically analyzed. The gray value difference is calculated by the difference of the gray values ​​of adjacent pixels, and the uniformity is calculated by the frequency distribution of gray values. The average variation of the gray value difference is then calculated. The spectral feature component is formed by horizontally splicing together the amplitude of grayscale and the uniformity of grayscale distribution. Next, considering the spatial relationship between the grid and its adjacent grids, the numerical difference between corresponding boundary pixels is calculated. Specifically, the difference in band reflectance values ​​of corresponding pixels at the boundaries of adjacent grids is calculated, and the difference results are averaged to obtain the boundary variation amplitude. Simultaneously, the proportion of pixels whose difference between the grid and its adjacent grids is less than a preset threshold is counted as a spatial continuity indicator. The boundary variation amplitude is then horizontally spliced ​​together to form a spatial structure feature component. The threshold is set by calculating the average and standard deviation of all difference results between the current grid and its adjacent grids, plus double the standard deviation. The above spectral feature components, texture feature components, and spatial structure feature components are spliced ​​together in a preset order to form a multidimensional feature vector of a unified dimension. and the spatial location coordinates of the remote sensing sample unit , Time information and tag status As basic attributes, they form a quintuple, with the specific expression being: , in, Indicates the first One remote sensing sample unit; , Indicates the first Spatial location coordinates of a remote sensing sample unit; Indicates the first Remote sensing data acquisition time for each remote sensing sample unit; Indicates the first The multidimensional feature vector corresponding to each remote sensing sample unit; Indicates the first The current label status of each remote sensing sample unit is derived from the category labels in the manually labeled sample data.

[0037] The technical objective of the above steps is to transform the basic sample dataset into computable, unified data units and establish the index foundation for all subsequent calculations. Then, the multidimensional feature vectors... This serves as the sole input feature set for subsequent algorithms. Based on this, and according to the label state... The entire set of remote sensing sample units is divided into two subsets: the labeled sample set and the subset. This refers to samples that already have manually labeled tags or historically highly reliable tags, among which... Indicates the first Each remote sensing sample unit, i.e., a labeled sample , Indicates the first The current label status of each remote sensing sample unit, with labeled samples serving as source nodes for label propagation; This indicates that the label is empty, meaning it is in a labeled state; the unlabeled sample set: This refers to samples that do not yet have labels or whose labels are unconfirmed. Indicates the first One remote sensing sample unit, i.e., an unlabeled sample , Indicates the first The current label status of each remote sensing sample unit, with unlabeled samples as objects to be inferred.

[0038] Further, we move into the weak label propagation calculation stage. The core of this stage lies in addressing the problem of traditional label propagation neglecting the coupling of space, time, and propagation path. Therefore, a spatiotemporally coupled nonlinear attenuation propagation model is constructed for any unlabeled sample. Calculate its value from the labeled samples The received label propagation weight is specifically implemented using the following formula: , in, Indicates labeled samples For unlabeled samples The label propagation weight, i.e., propagation reliability; Indicates labeled samples and unlabeled samples The normalized spatial distance is obtained by using the maximum value method on labeled samples. and unlabeled samples The actual spatial distance is obtained by normalizing the actual spatial distance, which is calculated by Euclidean distance from the spatial location coordinates of the remote sensing sample unit. Indicates labeled samples and unlabeled samples The normalized time difference is calculated by using the maximum value method for labeled samples. and unlabeled samples The difference in acquisition time between the two is obtained by normalization. The spatial attenuation coefficient is used to control the influence of normalized spatial distance on the label propagation weight. The minimum Euclidean distance between a labeled sample and all its neighboring samples in the labeled sample set is calculated as the minimum neighbor Euclidean distance. The average of the minimum neighbor Euclidean distances of all labeled samples is taken and normalized using the maximum and minimum value method to obtain the average neighborhood distance. The reciprocal of the square of the average neighborhood distance is used as the spatial attenuation coefficient. The time decay coefficient is used to control the impact of normalized time difference on label propagation weight. The time interval between two consecutive remote sensing images of the same area is statistically analyzed. All adjacent time intervals are averaged and normalized using the maximum and minimum value method to obtain the average update cycle. The reciprocal of the square of the average update cycle is used as the time decay coefficient. It is an exponential term, determined based on the Gaussian radial basis function (RBF kernel) extended model, used to describe the stronger the interaction between samples that are closer in space and time. Indicates labeled samples and unlabeled samples The angle between the multidimensional feature vectors is obtained by cosine similarity calculation. Represents the cosine similarity between multidimensional feature vectors of samples; Indicates a non-negative weighting factor, used to... Mapped to This reflects the degree of similarity between two samples at the spectral, texture, and semantic levels. Indicates the propagation path length, used to describe the path length from the labeled sample. Propagation to unlabeled samples is achieved through successive, layer-by-layer neighborhood expansion. The minimum number of propagation layers required, namely, the layer-by-layer neighborhood expansion propagation, i.e., the first layer propagation: only allowing direct neighbor samples to propagate, the second layer propagation: allowing neighbor samples of neighboring regions to propagate, and so on, are completed in sequence; The path attenuation coefficient is used to control the impact of propagation path length on propagation reliability. Samples with real labels are selected from the historical sample database as validation samples. Labels are propagated according to different propagation path lengths, and the proportion of predicted labels matching the real labels under each propagation path length is calculated to obtain the corresponding propagation accuracy. Using the propagation accuracy of the direct neighborhood propagation path as a benchmark, the percentage decrease in propagation accuracy under different propagation path lengths is calculated. The percentage decreases for all propagation path lengths are then normalized using the averaging and maximum / minimum method to obtain the path attenuation coefficient. In practical applications, a single calibration can be performed based on the propagation results during the initialization phase of the sample library, with the value range set to... ; It is a path penalty term, determined based on the path decay model in the graph propagation algorithm. It is used to describe that the longer the propagation path, the more obvious the error accumulation, and therefore it is necessary to reduce its impact.

[0039] The formula is formed as follows: First, based on the spatial continuity of remotely sensed ground features, an extended model based on Gaussian radial basis function (RBF kernel) is introduced to normalize the spatial distance between samples. As an influencing factor, the propagation weight of samples with greater spatial distance is reduced; at the same time, considering the characteristics of remotely sensed ground features changing over time, the normalized time difference is... By incorporating samples into the same attenuation model, the propagation impact is reduced for samples with large time differences. Secondly, to ensure that the propagation process conforms to the consistency of ground features, cosine similarity between multidimensional feature vectors of samples is introduced. The labels are then converted into non-negative weighting factors through normalization to enhance the label transfer capability between samples with similar spectral, texture, and semantic features. Finally, to address the error accumulation problem during multiple propagations of weak labels, a propagation path length is introduced. The uncertainty caused by long-distance multi-hop propagation is limited by a decreasing function.

[0040] Furthermore, for each unlabeled sample Iterate through all labeled samples Calculate the label propagation weight The propagation weights of tags belonging to the same category are accumulated to obtain the category propagation strength, calculated using the following formula: , in, Indicates category The intensity of transmission; This represents a set of categories, determined based on the category labels in the labeled samples. Indicates labeled samples The category. Select to make Largest category as a candidate weak label Samples without labeled numbers As the weakly labeled sample number, and as the first One remote sensing sample to be expanded, and the corresponding category at this time. Below are the labeled samples. For unlabeled samples The label propagation weight is denoted as The labeled samples are used as one of the input variables for subsequent credibility calculations. Indicates supported categories This step involves selecting any one of the labeled samples from all sources. The technical objective of this step is to fuse multi-source propagation information into a single-label output while retaining propagation strength information for subsequent evaluation.

[0041] S2. Based on the label propagation weights corresponding to the candidate weak labels, and combined with remote sensing sample units, construct a multi-source consistency-driven credibility evolution model to calculate the final credibility of the weak label samples; compare the final credibility with the set screening threshold, and add the weak label samples that meet the credibility requirements to the remote sensing sample library.

[0042] This is followed by the weak label credibility evolution stage. The problem to be solved in this stage is: how to determine the reliability of the above propagation results and prevent erroneous labels from entering the remote sensing sample database. Therefore, a multi-source consistency-driven credibility evolution model is constructed, which incorporates the label propagation weights from the previous stage. As one of the core inputs, it combines spatial consistency and temporal stability for unified calculation, and its expression is: , in, Indicates the first The final credibility of a weakly labeled sample; This indicates the number of labeled samples participating in the current confidence calculation, i.e., when... Maximum supported categories The total number of all labeled samples; Indicates the corresponding category Below are the labeled samples. For unlabeled samples Tag propagation weight; Indicates labeled samples and unlabeled samples Normalized spatial distance; Indicates the spatial distance attenuation factor; It is the spatial propagation consistency factor, which is the evaluation result of spatial consistency of propagation sources and is used to describe all support categories. The labeled sample pairs Spatial consistency of weakly labeled samples; Indicates the first The degree of temporal change of the weakly labeled sample, i.e., the... The degree of feature change of a weakly labeled sample is determined based on the multidimensional feature vector corresponding to that sample. Specifically, the multidimensional feature vectors corresponding to the same spatial location of the sample at different time points are obtained, and the feature differences between adjacent time points are accumulated to obtain the change of the sample in the time dimension, which is used as... ; It is a change sensitivity coefficient used to control the rate at which changes over time affect the final confidence level, and it measures the changes in all samples. Perform mean calculation and use the reciprocal of the mean as... ; This represents the time transformation decay factor, or time stability, which describes how time changes lead to a decrease in the final credibility. Indicates the first The number of times a weakly labeled sample maintains the same class status in multi-temporal remote sensing data, i.e., the number of times the weakly labeled sample maintains the same class status. The number of times a weakly labeled sample has remained in the same category throughout its history; It is a time normalization scale used to limit Unlimited growth leads to an excessive increase in the final credibility, which is set to 10 here; This represents the stability enhancement factor, or historical reliability, used to describe the enhancing effect of long-term consistency on the credibility of weak labels.

[0043] The formula is formed as follows: according to category Label propagation weight Weighted statistics were performed on all labeled samples, and a spatial distance attenuation factor was introduced. First, the spatial consistency evaluation results of the propagation source are obtained, which reflect the reliability of the spatial distribution of samples supporting the weak label. Second, the degree of variation of sample features is calculated based on the multidimensional feature vector. Using the exponential decay model The weakening effect of land cover changes on label credibility is characterized; finally, the number of times the sample maintains the same category status based on its history is analyzed. Using a logarithmic growth model This study characterizes the enhancing effect of long-term stable information on tag credibility. It integrates three influencing factors—spatial consistency evaluation results of the propagation source, temporal stability, and historical reliability—by multiplying and fusing them to form a comprehensive weak tag credibility evaluation formula. This enables dynamic quantitative assessment of the reliability of extended sample labels.

[0044] In obtaining the first Final credibility of weakly labeled samples Then, the final credibility will be determined. With preset filtering threshold A comparison is made to determine whether the weakly labeled sample should be expanded into the remote sensing sample library. At that time, the first Each weakly labeled sample and its corresponding weakly labeled category is written into the remote sensing sample database; when If the label is weak, the sample is discarded or retained in the candidate sample set for later re-evaluation. The selection threshold is... Based on the number of labeled samples of the same category in the current remote sensing sample database and the distribution of weak label confidence, specifically, the average confidence value of samples of the same category that have entered the remote sensing sample database is first calculated. and will The reciprocal of the value is used as the filtering threshold for the current category. This method enables automatic filtering of weak tags based on credibility constraints.

[0045] The data collected in S1 and S2, as well as the intermediate data obtained from calculation, are stored in the historical sample database.

[0046] In summary, a method for weak label expansion and filtering for constructing remote sensing sample libraries has been completed.

[0047] The order of the embodiments is for illustrative purposes only and does not represent the superiority or inferiority of the embodiments. The processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0048] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0049] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for weak label expansion and filtering for remote sensing sample database construction, characterized in that, Includes the following steps: S1. Acquire multi-source remote sensing data and existing manually labeled sample data. Perform spatial benchmark unification, grid division, and feature construction on the remote sensing image data in the multi-source remote sensing data. Combine with the existing manually labeled sample data to form remote sensing sample units containing spatial location coordinates, temporal information, multi-dimensional feature vectors, and label status. Divide the remote sensing sample units according to the label status to obtain a set of labeled samples and a set of unlabeled samples. Based on the remote sensing sample units, calculate the label propagation weight of labeled samples to unlabeled samples using a spatiotemporal coupled nonlinear attenuation propagation model. Based on the label propagation weight, accumulate the label propagation weights belonging to the same category to obtain the category propagation intensity. Based on the category propagation strength, candidate weak labels are generated for unlabeled samples, and the label propagation weights corresponding to the candidate weak labels are obtained; S2. Based on the label propagation weights corresponding to the candidate weak labels, and combined with remote sensing sample units, construct a multi-source consistency-driven credibility evolution model to calculate the final credibility of the weak label samples; compare the final credibility with the set screening threshold, and add the weak label samples that meet the credibility requirements to the remote sensing sample library.

2. The weak label expansion and screening method for remote sensing sample database construction according to claim 1, characterized in that, S1 specifically includes: In the spatiotemporal coupled nonlinear decay propagation model, based on the spatial coordinates of labeled and unlabeled samples, the normalized spatial distance between them is calculated using Euclidean distance and normalization. Based on temporal information, the difference in acquisition time between labeled and unlabeled samples is normalized to calculate the normalized time difference. Based on the multidimensional feature vectors of labeled and unlabeled samples, the angle between their multidimensional feature vectors is calculated using cosine similarity.

3. The weak label expansion and screening method for remote sensing sample database construction according to claim 2, characterized in that, S1 specifically includes: Based on normalized spatial distance and normalized time difference, the exponential term is calculated by extending the model through Gaussian radial basis function; based on the angle between the multidimensional feature vectors of labeled and unlabeled samples, the non-negative weight factor is calculated through cosine similarity calculation and normalization; the propagation path length is introduced, and the path penalty term is calculated through a decreasing function.

4. The weak label expansion and screening method for remote sensing sample database construction according to claim 3, characterized in that, S1 specifically includes: Multiplying the exponential term, the non-negative weight factor, and the path penalty term yields the label propagation weight of labeled samples on unlabeled samples. The calculation formula is as follows: in, Indicates labeled samples For unlabeled samples Tag propagation weight; Indicates labeled samples and unlabeled samples Normalized spatial distance; Indicates labeled samples and unlabeled samples Normalized time difference; Indicates the spatial attenuation coefficient; Indicates the time decay coefficient; It is an exponential term; Indicates labeled samples and unlabeled samples The angle between the multidimensional feature vectors; Indicates a non-negative weighting factor; Indicates the propagation path length; Indicates the path attenuation coefficient; It is a path penalty item.

5. The weak label expansion and screening method for remote sensing sample database construction according to claim 4, characterized in that, S1 specifically includes: Based on the category propagation strength, the category with the greatest category propagation strength is selected as the candidate weak label, and the corresponding unlabeled sample number is used as the weak label sample number; based on the label propagation weight, the label propagation weight of the labeled sample to the unlabeled sample corresponding to the candidate weak label is obtained.

6. The weak label expansion and screening method for remote sensing sample database construction according to claim 1, characterized in that, S2 specifically includes: In the multi-source consistency-driven credibility evolution model, the spatial distance decay factor is calculated based on the normalized spatial distance between labeled and unlabeled samples corresponding to candidate weak labels plus the reciprocal of 1; based on the multi-dimensional feature vector, the feature differences between adjacent time nodes of weak label samples are accumulated to obtain the temporal change degree of weak label samples.

7. The weak label expansion and screening method for remote sensing sample database construction according to claim 6, characterized in that, S2 specifically includes: Based on the spatial distance decay factor, the propagation weights of the candidate weak labels are weighted statistically to obtain the spatial propagation consistency factor; based on the degree of temporal change of the weak label samples, the time transformation decay factor is calculated through the exponential decay model.

8. The weak label expansion and screening method for remote sensing sample database construction according to claim 7, characterized in that, S2 specifically includes: The number of times a weakly labeled sample has maintained the same class status in history is introduced, and a stability enhancement factor is calculated using a logarithmic function. The spatial propagation consistency factor, the time transformation decay factor, and the stability enhancement factor are multiplied together to obtain the final credibility of the weakly labeled sample.

9. The weak label expansion and screening method for remote sensing sample database construction according to claim 8, characterized in that, S2 specifically includes: When the final confidence level is greater than or equal to the screening threshold, the weakly labeled sample and its corresponding weakly labeled category are written into the remote sensing sample library; when the final confidence level is less than the screening threshold, the weakly labeled sample is discarded or retained in the candidate sample set for subsequent re-evaluation.

Images