Ground cover remote sensing image training sample collection method and related equipment

Abstract

The present application belongs to the technical field of remote sensing image intelligent interpretation and machine learning, and relates to a land cover remote sensing image training sample collection method and related equipment, wherein the method comprises the following steps: according to a land cover category set, extracting a plurality of candidate features from a remote sensing image, calculating feature correlation, eliminating redundant features, and obtaining an optimal feature subset based on forward iteration selection of JM distance; obtaining an initial sample set by using land cover classification data; obtaining a robust center and a robust covariance structure of a target land class in a feature space by using robust global statistical modeling; performing heterogeneity measurement based on a regularized Mahalanobis distance on the initial sample set to obtain a regularized Mahalanobis distance; and performing denoising on the initial sample set by using an adaptive threshold based on an empirical distribution according to the regularized Mahalanobis distance to obtain a high-purity training sample set, so that the generation efficiency of land cover remote sensing image training sample data is improved, and the production cost and label noise problem are reduced.

Description

Technical Field

[0001] This invention relates to intelligent interpretation and machine learning technology of remote sensing images, and in particular to a method and related equipment for collecting training samples of land cover remote sensing images. Background Technology

[0002] Currently, remote sensing image classification mainly employs supervised methods such as deep learning. However, its performance heavily relies on massive amounts of high-quality training samples with accurate labels. Acquiring these samples (especially historical remote sensing image samples over long time series) is extremely costly, becoming a major bottleneck restricting the large-scale application of this technology.

[0003] Existing technologies and their problems include: cost and efficiency issues. Manual annotation is costly and almost impractical for remote sensing data covering tens of thousands of square kilometers with continuous time series. Even annotating only a small number of samples requires a significant investment of manpower and expertise. Furthermore, directly using existing products such as GlobeLand30 as training samples leads to severe label noise problems, mainly in the following two categories: (1) Inherent classification error: It is difficult for any automatic classification product to achieve 100% accuracy, especially in boundary areas, ecotones and mixed pixel areas, where labeling errors are more common; (2) Temporal misalignment error: When the production year of the existing product (such as 2000 / 2010 / 2020 of GlobeLand30) is inconsistent with the acquisition year of the image to be classified (such as other years between 2000 and 2025), the actual changes in land features will cause the label to fail. For example, an area in the 2015 image was marked as farmland in the 2010 version of GlobeLand30, but it was actually converted from farmland to forest in 2015. In this case, the label is incorrect.

[0004] Feature redundancy and insufficient separability: Remote sensing images can extract dozens of spectral and texture features, but existing methods either use a single or a few empirical indices (such as NDVI) or use all features without selection. This leads to feature redundancy and poor inter-class separability. Feature redundancy can cause strong correlations between multiple features, resulting in high computational cost and model instability. Poor inter-class separability means that fixed feature combinations are difficult to achieve optimal differentiation on all surface pairs.

[0005] The method has poor robustness. Active / semi-supervised learning still requires a high-quality initial sample set. In complex scenarios with multiple classes and extreme class imbalance (such as the coexistence of large water bodies and small artificial land cover), it is easy to get trapped in local optima. Mislabels will spread and amplify with iteration, making it difficult to obtain a stable and reliable final sample set.

[0006] Lacking a global statistical perspective, existing simple filtering methods (such as removing boundary pixels and small-area patches) only utilize local spatial information and fail to identify statistically abnormal samples from the perspective of global feature spatial distribution, resulting in a large number of spectrally abnormal but spatially normal noise samples that cannot be effectively removed. Summary of the Invention

[0007] In order to at least solve one of the technical problems existing in the prior art, the present invention provides a method and related equipment for collecting training samples of land cover remote sensing images.

[0008] One aspect of the present invention provides a method for acquiring training samples from land cover remote sensing images, comprising: In response to a classification request, determine the set of categories for land cover classification; Based on the land cover classification category set, multiple candidate features for each target land cover category are extracted from remote sensing images, a candidate feature pool is constructed and feature correlation is calculated, and redundant features are eliminated based on correlation; a forward iterative feature selection strategy based on JM distance is adopted to obtain the optimal feature subset. An initial sample set is obtained based on the pre-defined land cover classification data for the target area; Robust global statistics are applied to the target land class in the initial sample set to obtain the robust center and robust covariance structure of the target land class in the feature space; Based on the robust center and robust covariance structure, the heterogeneity of the initial sample set is measured by regularized Mahalanobis distance to obtain the regularized Mahalanobis distance. Based on the regularized Mahalanobis distance, the initial sample set is denoised using an adaptive threshold based on the empirical distribution to obtain a high-purity training sample set.

[0009] According to the land cover remote sensing image training sample acquisition method, multiple candidate features for each target land cover category are extracted from the remote sensing image based on the land cover classification category set. A candidate feature pool is constructed, and feature correlation is calculated. Redundant features are removed based on correlation. A forward iterative feature selection strategy based on JM distance is adopted to obtain the optimal feature subset, including: Based on the set of land cover classification categories, at least one of multiple spectral features and multiple texture features for each target land cover type is extracted from remote sensing images to obtain a candidate feature pool. For any two candidate features in the candidate feature pool, the Pearson correlation coefficient r is used to generate a correlation matrix. The correlation matrix includes redundant feature pairs. Redundant feature elimination is obtained by retaining the features with greater median difference and less inter-class overlap based on the distribution of the two candidate features in different types of box plots. A forward iterative feature selection strategy based on JM distance is adopted, which involves starting from an empty set and adding features that maximize the minimum JM distance of all class pairs until the minimum JM distance reaches a preset threshold, thus obtaining the optimal feature subset.

[0010] According to the land cover remote sensing image training sample acquisition method, the initial sample set is obtained based on the target area using preset land cover classification data, including: Obtain preset land cover classification data based on the target area and surface remote sensing images, and generate a binary mask of the preset land cover classification data based on the target land type. The binary mask is overlaid on the surface remote sensing image to extract the feature vector of the pixels including the target land type; An initial sample set is generated based on the feature vectors of the pixels of the target land type and the initial labels of the preset land cover classification data.

[0011] According to the land cover remote sensing image training sample acquisition method, the target land cover types in the initial sample set are processed using robust global statistics to obtain the robust centers and robust covariance structures of the target land cover types in the feature space, including: Determine the median vector using the median of each feature dimension in the initial sample set, and perform the following iterative processing using the median vector as the initial robust center: Calculate the Euclidean distance from each sample to the current mean robust center, and then calculate the robust scaling estimate based on the Euclidean distance. for: ; ; in, This indicates the calculation of the median. Indicates the first sample in the initial sample set. One sample, Indicates the first The sample at the th The Euclidean distance to the mean robust center at the next iteration. Indicates the first The mean robust center at the next iteration Used for robust scaling estimation It becomes an unbiased estimator of the standard deviation; Based on robust scaling estimation Calculate Tukey weights for: ; in, It is a constant; Based on Tukey weights The updated mean robust center is: ; If the iteration satisfies If the iteration stops, then the iteration stops. To preset the convergence accuracy; Furthermore, it also includes regularizing the initial sample set using Ledoit-Wolf contraction estimation to obtain a robust covariance structure. for: ; in, It is the covariance matrix calculated based on weighted samples. It is the dimension of the feature space. It is the trace of the matrix. It is the identity matrix. It is the shrinkage parameter.

[0012] According to the aforementioned method for collecting training samples of land cover remote sensing images, the initial sample set is subjected to a heterogeneity measure based on regularized Mahalanobis distance, based on robust centers and robust covariance structures, to obtain the regularized Mahalanobis distance, including: Calculate the regularized Mahalanobis distance of the robust center for each sample in the initial sample set. for: ; Among them, the regularized Mahalanobis distance Used to characterize the probability that a sample is labeled noise.

[0013] According to the aforementioned method for acquiring training samples of land cover remote sensing images, the initial sample set is denoised using an adaptive threshold based on empirical distribution according to regularized Mahalanobis distance to obtain a high-purity training sample set, including: Calculate the lower quartile of the regularized Mahalanobis distance for all samples in the initial sample set. Upper quartiles and interquartile range Interquartile range ; The initial sample set is processed by removing samples based on a removal threshold and regularized Mahalanobis distance to obtain a high-purity training sample set. The removal threshold is... The calculation method is as follows: ; in, To control the stringency of noise reduction.

[0014] According to the aforementioned method for collecting training samples of land cover remote sensing images, the method further includes: The high-purity training sample set is preprocessed and augmented, including randomly cropping the high-purity training sample set after cleaning each target land type to obtain sample blocks of fixed size, and then performing random rotation, horizontal and vertical flipping and brightness scaling on the sample blocks to obtain the final surface remote sensing image training sample set.

[0015] Another aspect of the present invention provides a land cover remote sensing image training sample acquisition device, comprising: The first module is used to determine the set of categories for land cover classification based on the classification request; The second module is used to extract multiple candidate features for each target land cover category from remote sensing images based on the land cover classification category set, construct a candidate feature pool and calculate feature correlation, and remove redundant features based on correlation; a forward iterative feature selection strategy based on JM distance is adopted to obtain the optimal feature subset. The third module is used to obtain an initial sample set based on the preset land cover classification data of the target area; The fourth module is used to process the target land class in the initial sample set using robust global statistics to obtain the robust center and robust covariance structure of the target land class in the feature space. The fifth module is used to measure the heterogeneity of the initial sample set based on the robust center and robust covariance structure, and obtain the regularized Mahalanobis distance. The sixth module is used to denoise the initial sample set using an adaptive threshold based on an empirical distribution, based on the regularized Mahalanobis distance, to obtain a high-purity training sample set.

[0016] Another aspect of the present invention provides an electronic device, including a processor and a memory; The memory is used to store programs; The processor executes the program to implement the method as described above.

[0017] This invention also discloses a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device can read the computer instructions from the computer-readable storage medium and execute the computer instructions, causing the computer device to perform the methods described above.

[0018] The beneficial effects of this invention are as follows: the entire process requires no human-computer interaction or manual annotation, and is fully automated from initial sample acquisition to noise cleaning, easily handling massive amounts of historical remote sensing data; it employs robust statistical methods (IRLS, Mahalanobis distance) instead of simple local filtering, effectively identifying and eliminating label noise caused by product errors and time misalignments; both the parameters for robust mean estimation (Tukey function) and the final elimination threshold (Tukey fence method) are adaptively calculated based on the data distribution of the current image and the current category, avoiding the subjectivity and incompatibility caused by manually setting parameters; the binary classification strategy decomposes multi-class problems into multiple binary problems, effectively mitigating the interference of class imbalance and spectral overlap, allowing the GSH detector to work more accurately in each binary space. The embodiments of this invention have been applied to images from six different years and seasons, demonstrating its strong cross-condition robustness. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the land cover remote sensing image training sample acquisition process according to an embodiment of the present invention; Figure 2 This is a flowchart of an automatic acquisition method for training samples of remote sensing images based on feature optimization and global statistical heterogeneity detection according to an embodiment of the present invention. Figure 3 This is a correlation diagram between features in an embodiment of the present invention; Figure 4 This is a comparison chart of JM distance analysis for stepwise feature selection under a relevance threshold of R=0.7 according to an embodiment of the present invention, wherein (a) is the minimum JM distance and step gain for each category under a relevance threshold of R=0.7; (b) is the step JM distance distribution for all target categories under a relevance threshold of R=0.7; and (c) is the JM distance from the target category to other categories in the final feature selection under a relevance threshold of R=0.7. Figure 5 This is a comparison chart of the JM distance analysis for stepwise feature selection under a relevance threshold of R=0.85 according to an embodiment of the present invention. (a) shows the minimum JM distance and step gain for each category under a relevance threshold of R=0.85; (b) shows the step JM distance distribution for all target categories under a relevance threshold of R=0.85; and (c) shows the JM distance from the target category to other categories in the final feature selection under a relevance threshold of R=0.85. Figure 6This is a comparison chart of the JM distance analysis for stepwise feature selection under a relevance threshold of R=0.9 according to an embodiment of the present invention. (a) shows the minimum JM distance and step gain for each category under a relevance threshold of R=0.9; (b) shows the step JM distance distribution for all target categories under a relevance threshold of R=0.9; and (c) shows the JM distance from the target category to other categories in the final feature selection under a relevance threshold of R=0.9. Figure 7 This is a training sample collection result map of land features at different times according to an embodiment of the present invention, including the initial training samples of water bodies, vegetation, artificial land surface (CL) and cultivated land (AS) in a certain area. Figure 8This invention presents the impact of different values ​​of parameter k on the number of refined samples for four land types, the distribution of RMD when the optimal k value is taken, and the accuracy assessment of sample collection results. (a1), (b1), and (c1) respectively represent the curves showing the change in the number and abandonment rate of water samples with the k value (IQR multiplier) (optimal k=1.00), the histogram of the frequency distribution of regularized Mahalanobis distance (RMD) of water samples, the screening threshold (threshold=4.379), and the bar chart of water classification performance evaluation indicators (including precision, recall, and F1 score). Score), Overall Precision (OA), Intersection over Union (IoU); (a2), (b2), and (c2) respectively represent the curves of the number and abandonment rate of vegetation samples as a function of the k value (IQR multiplier) (optimal k=1.30), the histogram of the frequency distribution of the regularized Mahalanobis distance (RMD) of vegetation samples, the screening threshold (threshold=4.350), and the bar chart of vegetation classification performance evaluation indicators (including precision, recall, F1 score). Score), Overall Precision (OA), Intersection over Union (IoU); (a3), (b3), and (c3) respectively represent the curves of the number and abandonment rate of abandoned farmland samples as a function of the k value (IQR multiplier) (optimal k=1.30), the histogram of the frequency distribution of the regularized Mahalanobis distance (RMD) of farmland samples, the screening threshold (threshold=4.719), and the bar chart of farmland classification performance evaluation indicators (including precision, recall, F1 score). Score), Overall Precision (OA), Intersection over Union (IoU)); (a4), (b4) and (c4) respectively represent the curves of the number and abandonment rate of artificial surface samples as a function of k value (IQR multiplier) (optimal k=1.40), the histogram of the frequency distribution of regularized Mahalanobis distance (RMD) of artificial surface samples and the screening threshold (threshold=4.350), and the bar chart of artificial surface classification performance evaluation indicators (including precision, recall, F1 score, overall precision (OA), intersection over union (IoU)). Figure 9 These are water body overall accuracy and loss curves according to an embodiment of the present invention, wherein (a) is a water body overall accuracy curve and (b) is a water body loss curve. Figure 10 These are forest overall accuracy and loss curves according to an embodiment of the present invention, wherein (a) is a forest overall accuracy map and (b) is a forest loss curve map; Figure 11 These are the overall accuracy and loss curves of cultivated land according to an embodiment of the present invention, wherein (a) is the overall accuracy of cultivated land and (b) is the loss curve of cultivated land; Figure 12These are the overall accuracy and loss curves of the artificial land surface according to an embodiment of the present invention, wherein (a) is the overall accuracy map of the artificial land surface and (b) is the loss curve of the artificial land surface; Figure 13 This is a schematic diagram of a land cover remote sensing image training sample acquisition device according to an embodiment of the present invention. Detailed Implementation

[0020] The embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings. Throughout the description, the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions. In the following description, suffixes such as "module," "part," or "unit" used to denote elements are used only for the purpose of illustrative purposes and have no specific meaning in themselves. Therefore, "module," "part," or "unit" can be used interchangeably. Terms such as "first," "second," etc., are used only to distinguish technical features and should not be construed as indicating or implying relative importance, or implicitly indicating the number of indicated technical features, or implicitly indicating the sequential relationship of the indicated technical features. In the following description, the consecutive reference numerals for method steps are for ease of review and understanding. Adjusting the implementation order of steps, in conjunction with the overall technical solution of the present invention and the logical relationship between the various steps, will not affect the technical effect achieved by the technical solution of the present invention. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0021] Please combine Figure 1 and Figure 2 ,in Figure 1 This is a schematic diagram of the process for collecting training samples for land cover remote sensing images, including but not limited to steps S100~S600: S100, in response to the classification request, determines the set of categories for land cover classification.

[0022] In some embodiments, a target classification system is determined based on application requirements, and a set of categories for land cover classification is determined (determining the target classification system), such as water bodies, vegetation, cultivated land, artificial land, bare land, etc.

[0023] S200 extracts multiple candidate features for each target land cover category from remote sensing images based on the land cover classification category set, constructs a candidate feature pool and calculates feature correlation, and removes redundant features based on correlation; and adopts a forward iterative feature selection strategy based on JM distance to obtain the optimal feature subset.

[0024] In some embodiments, based on the set of land cover classification categories, at least one of multiple spectral features and multiple texture features for each target land cover type is extracted from remote sensing images to obtain a candidate feature pool. A correlation matrix is ​​generated by applying the Pearson correlation coefficient r to any two candidate features in the candidate feature pool. The correlation matrix includes redundant feature pairs. Redundant feature removal is achieved by retaining features with larger median differences and less inter-class overlap, based on the distribution of the two candidate features in different types of box plots. A forward iterative feature selection strategy based on JM distance is adopted, starting from an empty set and adding features that maximize the minimum JM distance of all class pairs each time, until the minimum JM distance reaches a preset threshold, thus obtaining the optimal feature subset.

[0025] It should be noted that each target land category is obtained through the above embodiment. Determine an optimal feature subspace This feature subspace will be used for subsequent initial sample extraction and GSH detection.

[0026] For example, a candidate feature pool is constructed to extract multiple candidate features from remote sensing images, referencing Figure 3 Specifically, it includes: Spectral characteristics: Enhanced Normalized Differential Impermeable Surface Index (ENDII), Normalized Differential Water Index (NDWI), Normalized Differential Building Index (NDBI), Normalized Differential Soil Index (NDSI), Near Infrared Reflectance Vegetation Index (NIRV), Normalized Differential Vegetation Index (NDVI), Normalized Differential Bare Land Index (NDBLI), Multiband Combined Water Index (MBCWI), etc.

[0027] Texture features: correlation coefficient (CORR), angular second moment (ASM) calculated based on gray-level co-occurrence matrix (GLCM), contrast (CON), inverse moment (IDM), mean (Mean), entropy (ENT), heterogeneity (DISS), variance (VAR), and texture mean (TM).

[0028] For example, quantitatively assessing the correlation between features includes calculating the Pearson correlation coefficient r between all candidate features, generating a correlation matrix, and identifying highly correlated feature pairs. For each pair of highly correlated features, they are retained or removed based on their inter-class distinguishing ability across different land cover types: the box plot distributions of the two can be compared across different types, and features with greater median differences and less inter-class overlap are retained.

[0029] For example, the forward iterative selection based on the Jeffries-Matusita (JM) distance uses the Jeffries-Matusita (JM) distance to quantitatively evaluate the inter-class separability of feature combinations. The JM distance ranges from 0 to 2, with a larger value indicating that the two classes are easier to distinguish (>1.8 is considered excellent). Specifically, the forward iterative selection strategy adopted by the JM distance is as follows: start from an empty set; add a feature at each step that maximizes the minimum JM distance improvement for all class pairs; the iteration terminates when the minimum JM distance reaches a preset threshold (e.g., 1.9) or when adding more features no longer significantly improves the result.

[0030] S300: Obtain an initial sample set based on preset land cover classification data for the target area.

[0031] In some embodiments, preset land cover classification data is obtained based on the target area; a binary mask of the target land class is generated based on the preset land cover classification data; the binary mask is overlaid with the land remote sensing image to extract feature vectors including the target land class; and an initial sample set is generated based on the feature vectors of the target land class and the initial labels of the preset land cover classification data.

[0032] It is understood that the preset land cover classification data in this embodiment already includes land cover classification products such as GlobeLand30 and ESACCI-LC.

[0033] For example, for a target classification system that includes For each target land category The steps for obtaining the initial sample set are as follows: (1) Generate a binary mask based on the land cover classification product. Among them, The pixel is labeled 1, and all other pixels (as background) are labeled 0. The target classification system includes... Land category: .

[0034] (2) Overlay the mask onto the remote sensing image and extract the feature vectors corresponding to all pixels with a mask value of 1.

[0035] (3) The feature vector may include: spectral features (such as pixel values ​​of each band and band combination index values), texture features (such as second-order statistics of gray-level co-occurrence matrix), etc.

[0036] (4) Each sample contains its feature vector. An initial label from a land cover product This constitutes the initial sample set. .

[0037] It is understandable that the initial sample set acquisition step is performed independently for each land type, and there is no sequential dependency between the initial sample acquisition of different land types. The sample refinement result of any land type will not affect the extraction of initial samples of other land types.

[0038] It should be noted that in this embodiment, steps S400 to S600 involve refining the initial samples for each land type using a Global Statistical Heterogeneity Detector (GSH detector). S400 uses robust global statistics to process the target land class in the initial sample set, obtaining the robust center and robust covariance structure of the target land class in the feature space.

[0039] In some embodiments, the median vector is determined using the median of each feature dimension in the initial sample set, i.e., let... Using the median vector of the sample set (the median for each feature dimension) as the initial robust center, perform the following iterative process: Calculate the Euclidean distance from each sample to the current mean robust center, and then calculate the robust scaling estimate based on the Euclidean distance. for: ; ; in, This indicates the calculation of the median. Indicates the first sample in the initial sample set. One sample, Indicates the first The sample at the th The Euclidean distance to the mean robust center at the next iteration. Indicates the first The mean robust center at the next iteration Used for robust scaling estimation It becomes an unbiased estimator of the standard deviation; Based on robust scaling estimation Calculate Tukey weights for: ; in, It is a constant; Based on Tukey weights The updated mean robust center is: ; If the iteration satisfies If the iteration stops, then the iteration stops. To preset the convergence accuracy (e.g.) ).

[0040] It should be noted that this embodiment estimates the target land type. The robust centers and robust covariance structure in the feature space ensure that the estimation results are not affected by mislabeled (outlier) points in the initial sample set.

[0041] This embodiment uses the Iterative Reweighted Least Squares (IRLS) algorithm combined with Tukey's Biweight function. The core idea of ​​this algorithm is to assign different weights to samples based on their distance from the current mean in each iteration. Samples that are farther away (more likely to be noise) have lower weights, while samples that are very close have weights close to 1. The mean obtained after the iteration converges is the robust estimate.

[0042] Furthermore, it also includes regularizing the initial sample set using Ledoit-Wolf contraction estimation to obtain a robust covariance structure. for: ; in, It is the covariance matrix calculated based on weighted samples. It is the dimension of the feature space. It is the trace of the matrix. It is the identity matrix. It is the shrinkage parameter.

[0043] It is understandable that directly calculating the sample covariance matrix may be unstable in high-dimensional, small-sample cases, and the presence of extreme outliers can lead to covariance estimation inflation. Therefore, this embodiment uses a regularized covariance matrix. The estimation is regularized using the Ledoit-Wolf contraction estimation method described above, which ensures the invertibility and numerical stability of the covariance matrix, making the subsequent Mahalanobis distance calculation reliable in high-dimensional feature spaces.

[0044] S500, based on robust centers and robust covariance structures, performs a heterogeneity measure on the initial sample set based on regularized Mahalanobis distance to obtain the regularized Mahalanobis distance.

[0045] In some embodiments, the regularized Mahalanobis distance of the robust center is calculated for each sample in the initial sample set. for: ; Among them, the regularized Mahalanobis distance Used to characterize the probability that a sample is labeled noise.

[0046] It's important to note that RMD represents the "abnormality" of a sample after considering feature correlation and scale. If a sample is a "pure" target land cover sample, it will be close to the robust center, resulting in a smaller RMD value; if a sample is mislabeled (e.g., incorrectly labeling a pixel that should be farmland as water), it will be far from the robust center of the water land cover in the feature space, resulting in a larger RMD value. The magnitude of the RMD value directly reflects the probability that the sample is labeled noise.

[0047] S600, based on the regularized Mahalanobis distance, uses an adaptive threshold based on the empirical distribution to denoise the initial sample set, resulting in a high-purity training sample set.

[0048] The adaptive threshold based on empirical distribution in this embodiment adopts Tukey's fence method.

[0049] In some embodiments, the adaptive threshold based on the empirical distribution includes calculating the lower quartile of the regularized Mahalanobis distance for all samples in the initial sample set. Upper quartiles and interquartile range Interquartile range ; The initial sample set is processed by removing samples based on a removal threshold and regularized Mahalanobis distance to obtain a high-purity training sample set. The removal threshold is... The calculation method is as follows: ; in, To control the stringency of noise reduction, such as Moderately stringent (identifies mild outliers). This is very stringent (only identifying extreme outliers). The invention is set to... ,and It can be flexibly adjusted according to the actual data distribution.

[0050] In another embodiment, the sample output and classification application includes randomly cropping the cleaned sample set for each land type to generate fixed-size sample blocks, and applying data augmentation processing (including random rotation, horizontal and vertical flipping, brightness scaling, etc.) to finally construct a complete training dataset. This dataset can be directly used to train any classifier, including deep learning models such as support vector machines, random forests, fully convolutional networks (FCN), and U-Net. Since the samples have been efficiently cleaned by the GSH detector, no additional sample cleaning or active learning iteration is required during model training, enabling one-step training and high-precision convergence.

[0051] To verify the effectiveness of the automatic training sample refinement method based on Global Statistical Heterogeneity Detection (GSH) proposed in this invention, a wetland area was selected as the study area. A fully automated remote sensing imagery acquisition experiment was conducted using Landsat satellite imagery from six time periods (2000, 2005, 2010, 2015, 2020, and 2025, including Landsat 5 TM, Landsat 7 ETM+, and Landsat 8 OLI, four scenes per period, totaling 24 scenes). This data was then used to train the MS-FCN classifier. The specific experimental setup, methods, and results are as follows.

[0052] refer to Figure 3 , Figure 3 This is a correlation diagram of features in an embodiment of the present invention. From 16 candidate features, highly correlated features were eliminated through correlation analysis, and then the optimal features for each land type were selected through forward iteration using JM distance. The correlation and separability analyses are as follows: Figures 4-6 As shown, where Figures 4-6 The JM distance analysis of stepwise feature selection under different relevance thresholds (R=0.7, 0.85, 0.9) is presented sequentially. For each threshold, three subplots are shown: (a) the minimum JM distance and step gain for each category; (b) the distribution of step JM distances for all target categories; and (c) the JM distances from the target category to other categories in the final feature selection. The optimal feature selection results are shown in Table 1.

[0053] Table 1. Optimized feature set for each land cover type and JM distance to other types.

[0054] In the sample collection section, this embodiment of the invention first utilizes the GlobeLand30 land cover classification product (three phases: 2000, 2010, and 2020) to acquire initial training samples of water bodies, vegetation, artificial land surfaces, and cultivated land in a specific wetland area, based on the principle of proximity at each time point. Then, using the initial sample set as a mask for remote sensing images, and based on the optimal feature sets for each category (see Table 1), feature images are calculated to construct feature vectors. Finally, a GSH detector is used to automatically remove noisy samples, resulting in a sample dataset for each category, as shown in Table 1. Figure 7 As shown.

[0055] The impact of different values ​​of parameter k for the four land types on the number of refined samples, the distribution of RMD when the optimal k value is taken, and the accuracy assessment of the sample collection results are as follows: Figure 8As shown, (a1), (b1), and (c1) represent, respectively, the curves of the number and abandonment rate of abandoned water samples as a function of the k value (IQR multiplier) (optimal k=1.00), the histogram of the frequency distribution of the regularized Mahalanobis distance (RMD) of water samples, the screening threshold (threshold=4.379), and the bar chart of water classification performance evaluation indicators (including precision, recall, F1 score, overall accuracy (OA), and intersection-over-union ratio (IoU)); (a2), (b2), and (c2) represent, respectively, the curves of the number and abandonment rate of abandoned vegetation samples as a function of the k value (IQR multiplier) (optimal k=1.30), the histogram of the frequency distribution of the regularized Mahalanobis distance (RMD) of vegetation samples, the screening threshold (threshold=4.350), and the bar chart of vegetation classification performance evaluation indicators (including precision, recall, F1 score, and intersection-over-union ratio (IoU)). Score), Overall Precision (OA), Intersection over Union (IoU); (a3), (b3), and (c3) respectively represent the curves of the number and abandonment rate of abandoned farmland samples as a function of the k value (IQR multiplier) (optimal k=1.30), the histogram of the frequency distribution of the regularized Mahalanobis distance (RMD) of farmland samples, the screening threshold (threshold=4.719), and the bar chart of farmland classification performance evaluation indicators (including precision, recall, F1 score). Score), Overall Precision (OA), Intersection over Union (IoU); (a4), (b4), and (c4) respectively represent the curves of the number and abandonment rate of abandoned artificial surface samples as a function of the k value (IQR multiplier) (optimal k=1.40), the histogram of the frequency distribution of the regularized Mahalanobis distance (RMD) of artificial surface samples, the screening threshold (threshold=4.350), and the bar chart of the classification performance evaluation indicators of artificial surface (including precision, recall, F1 score, overall precision (OA), and intersection over union (IoU)). Figure 8 It can be seen that the optimal k-value and RMD threshold selection vary for different land types. In this experiment, the optimal k-value for water bodies was 1.00 and the RMD threshold was 4.379; for vegetation, the optimal k-value was 1.30 and the RMD threshold was 4.350; for cultivated land, the optimal k-value was 1.30 and the RMD threshold was 4.719; and for artificial surfaces, the optimal k-value was 1.40 and the RMD threshold was 5.110. One advantage of this invention is its ability to adaptively determine the k-value and RMD threshold based on the actual distribution characteristics of the sampled data. Finally, the embodiments of this invention verified the adaptive sampling accuracy using real ground data. The overall accuracy for water bodies and vegetation was relatively high, while that for cultivated land and artificially covered cultivated land was relatively low, but the misclassification was zero, which is precisely the core issue that sample collection focuses on. Figure 9These are water body overall accuracy and loss curves according to an embodiment of the present invention, wherein (a) is a water body overall accuracy curve and (b) is a water body loss curve. Figure 10 These are forest overall accuracy and loss curves according to an embodiment of the present invention, wherein (a) is a forest overall accuracy map and (b) is a forest loss curve map; Figure 11 These are the overall accuracy and loss curves of cultivated land according to an embodiment of the present invention, wherein (a) is the overall accuracy of cultivated land and (b) is the loss curve of cultivated land; Figure 12 These are overall accuracy and loss curves of the artificial land surface according to an embodiment of the present invention, wherein (a) is an overall accuracy map of the artificial land surface, and (b) is a loss curve of the artificial land surface. (Refer to...) Figures 9-12 (Automatic removal of noisy samples using a GSH detector) Figure 7 (2005) It can be determined that the classification accuracy of the six phases of images obtained exceeds 90% for each type of land cover.

[0056] Figure 13 This is a schematic diagram of a land cover remote sensing image training sample acquisition device according to an embodiment of the present invention. The device includes a first module 1410, a second module 1420, a third module 1430, a fourth module 1440, a fifth module 1450, and a sixth module 1460.

[0057] The system comprises the following modules: The first module determines the set of land cover classification categories based on the classification request; the second module extracts multiple candidate features for each target land cover category from remote sensing imagery based on the land cover classification category set, constructs a candidate feature pool, calculates feature correlation, and removes redundant features based on correlation; a forward iterative feature selection strategy based on JM distance is used to obtain the optimal feature subset; the third module obtains an initial sample set based on preset land cover classification data for the target area; the fourth module processes the target land cover categories in the initial sample set using robust global statistics to obtain robust centers and robust covariance structures of the target land cover categories in the feature space; the fifth module performs a heterogeneity measurement based on regularized Mahalanobis distance on the initial sample set based on the robust centers and robust covariance structures to obtain the regularized Mahalanobis distance; and the sixth module performs denoising on the initial sample set using an adaptive threshold based on empirical distribution based on the regularized Mahalanobis distance to obtain a high-purity training sample set.

[0058] Exemplarily, with the cooperation of the first, second, third, fourth, fifth, and sixth modules in the device, the embodiment device can implement any of the aforementioned methods for acquiring training samples of land cover remote sensing images. Specifically, in response to a classification request, it determines the set of land cover classification categories; based on the set of land cover classification categories, it extracts multiple candidate features for each target land cover category from the remote sensing image, constructs a candidate feature pool, calculates feature correlation, and removes redundant features based on correlation; it employs a forward iterative feature selection strategy based on JM distance to obtain the optimal feature subset; it acquires an initial sample set based on preset land cover classification data for the target area; it processes the target land cover categories in the initial sample set using robust global statistics to obtain robust centers and robust covariance structures of the target land cover categories in the feature space; based on the robust centers and robust covariance structures, it performs a heterogeneity measurement based on regularized Mahalanobis distance on the initial sample set to obtain regularized Mahalanobis distance; and based on the regularized Mahalanobis distance, it uses an adaptive threshold based on empirical distribution to denoise the initial sample set, obtaining a high-purity training sample set. The beneficial effects of this invention are as follows: the entire process requires no human-computer interaction or manual annotation, and is fully automated from initial sample acquisition to noise cleaning, easily handling massive amounts of historical remote sensing data; it employs robust statistical methods (IRLS, Mahalanobis distance) instead of simple local filtering, effectively identifying and eliminating label noise caused by product errors and time misalignments; both the parameters for robust mean estimation (Tukey function) and the final elimination threshold (Tukey fence method) are adaptively calculated based on the data distribution of the current image and the current category, avoiding the subjectivity and incompatibility caused by manually setting parameters; the binary classification strategy decomposes multi-class problems into multiple binary problems, effectively mitigating the interference of class imbalance and spectral overlap, allowing the GSH detector to work more accurately in each binary space. The embodiments of this invention have been applied to images from six different years and seasons, demonstrating its strong cross-condition robustness.

[0059] This invention also provides an electronic device, which includes a processor and a memory; The memory stores the program; The processor executes a program to perform the aforementioned method for acquiring training samples of land cover remote sensing images; the electronic device has the function of carrying and running the software system for acquiring training samples of land cover remote sensing images provided in the embodiments of the present invention, such as a personal computer, minicomputer, main frame, workstation, network or distributed computing environment, standalone or integrated computer platform, or communicating with charged particle tools or other imaging devices, etc.

[0060] This invention also provides a computer-readable storage medium storing a program that is executed by a processor to implement the land cover remote sensing image training sample acquisition method described above.

[0061] In some alternative embodiments, the functions / operations mentioned in the block diagrams may not occur in the order shown in the operation diagrams. For example, depending on the functions / operations involved, two consecutively shown blocks may actually be executed substantially simultaneously, or the blocks may sometimes be executed in reverse order. Furthermore, the embodiments presented and described in the flowcharts of this invention are provided by way of example to provide a more comprehensive understanding of the technology. The disclosed methods are not limited to the operations and logic flows presented in the embodiments of this invention. Alternative embodiments are contemplated, in which the order of various operations is changed and sub-operations described as part of a larger operation are executed independently.

[0062] This invention also discloses a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device can read the computer instructions from the computer-readable storage medium and execute the computer instructions, causing the computer device to perform the aforementioned method for acquiring training samples of land cover remote sensing images.

[0063] Furthermore, although the invention has been described in the context of functional modules, it should be understood that, unless otherwise stated, one or more of the described functions and / or features may be integrated into a single physical device and / or software module, or one or more functions and / or features may be implemented in a separate physical device or software module. It is also understood that a detailed discussion of the actual implementation of each module is unnecessary for understanding the invention. Rather, considering the properties, functions, and internal relationships of the various functional modules in the apparatus disclosed in the embodiments of the invention, the actual implementation of the module will be understood within the scope of conventional skill of an engineer. Therefore, those skilled in the art can implement the invention as set forth in the claims using ordinary techniques without excessive experimentation. It is also understood that the specific concepts disclosed are merely illustrative and are not intended to limit the scope of the invention, which is determined by the full scope of the appended claims and their equivalents.

[0064] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0065] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can include, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0066] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0067] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0068] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0069] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

[0070] The above is a detailed description of the preferred embodiments of the present invention, but the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention, and these equivalent modifications or substitutions are all included within the scope defined by the claims of this application.

Claims

1. A method for collecting training samples from remote sensing images of land cover, characterized in that, include: In response to a classification request, determine the set of categories for land cover classification; Based on the land cover classification category set, multiple candidate features for each target land cover category are extracted from remote sensing images, a candidate feature pool is constructed and feature correlation is calculated, and redundant features are eliminated based on correlation; a forward iterative feature selection strategy based on JM distance is adopted to obtain the optimal feature subset. An initial sample set is obtained based on the pre-defined land cover classification data for the target area; Robust global statistics are applied to the target land class in the initial sample set to obtain the robust center and robust covariance structure of the target land class in the feature space; Based on the robust center and robust covariance structure, the heterogeneity of the initial sample set is measured by regularized Mahalanobis distance to obtain the regularized Mahalanobis distance. Based on the regularized Mahalanobis distance, the initial sample set is denoised using an adaptive threshold based on the empirical distribution to obtain a high-purity training sample set. The process of obtaining an initial sample set based on preset land cover classification data for the target area includes: Obtain preset land cover classification data based on the target area and surface remote sensing images, and generate a binary mask of the preset land cover classification data based on the target land type. The binary mask is overlaid on the surface remote sensing image to extract the feature vector of the pixels including the target land type; An initial sample set is generated based on the feature vectors of the pixels of the target land type and the initial labels of the preset land cover classification data; The process of applying robust global statistics to the target land class in the initial sample set to obtain the robust center and robust covariance structure of the target land class in the feature space includes: Determine the median vector using the median of each feature dimension in the initial sample set, and perform the following iterative processing using the median vector as the initial robust center: Calculate the Euclidean distance from each sample to the current mean robust center, and then calculate the robust scaling estimate based on the Euclidean distance. for: ； ； in, This indicates the calculation of the median. Represents the first sample in the initial sample set. One sample, Indicates the first The sample at the th The Euclidean distance to the mean robust center at the next iteration. Indicates the first The mean robust center at the next iteration Used for robust scaling estimation It becomes an unbiased estimator of the standard deviation; Based on robust scaling estimation Calculate Tukey weights for: ； in, It is a constant; Based on Tukey weights The updated mean robust center is: ； If the iteration satisfies If the iteration stops, then the iteration stops. To preset the convergence accuracy; Furthermore, it also includes regularizing the initial sample set using Ledoit-Wolf contraction estimation to obtain a robust covariance structure. for: ； in, It is the covariance matrix calculated based on weighted samples. It is the dimension of the feature space. It is the trace of the matrix. It is the identity matrix. It is the shrinkage parameter.

2. The method for collecting training samples of land cover remote sensing images according to claim 1, characterized in that, The method involves extracting multiple candidate features for each target land cover category from remote sensing images based on the category set for land cover classification, constructing a candidate feature pool, calculating feature correlation, and eliminating redundant features based on correlation. A forward iterative feature selection strategy based on JM distance is adopted to obtain the optimal feature subset, including: Based on the set of land cover classification categories, at least one of multiple spectral features and multiple texture features for each target land cover type is extracted from remote sensing images to obtain a candidate feature pool. For any two candidate features in the candidate feature pool, the Pearson correlation coefficient r is used to generate a correlation matrix. The correlation matrix includes redundant feature pairs. Redundant feature elimination is obtained by retaining the features with greater median difference and less inter-class overlap based on the distribution of the two candidate features in different types of box plots. A forward iterative feature selection strategy based on JM distance is adopted, which involves starting from an empty set and adding features that maximize the minimum JM distance of all class pairs until the minimum JM distance reaches a preset threshold, thus obtaining the optimal feature subset.

3. The method for collecting training samples of land cover remote sensing images according to claim 1, characterized in that, The process involves applying a heterogeneity measure based on regularized Mahalanobis distance to the initial sample set, based on robust centers and robust covariance structures, to obtain the regularized Mahalanobis distance, including: Calculate the regularized Mahalanobis distance of the robust center for each sample in the initial sample set. for: ； Among them, the regularized Mahalanobis distance Used to characterize the probability that a sample is labeled noise.

4. The method for collecting training samples of land cover remote sensing images according to claim 3, characterized in that, The process involves denoising the initial sample set using an adaptive threshold based on an empirical distribution, based on regularized Mahalanobis distance, to obtain a high-purity training sample set, including: Calculate the lower quartile of the regularized Mahalanobis distance for all samples in the initial sample set. Upper quartiles and interquartile range Interquartile range ; The initial sample set is processed by removing samples based on a removal threshold and regularized Mahalanobis distance to obtain a high-purity training sample set. The removal threshold is... The calculation method is as follows: ； in, To control the stringency of noise reduction.

5. The method for collecting training samples of land cover remote sensing images according to claim 1, characterized in that, The method further includes: The high-purity training sample set is preprocessed and augmented, including randomly cropping the high-purity training sample set after cleaning each target land type to obtain sample blocks of fixed size, and then performing random rotation, horizontal and vertical flipping and brightness scaling on the sample blocks to obtain the final surface remote sensing image training sample set.

6. A device for acquiring training samples of land cover remote sensing images, characterized in that, include: The first module is used to determine the set of categories for land cover classification based on the classification request; The second module is used to extract multiple candidate features for each target land cover category from remote sensing images based on the land cover classification category set, construct a candidate feature pool and calculate feature correlation, and remove redundant features based on correlation; a forward iterative feature selection strategy based on JM distance is adopted to obtain the optimal feature subset. The third module is used to obtain an initial sample set based on the preset land cover classification data of the target area; The fourth module is used to process the target land class in the initial sample set using robust global statistics to obtain the robust center and robust covariance structure of the target land class in the feature space. The fifth module is used to measure the heterogeneity of the initial sample set based on the robust center and robust covariance structure, and obtain the regularized Mahalanobis distance. The sixth module is used to denoise the initial sample set using an adaptive threshold based on an empirical distribution, based on the regularized Mahalanobis distance, to obtain a high-purity training sample set. The process of obtaining an initial sample set based on preset land cover classification data for the target area includes: Obtain preset land cover classification data based on the target area and surface remote sensing images, and generate a binary mask of the preset land cover classification data based on the target land type. The binary mask is overlaid on the surface remote sensing image to extract the feature vector of the pixels including the target land type; An initial sample set is generated based on the feature vectors of the pixels of the target land type and the initial labels of the preset land cover classification data; The process of applying robust global statistics to the target land class in the initial sample set to obtain the robust center and robust covariance structure of the target land class in the feature space includes: Determine the median vector using the median of each feature dimension in the initial sample set, and perform the following iterative processing using the median vector as the initial robust center: Calculate the Euclidean distance from each sample to the current mean robust center, and then calculate the robust scaling estimate based on the Euclidean distance. for: ； ； in, This indicates the calculation of the median. Indicates the first sample in the initial sample set. One sample, Indicates the first The sample at the th The Euclidean distance to the mean robust center at the next iteration. Indicates the first The mean robust center at the next iteration Used for robust scaling estimation It becomes an unbiased estimator of the standard deviation; Based on robust scaling estimation Calculate Tukey weights for: ； in, It is a constant; Based on Tukey weights The updated mean robust center is: ； If the iteration satisfies If the iteration stops, then the iteration stops. To preset the convergence accuracy; Furthermore, it also includes regularizing the initial sample set using Ledoit-Wolf contraction estimation to obtain a robust covariance structure. for: ； in, It is the covariance matrix calculated based on weighted samples. It is the dimension of the feature space. It is the trace of the matrix. It is the identity matrix. It is the shrinkage parameter.

7. An electronic device, characterized in that, Including the processor and memory; The memory is used to store programs; The processor executes the program to implement the method for collecting training samples of land cover remote sensing images as described in any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, The storage medium stores a program, which is executed by a processor to implement the method for collecting training samples of land cover remote sensing images as described in any one of claims 1-5.

Images