Riverbed Sediment Enhancement Method and System Integrating Confidence Learning and Superpixel Segmentation

Abstract

The present invention belongs to the technical field of riverbed sediment exploration, and discloses a riverbed sediment enhancement method and system that integrates confidence learning and superpixel segmentation. The method uses the superpixel segmentation method to divide the multi-beam backscattering intensity image; performs sample augmentation on the pixels after superpixel segmentation; and uses confidence learning for sample enhancement. Based on the on-site riverbed sediment sampling points, the present invention uses the superpixel segmentation method to divide the multi-beam backscattering intensity image into different superpixel regions, takes the center of the superpixel block as the representative of the superpixel, and uses confidence learning to remove the noise in the augmented dataset to achieve riverbed sediment sample enhancement.

Description

Technical Field

[0001] The present invention belongs to the technical field of riverbed sediment exploration, and particularly relates to a riverbed sediment enhancement method and system integrating confidence learning and superpixel segmentation. Background Art

[0002] Riverbed sediments, as key geographical elements carrying the benthic ecology of the basin and connecting areas such as rivers, lakes, deltas, and estuaries, rapid and effective underwater sediment detection and identification within the river basin are crucial for studying the distribution and migration process of river-inhabiting organisms, calculating the sediment-carrying capacity of water flow, the bed load sediment transport rate, analyzing and predicting the water flow forces of river evolution, and predicting the evolution trend of rivers. Riverbed sediments, as the sink and source of various substances in the basin, are of great significance in river water environmental protection, improvement of the basin ecological environment, and restoration of ecosystem functions.

[0003] Currently, riverbed sediment exploration technologies mainly include traditional on-site sampling technologies and acoustic remote sensing detection methods relying on acoustic media. The multi-beam echo sounding system can obtain continuous depth and backscatter intensity data over a large range. By comprehensively applying traditional on-site sampling and acoustic detection, it can characterize the spatial distribution and characteristics of riverbed sediments over a large range, and has become an essential basic tool for measuring the main biological and abiotic characteristics of the habitat environment.

[0004] Traditional underwater in-situ sampling technologies use direct sampling technologies such as box sampling, tube sampling, trawl sampling, underwater photography, grab sampling, and manual diving sampling to collect sediment samples. However, these technologies usually have problems such as complex operation, long time consumption, and difficult sampling. Limited by the sediment sampling method, the scarcity of sediment sampling has become a key problem hindering the further development of sediment classification. With the maturity of underwater detection technologies, the resolution of multi-beam sonar systems has been continuously improved, resulting in a gradually widening gap between the number of underwater on-site sampling samples obtained and the volume of acoustic data. The resulting model overfitting problem will affect the final classification accuracy. To address the above problems, the current mainstream solutions are: one is to divide the underwater acoustic image into multiple small-scale images, determine the sediment type of each small-scale image through direct sampling analysis, and thus obtain expanded samples of sediments for classification; the other is to obtain the sediment type of the fixed neighborhood of the sampling point based on the distribution of surface sediment sample data and historical sediment distribution, so that the number of samples meets the requirements for constructing the seabed sediment supervised classification model. The above two sample enhancement methods are both based on the principle of expanding from limited seabed sediment sampling points: regularly expanding outward from known sample points. This method is suitable for small-scale seas with fewer sediment types and can construct a classification model with higher accuracy.

[0005] In fact, the benthic environment of the riverbed is diverse, resulting in the spatial irregularity of the substrate variation. The prominent features are the ambiguity of the substrate categories and the multi-scale span of the substrate boundaries, which conflicts with the method of expanding samples using a fixed neighborhood. At present, the concept of superpixel segmentation has been introduced in the sample expansion of seabed substrate classification. For example, superpixel segmentation is performed on the multi-beam backscatter intensity image, and the samples are screened by combining the characteristics of the original sampling points and other sampling points in the expanded area; active learning based on voting entropy is used to screen the superpixel-expanded samples to solve the problems of few and unstable expanded samples. However, affected by the changes in river hydrodynamic conditions and the undulation of the riverbed substrate, sediment is constantly scoured and silted up, and the riverbed alluvial layer has high spatial heterogeneity and temporal variability in physical properties and biogeochemical processes. Using the above methods often leads to the possibility of multiple categories being mixed within the segmented pixel blocks, especially the interference of attribute-free noise signals, thus bringing difficulties to classification. Summary of the Invention

[0006] To overcome the problems in the related art, the disclosed embodiments of the present invention provide a riverbed substrate enhancement method and system that combines confidence learning and superpixel segmentation.

[0007] The technical solution is as follows: The riverbed substrate enhancement method that combines confidence learning and superpixel segmentation includes the following steps:

[0008] S1, dividing the multi-beam backscatter intensity image using the superpixel segmentation method;

[0009] S2, expanding the samples of the pixels after superpixel segmentation;

[0010] S3, enhancing the samples using confidence learning.

[0011] In step S1, dividing the multi-beam backscatter intensity image using the superpixel segmentation method includes:

[0012] S101, initializing the seed points and evenly distributing seed points in the backscatter image, with each superpixel having a size of ;

[0013] S102, optimizing the seed points, calculating the gradients of all pixels within the neighborhood of the initial seed points, and moving the seed points to the gradient within the neighborhood. The gradient represents the direction and magnitude of the fastest change in pixel values in the image. For a grayscale image, the pixel value range is [0, 255], and the theoretical range of the gradient magnitude is ;

[0014] S103, calculating the distance metrics, including color distance and spatial distance.

[0015] In step S103, a distance metric is calculated, and the formula is:

[0016] (1)

[0017] (2)

[0018] (3)

[0019] (4)

[0020] In the formula, is the color distance, is the spatial distance, is the distance metric, is the maximum spatial distance within the cluster, is the brightness value of the superpixel seed point, is the pixel point 's brightness. In the multi-beam backscatter intensity map, the brightness is the gray level; is the abscissa of the superpixel seed point, is the abscissa of the pixel point, is the ordinate of the superpixel seed, is the ordinate of the pixel point, is the maximum color distance, is the total number of pixels in the image, is the number of pre-segmented superpixels.

[0021] In step S2, sample augmentation is performed, including: when there is only one label within the superpixel block, the entire superpixel block is labeled according to the label; when there are multiple labels, the majority voting method is used to select the label mode of the labeled samples within the pixel block as the label of the superpixel block. If the majority voting is not satisfied, the superpixel is re-segmented;

[0022] Define 's value satisfies the following formula:

[0023] ;

[0024] Among them, is the parameter set for the first segmentation. If the majority voting is not satisfied, let , and re-segment until the majority voting method is satisfied and stop.

[0025] In step S3, sample enhancement is performed using confidence learning, including:

[0026] First, feature extraction is performed on the multi-beam water depth and backscatter intensity data;

[0027] Secondly, in combination with the on-site sediment sampling points, the samples after superpixel segmentation expansion are denoised by confidence learning.

[0028] Furthermore, denoising the labels includes:

[0029] S301, estimating the joint distribution of noisy labels and true labels ;

[0030] S302, data cleaning, setting the wrong label to , where is the wrong label, is the category, is the estimated probability, is the noisy label, is the category, is the given sample, is the model parameter; estimating the noisy labels through the non-diagonal elements, finding the noisy samples in the samples expanded based on superpixel segmentation, and cleaning the noisy samples;

[0031] S303, retraining the model, putting the samples with noise removed into the sediment classification model for retraining to obtain a new sediment classification result.

[0032] In step S301, estimating the joint distribution of noisy labels and true labels , including:

[0033] The estimated probability of cross-validation, the expression is:

[0034] (5)

[0035] In the formula, is the feature, is the number of folds of data cross-validation, is the fold model's output probability for the th category prediction;

[0036] In the count matrix, the diagonal captures the correct labels, and the non-diagonal captures the asymmetric wrong labels, the expression is:

[0037] (6)

[0038] (7)

[0039] (8)

[0040] In the formula, is the count matrix; is the noisy label of the sample, i.e., the label provided by the dataset value is the true label of the sample is in the dataset in which the samples are labeled as class subset is the average confidence threshold for each class is in the dataset in which the samples are labeled as class subset is the set of samples observed as the th class but actually the th class is the subset of samples with noise is the observable noisy label is the true noise-free label is the probability that the sample belongs to the label under the model parameters probability

[0041] The estimation formula for the joint distribution is as follows

[0042] (9)

[0043] In the formula is the number of samples that are actually in class among the samples labeled as number is the size of the sample set with the noisy label number is the probability that the true label is among the samples with the noisy label probability is the set of classes

[0044] Another object of the present invention is to provide a riverbed sediment sample enhancement system that integrates confidence learning and superpixel segmentation. The system implements the riverbed sediment enhancement method that integrates confidence learning and superpixel segmentation. The system includes

[0045] A superpixel segmentation module for dividing the multi-beam backscatter intensity image using the superpixel segmentation method

[0046] A sample augmentation module for augmenting the samples of the pixels after superpixel segmentation

[0047] A confidence learning module for enhancing the samples using confidence learning

[0048] Furthermore, the system is installed on a computer device, which includes at least one processor, a memory, and a computer program stored in the memory and executable on the at least one processor. When the processor executes the computer program, the functions in the above-mentioned riverbed sediment sample enhancement system that integrates confidence learning and superpixel segmentation are realized.

[0049] Furthermore, the system is installed on a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, the functions in the riverbed sediment sample enhancement system that integrates confidence learning and superpixel segmentation can be realized.

[0050] Combining all the above technical solutions, the beneficial effects of the present invention are as follows: The present invention provides a method for enhancing riverbed sediment classification samples by integrating superpixel segmentation and confidence learning. Based on on-site riverbed sediment sampling points, the multi-beam backscatter intensity image is segmented into different superpixel regions using the superpixel segmentation method. The center of the superpixel block is used as the representative of the superpixel, and confidence learning is used to remove the noise in the augmented dataset, realizing the enhancement of riverbed sediment samples.

[0051] The present invention reduces the cost of on-site sampling and improves the classification performance of the model. With the same amount of training data, the classification accuracy is improved, and a more accurate classification result of riverbed sediments is obtained. The present invention supports engineering decisions such as river ecological restoration and waterway dredging in precise sediment mapping; it provides a high-precision data basis for sediment migration and pollutant diffusion research in environmental monitoring optimization.

[0052] Existing methods at home and abroad mostly rely on on-site sampling to obtain the sediment types in the fixed neighborhood of the sampling points or segment underwater acoustic images into multiple small-scale images, and determine the sediment types of each small-scale image through direct sampling analysis, lacking an automated and highly robust sample enhancement scheme. The present invention combines superpixel segmentation and confidence learning for the first time: realizing an end-to-end automated process from multi-beam backscatter intensity image segmentation to sample noise cleaning based on on-site sampling samples.

[0053] Riverbed sediment sampling points are sparse and costly, and traditional on-site sampling methods are inefficient and costly. The present invention generates a large number of candidate samples based on superpixel segmentation, covering areas that are not directly sampled, and completes sample augmentation. High-quality data is extracted from the candidate samples through confidence learning for sample enhancement, thus effectively alleviating the problem of overfitting of small samples. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] The accompanying drawings herein are incorporated into the specification and constitute a part of this specification, showing embodiments consistent with the present disclosure, and are used together with the specification to explain the principles of the present disclosure;

[0055] Figure 1 It is the schematic diagram of the riverbed sediment enhancement method that integrates confidence learning and superpixel segmentation provided by an embodiment of the present invention;

[0056] Figure 2 It is the flowchart of the riverbed sediment enhancement method that integrates confidence learning and superpixel segmentation provided by an embodiment of the present invention;

[0057] Figure 3 It is the flowchart of superpixel segmentation sample augmentation provided by an embodiment of the present invention;

[0058] Figure 4 It is the effect diagram containing only one type of label in the superpixel segmentation sample augmentation strategy provided by an embodiment of the present invention;

[0059] Figure 5 It is the effect diagram of two types of labels with the same quantity in the superpixel segmentation sample augmentation strategy provided by an embodiment of the present invention;

[0060] Figure 6 It is the effect diagram of two types of labels with different quantities in the superpixel segmentation sample augmentation strategy provided by an embodiment of the present invention;

[0061] Figure 7 It is the augmentation result diagram of one type of label in the superpixel segmentation sample augmentation strategy provided by an embodiment of the present invention;

[0062] Figure 8 It is the augmentation result diagram of two types of labels with the same quantity in the superpixel segmentation sample augmentation strategy provided by an embodiment of the present invention;

[0063] Figure 9 It is the augmentation result diagram of two types of labels with different quantities in the superpixel segmentation sample augmentation strategy provided by an embodiment of the present invention;

[0064] Figure 10 It is the sample enhancement result diagram using neighborhood window augmentation provided by an embodiment of the present invention;

[0065] Figure 11 It is the sample enhancement result diagram based on superpixel segmentation and confidence learning in the present invention provided by an embodiment of the present invention;

[0066] Figure 12 It is the bathymetric value diagram in the multi-beam measurement area of a certain area provided by an embodiment of the present invention; the horizontal and vertical coordinates are projection coordinates;

[0067] Figure 13 It is the backscatter intensity value diagram in the multi-beam measurement area of a certain area provided by an embodiment of the present invention; among them, Figure 13 the geographical locations and categories of 18 on-site sampling points are marked, and the horizontal and vertical coordinates are projection coordinates;

[0068] Figure 14It is the substrate classification result graph without using the sample enhancement method of the present invention in the comparison of substrate classification results by the support vector machine classification algorithm provided by the embodiment of the present invention; the horizontal and vertical coordinates are projection coordinates;

[0069] Figure 15 It is the substrate classification result graph using the sample enhancement method of the present invention in the comparison of substrate classification results by the support vector machine classification algorithm provided by the embodiment of the present invention, and the horizontal and vertical coordinates are projection coordinates. Specific embodiments

[0070] To make the above objects, features and advantages of the present invention more obvious and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description in order to fully understand the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without departing from the connotation of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed below.

[0071] The innovation of the present invention lies in that: the sample enhancement method proposed by the present invention combines superpixel segmentation and confidence learning, constructs a sediment data set based on on-site sampling points, eliminates the noise in the sediment data set, makes full use of the information of on-site sampling points, effectively mines the information in unlabeled samples, and has a significant improvement in the accuracy of sample selection, thus improving the accuracy of substrate classification.

[0072] Example 1, as Figure 1 shown, the riverbed substrate enhancement method combining confidence learning and superpixel segmentation provided by the embodiment of the present invention combines confidence learning to remove noise from the substrate sampling sample points expanded by superpixel segmentation, effectively enhances the original sampling points, puts the samples with noise removed into the classification model for retraining, and obtains more accurate substrate classification results, providing a new solution for riverbed substrate classification sample enhancement.

[0073] As Figure 2 shown, it specifically includes:

[0074] S1, use the superpixel segmentation method to divide the multi-beam backscattering intensity image;

[0075] Superpixel segmentation uses color space information and spatial information to locally cluster the pixels of the multi-beam backscattering intensity image, generating compact and approximately uniform superpixels, which can better express the texture information between ground objects. At the same time, when there is a large amount of segmentation data, superpixel segmentation can obtain highly pure homogeneous regions. Specifically, it includes:

[0076] S101, initialize the seed points and evenly distribute them in the backscattering image A number of seed points, each superpixel having a size of ;

[0077] S102. Optimize the seed points, calculate the gradients of all pixels within the neighborhood of the initial seed points, and move the seed points to the gradients within the neighborhood. The gradient represents the direction and magnitude in which the pixel values in the image change most rapidly. For a grayscale image, the pixel value range is [0, 255], and the theoretical range of the gradient magnitude is ;

[0078] S103. Calculate the distance metrics, including color distance and spatial distance.

[0079] Calculate the distance metrics. The formula is:

[0080] (1)

[0081] (2)

[0082] (3)

[0083] (4)

[0084] In the formula, is the color distance, is the spatial distance, is the distance metric, is the maximum spatial distance within the cluster, is the brightness value of the superpixel seed point, is the pixel point The brightness, in the multi-beam backscattering intensity map, is the gray level; is the abscissa of the superpixel seed point, is the abscissa of the pixel point, is the ordinate of the superpixel seed, is the ordinate of the pixel point, is the maximum color distance, is the total number of pixels in the image, is the number of pre-segmented superpixels.

[0085] S2. Perform sample augmentation on the pixels after superpixel segmentation;

[0086] Combined with the positions of the on-site sampling points, mark the on-site sampling points in the multi-beam backscattering intensity map to obtain the association between the samples and the superpixel blocks. Set category labels for samples of each category to facilitate sample augmentation and enhancement.

[0087] The sample augmentation strategy is as follows: when there is only one label within a superpixel block, the entire superpixel block is labeled according to the label; when there are multiple labels, the majority voting method is adopted, and the label mode of the labeled samples within the pixel block is selected as the label of the superpixel block. If the majority voting is not satisfied, the superpixel is re-segmented.

[0088] Define The value of satisfies the following formula:

[0089] ;

[0090] Among them, is the parameter set for the first segmentation. If the majority voting is not satisfied, let , and re-segment until the majority voting method is satisfied and stop. As Figure 3 shown, it shows the flowchart of the sample augmentation strategy.

[0091] As Figure 4 is the effect diagram of only one type of label in the superpixel segmentation sample augmentation strategy provided by the embodiment of the present invention; Figure 5 is the effect diagram of two types of labels with the same number in the superpixel segmentation sample augmentation strategy provided by the embodiment of the present invention; Figure 6 is the effect diagram of two types of labels with different numbers in the superpixel segmentation sample augmentation strategy provided by the embodiment of the present invention; Figure 7 is the augmented result diagram of one type of label in the superpixel segmentation sample augmentation strategy provided by the embodiment of the present invention; Figure 8 is the augmented result diagram of two types of labels with the same number in the superpixel segmentation sample augmentation strategy provided by the embodiment of the present invention; Figure 9 is the augmented result diagram of two types of labels with different numbers in the superpixel segmentation sample augmentation strategy provided by the embodiment of the present invention; An example of superpixel augmentation is given. In the example, the multi-beam backscattering intensity image is segmented using the superpixel segmentation method, and the scale number n is 17. Figure 4 and Figure 6 give the augmented results after satisfying the majority voting. Figure 5 does not satisfy the majority voting. After adjusting the scale number n to 18, the adjacent sampling points fall within one superpixel block.

[0092] S3. Use confidence learning for sample enhancement.

[0093] Before performing sample enhancement on the augmented samples based on superpixel segmentation, first extract the features of the multi-beam water depth and backscattering intensity data, and then combine the on-site sediment sampling points to remove the noise labels from the samples augmented by superpixel segmentation through confidence learning. The steps of removing the noise labels are as follows.

[0094] S301. Estimate the joint distribution of the noise label and the true label ;

[0095] To estimate the joint distribution, cross-validation is used to estimate the labels of superpixel augmentation to obtain the class probabilities of each sample ; based on this probability, the count matrix is calculated and calibrated , and finally the joint distribution is estimated .

[0096] The estimated probability of cross-validation, the expression is:

[0097] (5)

[0098] In the formula, is the feature, is the number of folds of data cross-validation, is the th fold model for the output probability of class prediction;

[0099] In the count matrix, its diagonal captures the correct labels, and the non-diagonal captures the asymmetric wrong labels, and its calculation is as follows:

[0100] (6)

[0101] (7)

[0102] (8)

[0103] In the formula, is the count matrix; is the noisy label of the sample, that is, the label provided by the data set value, is the true label of the sample, is in the data set is labeled as class subset of samples; is the average confidence threshold for each class, is in the data set is labeled as class subset of samples, is observed as the th class, actually the th class set, is the subset of samples with noise, is the observable noisy label; is the true label without noise; is the probability that the sample belongs to the label under the model parameters ;

[0104] The estimation formula for the joint distribution is as follows:

[0105] (9)

[0106] where is the number of samples that are actually in category among the samples labeled as , is the size of the sample set with noisy label , is the probability that the true label is among the samples with noisy label , is the set of categories.

[0107] S302, Data cleaning, set the wrong label to , where is the wrong label, is the category, is the estimated probability, is the noisy label, is the category, is the given sample, is the model parameter; estimate the noisy label through the non - diagonal elements, find the noisy samples in the samples augmented based on super - pixel segmentation, and clean the noisy samples;

[0108] S303, Retrain the model, put the samples with noise removed into the substrate classification model for retraining to obtain a new substrate classification result.

[0109] Figure 10 shows the sample enhancement result diagram using neighborhood window augmentation, Figure 11 shows the sample enhancement result diagram based on super - pixel segmentation and confidence learning in the present invention. It can be seen from Figure 11 that the present invention can better combine the texture features of the image compared with neighborhood window augmentation.

[0110] As can be seen from the above embodiments, the role of the super - pixel segmentation algorithm of the present invention is to segment the multi - beam backscattering intensity image. After segmentation, sample augmentation is performed in combination with the positions of the on - site substrate sampling sample points; the role of confidence learning is to remove the noisy sample points from the samples augmented by super - pixel segmentation. The present invention combines confidence learning and super - pixel segmentation, which can segment to obtain highly pure homogeneous regions, and can also perform the functions of augmenting, denoising, and purifying rare sample sampling points.

[0111] The present invention divides the multi-beam backscattering intensity image based on superpixel segmentation, expands the segmentation result around the in-situ sediment sampling points, and realizes the effective expansion of the riverbed sediment samples; determines the noise samples in the samples expanded by superpixel segmentation through confidence learning, cleans the noise samples, and combines the samples with noise removed and substitutes them into the classification model for retraining to obtain a more accurate sediment classification result, providing important technical support for riverbed sediment classification.

[0112] Embodiment 2. The embodiment of the present invention provides a riverbed sediment sample enhancement system integrating confidence learning and superpixel segmentation, including:

[0113] A superpixel segmentation module, configured to divide the multi-beam backscattering intensity image using the superpixel segmentation method;

[0114] A sample expansion module, configured to expand the samples of the pixels after superpixel segmentation;

[0115] A confidence learning module, configured to perform sample enhancement using confidence learning.

[0116] In the above embodiments, the descriptions of the respective embodiments have their own emphases. For parts not detailed or recorded in a certain embodiment, reference may be made to the relevant descriptions of other embodiments.

[0117] The present invention takes the in-situ collected multi-beam data of a certain basin in a certain province as an example for effect comparison. Figures 12 - 13 Shows the multi-beam sounding and backscattering intensity result maps of an experimental area in a certain region. First, feature extraction is performed on the multi-beam sounding and backscattering intensity data. Secondly, the multi-beam backscattering intensity image is preprocessed by the superpixel segmentation algorithm, and the backscattering image is segmented into several irregular superpixel regions. If there are in-situ sampling samples, sample expansion is performed; subsequently, confidence learning is used to purify the expanded samples to remove the noise samples therein. Finally, the samples are put into the classification model to generate classification results. Figures 14 - 15 Shows the riverbed sediment classification result using the support vector machine classification algorithm, Figure 14 The total accuracy of the sediment classification in [reference] is 78.6%, and the Kappa coefficient is 0.704. Figure 15 The total accuracy of the sediment classification in [reference] is 79.8%, and the Kappa coefficient is 0.722. After using the sample enhancement method proposed by the present invention, the accuracy of the riverbed sediment classification result is higher.

[0118] The above is only a relatively preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any modification, equivalent replacement, and improvement made within the spirit and principle of the present invention by those skilled in the art within the technical scope disclosed by the present invention shall be covered by the protection scope of the present invention.

Claims

1. A method for enhancing riverbed sediment by integrating confidence learning and superpixel segmentation, characterized in that, The method includes the following steps: S1. Divide the multi-beam backscattering intensity image using the superpixel segmentation method; S2. Augment the samples of the pixels after superpixel segmentation; S3. Perform sample enhancement using confidence learning; In step S1, dividing the multi-beam backscattering intensity image using the superpixel segmentation method includes: S101, Initialize seed points, and evenly distribute H seed points in the backscatter image, with each superpixel having a size of N s ; S102, Optimize the seed points, calculate the gradients of all pixels within the neighborhood of the initial seed points, and move the seed points to the gradients within the neighborhood. The gradient represents the direction and magnitude in which the pixel values in the image change most rapidly. For a grayscale image, the range of its pixel values is [0, 255], and the theoretical range of the gradient magnitude is S103. Calculate the distance metrics, including color distance and spatial distance; In step S3, performing sample enhancement using confidence learning includes: First, extract features from the multi-beam water depth and backscattering intensity data; Second, in combination with the in-situ sediment sampling points, remove the noise labels from the samples after superpixel segmentation and augmentation through confidence learning.

2. The enhanced method for riverbed sediment based on the fusion of confidence learning and superpixel segmentation according to claim 1, wherein In step S103, the formula for calculating the distance metrics is: where d c is the color distance, d s is the spatial distance, D is the distance metric, N s is the maximum spatial distance within the cluster, l i is the brightness value of the superpixel seed point, l j is the brightness of pixel point j, and in the multi-beam backscatter intensity map, the brightness is the gray level; x i is the abscissa of the superpixel seed point, x j is the abscissa of the pixel point, y i is the ordinate of the superpixel seed, y j is the ordinate of the pixel point, N c is the maximum color distance, W is the total number of pixels in the image, and H is the number of pre-segmented superpixels.

3. The enhanced method for riverbed sediment based on the fusion of confidence learning and superpixel segmentation according to claim 2, characterized in that, In step S2, sample augmentation includes: when there is only one label within a superpixel block, label the entire superpixel block according to the label; when there are multiple labels, use the majority voting method to select the mode of the labels of the labeled samples within the pixel block as the label of the superpixel block. If the majority voting is not satisfied, re-segment the superpixel; Define that the value of H satisfies the following formula: H = 2 n , H ≤ W Among them, n is the parameter set for the first segmentation. If the majority voting is not satisfied, let H = 2 (n+1) , and re-segment until the majority voting method is satisfied and stop.

4. The enhanced method for riverbed sediment based on the integration of confidence learning and superpixel segmentation according to claim 1, wherein Removing the noise labels includes: S301, estimate the joint distribution of the noisy labels and the true labels S302, Data cleaning, set the error label as where is the error label, j is the category, is the estimated probability, is the noisy label, x k is the given sample, θ is the model parameter; estimate the noisy labels through the non - diagonal elements, find the noisy samples in the augmented samples based on super - pixel segmentation, and clean the noisy samples; S303. Retrain the model, put the samples with noise removed into the sediment classification model for retraining, and obtain a new sediment classification result.

5. The method for enhancing riverbed sediment based on the integration of confidence learning and superpixel segmentation according to claim 4, wherein In step S301, estimate the joint distribution of the noisy labels and the true labels including: The estimated probability of cross-validation, the expression is: where x is the feature, K is the number of folds for data cross-validation, and f k (x; θ) j is the output probability of the k-th fold model for predicting the j-th class of x; In the count matrix, the diagonal captures the correct labels, and the non-diagonal captures the asymmetric error labels, the expression is: Wherein, is the counting matrix; [i] is the noisy label of the sample, that is, the label provided by the data set value, [j] is the true label of the sample, is the subset of samples labeled as class i in the data set X; t j is the average confidence threshold for each class, is the subset of samples labeled as class j in the data set X, is the set observed as the i-th class but actually the j-th class, is the subset of samples with noise, is the observable noisy label; y * is the true label without noise; is the probability that the sample x belongs to the label under the parameter θ of the model; The estimation formula for the joint distribution is: Wherein, is the number of samples that actually belong to class j among the samples labeled as i, is the size of the sample set with noisy label i, is the probability that the true label is j among the samples with noisy label i, and M is the set of classes.

6. A riverbed sediment enhancement system integrating confidence learning and superpixel segmentation, characterized in that, The system implements the riverbed sediment enhancement method that combines confidence learning and superpixel segmentation according to any one of claims 1-5. The system includes: A superpixel segmentation module for dividing the multi-beam backscattering intensity image using the superpixel segmentation method; A sample augmentation module for augmenting the samples of the pixels after superpixel segmentation; A confidence learning module for performing sample enhancement using confidence learning.

7. The riverbed sediment enhancement system integrating confidence learning and superpixel segmentation according to claim 6, characterized in that, The system is installed on a computer device, and the computer device includes: at least one processor, a memory, and a computer program stored in the memory and executable on the at least one processor. When the processor executes the computer program, it realizes the functions in the above-mentioned riverbed sediment sample enhancement system that combines confidence learning and superpixel segmentation.

8. The riverbed sediment enhancement system integrating confidence learning and superpixel segmentation according to claim 6, characterized in that, The system is installed on a computer-readable storage medium, and the computer-readable storage medium stores a computer program. When the computer program is executed by a processor, it can realize the functions in the riverbed sediment sample enhancement system that combines confidence learning and superpixel segmentation.

Images