Multi-beam seabed sediment classification method based on multi-classifier decision fusion mechanism

Abstract

The invention discloses a multi-beam seabed sediment classification method based on a multi-classifier decision fusion mechanism, and the method comprises the following steps: carrying out the multi-scale texture feature extraction and topographic feature extraction of an input multi-beam backscattering intensity image and depth image respectively; performing precision analysis by using K-means clustering to screen acoustic features of an optimal scale, and obtaining acoustic features of high contribution degree in combination with a geographically weighted regression analysis method; carrying out superpixel segmentation on the multi-beam backscattering intensity image to obtain a sample-superpixel block, and realizing effective expansion of seabed sample points; taking half of the expanded sample set as a training set to construct a seabed sediment classification model with multiple classifiers; and creating a multi-classifier decision fusion mechanism model to predict the type of the seabed sediment by using the classification result of each model and the producer precision, and obtaining a seabed sediment prediction result map. The method has high prediction precision and can be applied to the fields of ocean engineering construction, resource management development, surveying and mapping and the like.

Description

Technical Field

[0001] The present invention relates to the fields of multi-beam seafloor sediment classification, ocean mapping, image processing and machine learning, and in particular to a multi-beam seafloor sediment classification method based on a multi-classifier decision fusion mechanism. Background Art

[0002] Full coverage and high efficiency of seabed detection and identification support the three-dimensional and thematic development of marine spatial geographic information. With the development of underwater bathymetric technology, acoustic bottom interpretation has become one of the core technologies of marine surveying and mapping. It cross-integrates acoustic signal processing, feature mining, qualitative and quantitative expression and evaluation of bottom types, and is the key basis for the construction of marine geographic spatial information. In essence, acoustic bottom interpretation has certain similarities with remote sensing image interpretation. Its basic content is to obtain the interpretation and expression of the actual situation of the seabed surface from the sound spectrum and image signals through pattern recognition or artificial intelligence methods.

[0003] Most of the research on multi-beam acoustic seabed sediment methods focuses on seabed roughness assessment, sediment type definition, benthic habitat detection and other applications in near-shallow sea scenarios. The main method is to use the backscattering intensity data and bathymetric data collected by the multi-beam bathymetry system as the main method, supplemented by underwater in-situ sampling technology, to obtain the corresponding features and build a classification model, thereby obtaining a high-precision seabed sediment classification map. It also forms an acoustic seabed sediment classification system with a series of technical methods including acoustic feature extraction and mining, feature optimization and dimensionality reduction, classifier selection and model construction, classification result mapping and accuracy evaluation. In recent years, domestic research on the classification of multi-beam acoustic seabed sediments has developed rapidly. Although corresponding progress has been made, it is still limited by the lack of feature characterization capabilities, the small number of samples due to the difficulty in sampling the seabed sediment, and the poor stability of the classification model in large-scale sea areas and complex water environments. Problems that need to be solved urgently. First, the acoustic features extracted from multi-beam data at a single scale have the phenomenon that the feature outliers of certain texture features are not obvious or the image edges are blurred, and the positioning accuracy is low; the introduction of terrain features to construct a multivariate feature space has the phenomenon of feature redundancy caused by the extremely low contribution of some features. Therefore, how to extract and optimize multi-scale acoustic features and how to obtain high-contribution features to optimize the multivariate feature space are urgent problems to be solved. Secondly, the traditional method of seabed sediment sample enhancement is to grid-expand the area where the sampling point is located. Currently, it mainly includes image segmentation method and fixed neighborhood method, which lacks theoretical basis, and the reliability of sampling points expanded in large-scale sea areas is not high. Therefore, a more reliable sample expansion method still needs to be further explored to ensure the prediction accuracy and confidence of the seabed sediment classification model. Finally, in the actual construction of various classification models of multi-beam seabed sediment, there will be a large difference in the overall classification accuracy of different classification models constructed by the same sediment data set, and the classification model with low accuracy has a strong ability to recognize a certain sediment. Therefore, what kind of classification model should be used to identify the sediment distribution in the research area needs to be solved urgently. Summary of the invention

[0004] In view of this, the present invention proposes a multi-beam seabed sediment classification method based on a multi-classifier decision fusion mechanism. The method has a reasonable design and solves the problem that the classification accuracy of various seabed sediment classification models for the same sea area is very different to a certain extent. By constructing a systematic multi-beam seabed sediment classification model, the classification accuracy and confidence are improved.

[0005] In order to achieve the above object, the present invention adopts the following technical solution: A multi-beam seafloor sediment classification method based on a multi-classifier decision fusion mechanism, comprising: Step 1: Multi-scale acoustic feature optimization evaluation: texture feature extraction and terrain feature extraction are performed on the input multi-beam backscatter intensity image and depth image through sliding windows of different scales. The K-means clustering method is used to perform precision analysis on the corresponding features of each scale to select the acoustic features of the optimal scale, and the geographically weighted regression analysis method is combined to obtain the acoustic features with high contribution. Step 2: Expand the effective sample points. Use the superpixel segmentation of simple linear clustering analysis to segment the multi-beam backscatter intensity image, combine it with the in-situ sampling points to obtain the sample-superpixel block, and use the high-contribution acoustic features corresponding to each pixel in the sample-superpixel block combined with the adaptive threshold method to effectively expand the seabed sample points and obtain the sample set. Step 3, multi-classification model construction, using half of the expanded sample set as the training set to construct a seabed sediment classification model with random forest, K nearest neighbor, decision tree, naive Bayes, BP neural network, and support vector machine classifiers; Step 4: Build a decision fusion mechanism. The other half is used as a validation set and combined with the confusion matrix to evaluate the accuracy of the model. Use the classification results of each model and the producer accuracy to create a multi-classifier decision fusion mechanism model to predict the seabed sediment type and obtain the seabed sediment prediction result map.

[0006] In step 1, the multi-scale acoustic features are preferably evaluated, including: Step 1.1, multi-scale acoustic feature extraction, using sliding windows of different sizes to extract the mean, dissimilarity, homogeneity, correlation, contrast, entropy, variance, and angular second-order moment texture features of the multi-beam backscatter intensity image, and extract the slope aspect, slope, plane curvature, profile curvature, and roughness terrain features of the multi-beam intensity image; Step 1.2, K-means clustering, combined with the types of seafloor sampling points, set its value as the K value, and input the acoustic features of different scales to perform unlabeled clustering of multi-scale features using K-means; Step 1.3, predict the label and determine it. In each iteration, according to the location of the same type of bottom sediment samples, obtain the labels of the surrounding 3×3 neighborhood, and count the labels with the largest number. The label type can be determined as the bottom sediment type. Step 1.4, clustering accuracy evaluation, take clustering accuracy P n The scale corresponding to the highest clustering result is taken as the optimal scale feature; ; In the formula, x i is the label after clustering; y i is the true label; n is the number of original samples; φ( x , y ) is the discriminant function, if x = y ,but φ =1, if x ≠y, then φ =0; Step 1.5, multi-scale acoustic feature evaluation, input the optimal scale features corresponding to the in-situ sampling points into the geographically weighted regression analysis model respectively, obtain the determination coefficient of the response, and normalize the determination coefficient to obtain the contribution of each feature, remove the feature types with low contribution, and finally obtain the acoustic features with high contribution.

[0007] In step 2, the effective sample points are expanded, including: Step 2.1, superpixel segmentation, using a simple linear iterative clustering superpixel segmentation method to segment the backscatter intensity image to obtain superpixel blocks with approximately uniform interior and compact connections; Step 2.2, sample-superpixel block acquisition, multiply the pre-made binary mask image with the segmented image to obtain the superpixel block with the original sample point; Step 2.3, reference threshold calculation, normalize each high contribution feature, calculate the feature distance between the original sample point and other points in the sample-superpixel block, obtain the distribution interval of the feature distance through the feature histogram, select the distribution interval with the largest frequency, and average all feature distances less than this interval to obtain the sample enhancement reference threshold; Step 2.4, true threshold calculation, obtains the true threshold by adjusting the reference thresholds in different sample-superpixel blocks, so as to balance the expansion quantity of each substrate sample and achieve effective expansion of substrate sample points.

[0008] In step 3, the multi-classification model is constructed, including: Step 3.1, training each classification model, using different machine learning algorithms to train multiple classification models, setting appropriate parameters for random forest, decision tree, K nearest neighbor, naive Bayes, support vector machine and BP neural network classifiers, and building each multi-beam seafloor sediment classification model in combination with training samples; Step 3.2: Each model predicts the classification results. Each model predicts the distribution of seafloor sediments based on the characteristics of the study area and outputs the classification results.

[0009] In step 4, the decision fusion mechanism is constructed, including: Step 4.1: majority voting. Integrate the prediction results of each model and obtain the probability of the same pixel being classified into different bottom types under different classification models through majority voting. Step 4.2, producer accuracy calculation, predict the category of the verification sample according to the obtained supervised classification model, and then obtain the number of samples correctly identified as this category, and compare it with the number of all samples actually belonging to this category, so as to obtain its producer accuracy, and then remove the producer accuracy with lower accuracy in each classification model and take the average as the weight of the bottom type decision; Step 4.3, calculate the decision value, use the calculated mean of producer precision for each category as the weight, adjust the classification decision probability for each category under majority voting, and thus obtain the decision value; Step 4.4, output of the bottom sediment type, the category with the highest decision threshold is used as the final predicted category of the pixel.

[0010] Beneficial technical effects brought by the present invention: The present invention discloses a multi-beam seabed sediment classification method based on a multi-classifier decision fusion mechanism, and systematically studies the multi-beam seabed sediment classification method from three perspectives: feature extraction, feature optimization, and classification model construction. Compared with the existing methods, the further improvement and optimization of the multi-scale feature extraction and optimization method and the sample enhancement method based on simple linear iterative clustering superpixel segmentation in this paper effectively improve the characterization ability of acoustic features for the sediment; the decision mechanism model for constructing a multi-classification model by integrating producer precision is fully utilized to give full play to the prediction ability of each classification model for different seabed sediments, and a higher prediction accuracy is obtained, which provides systematic support for the multi-beam seabed sediment classification system. The invention can play an important role in offshore marine engineering construction, marine resource management and development, and marine surveying and mapping. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 It is the multi-beam backscatter intensity image and bottom sampling distribution map in the present invention.

[0012] Figure 2 It is the multi-beam water depth image in the present invention.

[0013] Figure 3 This is the process of the multi-beam seafloor sediment classification method based on the multi-classifier decision fusion mechanism in the present invention.

[0014] Figure 4 This is the result diagram of multi-scale texture feature extraction in the present invention.

[0015] Figure 5 This is the multi-scale terrain feature extraction result map in the present invention.

[0016] Figure 6 This is the bottom sediment classification determination diagram in the present invention.

[0017] Figure 7 This is the superpixel segmentation effect diagram in the present invention.

[0018] Figure 8 This is the reference threshold sample expansion result in the present invention.

[0019] Fig. 9 This is the effective sample expansion result in the present invention.

[0020] Fig.10 Schematic diagram of the majority voting mechanism in the present invention.

[0021] Fig.11 This is a diagram of the seabed sediment classification results based on the multi-classifier decision fusion mechanism in the present invention. DETAILED DESCRIPTION

[0022] The present invention will be described in detail below with reference to the accompanying drawings and embodiments.

[0023] Here, the multi-beam backscatter intensity images, water depth images, and the locations and types of in-situ bottom sampling points collected by a certain country's geological survey in a certain sea area are used. The image resolution is 8m. According to the Folk classification nomenclature, the 60 groups of bottom sediments collected are divided into 8 categories: 11 groups of gravelly muddy sand (gmS), 14 groups of gravelly sand (gS), 5 groups of muddy sand (mS), 2 groups of muddy sandy gravel (msG), 4 groups of sand (S), 4 groups of sandy gravel (sG), 9 groups of gravelly muddy sand ((g)mS), and 11 groups of gravelly sand ((g)S). Figure 1 and Figure 2 The specific process is as follows Figure 3 As shown in FIG, the method comprises the following steps: step 1, multi-scale acoustic feature optimization evaluation, using sliding windows of different scales to extract texture features and terrain features from the input multi-beam backscatter intensity image and depth image, respectively, using the K-means clustering method to perform precision analysis on the corresponding features of each scale to select the acoustic features of the optimal scale, and combining the geographically weighted regression analysis method to obtain acoustic features with high contribution; step 2, effective sample point expansion, using the superpixel segmentation of the multi-beam backscatter intensity image by simple linear clustering analysis, combining the in-situ sampling points to obtain the sample-superpixel block, and through each image in the sample-superpixel block The high-contribution acoustic features corresponding to the elements are combined with the adaptive threshold method to achieve effective expansion of the seabed sample points and obtain the sample set; Step 3, multi-classification model construction, half of the expanded sample set is used as the training set to build the seabed sediment classification model with random forest, K nearest neighbor, decision tree, naive Bayes, BP neural network, and support vector machine classifiers; Step 4, decision fusion mechanism construction, the other half is used as the verification set combined with the confusion matrix to evaluate the accuracy of the model, and the classification results of each model and the producer accuracy are used to create a multi-classifier decision fusion mechanism model to predict the seabed sediment type and obtain the seabed sediment prediction result map.

[0024] Step 1: Multi-scale acoustic feature optimization evaluation.

[0025] Step 1.1, first set sliding windows of different sizes, mainly including 3×3, 5×5, 7×7, and 9×9, with the step size set to 1, extract the mean, dissimilarity, homogeneity, correlation, contrast, entropy, variance, angular second-order moment and other texture features of the multi-beam backscattering intensity image, extract the slope aspect, slope, plane curvature, profile curvature, roughness and other terrain features of the multi-beam intensity image, and the extraction results are shown in Figure 1. Figure 4 and Figure 5 As shown; Step 1.2, K-means clustering. Combined with the types of seafloor sampling points in the experimental area, the K value is set to 8. At the same time, the acoustic features of different scales (single features) are input and K-means is used to perform label-free clustering of multi-scale features. Step 1.3, predict the label to determine, K-means predicted labels are uncertain, that is, the label value will change with the change of the cluster center, and the corresponding substrate type is also unknown, so it is necessary to obtain the labels of the surrounding 3×3 neighborhood according to the location of the same type of substrate samples in each iteration, and count the labels with the largest number, then the label type can be determined as this type of substrate type, such as Figure 6 As shown; Step 1.4, clustering accuracy evaluation, using clustering accuracy P n To evaluate the clustering effect, for the same texture feature, eigenvalues ​​are extracted at different scales. If the eigenvalues ​​themselves are not very different, the clustering accuracy will be poor, and the corresponding features will not be suitable for seabed sediment prediction. Finally, the scale corresponding to the clustering result with the highest accuracy is taken as the optimal scale feature, mainly including 3×3 mean, 3×3 dissimilarity, 9×9 homogeneity, 7×7 correlation, 3×3 contrast, 3×3 entropy, 9×9 variance, 3×3 angular second-order moment, 7×7 aspect, 3×3 slope, 7×7 plane curvature, 9×9 profile curvature, 5×5 roughness and other features; ; In the formula, x i is the label after clustering; y i is the true label; n is the number of original samples; φ ( x , y ) is the discriminant function, if x = y ,but φ =1, if x ≠y, then φ =0.

[0026] Step 1.5, multi-scale acoustic feature evaluation, the optimal scale features corresponding to the in-situ sampling points are input into the geographically weighted regression analysis model respectively to obtain the determination coefficient of the response, and the determination coefficient is normalized to obtain the contribution of each feature, among which the corresponding contributions are obtained: intensity 17.15%, mean 8.92%, dissimilarity 8.90%, homogeneity 5.99%, correlation 6.54%, contrast 5.42%, entropy 7.82%, variance 4.16%, angular second moment 7.77%, water depth 8.56%, aspect 5.12%, slope 5.02%, plane curvature 2.80%, profile curvature 2.65%, roughness 3.19%, and the feature types with low contribution (<5%) are removed to finally obtain the acoustic features with high contribution.

[0027] Step 2: Expand the valid sample points.

[0028] Step 2.1, superpixel segmentation, use the simple linear iterative clustering superpixel segmentation method to segment the backscatter intensity image to obtain superpixel blocks with approximately uniform internal structure and compact connections. Set the number of superpixel blocks to 4000, as shown in Figure 7 As shown; Step 2.2, sample-superpixel block acquisition, multiply the pre-made binary mask image with the segmented image to obtain the superpixel block with the original sample point; Step 2.3, reference threshold calculation, normalize each high contribution feature, calculate the feature distance between the original sample point and other points in the sample-superpixel block, obtain the distribution interval of the feature distance through the feature histogram, select the distribution interval with the largest frequency, and average all feature distances less than this interval to obtain the sample enhancement reference threshold; Step 2.4, calculation of the real threshold. The number of various types of substrate in the actual sample may not be consistent. If the sample is expanded only by the reference threshold, a long-tail effect will occur, that is, the number of samples of a certain substrate is too large or too small, and the expansion result will be Figure 8 Therefore, it is necessary to obtain the true threshold by adjusting the reference threshold in different sample-superpixel blocks to balance the number of expansions of each substrate sample and achieve effective expansion of the substrate sample points. The 60 groups of sample points are expanded to 8269 groups. The expansion results are shown in Fig. 9 shown.

[0029] Step 3, the multi-classification model construction includes: Step 3.1, train each classification model. Use different machine learning algorithms to train multiple classification models. Randomly select 1 / 2 of the expanded sample points of each category, a total of 4136 groups as training samples, and the remaining 4133 groups of sample points as test samples. Appropriate parameters were set for classifiers such as random forest, decision tree, K nearest neighbor, naive Bayes, support vector machine and BP neural network, and various multi-beam seafloor sediment classification models were constructed in combination with training samples. The number of decision trees in the random forest was set to 100, the number of features for building the decision tree was set to 8, and the Gini coefficient was used as the basis for decision tree bifurcation; the K value of the K nearest neighbor was set to 8, the Euclidean distance was selected as the distance metric, the weight was set to equal weight, and the KD tree was selected as the nearest neighbor algorithm; the minimum parent node in the decision tree was set to 100, the minimum number of samples of the leaf node was set to 5, and the Gini impurity was selected as the node splitting standard; the probability distribution of each feature in the naive Bayes was set to kernel density estimation, the Gaussian kernel was selected, and the prior probability of each class was selected to be equal to the prior probability of all classes; the number of hidden layer neurons in the BP neural network was set to 10, the Levenberg-Marquardt algorithm was selected as the training function, the hyperbolic tangent function was selected as the activation function, and the number of iterations was selected to be 1000 times; the kernel function in the support vector machine was set to Gaussian kernel, and the scale parameter of the kernel function was automatically adjusted.

[0030] Step 3.2: Each model predicts the classification results. Each model predicts the distribution of seafloor sediments based on the characteristics of the study area and outputs the classification results.

[0031] Step 4, constructing the decision fusion mechanism, including: Step 4.1: majority voting. Integrate the prediction results of each model and obtain the probability of the same pixel being classified into different sediment types under different classification models through majority voting, such as Fig.10 As shown in the figure, after majority voting, the probability of the same pixel being predicted as substrate type 1 is 0.66, the probability of being predicted as substrate type 3 is 0.17, and the probability of being predicted as substrate type 7 is 0.17. ; Step 4.2, producer accuracy calculation, predict the category of the validation sample according to the obtained supervised classification model, and then get the number of samples correctly identified as this category, and compare it with the number of all samples actually belonging to this category, so as to obtain its producer accuracy, and remove the producer accuracy with lower accuracy in each classification model and take the average as the weight of the bottom sediment type decision, among which, the production accuracy means of gravelly muddy sand, gravelly sand, muddy sand, muddy sandy gravel, sand, sandy gravel, gravelly muddy sand, and gravel sand are 78.82%, 82.54%, 84.95%, 84.57%, 81.28%, 85.44%, 84.47%, and 87.28%, respectively; Step 4.3, calculate the decision value, use the calculated producer precision mean of each category as the weight, adjust the classification decision probability of each category under majority voting, and thus obtain the decision value. That is, for categories with the same decision probability, the model's recognition ability for these categories can be improved by multiplying them with the producer precision weight; Step 4.4, output of the bottom sediment type. The category with the highest decision threshold is used as the final predicted category of the pixel. The distribution of bottom sediment types is as follows: Fig.11 As shown in the figure, the accuracy of the proposed method was evaluated by combining it with the validation set, and an overall classification accuracy of 91.17% and a Kappa coefficient of 0.90 were obtained.

[0032] The above-mentioned embodiments only express several implementation methods of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the invention. It should be pointed out that for ordinary technicians in this field, several modifications and improvements can be made without departing from the concept of the present invention, all of which belong to the protection scope of the present invention. The protection scope of the present invention is given by the attached claims and any equivalent technical solutions thereof.

Claims

1. A multi-beam seafloor sediment classification method based on a multi-classifier decision fusion mechanism, characterized in that: The method includes: Step 1: Multi-scale acoustic feature optimization evaluation: texture feature extraction and terrain feature extraction are performed on the input multi-beam backscatter intensity image and depth image through sliding windows of different scales. The K-means clustering method is used to perform precision analysis on the corresponding features of each scale to select the acoustic features of the optimal scale, and the geographically weighted regression analysis method is combined to obtain the acoustic features with high contribution. Step 2: Expand the effective sample points. Use the superpixel segmentation of simple linear clustering analysis to segment the multi-beam backscatter intensity image, combine it with the in-situ sampling points to obtain the sample-superpixel block, and use the high-contribution acoustic features corresponding to each pixel in the sample-superpixel block combined with the adaptive threshold method to effectively expand the seabed sample points and obtain the sample set. Step 3, multi-classification model construction, using half of the expanded sample set as the training set to construct a seabed sediment classification model with random forest, K nearest neighbor, decision tree, naive Bayes, BP neural network, and support vector machine classifiers; Step 4: Build a decision fusion mechanism. The other half is used as a validation set and combined with the confusion matrix to evaluate the accuracy of the model. Use the classification results of each model and the producer accuracy to create a multi-classifier decision fusion mechanism model to predict the seabed sediment type and obtain the seabed sediment prediction result map.

2. The multi-beam seafloor sediment classification method based on a multi-classifier decision fusion mechanism according to claim 1 is characterized in that: In step 1, the multi-scale acoustic features are preferably evaluated, including: Step 1.1, multi-scale acoustic feature extraction, using sliding windows of different sizes to extract the mean, dissimilarity, homogeneity, correlation, contrast, entropy, variance, and angular second-order moment texture features of the multi-beam backscatter intensity image, and extract the slope aspect, slope, plane curvature, profile curvature, and roughness terrain features of the multi-beam intensity image; Step 1.2, K-means clustering, combined with the types of seafloor sampling points, set its value as the K value, and input the acoustic features of different scales to perform label-free clustering of multi-scale features using K-means; Step 1.3, predict the label and determine it. In each iteration, according to the location of the same type of bottom sediment samples, obtain the labels of the surrounding 3×3 neighborhood, and count the labels with the largest number. The label type can be determined as the bottom sediment type. Step 1.4, clustering accuracy evaluation, take clustering accuracy P n The scale corresponding to the highest clustering result is taken as the optimal scale feature; ； In the formula, x i is the label after clustering; y i is the true label; n is the number of original samples; φ ( x , y ) is the discriminant function, if x = y ,but φ =1, if x ≠y, then φ =0; Step 1.5, multi-scale acoustic feature evaluation, input the optimal scale features corresponding to the in-situ sampling points into the geographically weighted regression analysis model respectively, obtain the determination coefficient of the response, and normalize the determination coefficient to obtain the contribution of each feature, remove the feature types with low contribution, and finally obtain the acoustic features with high contribution.

3. The multi-beam seafloor sediment classification method based on a multi-classifier decision fusion mechanism according to claim 1 is characterized in that: In step 2, the effective sample points are expanded, including: Step 2.1, superpixel segmentation, using a simple linear iterative clustering superpixel segmentation method to segment the backscatter intensity image to obtain superpixel blocks with uniform interior and compact connections; Step 2.2, sample-superpixel block acquisition, multiply the pre-made binary mask image with the segmented image to obtain the superpixel block with the original sample point; Step 2.3, reference threshold calculation, normalize each high contribution feature, calculate the feature distance between the original sample point and other points in the sample-superpixel block, obtain the distribution interval of the feature distance through the feature histogram, select the distribution interval with the largest frequency, and average all feature distances less than this interval to obtain the sample enhancement reference threshold; Step 2.4, true threshold calculation, obtains the true threshold by adjusting the reference thresholds in different sample-superpixel blocks, so as to balance the expansion quantity of each substrate sample and achieve effective expansion of substrate sample points.

4. The multi-beam seafloor sediment classification method based on a multi-classifier decision fusion mechanism according to claim 1 is characterized in that: In step 3, the multi-classification model is constructed, including: Step 3.1, training each classification model, using different machine learning algorithms to train multiple classification models, setting appropriate parameters for random forest, decision tree, K nearest neighbor, naive Bayes, support vector machine and BP neural network classifiers, and building each multi-beam seafloor sediment classification model in combination with training samples; Step 3.2, each model predicts the classification results. Each model predicts the distribution of seabed sediments in combination with the characteristics of the study area and outputs the classification results.

5. The multi-beam seafloor sediment classification method based on a multi-classifier decision fusion mechanism according to claim 1 is characterized in that: In step 4, the decision fusion mechanism is constructed, including: Step 4.1: majority voting, integrate the prediction results of each model, and obtain the probability of the same pixel being classified into each sediment type under different classification models by majority voting; Step 4.2, producer accuracy calculation, predict the category of the verification sample according to the obtained supervised classification model, and then obtain the number of samples correctly identified as this category, and compare it with the number of all samples actually belonging to this category, so as to obtain its producer accuracy, and then remove the producer accuracy with lower accuracy in each classification model and take the average as the weight of the bottom type decision; Step 4.3, calculate the decision value, use the calculated mean of producer precision for each category as the weight, adjust the classification decision probability for each category under majority voting, and thus obtain the decision value; Step 4.4, output of the bottom sediment type, the category with the highest decision threshold is used as the final predicted category of the pixel.

Images