Deep sea manganese nodule coverage rate acoustic characterization method based on multi-classifier decision fusion

Abstract

The invention belongs to the technical field of deep sea mineral resource exploration, and discloses a deep sea manganese nodule coverage rate acoustic characterization method based on multi-classifier decision fusion, and the method comprises the steps: extracting 14 backscattering texture features and 4 submarine topographic features, carrying out the optimization of the extracted features through employing a Boruta algorithm, and constructing a feature set for the refined characterization of the seabed; an iterative robust estimation algorithm based on a sliding window is introduced, observation weights are dynamically adjusted through continuous smooth data in different directions, and gross errors in feature images are recognized and suppressed; in the model construction stage, a Stacking mechanism in ensemble learning is adopted, the advantages of a plurality of classifiers are fused to carry out decision hierarchy integration, and a spatial distribution diagram of the tuberculosis coverage rate is generated. According to the method, a Boruta feature selection algorithm is introduced, key features with the strongest tuberculosis coverage rate discrimination capability are screened out from original acoustic features, redundant interference is avoided, and the precision and prediction discrimination capability of the model are improved.

Description

Technical Field

[0001] This invention belongs to the field of deep-sea mineral resource exploration technology, and in particular relates to an acoustic characterization method for deep-sea manganese nodule coverage based on multi-classifier decision fusion. Background Technology

[0002] To gain a deeper understanding of the distribution characteristics of manganese nodules on the deep-sea seabed, obtaining high-quality seabed sediment information is crucial. With the development of the industry, survey methods for benthic mineral resources are constantly evolving. Direct sampling (such as using a grab bucket for sample acquisition) is the most direct and effective method for obtaining relevant seabed information. However, in the deep-sea environment, large-scale data acquisition through direct sampling faces significant challenges. On the one hand, data collection at each sampling point is time-consuming, resulting in low overall sampling efficiency; on the other hand, large-scale, high-density sampling may cause irreversible damage to the seabed ecosystem. The emergence of integrated optical towed bodies (near-bottom photography and video) provides a more efficient and refined solution for seabed sediment surveys. Continuous photography or video recording during the towed body's movement can obtain a large amount of high-definition sampling information. However, compared to shallow water environments, the physical characteristics of the deep sea limit the coverage area of ​​optical towed bodies (only a few meters). Their survey area is usually limited to a cross-section along the survey line, thus affecting the ability to conduct comprehensive seabed sediment surveys. Compared to the methods mentioned above, underwater acoustic methods offer a cost-effective and efficient means of seabed topographic measurement. Shipborne multibeam echo sounders (MBES) can cover a large area of ​​the seabed (generally 3-4 times the water depth). Because they can simultaneously record the seabed topography and backscatter intensity, they have the characteristics of high precision, high efficiency and full coverage. They have now become a mainstream choice for drawing seabed topographic maps, detecting seabed sediment types, detecting seabed plumes, and detecting benthic organisms.

[0003] Deep-sea manganese nodules are rich in various rare metals and are important mineral resources for alleviating the supply and demand imbalance of terrestrial resources and supporting energy transition. Besides their metallic content, manganese nodules are also considered core to the deep-sea ecosystem. They are among the few hard substrates in the typical large-scale scenario of soft siliceous clay in deep-sea basins. Studies suggest that the nodule coverage on the seabed is closely related to the abundance, spatial distribution, species richness, and community composition of benthic animals and microorganisms. Manganese nodule coverage is a key indicator for resource assessment, and high-precision nodule coverage prediction is crucial for resource evaluation and development. However, due to the complexity and spatial heterogeneity of the deep-sea environment, traditional surface sediment and manganese nodule detection mainly focuses on attribute classification, and quantifying resource assessment indicators such as the coverage of polymetallic nodules on the seabed remains challenging.

[0004] Based on the above analysis, the problems and shortcomings of the existing technology are as follows: (1) In the traditional method of large-scale data acquisition through direct sampling, the data collection time for each sampling point is long, the overall sampling efficiency is low, and large-scale high-density sampling may cause irreversible damage to the seabed ecological environment; while the coverage of integrated optical towed bodies is extremely limited, and their survey area is usually limited to the cross section along the survey line, thus affecting the ability to conduct full-coverage surveys of seabed sediments.

[0005] (2) Acoustic detection results are limited to binary discrimination and lack quantitative characterization of coverage. Shipborne multibeam echo sounders (MBES) have become the mainstream detection method, providing data on water depth and backscatter intensity over a wide range. However, most existing research in the field of seabed classification / resource exploration remains at the binary classification of "presence / absence," which cannot achieve quantitative analysis of coverage and thus cannot support refined resource reserve assessment and ecological environment protection. Moreover, existing classification modeling relies on a single model, resulting in poor stability. Existing studies mostly use a single classifier for prediction, and different algorithms show significant differences in performance under complex seabed environments, lacking stability and robustness. Single models rely excessively on specific assumptions or sample distributions, resulting in strong bias in prediction results and difficulty in adapting to the spatial heterogeneity of manganese nodule coverage.

[0006] (3) Insufficient feature construction and optimization. Existing methods have limitations in utilizing acoustic features, often neglecting the combined effect of multi-dimensional information such as backscatter texture and seabed topography; at the same time, feature selection relies heavily on human experience, which can easily introduce redundancy and noise, weakening the model's discriminative ability. The lack of a systematic feature optimization mechanism leads to poor stability and generalization of prediction results.

[0007] (4) Noise and gross errors are difficult to suppress effectively. In the deep-sea environment, there are ship noise, reverberation, bubbles and biological activity, which often cause gross errors and outliers in the observation of water depth and backscatter intensity. Existing methods are mostly unidirectional or local processing, lack robustness, and are difficult to completely eliminate the impact of gross errors on prediction. Summary of the Invention

[0008] To overcome the problems existing in related technologies, the present invention discloses an acoustic characterization method for the coverage of deep-sea manganese nodules based on multi-classifier decision fusion, the technical solution of which is as follows: This invention is implemented as follows: an acoustic characterization method for deep-sea manganese nodule coverage based on multi-classifier decision fusion, comprising the following steps: S1, based on the Kongsberg EM122 multibeam echo sounder, acquires water depth and backscatter intensity data. It combines this with a 6000-meter-class integrated optical towed body to obtain near-bottom images. The near-bottom optical towed body is positioned using an ultra-short baseline (USBL) combined inertial navigation system (INS) and a Doppler velocity recorder (DVL) to locate underwater optical sampling points. The shipborne multibeam echo sounder acquires a precise underwater terrain model via a Global Navigation Satellite System (GNSS) combined inertial navigation system (INS). By constructing the sampling points and the ship-based terrain using the same projection method, water depth and backscatter intensity data at the sampling point locations are obtained. S2, by extracting 14 backscatter texture features and 4 seabed topographic features, the Boruta algorithm is used to optimize the extracted features and construct a feature set that finely represents the seabed. S3 introduces an iterative robust estimation algorithm based on a sliding window, which dynamically adjusts the observation weights by continuously smoothing data in different directions to identify and suppress gross errors in the feature image; S4. In the model building stage, the Stacking mechanism in ensemble learning is adopted to integrate the advantages of multiple classifiers at the decision level and generate a spatial distribution map of nodule coverage.

[0009] The base classifiers are trained using a k-fold cross-validation strategy to ensure the robustness of the predicted meta-features. Each base classifier generates out-of-fold results through cross-prediction on the training set. These results are then used as new features input to the secondary classifier (Gradient Boosting), thereby achieving non-linear fusion at the decision layer. The final output is a discrete coverage level (low, medium, high), presenting the spatial distribution of deep-sea manganese nodules in the form of a classification map.

[0010] In step S2, the backscattering texture features include water depth, slope, curvature, slope aspect, and backscattering intensity. , , , , , , , Standard deviation, kurtosis, skewness, energy, entropy, and seafloor topographic features include depth, slope, aspect, and planar curvature.

[0011] In step S2, the Boruta algorithm is used to optimize the features, including: The Boruta algorithm expands the feature space by adding random attributes, thereby eliminating the correspondence between sample attribute values ​​and labels; it then uses the expanded feature space to perform classification tests and calculates the importance index of all attributes; finally, it uses the importance index of shadow attributes as a reference to evaluate the importance of each attribute. The Boruta algorithm is used to evaluate the importance of each feature variable, including: (1) Initialization and data preparation: The dataset is ,in, It is the first One characteristic, It is the total number of features; the dataset also contains the target variable. , used to represent the label of each sample; (2) Generating image features: The Boruta algorithm generates a set of random "image features" to simulate noise, thereby determining whether the real features are important; for each original feature Generate an image feature Image features The value is randomly shuffled The sample labels in the data are obtained and represent only noise; (3) Merging original features and image features: Merging original features and image features The features are merged to form a dataset containing both the original features and image features. Use the merged feature set Train a random forest model to estimate the importance of each feature; the random forest model outputs an importance score for each feature and compares the importance of the original features with that of the image features.

[0012] Furthermore, the Boruta algorithm uses the Gini index to measure the importance of features; For each feature Train a random forest model and calculate its importance to determine the target variable. The contribution of the prediction; the random forest model evaluates the importance of features based on the contribution of each feature.

[0013] Furthermore, the importance of comparing raw features with image features in random forest models includes: Calculate the maximum importance of all image features : ; For each original feature , will be important and Compare; if If the feature is considered important, then it is considered important; otherwise, it is considered useless. The Boruta algorithm runs multiple times, iterating through the image feature importance comparison steps until the classification of each feature is stable.

[0014] In step S3, the iterative robust estimation based on the sliding window includes: (1) Slide a window of a specified size on the feature image; (2) In each window, local pixel sequences are extracted along multiple directions (including horizontal, vertical and diagonal directions); by constructing one-dimensional pixel chains in different directions, the spatial correlation of acoustic scattering signals in the nodule coverage area is utilized to enhance the sensitivity of the feature extraction process to spatial texture and nodule edge structure.

[0015] (3) Based on the Huber loss function, weights are assigned to the observation points in the pixel sequence of each direction, and the local regression model parameters of the direction are updated by iterative weighted least squares until convergence, so as to obtain the robust estimation result of the direction; robust regression is performed on the sequence data of each direction, and the Huber function is used to reduce the gross error caused by noise pulses, deep water bubbles or local abnormal reflections, so as to make the estimation results of the tuberculosis coverage related features more stable.

[0016] (4) Calculate the directional weights dynamically based on the robust standard deviation of the residuals in each direction, and perform weighted fusion of the estimation results in different directions to obtain the final robust estimate of the center pixel of the window; (5) Move the window to the next position and repeat steps (1)-(4) until the entire feature image is traversed; through multi-directional weighted residuals and iterative optimization process, robust parameter estimates that are insensitive to noise and gross errors are obtained.

[0017] The iterative robust estimation algorithm based on sliding windows: The core idea is to use robust estimation (Huber) to fit a local model in each window, predict the estimated value of the window center point and replace the center pixel. (1) For pixels Define a local sliding window with a fixed size of 5×5 and a step size of 1 pixel (processed pixel by pixel) to ensure that the boundary prediction is concentrated and the coverage is complete. (2) Extract the 1D pixel chain containing the center pixel (length is the same as the window size, which is 5 pixel units in this method), along... For each chain, a robust regression is performed in four directions using the Huber function as the kernel function to obtain the center estimate for that direction. and the residual scale in that direction .

[0018] (3) Take a weighted average of the four estimates obtained from the above multiple directions as the final output; the expression is: ; in, For in position Location, along direction The weighting coefficients, It is a very small positive number (regularization term) to prevent the denominator from being zero; For in position The final estimated value at that location, To normalize the weights, the sum of the weights in all directions is guaranteed to be 1.

[0019] In step S4, the multi-classifier decision fusion based on the Stacking mechanism includes: Five supervised classification algorithms were selected as first-layer classifiers, and each was trained and predicted independently on the data. The prediction results of the first-layer models were used to evaluate the individual performance of each model and were passed as input features to the second-layer model. A performance screening threshold was set. If the prediction accuracy of a first-layer classifier was lower than the threshold, the prediction result was not passed to the meta-learning model of the second layer. A gradient boosting classifier was used as the final decision model to integrate the prediction information of the first-layer classifiers.

[0020] Furthermore, in spatial prediction of tuberculosis coverage, the first layer of the Stacking ensemble learning method classifies each pixel separately, resulting in several independent prediction results. After filtering out low-precision models through a performance screening mechanism, the remaining prediction results are combined to form a new feature set, which is then input into the meta-model for final decision-making. This decision-making process is based on the majority voting principle to ensure that the final category determination integrates the advantages of different classifiers.

[0021] Furthermore, supervised classification algorithms include random forests, decision trees, backpropagation networks, support vector machines, and K-nearest neighbors classifiers.

[0022] Another objective of this invention is to provide an acoustic characterization system for deep-sea manganese nodule coverage based on multi-classifier decision fusion. This system is used to regulate the aforementioned acoustic characterization method for deep-sea manganese nodule coverage. The system includes: The data acquisition module is used to acquire water depth and backscatter intensity data based on the Kongsberg EM122 multibeam echo sounder system. It is combined with a 6000-meter-class integrated optical towed body to acquire near-bottom images. The positioning of the near-bottom optical towed body is obtained by using an ultra-short baseline USBL combined inertial navigation system (INS) and a Doppler velocity recorder (DVL) to obtain the location of underwater optical sampling points. The shipborne multibeam echo sounder acquires a precise underwater terrain model through a Global Navigation Satellite System (GNSS) combined inertial navigation system (INS). By constructing the sampling points and the ship-based terrain using the same projection method, the water depth and backscatter intensity data at the location of the sampling points are acquired. The feature optimization module extracts 14 backscatter texture features and 4 seabed topographic features, and uses the Boruta algorithm to optimize the extracted features, thereby constructing a refined feature set representing the seabed. The iterative robust estimation module introduces an iterative robust estimation algorithm based on a sliding window. By dynamically adjusting the observation weights through continuously smoothed data in different directions, it identifies and suppresses gross errors in the feature image. The multi-classifier decision fusion module adopts the Stacking mechanism in ensemble learning to integrate the advantages of multiple classifiers at the decision level and generate a spatial distribution map of nodule coverage.

[0023] Combining all the above technical solutions, the beneficial effects of this invention are as follows: First, this invention focuses on high-precision, continuous spatial prediction of deep-sea polymetallic nodule coverage. It constructs a prediction framework integrating multi-source data and robust modeling, with a particular emphasis on improving the spatial interpretation capability of manganese nodule coverage. This invention is based on water depth and backscatter intensity data acquired by the Kongsberg EM122 multibeam system, combined with near-bottom images acquired by a 6000-meter-level integrated optical towed body. It extracts 14 backscatter texture features and 4 seabed topographic features, and performs feature optimization on the extracted features to construct a refined feature set representing the seabed. A robust estimation method is introduced to identify and suppress gross errors in the feature images, improving data robustness and usability. In the model building stage, the Stacking mechanism from ensemble learning is used to integrate the advantages of multiple classifiers at the decision-making level, generating a spatial distribution map of nodule coverage.

[0024] Secondly, this invention introduces the Boruta feature selection algorithm to select the key features with the strongest discriminative ability for nodule coverage from the original acoustic features, thereby improving the model's accuracy and generalization ability. In the task of predicting nodule coverage in the deep sea, the Boruta feature selection algorithm can improve feature effectiveness, avoid redundant interference, and enhance the model's predictive and discriminative ability, providing a solid feature foundation for subsequent manganese nodule coverage prediction. Compared with the single classifier model optimized based on Boruta feature selection and robust estimation, the Stacking method improves prediction accuracy by 0.38% (compared to RF), 36.21% (compared to SVM), 12.49% (compared to BP neural network), and 4.32% (compared to KNN), respectively, further verifying the effectiveness and applicability of the ensemble strategy in predicting manganese nodule distribution in complex deep-sea environments.

[0025] Third, this invention enables high-precision quantitative characterization of deep-sea manganese nodule coverage under large-scale, non-contact conditions, providing an efficient and low-cost solution for international seabed resource exploration and development. Its application can not only significantly improve exploration efficiency and reduce sampling costs, but also support collaborative decision-making for deep-sea mineral resource reserve assessment and ecological protection. This invention is the first to achieve multi-level quantitative prediction of deep-sea manganese nodule coverage and proposes a systematic framework integrating Boruta feature selection, robust estimation, and Stacking ensemble learning.

[0026] Fourth, this invention effectively solves the problem of quantitative coverage prediction in high-noise environments by eliminating gross errors through robust estimation, selecting key features using Boruta fusion, and employing multi-classifier fusion, achieving a goal that was previously unattainable. Traditional methods generally rely on a single classifier or empirical feature selection, depending on specific model assumptions or subjective judgments, leading to significant prediction biases. This invention, by introducing multi-classifier stacking fusion and data-driven Boruta feature optimization, breaks through the limitations of single models and empirical judgments, overcomes traditional technical biases, and achieves more objective, robust, and universally applicable coverage prediction. Attached Figure Description

[0027] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure; Figure 1 This is a schematic diagram of an iterative robust estimation algorithm based on a sliding window provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the acoustic characterization method for deep-sea manganese nodule coverage based on multi-classifier decision fusion provided in this embodiment of the invention. Figure 3 This is an importance score diagram of different features provided in the embodiments of the present invention; Figure 4 This is a correlation diagram of the coverage characteristics of manganese nodules provided in an embodiment of the present invention; Figure 5(a) is an RF diagram of the predicted tuberculosis coverage value provided in the embodiment of the present invention; Figure 5(b) is an SVM diagram of the predicted tuberculosis coverage provided in the embodiment of the present invention; Figure 5(c) is a BP plot of the predicted tuberculosis coverage value provided in the embodiment of the present invention; Figure 5(d) is a KNN diagram of the predicted tuberculosis coverage value provided in the embodiment of the present invention; Figure 5(e) is a DT graph of the predicted tuberculosis coverage rate provided in the embodiment of the present invention; Figure 6(a) is an RF diagram of the tuberculosis coverage prediction value after robust estimation provided by the embodiment of the present invention; Figure 6(b) is an SVM diagram of the tuberculosis coverage prediction value after robust estimation provided by the embodiment of the present invention; Figure 6(c) is a BP plot of the tuberculosis coverage prediction value after robust estimation provided by the embodiment of the present invention; Figure 6(d) is a KNN diagram of the tuberculosis coverage prediction value after robust estimation provided by the embodiment of the present invention; Figure 6(e) is a DT plot of the tuberculosis coverage prediction value after robust estimation provided by the embodiment of the present invention; Figure 7 This is a graph showing the tuberculosis coverage prediction results using the Stacking mechanism provided in this embodiment of the invention. Detailed Implementation

[0028] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, 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 to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0029] Example 1: The methodology of this invention is mainly divided into two parts. The first part focuses on feature-oriented correlation analysis, specifically exploring feature mining and selection of acoustic data collected by MBES. To ensure the stability and accuracy of the subsequent classification process, this invention introduces a robust estimation method to remove gross errors in the data, thereby improving the quality of the feature data, based on the optimized features.

[0030] The second part focuses on estimating tuberculosis coverage. This invention first uses five classification methods for preliminary classification, and then combines this with a Stacking ensemble learning mechanism to integrate the advantages of different models, thereby improving the robustness and generalization ability of the prediction model. The overall technical process is as follows: Figure 1 As shown.

[0031] like Figure 2 As shown, the acoustic characterization method for deep-sea manganese nodule coverage based on multi-classifier decision fusion provided in this embodiment of the invention specifically includes the following steps: S1, based on the Kongsberg EM122 multibeam echo sounder, acquires water depth and backscatter intensity data. It combines this with a 6000-meter-class integrated optical towed body to obtain near-bottom images. The near-bottom optical towed body is positioned using an ultra-short baseline (USBL) combined inertial navigation system (INS) and a Doppler velocity recorder (DVL) to locate underwater optical sampling points. The shipborne multibeam echo sounder acquires a precise underwater terrain model via a Global Navigation Satellite System (GNSS) combined inertial navigation system (INS). By constructing the sampling points and the ship-based terrain using the same projection method, water depth and backscatter intensity data at the sampling point locations are obtained. Different features exhibit varying performance in tasks characterizing the spatial distribution of manganese nodules at different coverage rates. Therefore, multi-dimensional feature extraction is necessary to improve the accuracy and reliability of predictions. Backscattering intensity is closely related to the physical properties of seafloor sediments, such as roughness, grain size, porosity, saturation, and incident angle. Simultaneously, seafloor topography alters the hydrodynamic model within a region during nodule mineralization, thus influencing the spatial distribution pattern of nodules. Effectively integrating this information will significantly improve the accuracy and interpretability of nodule spatial distribution predictions. Therefore, before using MBES data for training and prediction of nodule spatial distribution information, systematic feature extraction is required. This invention has accumulated 18 dimensions of features, including 14 dimensions based on multibeam backscattering intensity and 4 dimensions based on seafloor topography. However, due to the complexity of the deep-sea environment, underwater acoustic signals are interfered with by various factors such as seafloor reverberation, ship noise, and biological activity noise during long-distance propagation, resulting in a large amount of noise in the data. Therefore, before feature extraction, all MBES data underwent manual noise removal, a step implemented using Caris HIPS 11.4 software. Table 1 details the extracted feature factors.

[0032] (1) Backscattering intensity characteristics: Backscattering intensity characteristics mainly include backscattering angle response curve characteristics and backscattering intensity image characteristics. This invention mainly focuses on backscattering intensity image characteristics, that is, sonar image texture characteristics. Texture characteristics can effectively distinguish the spatial distribution information of different coverage rates of nodules.

[0033] (2) Seafloor topographic features: MBES can acquire high-quality water depth data and seafloor topographic information, and features can be extracted from the processed DEM data. This invention extracts four seafloor topographic features from the water depth raster data, including depth, slope, aspect, and plane curvature.

[0034] Table 1. Summary of texture and terrain features extracted based on multibeam backscattering intensity and depth sounding data, and calculation methods for each feature.

[0035] S2, by extracting 14 backscatter texture features and 4 seabed topographic features, the Boruta algorithm is used to optimize the extracted features and construct a feature set that finely represents the seabed. Backscattering texture features include water depth, slope, curvature, slope aspect, and backscattering intensity. , , , , , , , Standard deviation, kurtosis, skewness, energy, entropy, and seafloor topographic features include depth, slope, aspect, and planar curvature.

[0036] In learning tasks, feature selection aims to filter out the most discriminative feature variables from all attributes for a specific task, thereby constructing a concise and efficient feature space. This process effectively eliminates redundant and erroneous features, avoiding the adverse effects of high-dimensional data on classifier performance. In this invention, feature engineering is particularly crucial for predicting the spatial distribution of manganese nodule coverage. The extracted acoustic feature datasets often suffer from high noise and nonlinearity; excessively high dimensionality or inaccurate features can weaken the performance of subsequent classifiers. Feature selection and optimization strongly influence the design and performance of the classifier. This invention selects the Boruta algorithm for feature selection, extracting the most representative feature variables to provide a solid data foundation for accurately predicting the spatial distribution of deep-sea manganese nodules.

[0037] The Boruta algorithm is a "fully relevant" feature selection method that identifies all predictor variables that may be relevant to classification. It is a wrapper-style supervised feature selection algorithm built on a random forest model. It expands the feature space by adding random attributes to eliminate the correspondence between sample attribute values ​​and labels. Then, it uses this expanded feature space for classification testing and calculates the importance index of all attributes. Due to random fluctuations, the importance index of shadow attributes may be non-zero. Using the importance index of shadow attributes as a reference can assess the importance of each attribute. Since the distribution level of the importance index varies with the randomness of the classifier and is related to the presence of unimportant features and the specific realization of shadow attributes, the process of assigning random values ​​to shadow attributes needs to be repeated multiple times to obtain statistically significant results. The steps for evaluating the importance of each feature variable using the Boruta algorithm are as follows: First, initialization and data preparation are required. The dataset is... ,in, It is the first One characteristic, It is the total number of features; the dataset also contains the target variable. , used to represent the label of each sample.

[0038] Next, image features need to be generated. Unlike traditional feature selection methods, Boruta generates a set of random "image features" to simulate noise, thereby determining whether the true features are important enough. For each original feature... Generate an image feature Its value is obtained by randomly shuffling The image features are obtained from sample labels. Therefore, the generated image features are informationless and should only represent noise.

[0039] Next, the original features and image features will be merged, and the original features will be merged... and image features The features are merged to form a dataset containing both the original features and image features. Use the merged feature set Train a random forest model to estimate the importance of each feature. Typically, Boruta uses the Gini index to measure feature importance. For each feature... This invention trains a random forest and calculates its importance (i.e., its importance to the target variable). (Contribution to prediction). Random forest models evaluate feature importance based on the contribution of each feature.

[0040] (1); In the Gini index, It is the number of categories. It is the first in this node The proportion of class samples.

[0041] For each feature Random forest is used to calculate their importance. Here, The model evaluates the features relative to the target. Gini impurity, i.e., a measure of contribution.

[0042] (2); The random forest outputs an importance score for each feature, and compares the importance of the original features with that of the image features. The specific steps are as follows: Calculate the maximum importance of all image features. : (3); For each original feature Its importance and Compare. If If a feature is deemed important, it is considered important; otherwise, it is considered useless. This process is iterated until the classification ("important" or "useless") of each feature stabilizes. Typically, Boruta is run multiple times to ensure the stability and accuracy of feature selection. Only important variables are retained for subsequent spatial distribution prediction of manganese nodule coverage.

[0043] S3 introduces an iterative robust estimation algorithm based on a sliding window, which dynamically adjusts the observation weights by continuously smoothing data in different directions to identify and suppress gross errors in the feature image; In deep-sea multibeam bathymetry, seabed reverberation and environmental noise inevitably produce gross errors in water depth and backscatter intensity observations. In modern measurement adjustment theory, if the gross errors manifest as significant discrepancies between the prior random model and the actual model, they can be classified as part of the random model, which can be interpreted as a variance inflation model. The underlying principle is to continuously adjust the weights or variances of the observations based on iterative adjustment results, ultimately causing the weights of observations containing gross errors to approach zero or the variance to approach infinity. This ensures that the estimated parameters are less affected by model errors, especially gross errors. Robust estimation is an effective method to address this problem. In existing techniques, Zhang et al. grouped BS data according to the incident angle and applied a sliding window to each group along the track direction using continuous pings for robust estimation, thus handling track inhomogeneities. While this method effectively removes gross errors from each group, due to the limitations of the grouping strategy, its gross error removal process mainly targets a single direction, neglecting harmful data interference from multiple directions in the surrounding area. Therefore, this invention further considers the potential impact of seabed gross errors along and across the course. To this end, this invention proposes an iterative robust estimation algorithm based on a sliding window. This method continuously smooths MBES observation data in different directions and dynamically adjusts the observation weights to reduce the impact of local outliers on the final estimation results, thereby achieving more comprehensive gross error removal.

[0044] The iterative robust estimation algorithm based on sliding windows: The core idea is to use robust estimation (Huber) to fit a local model in each window, predict the estimated value of the window center point and replace the center pixel. (1) For pixels Define a local sliding window with a fixed size of 5×5 and a step size of 1 pixel (processed pixel by pixel) to ensure that the boundary prediction is concentrated and the coverage is complete. (2) Extract the 1D pixel chain containing the center pixel (length is the same as the window size, which is 5 pixel units in this method), along... For each chain, a robust regression is performed in four directions using the Huber function as the kernel function to obtain the center estimate for that direction. and the residual scale in that direction .

[0045] (3) Take a weighted average of the four estimates obtained from the above multiple directions as the final output; the expression is: ; In the formula, For in position Location, along direction The weighting coefficients, It is a very small positive number (regularization term) to prevent the denominator from being zero; For in position The final estimated value at that location, To normalize the weights, the sum of the weights in all directions is guaranteed to be 1.

[0046] like Figure 1 As shown, in robust estimation, the robust estimation method adopted in this invention is Iteratively Reweighted Least Squares (IRLS). "Outlier removal" is not an independent step completely separate from the main process, but is implemented gradually through a weighting function during the iterative fitting process. Outliers are not hard-removed points during multiple iterations in the Huber function, but rather soft-removed. Points with residuals exceeding a threshold are not directly discarded, but assigned a very small weight.

[0047] With the development and progress of modern measurement adjustment theory, various robust estimation methods have been developed to handle potential outliers. When some observations contain gross errors, robust estimation is superior to least squares estimation. The fundamental difference between robust estimation and classical estimation theory lies in that robust estimation is based on the actual distribution pattern of the observed data, rather than on some ideal distribution pattern. This invention uses the M-estimation theory proposed by Huber in 1966, a method introduced into the measurement community by Krarup and Kubik of Denmark in 1980. First, the parameters of the regression model are initialized, and an initial parameter estimate can be obtained using the least squares (OLS) method. Let the model parameters be: (4); In the formula, It is the number of features of the model. It is the first The actual observed values ​​of each data point That is the corresponding predicted value.

[0048] In the initialization phase, this invention obtains the initial parameters by minimizing the traditional squared error: (5); In the formula, It is the first The feature vector of each data point.

[0049] The second step is to calculate the residual at each point within the sliding window. The residual is the difference between the actual value and the predicted value. It reflects the model's fitting error to the data. For each data point, the residual is calculated based on the current model's prediction.

[0050] (6); In the formula, It is the parameter estimate for the current iteration.

[0051] The third step is to introduce the Huber loss function to calculate the weighted residuals for each observation within the sliding window, assigning a weight to each residual. The Huber loss function calculates the weighted residuals as follows: (7); In the formula, It is the threshold that determines when to switch from squared loss to linear loss.

[0052] The fourth step is to update the weight matrix; the weighted residuals are used to update the weights of the regression model. Specifically, the weight matrix... It is a diagonal matrix, where each element corresponds to the weighted residual of each sample: (8); The fifth step is to minimize the weighted loss. Next, the parameters are updated using weighted least squares. The optimization objective is to minimize the weighted loss function: (9); This optimization problem can be solved analytically or numerically (such as gradient descent). For weighted least squares, the analytical solution is: (10); The sixth step is an iterative update process that continues until the parameters are updated. The change is small enough, that is: (11); The above process is repeated at each position of the sliding window. Through weighted residuals and iterative optimization, Huber robust estimation can effectively reduce the impact of outliers and obtain robust parameter estimates. In this section, the invention estimates the gross errors in the processed MBES depth and intensity data. To address this, robust estimation is used to remove gross errors from the features extracted from the MBES depth and intensity data through a sliding window, ensuring the purity of the extracted features.

[0053] S4. In the model building stage, the Stacking mechanism in ensemble learning is adopted to integrate the advantages of multiple classifiers at the decision level and generate a spatial distribution map of nodule coverage.

[0054] The base classifiers are trained using a k-fold cross-validation strategy to ensure the robustness of the predicted meta-features. Each base classifier generates out-of-fold results through cross-prediction on the training set. These results are then used as new features input to the secondary classifier (Gradient Boosting), thereby achieving non-linear fusion at the decision layer. The final output is a discrete coverage level (low, medium, high), presenting the spatial distribution of deep-sea manganese nodules in the form of a classification map.

[0055] Machine learning has demonstrated strong applicability in seabed sediment classification and spatial distribution prediction of manganese nodules, capable of establishing complex nonlinear mapping relationships based on input feature data. However, different machine learning algorithms, due to variations in modeling methods, assumptions, and parameter selection, may produce different prediction results even with the same dataset. A single model is often limited by specific assumptions, potentially performing well in some areas but poorly in others. Multi-model approaches can help reduce prediction errors, better generalize models across broad geographical areas or more complex feature spaces, and improve confidence in locations with good spatial consistency. Therefore, to enhance prediction stability and generalization ability, this invention introduces a multi-classifier decision fusion method based on a stacking mechanism to better address the spatial variability of seabed sediment features within complex geographical regions.

[0056] Stacking is an efficient ensemble learning method that trains a meta-model by using the predictions of multiple base classifiers as input, further optimizing the final decision. In the field of remote sensing, stacking-based ensemble learning methods have achieved satisfactory results. In this invention, five classic supervised classification algorithms—Random Forest (RF), Decision Tree (DT), Backpropagation (BP), Support Vector Machine (SVM), and K-Nearest Neighbors (KNN)—are selected as the first-layer classifiers, each trained and used independently for prediction on the data. The predictions of the first-layer models are not only used to evaluate the individual performance of each model but also passed as input features to the second-layer meta-model to improve the overall classification accuracy and stability. This invention also considers the performance limitations of models on certain datasets. A performance screening threshold is set: if the prediction accuracy of a first-layer classifier is below 60%, its prediction result is not passed to the second-layer meta-model to avoid interference from the erroneous predictions of low-performance models in the final decision.

[0057] The choice of meta-model determines the final prediction result and has a significant impact on resource evaluation. In this invention, a Gradient Boosting Classifier (GBC) is used as the final decision model. GBC can construct a more robust classification boundary by iteratively optimizing the residuals and further integrate the prediction information of the first-layer classifier, thereby improving the overall performance of the model.

[0058] In spatial prediction of tuberculosis coverage, the first layer of the Stacking ensemble learning method classifies each pixel individually, yielding several independent predictions. After a performance screening mechanism eliminates low-precision models, the remaining predictions are combined to form a new feature set, which is then input into the meta-model for final decision-making. This decision-making process is based on majority voting, ensuring that the final class determination integrates the strengths of different classifiers and improves spatial consistency.

[0059] Example 2, the acoustic characterization system for deep-sea manganese nodule coverage based on multi-classifier decision fusion provided in this embodiment of the invention includes: The data acquisition module is used to acquire water depth and backscatter intensity data based on the Kongsberg EM122 multibeam echo sounder system. It is combined with a 6000-meter-class integrated optical towed body to acquire near-bottom images. The positioning of the near-bottom optical towed body is obtained by using an ultra-short baseline USBL combined inertial navigation system (INS) and a Doppler velocity recorder (DVL) to obtain the location of underwater optical sampling points. The shipborne multibeam echo sounder acquires a precise underwater terrain model through a Global Navigation Satellite System (GNSS) combined inertial navigation system (INS). By constructing the sampling points and the ship-based terrain using the same projection method, the water depth and backscatter intensity data at the location of the sampling points are acquired. The feature optimization module extracts 14 backscatter texture features and 4 seabed topographic features, and uses the Boruta algorithm to optimize the extracted features, thereby constructing a refined feature set representing the seabed. The iterative robust estimation module introduces an iterative robust estimation algorithm based on a sliding window. By dynamically adjusting the observation weights through continuously smoothed data in different directions, it identifies and suppresses gross errors in the feature image. The multi-classifier decision fusion module adopts the Stacking mechanism in ensemble learning to integrate the advantages of multiple classifiers at the decision level and generate a spatial distribution map of nodule coverage.

[0060] To further demonstrate the positive effects of the above embodiments, the present invention conducts the following experiments based on the above technical solutions.

[0061] 1. Experimental Data Introduction: The area to be analyzed is located in the western part of the Clarion-Clipperton Zone (CCZ) in the eastern Pacific Ocean, approximately 1150 km from Honolulu, Hawaii, USA. The CCZ is bounded by the Clipperton Fault to the south, the Clarion Fault to the north, the East Pacific Uplift to the east, and the Line Islands to the west. It is an intermediate block of the Pacific Plate, with Mesozoic strata overlying an oceanic basaltic basement. The western part of the CCZ is primarily a deep-sea plain environment, receiving abundant sediments from the ocean surface, including biological remains, terrigenous clastic materials, and cosmic dust. These sediments gradually accumulated over a long geological history, forming sedimentary strata of varying thicknesses and types, overlying the oceanic crust basement and fault structures.

[0062] The area to be analyzed is approximately 1181 km². 2 The geological age of the basement formation is Late Cretaceous (approximately 95-65 Ma). The low positive gravity background of the gravity anomaly in the region indicates a high-density mantle anomaly and seafloor uplift, with a water depth range of 4574-5451 m. In addition to multibeam bathymetry data, in-situ observation data of the seafloor were also collected. Using an LH-GT6000G, a 6000-meter integrated optical towed body, a deep-sea optical towed body survey line was acquired, and the seafloor surface matrix was captured using underwater high-definition imaging. The raw optical images were screened, removing images of poor quality, and the coverage index of manganese nodules was obtained through analysis of the filtered images. The optical towed body used a combination of Global Positioning System (GPS) and Ultra-Short Baseline Acoustic Navigation (USBL) for navigation and positioning. Each system has its own limitations, which have a cumulative effect on the total navigation error; the positional offset is a small error that is difficult to quantify. In addition to the optical data collected by the optical towed body, a box-type sampler, pulled by a geological cable, penetrated the seabed under its own weight. Upon contact with the seabed, the bottom shovel was released, and the sampler grabbed and collected samples, recording the underlying matrix on-site. A total of 16 box-type sampling points and 4753 optical towed body survey line sampling points were collected.

[0063] 2. Experimental results and analysis; 2.1 Feature extraction and evaluation of optimization results; 2.1.1 Feature Extraction and Optimization Results: Based on multibeam bathymetry and backscatter intensity data, this invention extracted 14 backscatter intensity features and 4 terrain features with the same resolution (i.e., 150m resolution). Table 1 shows all feature types and abbreviations, and the calculation methods for each feature are presented mathematically. During the Boruta feature optimization process, an importance score for each feature is calculated. The specific scores are visualized using a bar chart, such as... Figure 3 As shown.

[0064] The Boruta algorithm was run with maxRuns=500 iterations and a p-value of 0.05. According to Boruta analysis, 16 of the 18 predictor variables initially included in the model were considered important, none were considered weak features, and the features contrast and variance were deemed invalid. This invention removes the contrast and variance features, a choice that minimizes the number of predictor variables while maintaining high model performance.

[0065] 2.1.2 Evaluation of preferred features: Figure 4 Correlation analysis of different features shows that irrelevant features have been filtered out, and all remaining features are correlated with the target variable or with each other.

[0066] Analysis reveals the dominant role of topographic framework features. Water depth plays a crucial role as the primary predictor, while topographic curvature and aspect rank as the second and third most important indicators. By regulating local hydrodynamic intensity, they create sediment trapping "traps" in micro-topographic uplift areas (positive curvature), explaining the strong spatial coupling between nodule enrichment zones and seamount slopes observed. These topographic framework features further confirm that seafloor topography controls nodule distribution patterns through a topographic-hydrodynamic coupling mechanism, supporting the "topographic pump" mineralization hypothesis. Unlike traditional sediment classification methods, backscattering intensity is not as important as other features in traditional sediment classification. This invention argues that this is because nodules on the seafloor exhibit characteristics of being buried or partially buried by sediments, while backscattering intensity focuses more on reflecting the physical properties of the sediment surface, thus significantly increasing the difficulty of exploration.

[0067] Simply relying on the correlation between features and categories to evaluate feature effectiveness often suffers from strong subjectivity and limited discriminative ability. To further assess the change in model separability before and after feature selection, the Jeffries-Matusita (JM) distance index is used to compare and analyze the separation degree between different categories. Therefore, this invention further introduces the Jeffries-Matusita (JM) index as a separability metric to quantitatively evaluate the feature set before and after feature selection. The JM distance, based on the assumption of a normal distribution of data, yields the separation degree between different categories and is widely used in pattern recognition and feature selection. It measures the separation degree between categories by calculating the Bhattacharyya distance, and its value is normalized to 0. Between them, the closer the values ​​are This indicates that the stronger the separability between categories.

[0068] Table 2 JM distance matrix for different tuberculosis coverage rates

[0069] Table 3. JM distance matrix for different tuberculous coverage rates after feature selection

[0070] Table 2 shows the JM distance matrix between different tuberculosis coverage rates without feature selection. The results show that the covariance matrix is ​​singular due to highly redundant or linearly correlated features in the original features, failing to effectively consider inter-class separation. In contrast, Table 3 shows the JM distance matrix after feature selection using the Boruta algorithm. The JM distance between all classes is greater than 1, and the JM distance between multiple class pairs is close to... This indicates that the feature set filtered by Boruta significantly improves the discriminative power of the feature set and effectively enhances the separability between different categories.

[0071] In summary, the Boruta feature selection algorithm can improve feature effectiveness, avoid redundant interference, and enhance the model's predictive and discriminative capabilities in the task of predicting deep-sea nodule coverage, providing a solid feature foundation for subsequent manganese nodule coverage prediction.

[0072] 2.2 Tuberculosis coverage prediction results; 2.2.1 Comparison of Feature Optimization Performance from a Feature Perspective: Based on the natural distribution characteristics of coverage, regions were divided into three categories: low coverage (0-20%), medium coverage (20-40%), and high coverage (>40%). Subsequently, five mainstream classifier models were constructed, including Random Forest (RF), Support Vector Machine (SVM), Backpropagation Neural Network (BP), K-Nearest Neighbors (KNN), and Decision Tree (DT). The performance of different models in the coverage classification task was compared and analyzed from multiple dimensions, including overall accuracy (OA), Kappa coefficient, and user accuracy (UA) and producer accuracy (PA) for each category. The results are shown in Table 4.

[0073] Table 4. Prediction results of tuberculosis coverage using different classifiers

[0074] For coverage prediction, as shown in Figures 5(a)-5(e), the RF model exhibits a relatively smooth and coherent spatial distribution, especially in the central and eastern regions of the analysis area, where the boundaries between high-coverage (yellow) and low-coverage (blue) areas are clearly defined. It effectively captures spatial heterogeneity and the transition zones between coverage levels. The results obtained by DT are somewhat similar to RF in spatial distribution, but the texture is coarser. Although it can distinguish between high and low coverage areas well, the class boundary processing is rather abrupt, exhibiting the characteristics of a decision tree's "hard segmentation" in the feature space, making it difficult to finely characterize classification boundaries in complex environments. The BP neural network model, while maintaining the overall spatial structure, introduces more detailed information. It is somewhat similar to the spatial pattern of the RF model, but there is some overestimation in the medium-coverage area (green). This model performs well in identifying local anomalies, but there is a certain risk of misclassification at the coverage boundaries. The spatial distribution predicted by SVM and KNN models is relatively scattered, and the classification results show strong pixelation characteristics, poor spatial continuity, and high spatial noise. They also have weak generalization ability in complex seabed topography, resulting in unsatisfactory classification performance.

[0075] To further improve the robustness and generalization ability of the tuberculosis coverage classification model, this invention introduces the Boruta feature selection method to evaluate and simplify the importance of the original features. After completing the Boruta feature selection, this invention further conducted a comparative experiment on the classification accuracy of manganese tuberculosis coverage to evaluate the impact of feature optimization on model performance. Comparing the prediction performance of five mainstream classifiers before (see Table 4) and after (see Table 5) feature selection, the positive effects of Boruta selection in multiple dimensions can be clearly observed.

[0076] Overall, as shown in Table 5, Boruta feature selection significantly improved the overall accuracy (OA) and Kappa coefficient of most classifiers. Specifically, the Kappa coefficient of the RF model increased from 0.8808 to 0.9014, and the overall accuracy improved from 92.16% to 93.54%, further solidifying its dominant position in spatial prediction. This model improved both the UA and PA metrics for coverage across the three classes, indicating stronger balance and stability across various discriminative tasks. Classification results show that after feature selection, the Kappa coefficients of all classifiers improved, demonstrating enhanced model stability and discriminative ability.

[0077] Meanwhile, the overall classification accuracy of the other classifiers, except for the BP neural network, has also been improved to varying degrees. Although the accuracy of the BP neural network has not changed significantly, its performance has not decreased significantly, indicating that Boruta screening has a positive effect on the overall robustness of the model.

[0078] Table 5. Prediction results of tuberculosis coverage by different classifiers after using Boruta feature selection.

[0079] 2.2.2 Comparison of Accuracy from the Perspective of Feature Error Interference: This experiment focuses on the application of robust estimation theory in the prediction scenario of evaluation indicators for deep-sea nodule resources, particularly its effectiveness in removing gross errors. As mentioned earlier, this invention employs the M-estimation method in robust estimation, selecting the Huber function as the ρ function for gross error removal. In specific implementation, this invention uses a sliding window of size 7 for robust estimation and demonstrates the effect of this method on nodule coverage prediction, as shown in Table 6.

[0080] Table 6. Prediction results of tuberculosis coverage rate by different classifiers after robust estimation.

[0081] Figures 6(a)-6(e) show the spatial distribution of manganese nodule coverage predicted by five typical classification algorithms after robust estimation of the original feature set. Comparing the previous set of classification results based on the unrobustly estimated feature set, it can be seen that after robust estimation, the recognition ability of each model for different coverage areas is improved, especially in the spatial continuity and boundary recognition of medium coverage (20-40%) areas. Specifically, the distribution of medium coverage areas (green) in each classification scheme is more coherent, significantly better than the fragmented medium coverage distribution trend before processing.

[0082] Meanwhile, quantitative evaluation results show that almost all classification models achieved varying degrees of accuracy improvement after introducing robustness: a 1.97% improvement compared to Random Forest (RF); a 2.85% improvement compared to Support Vector Machine (SVM); a 1.51% improvement compared to Backpropagation (BP) Neural Network; and a 2.55% improvement compared to Decision Tree (DT). Compared to the nodule coverage prediction results after feature selection using Boruta, robustness estimation also brought significant accuracy improvements, with a 1.86% improvement compared to SVM, a 1.55% improvement compared to BP Neural Network, and a 0.83% improvement compared to DT.

[0083] The experimental results above demonstrate that robust estimation enhances the model's anti-interference and generalization capabilities at the feature level, effectively eliminating gross errors in the prediction process of deep-sea manganese nodule resource evaluation indicators. These gross errors often manifest as discrete outliers in the prediction map. Due to the complex deep-sea environment and the resolution limitations of multibeam systems, such gross errors are difficult to completely avoid in actual exploration. However, after robust processing using the M-estimation method, the discrete noise in the nodule coverage prediction map is significantly reduced, as observed in the image quality. Furthermore, the prediction accuracy of almost all classifiers is improved to varying degrees in terms of quantitative evaluation indicators, further validating the effectiveness of robust estimation in deep-sea applications.

[0084] 2.2.3 Comparison of the proposed method's effectiveness from a classification perspective: This invention performs spatial prediction analysis of manganese nodule coverage based on multiple mainstream classifiers. Experimental results show significant differences in classification performance among different classifiers. For example, Random Forest (RF) and Decision Tree (DT) exhibit high prediction accuracy and demonstrate good classification ability. However, some models, such as Support Vector Machine (SVM), perform relatively poorly in the current task. Furthermore, inconsistencies between prediction consistency and accuracy were observed in some classifiers: for example, while K-Nearest Neighbors (KNN) performs well in overall accuracy, its classification images show significant fragmentation; while BP neural network, although slightly less accurate overall, shows high consistency with most other models, exhibiting a more stable spatial pattern. To overcome the limitations of a single classifier and enhance the model's adaptability to complex data, a Stacking ensemble learning method is further introduced after completing the coverage prediction of a single classification model, aiming to integrate the advantages of multiple models to improve the overall model's generalization ability and robustness.

[0085] In the task of predicting tuberculosis coverage, the Stacking ensemble learning method significantly improves classification performance. Specifically, as... Figure 7As shown in the figure, the low-coverage areas (blue) in the prediction map exhibit higher internal consistency, smoother distribution boundaries, and a significant reduction in outlier discrete points, demonstrating strong noise robustness. The medium-coverage areas (green) show a more structured distribution with significantly enhanced spatial clustering features, effectively capturing the transitional zone characteristics of manganese nodule coverage. The high-coverage areas (yellow) have relatively clear boundaries, and their spatial coherence is superior to the previous two methods. As shown in Table 7, the prediction results based on the ensemble model achieve a total accuracy of 93.21% and a Kappa coefficient of 89.62%. Compared to using all features for prediction, the Stacking method improves prediction accuracy by 2.35% (compared to RF), 39.06% (compared to SVM), 14.00% (compared to BP neural network), 7.33% (compared to DT), and 3.69% (compared to KNN). Compared to the single classifier model optimized based on Boruta feature selection and robust estimation, the Stacking method improves the prediction accuracy by 0.38% (compared to RF), 36.21% (compared to SVM), 12.49% (compared to BP neural network) and 4.32% (compared to KNN), respectively, further verifying the effectiveness and applicability of the ensemble strategy in predicting the distribution of manganese nodules in complex deep-sea environments.

[0086] Table 7. Confusion matrix of nodule coverage using the Stacking mechanism.

[0087] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention and within the spirit and principles of the present invention should be covered within the scope of protection of the present invention.

Claims

1. An acoustic characterization method for deep-sea manganese nodule coverage based on multi-classifier decision fusion, characterized in that, The method includes the following steps: S1, based on the Kongsberg EM122 multibeam echo sounder, acquires water depth and backscatter intensity data. It combines a 6000-meter-class integrated optical towed body to acquire near-bottom images. The positioning of the near-bottom optical towed body uses an ultra-short baseline (USBL) combined inertial navigation system (INS) and a Doppler velocity recorder (DVL) to obtain the location of underwater optical sampling points. The shipborne multibeam echo sounder acquires a precise underwater terrain model through a Global Navigation Satellite System (GNSS) combined inertial navigation system (INS). By constructing the sampling points and the ship-based terrain using the same projection method, water depth and backscatter intensity data at the location of the sampling points are obtained. S2, by extracting backscattered texture features and seabed topographic features, the Boruta algorithm is used to optimize the extracted features and construct a set of features that represent the seabed in a refined manner. S3 introduces an iterative robust estimation algorithm based on a sliding window, which dynamically adjusts the observation weights by continuously smoothing data in different directions to identify and suppress gross errors in the feature image; S4. In the model building stage, the Stacking mechanism in ensemble learning is adopted to integrate the advantages of multiple classifiers at the decision level and generate a spatial distribution map of nodule coverage.

2. The acoustic characterization method for deep-sea manganese nodule coverage based on multi-classifier decision fusion according to claim 1, characterized in that, In step S2, the backscattering texture features include water depth, slope, curvature, slope aspect, and backscattering intensity. , , , , , , , Standard deviation, kurtosis, skewness, energy, entropy; Submarine topographic features include depth, slope, aspect, and planar curvature.

3. The acoustic characterization method for deep-sea manganese nodule coverage based on multi-classifier decision fusion according to claim 1, characterized in that, In step S2, the Boruta algorithm is used to optimize the extracted features, including: the Boruta algorithm expands the feature space by supplementing random attributes to eliminate the correspondence between sample attribute values ​​and labels; the expanded feature space is used to perform classification tests and calculate the importance index of all attributes; the importance index of shadow attributes is used as a reference to evaluate the importance of each attribute. The Boruta algorithm is used to evaluate the importance of each feature variable, including: (1) Initialization and data preparation: The dataset is ,in, It is the first One characteristic, It is the total number of features; the dataset also contains the target variable. , used to represent the label of each sample; (2) Generating image features: The Boruta algorithm generates a set of random image features to simulate noise, thereby determining whether the real features are important; for each original feature Generate an image feature Image features The value is randomly shuffled The sample labels in the data are obtained and represent only noise; (3) Merging original features and image features: Merging original features and image features The features are merged to form a dataset containing both the original features and image features. Use the merged feature set Train a random forest model to estimate the importance of each feature; the random forest model outputs an importance score for each feature and compares the importance of the original features with that of the image features.

4. The acoustic characterization method for deep-sea manganese nodule coverage based on multi-classifier decision fusion according to claim 3, characterized in that, The Boruta algorithm uses the Gini index to measure the importance of features; For each feature Train a random forest model and calculate its importance to determine the target variable. The contribution of the prediction; the random forest model evaluates the importance of features based on the contribution of each feature.

5. The acoustic characterization method for deep-sea manganese nodule coverage based on multi-classifier decision fusion according to claim 3, characterized in that, The importance of comparing raw features with image features in random forest models includes: Calculate the maximum importance of all image features : ； For each original feature , will be important and Compare; if If the feature is positive, then it is important; otherwise, it is useless. The Boruta algorithm runs multiple times, iterating through the image feature importance comparison steps until the classification of each feature is stable.

6. The acoustic characterization method for deep-sea manganese nodule coverage based on multi-classifier decision fusion according to claim 1, characterized in that, In step S3, the iterative robust estimation algorithm based on the sliding window includes: (1) Slide a window of a specified size on the feature image; (2) In each window, local pixel sequences are extracted along the horizontal, vertical and diagonal directions respectively; by constructing one-dimensional pixel chains in different directions, the spatial correlation of acoustic scattering signals in the nodule coverage area is utilized to enhance the sensitivity of the feature extraction process to spatial texture and nodule edge structure. (3) Based on the Huber loss function, weights are assigned to the observation points in the pixel sequence of each direction, and the local regression model parameters of the direction are updated by iterative weighted least squares until convergence, so as to obtain the robust estimation result of the direction; robust regression is performed on the sequence data of each direction, and the Huber function is used to reduce the gross error caused by noise pulses, deep water bubbles or local abnormal reflections, so as to make the estimation results of the tuberculosis coverage related features more stable. (4) Calculate the directional weights dynamically based on the robust standard deviation of the residuals in each direction, and perform weighted fusion of the estimation results in different directions to obtain the final robust estimate of the center pixel of the window; (5) Move the window to the next position and repeat steps (1)-(4) until the entire feature image is traversed; through multi-directional weighted residuals and iterative optimization process, robust parameter estimates that are insensitive to noise and gross errors are obtained.

7. The acoustic characterization method for deep-sea manganese nodule coverage based on multi-classifier decision fusion according to claim 6, characterized in that, The iterative robust estimation algorithm based on sliding windows uses a robust estimator Huber to fit a local model in each window, predicts the estimated value of the window center point, and replaces the center pixel. The specific steps are as follows: (1) For pixels Define a local sliding window with a fixed size of 5×5 pixels and a step size of 1 pixel, and process it pixel by pixel to ensure that the boundary prediction is concentrated and the coverage is complete; (2) Extract the 1D pixel chain containing the center pixel, along For each chain, a robust regression is performed in four directions using the Huber function as the kernel function to obtain the center estimate for that direction. and the residual scale in that direction ; (3) Take a weighted average of the four estimates obtained from the above multiple directions as the final output; the expression is: ； In the formula, For in position Location, along direction The weighting coefficients, The number must be positive to prevent the denominator from being zero; For in position The final estimated value at that location, To normalize the weights, the sum of the weights in all directions is guaranteed to be 1.

8. The acoustic characterization method for deep-sea manganese nodule coverage based on multi-classifier decision fusion according to claim 1, characterized in that, In step S4, the multi-classifier decision fusion based on the Stacking mechanism includes: Five supervised classification algorithms were selected as first-layer classifiers, and each was trained and predicted independently on the data. The prediction results of the first-layer models were used to evaluate the individual performance of each model and were passed as input features to the second-layer model. A performance screening threshold was set. If the prediction accuracy of a first-layer classifier was lower than the threshold, the prediction result was not passed to the meta-learning model of the second layer. A gradient boosting classifier was used as the final decision model to integrate the prediction information of the first-layer classifiers.

9. The acoustic characterization method for deep-sea manganese nodule coverage based on multi-classifier decision fusion according to claim 7, characterized in that, In spatial prediction of tuberculosis coverage, the first layer of the Stacking ensemble learning method classifies each pixel separately, resulting in several independent predictions. After filtering out low-precision models through a performance screening mechanism, the remaining prediction results are combined to form a new feature set, which is then input into the meta-model for final decision-making. The decision-making process is based on the majority voting principle, ensuring that the final category determination takes into account the advantages of different classifiers.

10. The acoustic characterization method for deep-sea manganese nodule coverage based on multi-classifier decision fusion according to claim 7, characterized in that, Supervised classification algorithms include random forests, decision trees, backpropagation (BP) networks, support vector machines, and K-nearest neighbors classifiers.

Images