Multi-source data seabed sediment classification method based on GA-XGBoost

Abstract

The multi-source data seabed sediment classification method based on GA-XGBoost belongs to the field of seabed sediment classification technology, including preprocessing multispectral data and multibeam bathymetric data and aligning the geographic coordinate system and spatial resolution, extracting multidimensional spectral features and terrain feature parameters from the processed data, using the XGBoost algorithm to calculate the feature importance score and sorting from high to low, and using the forward stepwise feature selection technology to establish a baseline model through the two most important features, and sequentially adding 1 feature variable to the XGBoost model training and testing, calculating the overall classification accuracy, and screening the optimal feature subset. Then, a GA-XGBoost based on genetic algorithm optimization of hyperparameters is constructed, and the model is trained and tested using the screened feature data set. The present invention improves the accuracy of seabed sediment classification and improves the performance and generalization ability of the classification model.

Description

Technical Field

[0001] The present invention designs a multi-source data fusion seabed sediment classification method, specifically relates to a multi-source data seabed sediment classification method based on GA-XGBoost with forward stepwise feature selection, and belongs to the technical field of multi-source data fusion and seabed sediment classification. Background Art

[0002] As the foundation of the marine ecosystem, the seabed sediment plays a vital role in maintaining marine biodiversity and ecological balance. Its classification not only helps us to understand the marine geological process more deeply, but also has significant economic and scientific research value for the exploration, development and protection of marine resources. Since the properties of seabed sediments can be reflected by the texture information characteristics and spectral information of the seabed sediment, the seabed sediment is classified based on multispectral data and multibeam bathymetric data, combined with data preprocessing and machine learning algorithms. The seabed ecological environment is monitored by detecting the seabed sediment, which assists in the identification, assessment and protection of marine biological communities.

[0003] Domestic and foreign researchers have conducted a lot of research on the classification of seabed sediments, which can be roughly divided into four categories. (1) In-situ sampling method: In-situ sampling generally uses specially designed sediment sampling instruments to place on the seabed. The surrounding environment does not change, and the temperature, pressure and geotechnical conditions of the sediments on the seabed are maintained to the greatest extent. The disturbance of sediments during sampling and transportation is avoided, and the sediment characteristics are directly measured. However, there are limitations such as high cost and limited spatial sampling. (2) Multi-beam acoustic sediment classification: Multi-backscatter intensity data has a strong correlation with the physical properties of the seabed sediment, such as roughness, sediment particle size, porosity, saturation and incident angle. Through a series of processing such as propagation loss correction, sound line bending correction, Lambert's law correction, sound illumination area correction, central beam area signal correction and filtering of the multi-beam backscatter intensity data, the seabed sediment type is reflected according to the backscatter intensity value range. However, the complex algorithms and models of multi-beam make it difficult to interpret acoustic data and highly dependent on professional knowledge. (3) Airborne LiDAR bathymetry (ALB) and optical bottom classification using multi-temporal data fusion: Airborne LiDAR bathymetry can obtain water depth data and laser pulse echo data containing bottom information; the bottom of the seabed based on airborne LiDAR bathymetry waveform data and combined with water depth derived data. Multi-temporal data fusion refers to the fusion of remote sensing data acquired at different times to capture the changes in the bottom of the seabed over time. However, the propagation of light in water is affected by factors such as water clarity and turbidity, which affects the visibility of bottom characteristics. (4) Bottom classification using multi-feature extraction from multi-source data. Different features are extracted by fusing optical and acoustic data. Optical remote sensing data can provide high-resolution surface coverage information, while acoustic data can provide detailed information on the seabed topography and bottom. Different machine learning classification algorithms, such as support vector machine (SVM), BP neural network, random forest (RF), etc., are combined to improve the accuracy of bottom classification. Summary of the invention

[0004] The purpose of the present invention is to use WorldView-2 multispectral data and multibeam bathymetric data to provide a multi-source data seabed sediment classification method based on forward stepwise feature selection and GA-XGBoost to improve the accuracy of seabed sediment classification.

[0005] The multi-source data seabed sediment classification method based on GA-XGBoost is characterized by comprising the following steps:

[0006] (0) Obtain multispectral images, multi-beam measured water depth data and measured bottom samples in the study area;

[0007] (1) Data preprocessing:

[0008] Perform radiation correction and atmospheric correction preprocessing on multispectral images;

[0009] The images are separated into water and land according to the normalized difference water index method;

[0010] Spatial scale registration of bathymetric data and spectral images;

[0011] (2) Multi-feature extraction and selection:

[0012] (2.1) Spectral feature extraction: retain the radiance values ​​of 6 bands as features, namely CoastalBule, Blue, Green, Yellow, Red, and RedEdge; then calculate the normalized vegetation index NDVI, ratio vegetation index SR, visible atmospheric impedance index VARI, and enhanced vegetation index based on the radiance values ​​of the corresponding bands, for a total of 10 spectral features;

[0013] (2.2) Terrain feature extraction: Multi-beam measured water depth data depth is used as the basic terrain data to extract relevant terrain feature data - slope, aspect, curvature and roughness;

[0014] (2.3) Feature vector extraction and feature selection: Arrange the extracted features in a layer structure, extract the feature data and substrate categories of the same coordinate position in each layer in units of 1×1 pixels, normalize them layer by layer, form feature vectors, and draw a box plot of the feature vector distribution;

[0015] (2.4) Select the optimal feature subset, including importance sorting and classification accuracy analysis, and use XGBoost to evaluate the importance of features. During the evaluation process, the average information gain of the features is calculated by comparing the changes in the objective function before and after the split; the final importance score of the feature is obtained by accumulating and averaging the gain values ​​of each feature; and the feature importance scores are sorted from high to low;

[0016] In order to determine the optimal feature vector subset combination, the forward stepwise feature selection method is adopted. To improve the computational efficiency, the two most important features are used as the initial input of XGBoost, and one feature variable is added to the XGBoost model training and testing in turn according to the feature importance score; each time a dimension of feature is added, the change in classification accuracy is observed; when the classification accuracy change is no longer significant, the current dimension feature set is considered to be optimal, and the feature selection is completed; the number of selected features is recorded as N;

[0017] (3) GA-XGBoost classification model construction:

[0018] The feature vector subset is divided into multiple subsets, each subset is used as a validation set in turn, and other subsets are used as training sets; the GA genetic algorithm represents the XGBoost hyperparameter combination to be optimized as a "chromosome", inputs a randomly generated hyperparameter combination, trains the XGBoost model with each hyperparameter combination, and calculates the model accuracy index as the fitness value on the validation set, generates a new generation of hyperparameter combinations through selection, crossover, and mutation operations, repeats the above operations until the preset number of iterations is reached, and outputs the chromosome with the highest fitness value as the optimal hyperparameter combination of the optimized XGBoost algorithm, and obtains the best parameter GA-XGBoost model;

[0019] (4) Classification of seabed sediments:

[0020] All feature vectors of other study areas are extracted and input into the trained GA-XGBoost model to classify the feature vectors of the entire study area.

[0021] In step (1), the radiation correction and atmospheric correction preprocessing is to preprocess the multispectral image using ENVI software;

[0022] The water-land separation according to the normalized difference water index method is performed by setting a threshold , Value not less than The area is a water body, otherwise it is a non-water body. The formula is as follows:

[0023] ,

[0024] In the formula, is the radiation brightness value of the green light band; is the radiation brightness value in the near-infrared band; is the normalized difference water index; is the threshold of the normalized difference water index method, .

[0025] In step (0), the multispectral image adopts the WorldView-2 image; in step (1), the spatial scale alignment of the acoustic data and the optical remote sensing data is performed by resampling and coordinate conversion of the multi-beam measured water depth data with reference to the 1.85 m resolution of the WorldView-2 image, so as to unify the coordinate system and resolution of the two data sources.

[0026] In step (2.1) of spectral feature extraction, the calculation method of 10-dimensional spectral features is shown in Table 1:

[0027] Table 1 Spectral characteristic index calculation formula

[0028] ,

[0029] Among them, R C Represents the CoastalBule radiation brightness value, R B Indicates the Bule radiation brightness value, R G Indicates the Green radiation brightness value, R Y Represents the Yellow radiation brightness value, R R Indicates the Red radiation brightness value, R RE Represents the RedEdge radiation brightness value, R NIR1 Indicates the NearInfrared1 radiation brightness value, R NIR2 Represents the NearInfrared2 radiation brightness value; C1=6, C2=7.5, L=1, G=2.5.

[0030] In step (2.2) terrain feature extraction, the formulas for each feature index are shown in Table 2:

[0031] Table 2 Calculation formula of terrain characteristic index

[0032] ,

[0033] in, represents the spatial coordinates, is the target point elevation, The elevation of a certain adjacent grid point, is the horizontal distance between this point and its adjacent point, is the normal of the 3D surface x , y Directional component; slope i is the slope between this point and the adjacent point; arctan 2 (Ny , Nx) For calculation The arc tangent of the argument.

[0034] In step (2.4), the specific formula for calculating the average information gain is:

[0035] ,

[0036] In the formula, Indicates j The node of the feature, j ∈1,…, N ; represents the number of all trees, Indicates The number of non-leaf nodes in a tree, Indicates The tree The partitioning features of non-leaf nodes, so , is the indicator function; is the hyperparameter of the regularization term, , Respectively indicate that they fall in The tree The sum of the first-order derivative and the second-order derivative of all samples on non-leaf nodes, Respectively indicate that they fall in The first The sum of the first-order derivatives on the left and right nodes of non-leaf nodes. Similarly, Respectively indicate that they fall in The first The sum of the second-order derivatives on the left and right nodes of non-leaf nodes, so:

[0037] ,

[0038] .

[0039] In step (2.4), the gain values ​​of each feature are accumulated and averaged to obtain the final importance score of the feature as follows:

[0040] , j ∈1,…, N .

[0041] In step (2.4), each time a dimension of features is added, the change in classification accuracy is observed. If the classification accuracy drops by more than 1% after adding the new features, or the increase compared with the previous highest accuracy does not exceed 0.5%, the current dimensional feature set is considered to be optimal, and feature selection is completed.

[0042] The accuracy performance indicator selected in step (3) calculates the fitness value as follows:

[0043] ,

[0044] in, s is the number of iterations of the GA genetic algorithm, and s The chromosome with the highest output fitness value in the iteration is used as the optimal hyperparameter combination of the optimized XGBoost algorithm to obtain the optimal parameter GA-XGBoost model.

[0045] In step (3), the expression of the classification model can be summarized as:

[0046] ,

[0047] Where: C is the CA-XGBoost classification algorithm, Feuture1— are the N features determined in step (2.4).

[0048] The present invention adopts a multi-source data fusion method of multispectral imaging, multi-beam measured water depth data combined with a small amount of measured bottom samples to improve the accuracy of seabed bottom classification; and redesigns the feature selection technology to remove irrelevant and redundant features, simplify the feature space, and improve the performance of the classification model; and adopts GA in combination with XGBoost to improve the performance and generalization ability of the classification model. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 This is the overall technical route of the proposed method.

[0050] Figure 2 Box plot of eigenvectors.

[0051] Figure 3 is the feature importance score and classification accuracy.

[0052] Figure 4 Flowchart for constructing the GA-XGBoost seabed sediment classification model.

[0053] Figure 5 Confusion matrix for GA-XGBoost test set classification.

[0054] Figure 6 Classification results of GA-XGBoost applied to the entire study area. DETAILED DESCRIPTION

[0055] The multi-source data seabed sediment classification method based on GA-XGBoost of the present invention mainly conducts the following research:

[0056] (1) The WorldView-2 multispectral remote sensing data was radiometrically and atmospherically corrected using ENVI software, and the normalized difference water index was used to separate water and land. The multibeam bathymetric data was converted to a multispectral coordinate system and the resolution and image size were unified.

[0057] (2) Feature extraction and selection. Spectral features and terrain features were extracted to form a data set, input into XGBoost to calculate the feature importance score and sort from high to low, and the optimal feature vector subset was determined by the forward stepwise feature selection method.

[0058] (3) Construction and application of classification model. The optimal feature vector subset was used as the data set to construct the sediment classification model. The sediment classification effects of four different machine learning algorithms, BP, GA-BP, XGBoost and GA-XGBoost, were compared.

[0059] Specifically, Figure 1 As shown, the multi-source data seabed sediment classification method based on GA-XGBoost is characterized by comprising the following steps:

[0060] (0) Obtain multispectral images, multi-beam measured water depth data and measured bottom samples in the study area;

[0061] (1) Data preprocessing:

[0062] ENVI software is used to perform radiation correction and atmospheric correction preprocessing on multispectral images to provide reflectance data that is closer to the actual conditions of the ground and sea surface. The water and land are separated according to the Normalized Difference Water Index (NDWI) method, and a threshold is set. , Value not less than The area is a water body, otherwise it is a non-water body. The formula is as follows:

[0063] ,

[0064] In the formula, The radiation brightness value of the green light band; The radiation brightness value of the near infrared band; is the normalized difference water index; is the threshold of the normalized difference water index method, In addition, the acoustic data and optical remote sensing data were spatially aligned, and the bathymetric data were resampled and coordinate transformed with the 1.85m resolution of the WorldView-2 image as a reference to unify the coordinate system and resolution of the two data sources.

[0065] (2) Multi-feature extraction and selection:

[0066] Spectral feature extraction. Since water absorbs almost all energy in the near-infrared band, the radiance values ​​of the first six bands are retained as features. In addition, the Normalized Difference Vegetation Index (NDVI), the Relative Vegetation Index (SR), the Visible Atmospheric Resistance Index (VARI), and the Enhanced Vegetation Index (EVI) are calculated based on the radiance values ​​of the corresponding bands, for a total of 10 spectral features. The characteristic index formulas are shown in Table 1:

[0067] Table 1 Spectral characteristic index calculation formula

[0068] .

[0069] Topographic feature extraction. The measured water depth presents rich texture information in the plane space. Compared with multispectral remote sensing images, it can be used to effectively characterize the surface morphology of different substrates and distinguish different types of landforms. In this paper, multi-beam measured water depth data (depth) is selected as the basic terrain data to extract relevant terrain feature data, and then calculate the slope, aspect, curvature and roughness. The characteristic index formula is shown in Table 2:

[0070] Table 2 Calculation formula of terrain characteristic index

[0071] .

[0072] Feature vector extraction and feature selection: There are limitations in using a single data source for bottom sediment classification. Therefore, a total of 15 dimensions are extracted by taking into account both multispectral features and terrain features. The extracted different features are arranged in a layer structure, and the feature data and bottom sediment categories at the same coordinate position of each layer are extracted in units of 1×1 pixels. The features are normalized layer by layer to form feature vectors. The feature vector distribution is plotted as a box plot, and the data distribution is observed.

[0073] Feature selection is the optimal feature subset, including importance ranking and classification accuracy analysis. XGBoost evaluates the importance of features. The specific formula for calculating the average information gain of recorded features by comparing the changes in the objective function before and after the split is:

[0074] ,

[0075] In the formula Indicates j The node of the feature, j ∈1,…, N ; represents the number of all trees, Indicates The number of non-leaf nodes in a tree, Indicates The tree The partitioning features of non-leaf nodes, so , is the indicator function; is the hyperparameter of the regularization term, , Respectively indicate that they fall in The tree The sum of the first-order derivative and the second-order derivative of all samples on non-leaf nodes, Respectively indicate that they fall in The first The sum of the first-order derivatives on the left and right nodes of non-leaf nodes. Similarly, Respectively indicate that they fall in The first The sum of the second-order derivatives on the left and right nodes of non-leaf nodes, so:

[0076] ,

[0077] .

[0078] The gain values ​​of each feature are accumulated and averaged to obtain the final importance score of the feature, which evaluates the contribution of each feature to the classification ability of the model:

[0079] .

[0080] In order to determine the optimal combination of feature vector subsets, the forward stepwise feature selection method is selected. To improve computational efficiency, the two most important features are used as the initial input of XGBoost, and one feature variable is added to the XGBoost model training and testing in turn according to the feature importance score; each time a dimension of feature is added, the change in classification accuracy is observed; when the classification accuracy change is no longer significant, the current dimensional feature set is considered to be optimal, and the feature selection is completed; the number of selected features is recorded as N.

[0081] (3) GA-XGBoost classification model construction:

[0082] Genetic Algorithm (GA): Genetic Algorithm is an optimization algorithm that simulates the biological evolution process. It searches for the optimal solution in the population of candidate solutions through operations such as selection, crossover (hybridization), and mutation. Genetic Algorithm is used to search for the best hyperparameter combination of XGBoost. XGBoost is a machine learning algorithm based on gradient boosted decision tree (GBDT). It generates a strong classification model by building and combining multiple decision tree models. The data set is pre-trained to generate a decision tree, and a second decision tree is added to fit the residual of the previous decision tree. Repeat the iteration until the termination condition is met, that is, an additive model composed of t sample base models. GA represents the XGBoost hyperparameter combination to be optimized as a "chromosome", uses XGBoost to train the model and evaluates its performance on the validation set, and selects the accuracy performance indicator to calculate the fitness value:

[0083] .

[0084] The chromosome with the highest fitness value is output as the optimal hyperparameter combination of the optimized XGBoost algorithm, and the optimal parameter GA-XGBoost model is obtained. The expression of the classification model can be summarized as:

[0085] ,

[0086] Where: C is the CA-XGBoost classification algorithm, Feuture1— are the N features determined after feature selection.

[0087] (4) Classification of seabed sediments:

[0088] All feature vectors of other study areas are extracted and input into the trained GA-XGBoost model to classify the feature vectors of each study area.

[0089] Specifically, multispectral images, multi-beam measured water depth data and measured bottom samples from other study areas are obtained, and after the same preprocessing, feature extraction and feature selection, a new data set is formed and input into the GA-XGBoost model trained in step (3), and the output of the model is used as the seabed bottom classification result.

[0090] Example

[0091] The data set selected by the present invention consists of WorldView-2 remote sensing data, field bottom type sampling data (seagrass, sand and coral reef) and multi-beam bathymetric data. After preprocessing, spectral features and terrain features are extracted and the optimal feature subset is selected for training and testing.

[0092] In order to verify the accuracy of the classification model, this paper selects the results of BP (Back Propagation neural network, BP), GA-BP (Genetic Algorithm- Back Propagation neural network, GA-BP) and XGBoost classification algorithms for accuracy comparison analysis, and evaluates the accuracy by constructing a confusion matrix to calculate the accuracy, F1-score and weighted F1 score indicators.

[0093] ,

[0094] ,

[0095] Where: is the weight of category i in the data set, represents the F1 value of category i.

[0096] The accuracy comparison of the four classification methods is shown in Table 3:

[0097] Table 3 Accuracy comparison of four classification methods

[0098] .

[0099] GA-XGBoost showed superior classification performance, outperforming the other three methods in all evaluation indicators, and the producer accuracy of the three categories was higher than that of other methods; the overall classification accuracy and weighted F1 score were both as high as 97.28%, and the confusion matrix Kappa coefficient was as high as 0.95. It can be seen that the GA-XGBoost algorithm can effectively capture data features and perform accurate classification when dealing with complex marine sediment classification problems.

Claims

1. Multi-source data seabed sediment classification method based on GA-XGBoost, The method is characterized by comprising the following steps: (0) Obtain multispectral images, multi-beam measured water depth data and measured bottom samples in the study area; (1) Data preprocessing: Perform radiation correction and atmospheric correction preprocessing on multispectral images; The images are separated into water and land according to the normalized difference water index method; Spatial scale registration of bathymetric data and spectral images; (2) Multi-feature extraction and selection: (2.1) Spectral feature extraction: retain the radiance values ​​of 6 bands as features, namely CoastalBule, Blue, Green, Yellow, Red, and RedEdge; then calculate the normalized vegetation index NDVI, ratio vegetation index SR, visible atmospheric impedance index VARI, and enhanced vegetation index based on the radiance values ​​of the corresponding bands, for a total of 10 spectral features; (2.2) Terrain feature extraction: Multi-beam measured water depth data depth is used as the basic terrain data to extract relevant terrain feature data - slope, aspect, curvature and roughness; (2.3) Feature vector extraction and feature selection: Arrange the extracted features in a layer structure, extract the feature data and substrate categories of the same coordinate position in each layer in units of 1×1 pixels, normalize them layer by layer, form feature vectors, and draw a box plot of the feature vector distribution; (2.4) Select the optimal feature subset, including importance ranking and classification accuracy analysis, and use XGBoost to evaluate the importance of the features. During the evaluation process, the average information gain of the features is calculated by comparing the changes in the objective function before and after the split; the final importance score of the feature is obtained by accumulating and averaging the gain values ​​of each feature; Sort the feature importance scores from high to low; In order to determine the optimal feature vector subset combination, the forward stepwise feature selection method is adopted. To improve the computational efficiency, the two most important features are used as the initial input of XGBoost, and one feature variable is added to the XGBoost model training and testing in turn according to the feature importance score; each time a dimension of feature is added, the change in classification accuracy is observed; when the classification accuracy change is no longer significant, the current dimension feature set is considered to be optimal, and the feature selection is completed; the number of selected features is recorded as N; (3) GA-XGBoost classification model construction: The feature vector subset is divided into multiple subsets, each subset is used as a validation set in turn, and other subsets are used as training sets; the GA genetic algorithm represents the XGBoost hyperparameter combination to be optimized as a "chromosome", inputs a randomly generated hyperparameter combination, trains the XGBoost model with each hyperparameter combination, and calculates the model accuracy index as the fitness value on the validation set, generates a new generation of hyperparameter combinations through selection, crossover, and mutation operations, repeats the above operations until the preset number of iterations is reached, and outputs the chromosome with the highest fitness value as the optimal hyperparameter combination of the optimized XGBoost algorithm, and obtains the best parameter GA-XGBoost model; (4) Classification of seabed sediments: All feature vectors of other study areas are extracted and input into the trained GA-XGBoost model to classify the feature vectors of each study area.

2. The multi-source data seabed sediment classification method based on GA-XGBoost as claimed in claim 1, characterized in that In step (1), the radiation correction and atmospheric correction preprocessing is to preprocess the multispectral image using ENVI software; The water-land separation according to the normalized difference water index method is performed by setting a threshold , Value not less than The area is a water body, otherwise it is a non-water body. The formula is as follows: ， In the formula, The radiation brightness value of the green light band; is the radiation brightness value in the near-infrared band; is the normalized difference water index; is the threshold of the normalized difference water index method, .

3. The multi-source data seabed sediment classification method based on GA-XGBoost as claimed in claim 1, characterized in that In step (0), the multispectral image is WorldView-2 image. In step (1), the acoustic data and optical remote sensing data are spatially aligned, with the 1.85 m resolution of WorldView-2 image as a reference, and the bathymetric data is resampled and coordinate converted to unify the coordinate system and resolution of the two data sources.

4. The multi-source data seabed sediment classification method based on GA-XGBoost as claimed in claim 1, characterized in that In step (2.1) of spectral feature extraction, the calculation method of 10-dimensional spectral features is shown in Table 1: Table 1 Spectral characteristic index calculation formula ， Among them, R C Represents the CoastalBule radiation brightness value, R B Indicates the Bule radiation brightness value, R G Indicates the Green radiation brightness value, R Y Represents the Yellow radiation brightness value, R R Indicates the Red radiation brightness value, R RE Represents the RedEdge radiation brightness value, R NIR1 Indicates the NearInfrared1 radiation brightness value, R NIR2 Represents the NearInfrared2 radiation brightness value; C1=6, C2=7.5, L=1, G=2.

5.

5. The multi-source data seabed sediment classification method based on GA-XGBoost as claimed in claim 1, characterized in that In step (2.2) terrain feature extraction, the formulas for each feature index are shown in Table 2: Table 2 Calculation formula of terrain characteristic index ， in, represents the spatial coordinates, is the target point elevation, The elevation of a certain adjacent grid point, is the horizontal distance between this point and its adjacent point, is the normal of the 3D surface x , y Directional component; slope i is the slope between this point and the adjacent point; arctan 2 (Ny , Nx) For calculation The arc tangent of the argument.

6. The multi-source data seabed sediment classification method based on GA-XGBoost as claimed in claim 1, characterized in that In step (2.4), the specific formula for calculating the average information gain is: ， In the formula, Indicates j The node of the feature, j ∈1,…, N ; represents the number of all trees, Indicates The number of non-leaf nodes in a tree, Indicates The tree The partitioning features of non-leaf nodes, so , is the indicator function; is the hyperparameter of the regularization term, , Respectively indicate that they fall in The tree The sum of the first-order derivative and the second-order derivative of all samples on non-leaf nodes, Respectively indicate that they fall in The first The sum of the first-order derivatives on the left and right nodes of non-leaf nodes. Similarly, Respectively indicate that they fall in The first The sum of the second-order derivatives on the left and right nodes of non-leaf nodes, so: ， 。 7. The multi-source data seabed sediment classification method based on GA-XGBoost as claimed in claim 1, characterized in that In step (2.4), the gain values ​​of each feature are accumulated and averaged to obtain the final importance score of the feature as follows: ， j ∈1,…, N 。 8. The multi-source data seabed sediment classification method based on GA-XGBoost as claimed in claim 1, characterized in that In step (2.4), each time a dimension of features is added, the change in classification accuracy is observed. If the classification accuracy drops by more than 1% after adding the new features, or the increase compared with the previous highest accuracy does not exceed 0.5%, the current dimensional feature set is considered to be optimal, and feature selection is completed.

9. The multi-source data seabed sediment classification method based on GA-XGBoost as claimed in claim 1, characterized in that In step (3), the accuracy performance index is selected to calculate the fitness value as follows: ， in, s is the number of iterations of the GA genetic algorithm, and s The chromosome with the highest output fitness value in the iteration is used as the optimal hyperparameter combination of the optimized XGBoost algorithm to obtain the optimal parameter GA-XGBoost model.

10. The multi-source data seabed sediment classification method based on GA-XGBoost as claimed in claim 1, characterized in that In step (3), the expression of the classification model can be summarized as: ， Where: C is the CA-XGBoost classification algorithm, Feuture1— are the N features determined in step (2.4).

Images