Shallow seabed classification method and system based on multi-scale multi-direction feature fusion

Abstract

The present application relates to the technical field of marine geological exploration, especially to a shallow seabed classification method and system based on multi-scale and multi-direction feature fusion, the method comprising: acquiring multi-source acoustic data of the shallow seabed, preprocessing the acquired multi-source acoustic data, constructing a spatial correlation model based on the preprocessed multi-source acoustic data, extracting seabed topographic features according to the water depth value of the MBES data, extracting multi-scale and multi-direction texture features by using the SSS image output by the spatial correlation model, optimizing the acquired topographic features and texture features, taking the core feature set as input, training a multi-class classifier and constructing a multi-model collaborative decision mechanism, outputting the seabed classification result and evaluating the classification accuracy through quantitative indicators, the present application takes the echo intensity image as the core, solves the problem of poor spatial consistency in traditional multi-source data registration, avoids classification errors caused by feature space misplacement, and improves the classification reliability.

Description

Technical Field

[0001] This invention relates to the field of marine geological exploration technology, and in particular to a shallow seabed sediment classification method and system based on multi-scale and multi-directional feature fusion. Background Technology

[0002] As a key transition zone between terrestrial and marine ecosystems, shallow island and reef areas are characterized by high ecological sensitivity and complex topographic structure. The type, composition and spatial distribution information of the seabed sediment are core basic data for marine resource development, ecological environment protection and marine engineering design. Currently, the mainstream technologies mainly rely on two types of acoustic equipment: multibeam echo sounder (MBES) and side-scan sonar (SSS). The complementarity of these two technologies makes it possible to accurately classify the seabed sediment.

[0003] However, existing seabed classification techniques based on MBES and SSS data still have significant limitations and are difficult to adapt to the complex environment of shallow island and reef areas. On the one hand, MBES is affected by the large beam footprints in shallow water areas, resulting in reduced spatial resolution and an inability to provide detailed seabed features. Although SSS has high texture resolution, its data positioning accuracy is easily affected by factors such as wind, waves, boat speed, and changes in the height of the trawler, making it difficult to guarantee spatial consistency. On the other hand, existing multi-source data fusion methods mostly remain at the level of simple spatial registration, failing to address the fundamental differences between MBES and SSS in terms of spatial resolution and acoustic directivity, leading to inaccuracies in classification results. Due to insufficient stability and reliability, some researchers have attempted to explore optimized paths for multi-source acoustic data fusion in recent years, such as combining airborne LiDAR and multispectral imagery to assist MBES data classification, or fusing on-site video data to enhance SSS texture recognition capabilities. However, the former has limited applicability in shallow island and reef areas with turbid water and cannot overcome the interference of water quality on optical data. The latter has failed to address the core issues of scale and directional feature differences in multi-source data and relies on additional on-site sampling data, increasing cost and operational complexity. At present, there is a need for shallow seabed sediment classification methods and systems based on multi-scale and multi-directional feature fusion. Summary of the Invention

[0004] To address the issues of low accuracy and insufficient reliability in seabed sediment classification for islands and reefs, this invention provides a shallow seabed sediment classification method and system based on multi-scale and multi-directional feature fusion.

[0005] Firstly, the shallow seabed sediment classification method based on multi-scale and multi-directional feature fusion provided by this invention adopts the following technical solution:

[0006] Shallow seabed sediment classification methods based on multi-scale and multi-directional feature fusion include:

[0007] Acquire multi-source acoustic data of shallow seabed sediments and preprocess the acquired multi-source acoustic data, including acquiring MBES data and SSS data;

[0008] A spatial correlation model is constructed based on preprocessed multi-source acoustic data, including determining key regions using echo intensity images in MBES data, and optimizing SSS data by local matching and feature point matching based on key regions.

[0009] Seabed topographic features are extracted based on water depth values ​​from MBES data, including three types of topographic features that reflect the characteristics of the seabed sediment: slope, aspect, and roughness.

[0010] Texture feature extraction is performed on the SSS image output by the spatial association model, including using the gray-level co-occurrence matrix, extracting texture features based on multi-scale and multi-directional parameters, and constructing a texture feature description vector;

[0011] The acquired terrain and texture features are optimized to obtain a core feature set, including weight calculation, sorting and filtering, and redundancy removal of different features using an improved Relief-F algorithm.

[0012] Using the core feature set as input, a multi-class classifier is trained and a multi-model collaborative decision-making mechanism is constructed. The bottom sediment classification results are output and the classification accuracy is evaluated by quantitative indicators.

[0013] Furthermore, the preprocessing of the acquired multi-source acoustic data includes converting MBES data and SSS data to the same projection coordinate system, using the polygonal region actually covered by the MBES data as the reference boundary, performing spatial inclusion cropping on the SSS data image based on the reference boundary, converting the water depth values ​​in all MBES data to a unified vertical reference, and performing geometric correction on the water depth values ​​of the SSS data based on the unified vertical reference. The geometric correction expression is:

[0014] ;

[0015] in, To correct the water depth values ​​in the SSS data, The raw water depth values ​​from the SSS data. Indicated as towing height, The slant range of the sonar signal. This represents the horizontal distance component.

[0016] Furthermore, the step of determining key regions using echo intensity images in MBES data includes extracting pixel gradient information of echo intensity images in MBES data using the Canny edge detection operator, setting significant edge regions based on the edge results extracted by the Canny edge detection operator, calculating the gray-level variance within a local window for the image after extracting pixel gradient information, setting an empirical threshold and determining high-variance regions based on the gray-level variance, and performing intersection operations between the high-variance regions and significant edge regions to generate an anchoring region mask.

[0017] Furthermore, the optimization of local matching and feature point matching of SSS data based on key regions includes defining MBES key regions based on anchored region masks, using the center coordinates of the MBES key regions as the origin of the SSS data image, obtaining multiple candidate matching windows in a preset neighborhood using the sliding window method, calculating the matching degree between each candidate matching window and the MBES key region template through normalized cross-correlation coefficients, and determining the initial corresponding region pair based on the cross-correlation coefficients. The matching degree calculation formula is as follows:

[0018] ;

[0019] Where A is the MBES critical region template, This represents the grayscale value of the template at the coordinates (x, y). The average grayscale value of the template. For candidate matching windows in the SSS image, This represents the grayscale value of the window at coordinates (x, y). This represents the average grayscale value of the window.

[0020] Furthermore, the optimization of local matching and feature point matching of SSS data based on key regions also includes using the SIFT algorithm to extract feature points in the initial corresponding region pairs, performing nearest neighbor matching on the extracted feature points based on Euclidean distance, eliminating mismatched points through the random consistency sampling algorithm to obtain a set of high-confidence matching point pairs, using cubic spline functions to fit the coordinate mapping relationship between MBES data and SSS data in the x and y directions respectively, using the set of high-confidence matching point pairs as control points, solving the fitting coefficients through the least squares method, and combining the fitting coefficients and the cubic spline function to obtain the spatial association model.

[0021] Furthermore, the extraction of seabed topographic features based on the water depth values ​​of MBES data includes: calculating the elevation change rate of the target pixel in the x-axis and y-axis directions using a difference algorithm based on the water depth values ​​of MBES data; converting the elevation change rate into an angle value using an arctangent function to obtain the slope; calculating the initial aspect value based on the elevation change rate using the arctangent function; adjusting the initial aspect value using the positive and negative signs of the elevation change rate in the x-axis and y-axis directions; and finally, fitting the water depth data of the seabed area covered by MBES data into a three-dimensional surface, calculating the ratio of the actual area of ​​the three-dimensional surface to the projected area of ​​the seabed area to obtain the roughness.

[0022] Furthermore, the extraction of texture features from the SSS image output by the spatial association model includes presetting multiple sliding windows of different sizes and multiple main directions of different angles as multi-scale parameters and multi-directional parameters, respectively. Gray-level compression processing is performed on the position-calibrated SSS image. For the combination of multi-scale and multi-directional parameters, the occurrence frequency of different gray-level combinations is counted within the corresponding sliding window according to the preset pixel spacing, and the occurrence frequency is normalized to obtain a probability value. A gray-level co-occurrence matrix is ​​constructed based on the probability value. Multiple texture features are extracted based on each gray-level co-occurrence matrix, and the multiple texture features are sorted to obtain a texture feature description vector.

[0023] Furthermore, the feature optimization of the acquired terrain and texture features includes introducing class prior probability and distance normalization mechanisms, calculating feature weights for terrain and texture features, calculating the nearest neighbors of all features of the same class and the nearest neighbors of different classes using Euclidean distance, calculating the feature difference between each feature based on the nearest neighbors of the same class and the nearest neighbors of different classes, iteratively updating the feature weights according to the feature differences, retaining high-contribution features by combining the weight threshold determined by cross-validation, and removing redundant features within the high-contribution features using the Pearson correlation coefficient. The feature difference formula is:

[0024] ;

[0025] in, For the feature samples to be processed, For category The prior probability, For category In and sample The j-th nearest sample, For the sample The true substrate category, The maximum value among all samples. The minimum value among all samples. For the sample Prior probability of substrate type.

[0026] Furthermore, the training of multi-class classifiers and the construction of a multi-model collaborative decision-making mechanism include selecting random forest, K-nearest neighbors, support vector machine, random undersampling enhancement algorithm, and wide neural network as multi-class classifiers, training each classifier separately using a 5-fold cross-validation mechanism, calculating the weight of each classifier using a normalization method based on the overall classification accuracy of each classifier on the validation fold of the 5-fold cross-validation, performing a weighted summation based on the weights of each classifier, and taking the class with the highest score after weighted summation as the final background classification result for the sample. The classification result expression is:

[0027] ;

[0028] in, For the number of classifiers, For the first The weights of each classifier For classifier Category The predicted probability.

[0029] Secondly, a shallow seabed sediment classification system based on multi-scale and multi-directional feature fusion includes:

[0030] The data acquisition module is configured to acquire multi-source acoustic data of shallow seabed sediment and preprocess the acquired multi-source acoustic data, including acquiring MBES data and SSS data.

[0031] The correlation model module is configured to: construct a spatial correlation model based on preprocessed multi-source acoustic data, including using echo intensity images in MBES data to determine key regions, and optimizing local matching and feature point matching of SSS data based on key regions;

[0032] The terrain feature module is configured to extract seabed terrain features based on the water depth values ​​of MBES data, including three types of terrain features that reflect the characteristics of the seabed sediment: slope, aspect, and roughness, calculated based on MBES data.

[0033] The texture feature module is configured to extract texture features from the SSS image output by the spatial association model, including using the gray-level co-occurrence matrix, extracting texture features based on multi-scale and multi-directional parameters, and constructing a texture feature description vector.

[0034] The optimization module is configured to: perform feature optimization on the acquired terrain and texture features to obtain a core feature set, including using an improved Relief-F algorithm to calculate weights, sort and filter different features, and remove redundancy.

[0035] The output module is configured to: take the core feature set as input, train a multi-class classifier and build a multi-model collaborative decision-making mechanism, output the sediment classification result and evaluate the classification accuracy through quantitative indicators.

[0036] In summary, the present invention has the following beneficial technical effects:

[0037] 1. This invention utilizes a unified coordinate system, unified regional boundaries, and unified elevation benchmark processing for multi-source acoustic data. Combined with dedicated geometric correction for side-scan sonar (SSS) data, it effectively eliminates spatial misalignment and elevation deviation caused by differences in acquisition principles between multibeam echo sounder (MBES) and SSS data. This ensures high consistency between the two types of data in spatial location, coverage, and elevation benchmark, avoiding subsequent feature mis-extraction caused by data deviation. It provides high-quality, highly consistent foundational data for subsequent classification, guaranteeing classification accuracy from the data source.

[0038] 2. This invention uses MBES echo intensity images as the core. It generates feature-stable anchor regions through edge enhancement, local variance filtering, and region intersection. Then, it combines local matching, feature point extraction, and mismatch removal to finally construct a nonlinear spatial association model. This achieves accurate spatial mapping between MBES and SSS data, solving the problem of poor spatial consistency in traditional multi-source data registration. It ensures that terrain features and texture features correspond accurately at the pixel scale, providing reliable geometric assurance for multi-source feature fusion, avoiding classification errors caused by feature space misalignment, and improving classification reliability.

[0039] 3. This invention extracts key features reflecting seabed topographic characteristics based on MBES data, and extracts multi-scale and multi-directional texture features based on SSS data, constructing a dual-dimensional feature system of topography and texture. This system comprehensively covers the topographic undulation characteristics and texture details of the seabed, effectively making up for the shortcomings of traditional single features in distinguishing seabeds with similar topography but different textures or similar textures but different topography. It significantly improves the separability of different seabed categories and provides sufficient feature support for high-precision classification.

[0040] 4. This invention employs an improved feature optimization algorithm. By introducing prior class probability and distance normalization mechanism, it accurately calculates feature contribution weights, selects high-value features and removes redundant information, reduces the interference of non-discriminating redundant features on the classification model, lowers the risk of model overfitting, ensures that the classification model focuses on the most critical core features for distinguishing the substrate, improves the model's classification stability and generalization ability, and further guarantees the reliability of the classification results.

[0041] 5. This invention constructs a multi-classifier collaborative decision-making mechanism, combining the advantages of different classifiers in scenarios such as high-dimensional features and imbalanced samples. By merging the output results of each classifier through weighted voting, this method avoids the performance limitations of traditional single classifiers in complex shallow island and reef environments, effectively addresses issues such as seabed mixing and sample imbalance, improves the ability to identify rare seabed types, and ensures that high-precision and high-reliability seabed classification results can be output in different marine environments. Attached Figure Description

[0042] Figure 1 This is a schematic diagram of the overall process of the shallow seabed sediment classification method based on multi-scale and multi-directional feature fusion according to an embodiment of the present invention.

[0043] Figure 2 These are the effect diagrams at various scales in the shallow seabed sediment classification method based on multi-scale and multi-directional feature fusion in this embodiment of the invention.

[0044] Figure 3 This is a diagram showing the performance of the RF classifier in the shallow seabed sediment classification method based on multi-scale and multi-directional feature fusion in an embodiment of the present invention.

[0045] Figure 4 This is a rendering of the RUSBoost algorithm used in the shallow seabed sediment classification method based on multi-scale and multi-directional feature fusion in this invention. Detailed Implementation

[0046] The present invention will be further described in detail below with reference to the accompanying drawings.

[0047] Example 1

[0048] Reference Figure 1 The shallow seabed sediment classification method based on multi-scale and multi-directional feature fusion in this embodiment includes:

[0049] Acquire multi-source acoustic data of shallow seabed sediments and preprocess the acquired multi-source acoustic data, including acquiring MBES data and SSS data;

[0050] A spatial correlation model is constructed based on preprocessed multi-source acoustic data, including determining key regions using echo intensity images in MBES data, and optimizing SSS data by local matching and feature point matching based on key regions.

[0051] Seabed topographic features are extracted based on water depth values ​​from MBES data, including three types of topographic features that reflect the characteristics of the seabed sediment: slope, aspect, and roughness.

[0052] Texture feature extraction is performed on the SSS image output by the spatial association model, including using the gray-level co-occurrence matrix, extracting texture features based on multi-scale and multi-directional parameters, and constructing a texture feature description vector;

[0053] The acquired terrain and texture features are optimized to obtain a core feature set, including weight calculation, sorting and filtering, and redundancy removal of different features using an improved Relief-F algorithm.

[0054] Using the core feature set as input, a multi-class classifier is trained and a multi-model collaborative decision-making mechanism is constructed. The bottom sediment classification results are output and the classification accuracy is evaluated by quantitative indicators.

[0055] Specifically, the shallow seabed sediment classification method based on multi-scale and multi-directional feature fusion includes the following steps:

[0056] like Figure 1 , Figure 2 As shown, S1, acquire multi-source acoustic data of shallow seabed sediment and preprocess the acquired multi-source acoustic data, including acquiring MBES data and SSS data;

[0057] In the target shallow island and reef survey area, a multibeam echo sounder (MBES) and side-scan sonar (SSS) equipment were deployed in a collaborative operation mode to conduct simultaneous data acquisition. MBES data acquisition involved transmitting multiple sonar signals to the seabed and receiving reflected signals. Two core types of data were recorded simultaneously: first, high-precision water depth data, i.e., the initial vertical distance from each sampling point on the seabed to the sea level; and second, echo intensity image data, i.e., grayscale images formed by quantizing the reflection intensity of sonar signals from different seabed sediments, with grayscale values ​​ranging from 0-255, reflecting differences in seabed reflectivity. During the acquisition process, through... The MBES, equipped with GPS and attitude sensors, simultaneously records the original spatial coordinates and equipment attitude parameters corresponding to each set of water depth and echo intensity data. SSS data acquisition involves the ship towing the SSS towed body along a pre-set survey line. The towed body emits high-frequency sonar signals to both sides of the seabed. After receiving the reflected signals, it records two types of core data: high-resolution sonar intensity image data (a continuous grayscale image formed by the intensity of the reflected signals from the seabed) and original water depth data (the initial vertical distance of the seabed sampling point calculated using the round-trip time of the sonar signal). During the acquisition process, the towed body's height is recorded by a pressure sensor. The slant range of the sonar signal is calculated by combining the propagation speed of the sonar signal and the round-trip time. Simultaneously record the original spatial coordinates and track information.

[0058] All collected MBES and SSS data must be uniformly converted to the WGS84-UTM projection coordinate system to ensure consistent planar position reference standards. If the original data uses a local coordinate system or other projection methods, the conversion must be completed using a seven-parameter coordinate transformation model. The formula for the seven-parameter coordinate transformation model is as follows:

[0059] ;

[0060] in,( , , ) represents the source coordinates of the original data, ( , , () represents the target coordinates in the unified coordinate system after transformation. This is a scale factor used to correct for scale differences between different coordinate systems. This is a rotation matrix used to align coordinates in spatial angles. , , As a translation vector, it represents the positional offset of the whole coordinates in the plane and vertical directions. This model can eliminate spatial misalignment problems caused by different coordinate references from different data sources.

[0061] Based on the coordinates of the MBES depth data sampling points after coordinate transformation, a polygon fitting algorithm (such as Delaunay triangulation) is used to generate the polygonal region actually covered by the MBES data. The vertices of this polygon are the outermost coordinates of the MBES data sampling points, ensuring complete encirclement of all valid MBES data. The polygonal region actually covered by MBES measurements is then used. As a baseline boundary, spatial inclusion clipping is performed on the SSS image. A spatial inclusion judgment algorithm is used; in this embodiment, the ray method is employed. The SSS sonar intensity image after coordinate transformation is judged pixel by pixel, retaining only the overlapping areas of the two data sets and removing invalid pixels outside the survey area to ensure strict consistency of the data range. The clipping formula is:

[0062] ;

[0063] in, This is the original side-scan sonar image. The cropped side-scan sonar image is consistent with the MBES data range. Using the baseline boundary, all MBES depth data are then converted to a unified vertical datum to ensure consistent reference standards for depth values. This process includes two parts: MBES depth value conversion and SSS depth value geometric correction. The MBES depth value vertical datum conversion calculates the instantaneous tide level at the time of MBES raw depth data acquisition based on the tidal harmonic constant provided by the tidal observation station in the survey area. The raw MBES depth values ​​are then subtracted from the instantaneous tide level, using the instantaneous sea level as the reference, to obtain unified MBES depth values ​​based on the theoretical lowest tide level. This ensures consistency of all MBES depth data on the vertical datum. The SSS depth value geometric correction is based on the towed body height of the raw SSS depth data. The distance from the bottom of the towed body to the sea level and the sonar signal slant range (i.e., the distance from the bottom of the towed body to the sea level) are significantly affected, resulting in a deviation. This deviation needs to be eliminated using a geometric correction formula to obtain a corrected SSS depth value based on the theoretical lowest tide level. For SSS data, due to the height of the towed body during its acquisition process... Slope distance This will introduce geometric deviations, which need to be corrected geometrically. The correction formula is as follows:

[0064] ;

[0065] in, To correct the water depth values ​​in the SSS data, The raw water depth values ​​from the SSS data. Indicated as towing height, The slant range of the sonar signal. The horizontal distance component is used for correction. This correction eliminates the influence of tow body height and slant distance on the accuracy of SSS water depth data, ensuring that MBES and SSS data are consistent in the elevation dimension.

[0066] S2. Construct a spatial correlation model based on the preprocessed multi-source acoustic data, including using echo intensity images in MBES data to determine key regions, and optimizing SSS data by local matching and feature point matching based on key regions.

[0067] Using the preprocessed MBES echo intensity image as the core, the seabed topographic boundaries and transition zones between different seabed substrates in the MBES echo intensity image exhibit abrupt changes in grayscale values. Pixel gradient information needs to be extracted using the Canny edge detection operator. First, Gaussian filtering is applied to the MBES echo intensity image, using a 5×5 Gaussian kernel to smooth image noise, such as grayscale fluctuations caused by shallow-sea bubble interference. The expression for the Gaussian kernel function is:

[0068] ;

[0069] in, The standard deviation of the Gaussian kernel is used to balance noise reduction and edge preservation. The function represents the relative coordinates of pixels within the filtering window. It performs window-by-window convolution on the image to eliminate interference from high-frequency noise in edge extraction. For the filtered image, the Sobel operator is used to calculate the gray-level change rate along the horizontal x-axis and vertical y-axis, respectively. Then, the gradient intensity of each pixel is calculated using the gradient magnitude formula. The expression is:

[0070] ;

[0071] in, The grayscale change rate is represented by the x-axis. The gradient intensity represents the grayscale change rate along the y-axis. A larger gradient intensity indicates that the pixel is more likely to be a terrain or substrate boundary, i.e., an edge region. Non-maximum suppression is performed on the acquired gradient intensity image. First, the gradient direction of each pixel is determined; the gradient direction is the direction of the most dramatic grayscale change at that pixel, calculated using the grayscale change rate along the x and y axes. Along the gradient direction of each pixel, two adjacent pixels are selected as comparison objects. If the gradient direction is 0°, the left and right adjacent pixels are compared; if the gradient direction is 45°, the upper left and lower right adjacent pixels are compared. The gradient intensity of the current pixel is then... Compare the gradient intensity of the current pixel with the gradient intensity of the two comparison pixels. If the gradient intensity of the current pixel is greater than the gradient intensity of the two comparison pixels, the pixel is determined to be a local maximum pixel in the gradient direction and its gradient intensity value is retained. If the gradient intensity of the current pixel is less than or equal to the gradient intensity of any comparison pixel, the pixel is determined to be a redundant pixel that is not at the edge core and its gradient intensity value is set to 0 to be removed.

[0072] By statistically analyzing the actual sample grayscale distribution of MBES echo intensity images in the test area, a low threshold is adaptively determined. and in accordance with Determine the high threshold This ratio ensures that a high threshold accurately locks the edge core, while a low threshold covers the extended parts of the edge. Based on the dual thresholds, the gradient intensity image after non-maximum suppression is classified pixel-by-pixel. If the pixel gradient intensity... Pixels identified as having strong edges are those corresponding to the core region of seabed topography or substrate boundaries, exhibiting drastic and stable grayscale changes, and are directly retained. If the pixel gradient intensity... Pixels classified as non-edge pixels are often located in areas with flat gray levels within the substrate or areas of noise interference. Their gradient intensity values ​​are set to 0 to remove them. If the pixel gradient intensity satisfies... Pixels are identified as weak edge pixels. For all weak edge pixels, an 8-neighborhood connectivity check is performed to see if there are strong edge pixels among the eight neighboring pixels. If at least one strong edge pixel exists, the weak edge pixel is considered an "extension of a strong edge," and its gradient intensity value is retained. If no strong edge pixel exists, the weak edge pixel is considered an "isolated false edge," and its gradient intensity value is set to 0 to remove it. The retained strong edge pixels and connected weak edge pixels together constitute a significant edge region. The significant edge region determination function is:

[0073] ;

[0074] in, Indicates whether a pixel belongs to a significant edge region. The threshold for edge saliency is used to ensure that the anchored area has both obvious edge intensity features and sufficient texture variation information, thereby enhancing the stability and representativeness of subsequent anchoring.

[0075] The edge-enhanced image is traversed using a 5×5 sliding window, and the local variance of the gray values ​​of all pixels within each window is calculated. An empirical variance threshold was set by statistically analyzing known sediment samples from the test area. In this embodiment, the average variance of the fine sand substrate is 1.5 times that of the original sample, retaining all... The pixels corresponding to the window form a high-variance region. A spatial intersection operation is performed between the salient edge region and the high-variance region, retaining only the pixels that belong to both the salient edge region and the high-variance region, thus generating an anchoring region mask. The expression for the anchoring region mask is:

[0076] ;

[0077] in, For binary masks of salient edge regions, For high-variance regions, a binary mask is used. This represents a pixel-wise logical AND operation. The set of pixels with a value of 1 in the mask is the MBES key region. Using the MBES key region as a template, in the preprocessed SSS sonar intensity image (with coordinate unification and region cropping completed), the optimal SSS matching region most similar to the template is found through a sliding window and normalized cross-correlation coefficient (NCC), forming an initial corresponding region pair, and anchoring the region mask. Extract the WGS84-UTM coordinates of all cells with a median value of 1. The spatial center coordinates of the key area were calculated using the mean method. The expression is: Where N is the total number of pixels in the key region of MBES. Since the SSS coordinates are consistent with the MBES coordinate reference after preprocessing, The search center is directly mapped to the SSS image. Combined with the maximum offset caused by the attitude fluctuation of the SSS tow body in shallow sea environment, a square search neighborhood of 30×30 pixels is set so that the search neighborhood can completely cover the possible offset range and ensure that the optimal matching area is not missed.

[0078] Then, the grayscale matrix of the key MBES region is extracted as the matching template A. Within the SSS search neighborhood, a sliding window of the same size as the template is moved pixel by pixel. Each move generates an SSS candidate matching window B. For each candidate window B, its similarity to the MBES template A is calculated using the cross-correlation coefficient (NCC) formula. The matching degree calculation formula is as follows:

[0079] ;

[0080] Where A is the MBES critical region template, This represents the grayscale value of the template at the coordinates (x, y). The average grayscale value of the template. For candidate matching windows in the SSS image, This represents the grayscale value of the window at coordinates (x, y). Given the mean grayscale value of the window, iterate through all candidate windows and retain the window with the largest NCC value. The SSS image region corresponding to this window is the optimal matching region of SSS. The MBES key region and the optimal matching region of SSS together constitute the initial corresponding region pair.

[0081] The next step is to perform feature point matching and optimization on the MBES key region and the SSS optimal matching region in the initial region pairing. First, feature points in the initial corresponding region pair are extracted using the SIFT algorithm. Gaussian difference pyramids are constructed for the MBES key region and the SSS optimal matching region respectively. Multi-scale images are generated by convolution with Gaussian kernels of different standard deviations. Then, the difference between adjacent scale images is calculated, and local extrema in the pyramid are detected as potential feature points to ensure that the feature points have scale invariance. For potential feature points, their sub-pixel positions are accurately determined by fitting a quadratic function. Thresholds for contrast and edge response are set to remove unstable points with low contrast and edge response. In this embodiment, the thresholds for contrast and edge response are 0.03 and 10, respectively. Then, based on the gradient orientation histogram in the neighborhood of the feature point, the direction corresponding to the peak of the histogram is taken as the main direction of the feature point. Taking the feature point as the center, a 16×16 neighborhood is taken and divided into 4×4 sub-blocks. Gradient histograms of 8 directions are calculated for each sub-block to generate a 128-dimensional feature descriptor. The similarity is measured based on the Euclidean distance of the feature descriptors. For each feature point descriptor in the MBES key region... All descriptors in the optimal matching region of SSS In the process, the descriptor with the smallest distance is selected as the matching candidate. The Euclidean distance formula is:

[0082] ;

[0083] in, , Represented as descriptors and The smaller the distance between the k-th component, the higher the descriptor similarity. The Random Consensus Sampling (RANSAC) algorithm is used to remove mismatched points. Four pairs of initial matching points are randomly selected, and a fundamental matrix is ​​fitted to describe the geometric relationship between the two views. An error threshold is set. The error threshold is set to 1-2 pixels, determined based on the image resolution, and all matching points are counted to satisfy the fundamental matrix constraints. For the interior points, repeat the above process, and retain the set of interior points corresponding to the basis matrix with the largest number of interior points. This set is the high-confidence matching point pair set. .

[0084] Matching point pairs with high confidence Using control points, a cubic spline function is used to fit the nonlinear mapping relationship from SSS pixel coordinates to MBES pixel coordinates, generating a spatial correlation model. The cubic spline function is selected as the nonlinear spatial transformation model, fitting the coordinate mappings in the x and y directions respectively. The model expression is:

[0085] ;

[0086] ;

[0087] in, Let (m,n) be the SSS pixel coordinates, (m,n) be the mapped MBES pixel coordinates, and t be the normalized spatial location parameter. Mapping the SSS pixel coordinates to the [0,1] interval eliminates the influence of coordinate range differences. , , and These are represented as fitting coefficients for different x-directions. , , and These are represented as fitting coefficients in different y-directions.

[0088] Set of high-confidence matching points SSS pixel coordinates With corresponding MBES pixel coordinates ,in, Let M be the number of matching point pairs. Substitute this into the above model to construct an overdetermined system of equations. Minimize the sum of squared errors between the actual mapped coordinates and the model-predicted coordinates using the least squares method to obtain the fitting coefficients in the x and y directions. Calculate the root mean square error (RMSE) of all matching point pairs to verify the model's mapping accuracy. The RMSE expression is:

[0089] ;

[0090] in, To use a spatial association model to determine the pixel coordinates of MBES The SSS pixel coordinates obtained by reverse mapping are ultimately obtained by a cubic spline function that meets the accuracy requirements. , Together they form a spatial association model that can accurately map any SSS pixel coordinate to MBES pixel coordinate, achieving pixel-level spatial association between the two types of data.

[0091] S3. Extract seabed topographic features based on the water depth values ​​of MBES data, including calculating three types of topographic features that reflect the characteristics of the seabed sediment: slope, aspect, and roughness.

[0092] Feature extraction is performed based on preprocessed high-precision MBES depth data. The original MBES depth data is in the form of discrete sampling points. It is first converted into regular grid data and outliers are eliminated. Then, the inverse distance weighted interpolation algorithm is used to interpolate the discrete MBES depth sampling points into a regular grid with a resolution of 1m×1m, resulting in gridded depth data. ,in, The row and column numbers of the grid cells. For the water depth value corresponding to the pixel, the 3σ criterion is used to screen and remove outliers in the gridded water depth data, and then three types of terrain features, namely slope, aspect and roughness, are extracted.

[0093] The slope refers to the angle between the seabed slope and the horizontal plane, reflecting the degree of terrain tilt. The calculation process is based on the elevation change rate of the target pixel along the x and y axes, which is converted into angle values ​​through the arctangent function. A 3×3 sliding window central difference method is used to calculate the elevation change rate of each target pixel (x,y) along the x and y axes in the gridded water depth data. The slope is calculated based on the rate of change of elevation. The slope calculation formula is as follows:

[0094] ;

[0095] Where ∂H / ∂x and ∂H / ∂y are the rates of change of elevation in the horizontal and vertical directions, respectively. The arctangent function calculates the result in radians. Slope aspect refers to the azimuth angle corresponding to the direction of maximum slope (i.e., the direction of most dramatic terrain inclination), reflecting the orientation of the slope. For example, 0° represents true north, and 90° represents true east, with a range of 0°-360°. The calculation process is based on the sign and ratio of the elevation change rates of the x-axis and y-axis. Initial values ​​are calculated using the arctangent function and then adjusted to determine the initial slope aspect. The elevation changes are calculated using the ratio of the rates of change of the x-axis and y-axis, as shown in the following formula:

[0096] ;

[0097] Based on the sign of the elevation change rate along the x and y axes, the initial slope aspect value is adjusted to a standard azimuth angle of 0°-360°, according to the following rules:

[0098] 1. If This indicates that the water depth increases along the positive y-axis, meaning the slope slopes southwards, and eventually the slope... ;

[0099] 2. If This indicates that the water depth decreases along the positive y-axis, meaning the slope slopes northwards, and eventually slopes towards... ;

[0100] 3. If This indicates that there is no elevation change along the y-axis. At this time, it indicates that the water depth decreases along the positive x-axis, with the slope facing west. ;

[0101] when At that time, it indicates that the water depth increases along the positive x-axis, with the slope facing east. ;

[0102] 4. If The target pixel is a flat area with no tilt direction. .

[0103] Finally, roughness calculation is performed. Roughness refers to the complexity of the seabed topography. Coarse-grained seabed material, due to its uneven surface, has a significantly higher roughness value than fine-grained seabed material. This is achieved by fitting a three-dimensional seabed surface and comparing the ratio of the actual surface area to the horizontal projected area. For the gridded water depth data, a bicubic B-spline interpolation algorithm is used to fit the water depth data of the target seabed area into a continuous three-dimensional surface. The expression is:

[0104] ;

[0105] in, The B-spline fitting coefficients are used to minimize the sum of squared errors between the fitted surface and the actual water depth data using the least squares method. The projection of the target seabed region onto the horizontal plane is a regular rectangle, and the area is... It equals the number of grid cells in the region multiplied by the area of ​​a single grid cell, as shown in the following formula:

[0106] ;

[0107] Where K is the total number of grid cells in the target region. For the area of ​​a single grid, based on the fitted 3D surface The actual surface area of ​​the target region is calculated using surface integrals. The discretization calculation formula is as follows:

[0108] ;

[0109] in, The number of grid rows and columns in the target area. , To fit the partial derivatives of the surface along the x and y axes, roughness is defined as the ratio of the actual area of ​​the three-dimensional surface to its horizontal projected area, i.e.:

[0110] ;

[0111] in, The area of ​​a unit surface within the region. The calculated slope, aspect, and roughness are mapped one-to-one with the pixels of the gridded water depth data to form three types of terrain feature data.

[0112] S4. Extract texture features from the SSS image output by the spatial association model, including using the gray-level co-occurrence matrix, extracting texture features based on multi-scale and multi-directional parameters, and constructing a texture feature description vector.

[0113] The position-calibrated SSS image, based on the spatial correlation model output, achieves pixel-level spatial alignment with the MBES data, eliminating spatial misalignment caused by device posture and environmental interference. Although the position-calibrated SSS image has achieved spatial alignment, its original grayscale value range may fluctuate due to sonar signal strength. Therefore, grayscale normalization is performed on it, traversing all pixels of the position-calibrated SSS image to obtain the minimum value of the original grayscale. With the maximum value A linear transformation is used to map the original grayscale values ​​to a standard range of 0-255, eliminating the influence of signal strength fluctuations in different regions. The standardization formula is as follows:

[0114] ;

[0115] in, This represents the original grayscale value of pixel (x,y) in the SSS image after position calibration. To ensure that all pixel grayscale values ​​are within the same order of magnitude for subsequent processing, the standardized grayscale values ​​are used. The minimum value of the original grayscale value. This represents the maximum value of the original grayscale value.

[0116] Three sliding window sizes of 3×3, 5×5, and 7×7 were selected as multi-scale parameters, each corresponding to the substrate texture response requirements. The sliding window traversed the SSS image pixel by pixel to ensure that each pixel could be analyzed at different scales. Then, four main directions of 0°, 45°, 90°, and 135° were selected as multi-directional parameters to comprehensively cover the spatial orientation differences of the substrate texture. Each directional parameter acted independently on the sliding window. Afterward, grayscale compression was performed on the position-calibrated SSS image to reduce computational complexity and retain key texture information. Considering the need to distinguish substrate textures in shallow island and reef areas, 16 levels were selected as the compression target. The standardized grayscale value range was divided into 16 intervals at equal intervals, with each interval corresponding to a compressed grayscale level. , The specific division formula is as follows:

[0117] ;

[0118] in, The function is rounded down. Then, a gray-level co-occurrence matrix (GLCM) is constructed. The GLCM describes the probability of gray-level combinations occurring between two pixels at a certain distance and in a certain direction in an image, quantifying the spatial distribution of texture. Combining the SSS image resolution and the background texture period, a pixel spacing of 1 is selected as the statistical interval. The image is traversed using a sliding window at the current scale, and for each pixel within the window, the gray-level co-occurrence matrix is ​​calculated according to the set direction. Given a pixel spacing d=1, count the occurrences of gray level combinations (i,j) of all target pixels and their neighboring pixels. Where d represents the pixel spacing. The direction parameter is represented by i, which represents the gray level of the target pixel, and j, which represents the gray level of the neighboring pixels. The number of occurrences is also specified. Divide by the total number of times all gray levels are combined. The probability of obtaining a combination of gray levels ， The normalization formula is as follows:

[0119] ;

[0120] in, Let be the number of occurrences of the gray-level combination (i,j) of the neighboring pixels. The total number of all gray level combinations, based on probability. Construct a gray-level co-occurrence matrix with dimensions of compressed gray levels × compressed gray levels. The matrix elements are Each combination of scale and direction corresponds to one GLCM matrix, generating a total of 3 (scale) × 4 (direction) = 12 GLCM matrices. Eight statistical features are extracted from each GLCM matrix: mean, variance, homogeneity, contrast, correlation, angular second moment, entropy, and dissimilarity. Homogeneity represents the degree of similarity among local image regions, calculated using the following formula:

[0121] ;

[0122] The second moment of an angle represents the regularity or uniformity of the texture, and its calculation formula is:

[0123] ;

[0124] Entropy reflects the information complexity of a texture. The more complex the texture (such as a mixture of shells and sand), the greater the entropy value; the simpler the texture (such as pure fine sand), the smaller the entropy value. The calculation formula is:

[0125] ;

[0126] Difference represents the absolute difference between grayscale values, and the calculation formula is:

[0127] ;

[0128] Texture features extracted from all scale and orientation combinations are integrated in a preset order. For each sorted scale and orientation combination, its corresponding eight statistical feature values ​​(mean, variance, homogeneity, contrast, correlation, second moment, entropy, and variability) are concatenated to form a texture feature description vector. .

[0129] S5. Perform feature optimization on the acquired terrain and texture features to obtain the core feature set, including using the improved Relief-F algorithm to calculate weights, sort and filter different features, and remove redundancy.

[0130] Based on the extracted terrain and texture feature descriptor vectors, the two types of features are concatenated in the order of terrain features first, followed by texture features, to form a multidimensional initial feature space. Among them, terrain features are represented as Since the dimensions of terrain features and texture features differ significantly, Z-score normalization is used to eliminate dimensional interference, resulting in a normalized feature space. Then, sampling points of known seabed types were obtained in the target shallow island and reef survey area. Each sampling point corresponds to a set of coordinates and a seabed type label. Based on the sampling point coordinates (X,Y), the data was analyzed in the standardized feature space. In the process, multidimensional feature values ​​of corresponding pixels are extracted to form a labeled sample set: ,in, Let be the multidimensional feature vector of the i-th sample. Let N be the total number of samples, and the percentage of samples in each category be counted as the prior probability of the category. ,in, For category The number of samples, This is used to balance the contribution of different classes of samples to the weight calculation in subsequent improvements to the Relief-F algorithm, and to prevent the feature differences of minority class samples from being masked by the majority class.

[0131] The traditional Relief-F algorithm does not consider imbalanced class distribution and differences in feature dimensions. This step improves the algorithm by introducing prior class probabilities and distance normalization mechanisms to calculate the contribution weight of each feature to sediment classification. The initial weights of all features are set to 0, i.e., a weight vector. For each sample in the sample set... The nearest neighbor samples of the same class and the nearest neighbor samples of different classes are searched using Euclidean distance. With sample Euclidean distance Used to measure the feature similarity between two samples, in relation to Among samples of the same class, find the sample with the smallest Euclidean distance and denote it as the nearest neighbor of the same class. For each category and For samples of different categories, find the sample with the smallest Euclidean distance for each category, and denote it as the nearest neighbor of that category. To eliminate the interference of differences in feature dimensions on the distance calculation, for each feature dimension k, calculate the sample... The difference in normalized features between the sample and its nearest neighbor sample in this dimension. Nearest neighbor of the same kind The normalized difference of the k-th dimension feature is given by the following formula:

[0132] ;

[0133] in, These are objects to be classified, among which, This represents the N feature values ​​of the i-th sample. Indicates and Nearest neighbor samples of the same type express and Differences in features, followed by calculation of differences between classes, samples nearest neighbors of different class l The normalized difference of the k-th dimension feature is given by the following formula:

[0134] ;

[0135] in, Indicates and Nearest neighbor samples of different classes express and The difference values ​​on the features, combined with the prior probability of the category. For each sample Update the weights of each feature dimension k according to the following formula, and iteratively strengthen the weights of high-contribution features:

[0136] ;

[0137] in, This represents the weighted sum of differences in the k-th dimension features between different classes of samples. The larger the weight, the better. To represent the difference in the k-th dimension of a feature among similar samples, a smaller value indicates that the feature is more stable among similar samples. The sample set is iterated T times. After the iteration terminates, all feature weights are normalized so that the weight values ​​are mapped to the range of 0-1, resulting in the normalized weights. .

[0138] The multidimensional features are sorted from largest to smallest according to their normalized weights to obtain the sorted feature list. ,in, The feature with the highest weight. For features with the minimum weight, 5-fold cross-validation is used, and the optimal weight threshold is determined by the classification accuracy. Specifically:

[0139] 1. Randomly divide the sample set S into 5 mutually exclusive subsets. Each time, take 4 subsets as the training set and 1 subset as the validation set.

[0140] 2. The sorted feature list The first m features are selected sequentially to form a feature subset. A random forest (RF) classifier is trained on the training set and the overall classification accuracy (OA) is calculated on the validation set.

[0141] 3. Repeat steps 1-2 a total of 5 times, calculating the average OA for each m, and finding the minimum m value that maximizes the average OA, denoted as . ;

[0142] 4. Take the sorted number 1. The normalized weights of each feature are used as the weight threshold. ,Right now .

[0143] Then, features with normalized weights greater than the weight threshold are retained to obtain a subset of high-contribution features. Next, for any two features in the subset of high-contribution features, the Pearson correlation coefficient between them is calculated, and a redundancy threshold is set. If the absolute value of the correlation coefficient between two features is greater than the redundancy threshold, it is determined to be a highly redundant feature. For highly redundant feature pairs, their normalized weights are compared, and the feature with the smaller weight is removed. The comparison operation is repeated until the absolute value of the correlation coefficient between any two features in the subset of high-contribution features is less than the redundancy threshold. After the above weight calculation, sorting and filtering and redundancy removal, the core feature set is finally obtained.

[0144] S6. Using the core feature set as input, train multi-class classifiers and build a multi-model collaborative decision-making mechanism to output the sediment classification results and evaluate the classification accuracy through quantitative indicators.

[0145] Using the core feature set as input, the system trains multi-class classifiers, constructs a multi-model collaborative decision-making mechanism, outputs the sediment classification results, and uses quantitative indicators to evaluate the classification accuracy. The specific process is as follows:

[0146] Zero-mean standardization is performed on the core feature set to eliminate the dimensional differences of features in different dimensions; the samples containing the sediment category label are divided into training set and test set in a 6:4 ratio, and the sediment category label is obtained by on-site grab sampling or underwater observation;

[0147] like Figure 3 , Figure 4 As shown, Random Forest (RF), K Nearest Neighbors (KNN), Support Vector Machine (SVM), Random Undersampling Boosting Algorithm (RUSBoost), and Wide Neural Network (BLS) are selected as multi-class classifiers. On the training set, a 5-fold cross-validation mechanism is used to train each classifier separately. The key parameters of each classifier are optimized through grid search to ensure that each classifier achieves optimal performance in its advantageous application scenario.

[0148] Based on the overall classification accuracy (OA) of each classifier on the validation fold of the 5-fold cross-validation, a normalization method is used to calculate the weight of each classifier. The weight value is positively correlated with the overall classification accuracy of the classifier, and the sum of the weights of all classifiers is 1. The test set is input into each trained classifier to obtain the class probability vector output by each classifier. Based on the weights of each classifier, a weighted summation is performed on the class probability vectors of the same test sample. The class with the highest score after weighted summation is taken as the final subspecies classification result for that sample. The expression for the classification result is:

[0149] ;

[0150] in, The number of classifiers (5 in this invention). For the first The weights of each classifier For classifier Category The system predicts probabilities and outputs a seabed sediment classification map for shallow island and reef areas. The classification map uses color coding to distinguish different sediment types. At the same time, it calculates two quantitative indicators, the overall classification accuracy (OA) and the Kappa coefficient, to evaluate the classification accuracy. The overall classification accuracy is the proportion of correctly classified samples to the total number of samples. The Kappa coefficient is used to eliminate the interference of random classification on the accuracy evaluation and to measure the consistency between the classification results and the true sediment categories. Finally, the system outputs the results.

[0151] Example 2

[0152] The difference between this embodiment and Embodiment 1 is that this embodiment provides a shallow seabed sediment classification system based on multi-scale, multi-directional feature fusion, including:

[0153] The data acquisition module is configured to acquire multi-source acoustic data of shallow seabed sediment and preprocess the acquired multi-source acoustic data, including acquiring MBES data and SSS data.

[0154] The correlation model module is configured to: construct a spatial correlation model based on preprocessed multi-source acoustic data, including using echo intensity images in MBES data to determine key regions, and optimizing local matching and feature point matching of SSS data based on key regions;

[0155] The terrain feature module is configured to extract seabed terrain features based on the water depth values ​​of MBES data, including three types of terrain features that reflect the characteristics of the seabed sediment: slope, aspect, and roughness, calculated based on MBES data.

[0156] The texture feature module is configured to extract texture features from the SSS image output by the spatial association model, including using the gray-level co-occurrence matrix, extracting texture features based on multi-scale and multi-directional parameters, and constructing a texture feature description vector.

[0157] The optimization module is configured to: perform feature optimization on the acquired terrain and texture features to obtain a core feature set, including using an improved Relief-F algorithm to calculate weights, sort and filter different features, and remove redundancy.

[0158] The output module is configured to: take the core feature set as input, train a multi-class classifier and build a multi-model collaborative decision-making mechanism, output the sediment classification result and evaluate the classification accuracy through quantitative indicators.

[0159] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A shallow seabed sediment classification method based on multi-scale and multi-directional feature fusion, characterized in that, include: Acquire multi-source acoustic data of shallow seabed sediments and preprocess the acquired multi-source acoustic data, including acquiring MBES data and SSS data; A spatial correlation model is constructed based on preprocessed multi-source acoustic data, including determining key regions using echo intensity images in MBES data, and optimizing SSS data by local matching and feature point matching based on key regions. The method of determining key regions using echo intensity images from MBES data includes: extracting pixel gradient information from the echo intensity images in MBES data using the Canny edge detection operator; performing non-maximum suppression on the image after extracting pixel gradient information; performing edge classification on the image after non-maximum suppression based on adaptive dual thresholds; retaining weak edge pixels connected to strong edge pixels by judging 8-neighborhood connectivity to generate significant edge regions; traversing the significant edge regions using a sliding window of a set size; calculating the gray-level variance within the local window; setting an empirical variance threshold based on the statistical characteristics of gray-level distribution of the substrate sample; determining high variance regions based on the comparison result of the gray-level variance and the empirical variance threshold; and performing spatial intersection operation between the significant edge regions and the high variance regions to generate an anchoring region mask to determine key regions. The optimization of local matching and feature point matching of SSS data based on key regions includes defining MBES key regions based on anchored region masks, using the center coordinates of the MBES key regions as the origin of the SSS data image, obtaining multiple candidate matching windows in a preset neighborhood using the sliding window method, calculating the matching degree between each candidate matching window and the MBES key region template through normalized cross-correlation coefficients, and determining the initial corresponding region pairs based on the cross-correlation coefficients. The matching degree calculation formula is as follows: ， Where A is the MBES critical region template, This represents the grayscale value of the template at the coordinates (x, y). The average grayscale value of the template. For candidate matching windows in the SSS image, This represents the grayscale value of the window at coordinates (x, y). The average grayscale value of the window; Seabed topographic features are extracted based on water depth values ​​from MBES data, including three types of topographic features that reflect the characteristics of the seabed sediment: slope, aspect, and roughness. The process of extracting seabed topographic features based on the water depth values ​​from MBES data includes: calculating the elevation change rate of target pixels in the x-axis and y-axis directions using a difference algorithm based on the water depth values ​​from MBES data; converting the elevation change rate into an angle value using an arctangent function to obtain the slope; calculating the initial aspect value based on the elevation change rate using the arctangent function; adjusting the initial aspect value using the positive and negative signs of the elevation change rate in the x-axis and y-axis directions; and finally, fitting the water depth data of the seabed area covered by MBES data into a three-dimensional surface and calculating the ratio of the actual area of ​​the three-dimensional surface to the projected area of ​​the seabed area to obtain the roughness. Texture feature extraction is performed on the SSS image output by the spatial association model, including using the gray-level co-occurrence matrix, extracting texture features based on multi-scale and multi-directional parameters, and constructing a texture feature description vector; The acquired terrain and texture features are optimized to obtain a core feature set, including weight calculation, sorting and filtering, and redundancy removal of different features using an improved Relief-F algorithm. The feature optimization of the acquired terrain and texture features includes introducing class prior probability and distance normalization mechanisms, calculating feature weights for terrain and texture features, calculating the nearest neighbors of all features of the same class and the nearest neighbors of different classes using Euclidean distance, calculating the feature difference between each feature and the nearest neighbors of the same class and the nearest neighbors of different classes, iteratively updating the feature weights according to the feature differences, retaining high-contribution features by combining the weight threshold determined by cross-validation, and removing redundant features within the high-contribution features using the Pearson correlation coefficient. The feature difference formula is as follows: ， in, For the feature samples to be processed, For category The prior probability, For category In and sample The j-th nearest sample, For the sample The true substrate category, The maximum value among all samples. The minimum value among all samples. For the sample Prior probability of substrate type; The formula for iteratively updating the feature weights is as follows: ， in, This represents the weighted sum of differences in the k-th dimension features between different classes of samples. The larger the weight, the better. To represent the difference in the k-th dimension of a feature among similar samples, a smaller value indicates that the feature is more stable among similar samples. The sample set is iterated T times. After the iteration terminates, all feature weights are normalized so that the weight values ​​are mapped to the range of 0-1, resulting in the normalized weights. The multidimensional features are sorted from largest to smallest according to their normalized weights to obtain a sorted feature list. Using the core feature set as input, a multi-class classifier is trained and a multi-model collaborative decision-making mechanism is constructed. The bottom sediment classification results are output and the classification accuracy is evaluated by quantitative indicators.

2. The shallow seabed sediment classification method based on multi-scale and multi-directional feature fusion according to claim 1, characterized in that, The preprocessing of the acquired multi-source acoustic data includes converting MBES and SSS data to the same projection coordinate system, using the polygonal region actually covered by the MBES data as the reference boundary, performing spatial inclusion cropping on the SSS data image based on the reference boundary, converting the water depth values ​​in all MBES data to a unified vertical reference, and performing geometric correction on the water depth values ​​of the SSS data based on the unified vertical reference. The geometric correction expression is as follows: ， in, To correct the water depth values ​​in the SSS data, The raw water depth values ​​from the SSS data. Indicated as towing height, The slant range of the sonar signal. This represents the horizontal distance component.

3. The shallow seabed sediment classification method based on multi-scale and multi-directional feature fusion according to claim 1, characterized in that, The optimization of local matching and feature point matching of SSS data based on key regions also includes using the SIFT algorithm to extract feature points in the initial corresponding region pairs, performing nearest neighbor matching on the extracted feature points based on Euclidean distance, eliminating mismatched points through the random consistency sampling algorithm to obtain a set of high-confidence matching point pairs, using cubic spline functions to fit the coordinate mapping relationship between MBES data and SSS data in the x and y directions respectively, using the set of high-confidence matching point pairs as control points, solving the fitting coefficients through the least squares method, and combining the fitting coefficients and cubic spline functions to obtain the spatial association model.

4. The shallow seabed sediment classification method based on multi-scale and multi-directional feature fusion according to claim 1, characterized in that, The process of extracting texture features from the SSS image output by the spatial association model includes: presetting multiple sliding windows of different sizes and multiple main directions of different angles as multi-scale parameters and multi-directional parameters, respectively; performing gray-level compression processing on the position-calibrated SSS image; for the combination of multi-scale and multi-directional parameters, counting the occurrence frequency of different gray-level combinations within the corresponding sliding window according to the preset pixel spacing; normalizing the occurrence frequency to obtain probability values; constructing a gray-level co-occurrence matrix based on the probability values; extracting multiple texture features based on each gray-level co-occurrence matrix; and sorting the multiple texture features to obtain a texture feature description vector.

5. The shallow seabed sediment classification method based on multi-scale and multi-directional feature fusion according to claim 1, characterized in that, The training of multi-class classifiers and the construction of a multi-model collaborative decision-making mechanism include selecting random forest, K-nearest neighbors, support vector machine, random undersampling enhancement algorithm, and wide neural network as multi-class classifiers; training each classifier separately using a 5-fold cross-validation mechanism; calculating the weight of each classifier using a normalization method based on the overall classification accuracy of each classifier on the validation fold of the 5-fold cross-validation; performing a weighted summation based on the weights of each classifier; and taking the class with the highest score after weighted summation as the final subspecies classification result for the sample. The classification result expression is: ， in, For the number of classifiers, For the first The weights of each classifier For classifier Category The predicted probability.

6. A shallow seabed sediment classification system based on multi-scale, multi-directional feature fusion, executed according to the method of claim 1, characterized in that, include: The data acquisition module is configured to acquire multi-source acoustic data of shallow seabed sediment and preprocess the acquired multi-source acoustic data, including acquiring MBES data and SSS data. The correlation model module is configured to: construct a spatial correlation model based on preprocessed multi-source acoustic data, including using echo intensity images in MBES data to determine key regions, and optimizing local matching and feature point matching of SSS data based on key regions; The terrain feature module is configured to extract seabed terrain features based on the water depth values ​​of MBES data, including three types of terrain features that reflect the characteristics of the seabed sediment: slope, aspect, and roughness, calculated based on MBES data. The texture feature module is configured to extract texture features from the SSS image output by the spatial association model, including using the gray-level co-occurrence matrix, extracting texture features based on multi-scale and multi-directional parameters, and constructing a texture feature description vector. The optimization module is configured to: perform feature optimization on the acquired terrain and texture features to obtain a core feature set, including using an improved Relief-F algorithm to calculate weights, sort and filter different features, and remove redundancy. The output module is configured to: take the core feature set as input, train a multi-class classifier and build a multi-model collaborative decision-making mechanism, output the sediment classification result and evaluate the classification accuracy through quantitative indicators.

Images