Deep-sea terrain reconstruction method for ocean engineering surveying and mapping

Abstract

The invention relates to the technical field of oceanographic engineering surveying and mapping, and provides an oceanographic engineering surveying and mapping deep sea terrain reconstruction method, which comprises the steps of multi-modal data synchronous acquisition, frequency domain feature cross enhancement, dynamic convolution feature fusion, inverse projection terrain reconstruction and terrain credibility verification. According to the invention, an acousto-optic multi-mode fusion technology is adopted, and the advantages of acoustic penetrability and optical high resolution are combined, so that the complex terrain detail capturing capability is improved; the dynamic convolution network adaptively adjusts convolution kernel parameters through frequency domain energy, the multi-scale feature fusion effect is enhanced, and the accuracy of terrain texture expression is improved; according to the motion distortion compensation method, carrier attitude drift is corrected in real time through inertial navigation data, and the geometric fidelity of a reconstruction model is guaranteed. The method is suitable for a deep sea complex environment, and provides high-precision terrain reconstruction method support for ocean engineering surveying and mapping and ocean resource exploration.

Description

Technical Field

[0001] The present invention belongs to the technical field of ocean engineering surveying and mapping, and particularly relates to a method for reconstructing deep-sea terrain in ocean engineering surveying and mapping. Background Art

[0002] The reconstruction of the seabed terrain in ocean engineering surveying and mapping refers to the process of highly accurately restoring the three-dimensional spatial structure of the seabed terrain through multi-sensor data fusion. The reconstruction of the seabed terrain in ocean engineering surveying and mapping, especially the reconstruction of deep-sea terrain, is an important technical means to support ocean resource development, submarine pipeline laying, ocean environmental monitoring, and underwater engineering safety. With the extension of deep-sea exploration to complex terrain areas, higher requirements are put forward for the accuracy, anti-interference ability, and real-time performance of deep-sea terrain modeling. Traditional methods mainly rely on single-modal detection of multi-beam sonar or side-scan sonar, supplemented by an inertial navigation system for motion compensation. However, in the deep-sea turbid water environment, there are problems such as severe attenuation of acoustic signals and insufficient penetration of optical detection, resulting in blurred terrain details and significant multipath interference, making it difficult to meet the requirements of high-precision engineering surveying.

[0003] In the prior art, the physical limitations of single-modal detection result in incomplete capture of complex terrain features, and the complementary advantages of acoustic and optical data are not fully exploited; the cross-modal data synchronization accuracy is insufficient, and the motion distortion compensation model does not consider the multi-band phase response differences; the feature fusion network lacks the dynamic adaptation ability to acoustic-optic coupling features, and it is difficult to balance detail preservation and noise suppression; the terrain credibility assessment lacks a multi-physical-field joint criterion, and the filtering of abnormal points is likely to cause the loss of effective terrain information. Specifically, the terrain reconstruction method based on a single sensor has the inherent defect of multi-dimensional information loss. Although the sonar system has strong penetration, the resolution of micro-topographic undulations is insufficient; although laser detection can obtain high-precision surface information, it is limited by water turbidity and detection depth. At the same time, multi-sensor collaborative detection faces the problem of spatio-temporal registration. The time synchronization error of different-modal data is likely to cause feature misalignment, and the geometric distortion caused by the attitude drift of the moving carrier will further reduce the topological accuracy of the reconstruction model. At the data processing level, the traditional convolutional neural network has limited ability to fuse acoustic-optic heterogeneous features, and the fixed-scale convolutional kernel is difficult to adapt to the multi-scale feature expression of deep-sea terrain, resulting in the loss of high-frequency details and the distortion of low-frequency contours. In addition, the existing confidence assessment methods are mostly based on the determination of a single physical quantity threshold, lacking the correlation analysis of multi-modal energy attenuation, and it is difficult to effectively distinguish the true terrain reflection from the artifacts of waterborne suspended matter, and the accuracy of abnormal point cloud removal is insufficient.

[0004] These problems severely restrict the terrain reconstruction accuracy and reliability of deep-sea engineering surveying and mapping equipment under complex sea conditions. Designing a method for reconstructing deep-sea terrain in ocean engineering surveying and mapping to solve the above problems is of great significance. Summary of the Invention

[0005] To solve the problems existing in the background technology, the present invention provides a deep-sea terrain reconstruction method for ocean engineering surveying and mapping, which includes the following steps:

[0006] S1. Multi-modal data synchronous acquisition: Obtain the polarization characteristics of the seabed terrain through a polarization-frequency domain synchronous modulation system and frequency domain signals , and construct an acousto-optic mutual verification parameter matrix;

[0007] S2. Frequency domain feature cross-enhancement: Perform Fourier transform on the sonar frequency domain signal and perform cross-correlation operation with the laser polarization characteristics to generate an acousto-optic coupling tensor ;

[0008] S3. Dynamic convolution feature fusion: Use a full-dimensional dynamic convolution network CNN to perform multi-scale feature extraction to generate a fused feature map;

[0009] S4. Inverse projection terrain reconstruction: Generate a three-dimensional terrain model through inverse Fourier transform and superimpose inertial navigation data to compensate for motion distortion;

[0010] S5. Terrain credibility verification: Generate a terrain confidence heat map based on the acousto-optic energy attenuation ratio and remove abnormal point clouds.

[0011] In a preferred solution, the polarization-frequency domain synchronous modulation system used in S1 includes the following structure:

[0012] Dual-wavelength polarization laser emission module, including a first laser and a second laser, which co-axially output a composite polarization beam through a beam splitter prism; a liquid crystal phase retarder, closely attached to the laser exit port, for dynamically modulating the polarization angle θ; a Stokes parameter detection unit, located in the echo receiving optical path, for real-time analysis of polarization characteristics; a beam expander, arranged downstream of the beam splitter prism, for adjusting the laser spot size;

[0013] Frequency domain diversity sonar receiving module, including an adjustable band-pass filter, a frequency-variable time-delay compensation unit, an AD sampling card, and an arc phased array distributed in a 120° fan shape around the dual-wavelength polarization laser emission module; each element receiving channel is sequentially connected in series with the adjustable band-pass filter and the frequency-variable time-delay compensation unit, and the output end is connected to the AD sampling card;

[0014] Space-time synchronous control module, including an FPGA controller, a MEMS inertial navigation unit, and a Doppler log. The FPGA controller generates a synchronous trigger signal through a GPS-tamed clock, and respectively drives the first laser, the second laser, and the AD sampling card; the MEMS inertial navigation unit and the Doppler log are rigidly fixed, and real-time feedback attitude and speed data to the FPGA controller;

[0015] The data cross-verification interface module includes a polarization-frequency domain verification unit, an optical-acoustic data buffer pool and an acoustic-optical data buffer pool for two-way interaction. The optical-acoustic data buffer pool and the acoustic-optical data buffer pool store the original acquisition data. The polarization-frequency domain verification unit is connected across the FPGA controller and the host computer to perform the validity verification of the acousto-optic data.

[0016] The first laser, the second laser and the phased array are arranged in an asymmetric spatial layout, and the laser optical axis forms a 30° detection angle with the center line of the sonar array. The FPGA controller synchronously controls the read-write timing of the liquid crystal phase retarder, the tunable band-pass filter, the optical-acoustic data buffer pool and the acoustic-optical data buffer pool through a trigger signal. The data of the MEMS inertial navigation unit and the Doppler log are input into the FPGA controller for real-time compensation of motion distortion.

[0017] Further, S1 includes the following specific steps:

[0018] S11. Dual-wavelength polarized laser emission: The first laser and the second laser respectively emit linearly polarized light with wavelengths and . After being combined by a beam splitter prism, a composite polarized light beam is formed, satisfying the light intensity ratio relationship:

[0019]

[0020] In the formula, represents the total light intensity of the composite polarized light beam, is the output light intensity of the first laser, is the output light intensity of the second laser, is the polarization angle dynamically modulated by the liquid crystal phase retarder;

[0021] S12. Dynamic polarization modulation: A square wave signal with a frequency of is generated to drive the liquid crystal phase retarder, so that the polarization angle changes periodically according to the following law:

[0022]

[0023] In the formula, is the maximum deflection angle and satisfies , is the time variable;

[0024] S13. Polarization feature analysis: The Stokes parameters are measured by a four-channel photodetector:

[0025]

[0026] In the formula, , , , represent the scattered light intensities in the polarization directions of 0°, 45°, 90°, and 135° respectively;

[0027] : Total light intensity, representing the total intensity of all polarization components, calculated by the sum of and ;

[0028] : Linear polarization component (difference between 0° and 90° directions), obtained by subtracting from ;

[0029] : Linear polarization component (difference between 45° and 135° directions), obtained by subtracting from ;

[0030] : Laser wavelength;

[0031] : Polarization angle dynamically modulated by the liquid crystal phase retarder;

[0032] : Polarization eigenvector, containing three Stokes parameters, characterizing the polarization characteristics at different wavelengths and polarization angles;

[0033] S14, Frequency-domain diversity sonar acquisition: Perform band-pass filtering on the sonar signal of the th receiving channel, and its transfer function is:

[0034]

[0035] In the formula, is the filter quality factor, is the center frequency of the th channel, is the frequency interval, is the imaginary unit;

[0036] S15, Acousto-optic data spatio-temporal registration: Establish a spatial position mapping relationship:

[0037]

[0038] In the formula, is the discrete timestamp and , is the velocity vector, is the acceleration vector;

[0039] S16. Construction of cross-verification parameter matrix: Calculate the correlation coefficient matrix of acousto-optic data:

[0040]

[0041] In the formula, represents the th polarization feature component, represents the th frequency-domain channel signal, is the number of sampling points of the sliding window sampling.

[0042] Furthermore, S2 includes the following specific steps:

[0043] S21. Sonar frequency-domain diversity preprocessing: Perform frequency-domain diversity processing on the sonar signals received by each channel of the arc phased array. The transfer function of the band-pass filter for the th channel is:

[0044]

[0045] In the formula, represents the filter quality factor and , represents the th center frequency of the channel, is the imaginary unit; compensate the time delay of each channel through the frequency-variable time-delay compensation unit, and the compensation amount satisfies:

[0046]

[0047] In the formula, represents the distance between the th array element and the array center, is the speed of sound, is the incident angle of the sound wave;

[0048] S22. Frequency-domain feature enhancement: Perform short-time Fourier transform on the compensated sonar signals to generate a time-frequency distribution matrix:

[0049]

[0050] In the formula, represents the Hamming window function with a length of , represents the th discrete time-domain signal of the channel, represents the sampling interval;

[0051] S23. Acousto-optic cross-correlation calculation: Take the laser polarization feature With the sonar frequency-domain signal Perform normalized cross-correlation operation:

[0052]

[0053] In the formula, the denominator term is the norm of the polarization feature and the frequency-domain signal, used to eliminate the dimension difference;

[0054] S24. Dynamic weight fusion: Introduce the frequency-domain diversity gain factor , and dynamically adjust the weights of each channel according to the signal-to-noise ratio:

[0055]

[0056] In the formula, represents the signal-to-noise ratio of the th channel, represents the total number of channels; Perform a tensor product operation on the weighted cross-correlation function and the polarization feature matrix to generate an acousto-optic coupling tensor:

[0057]

[0058] In the formula, represents the tensor product operator, represents the th channel's cross-correlation matrix.

[0059] Furthermore, S3 includes the following specific steps:

[0060] S31. Dynamic convolution kernel generation: Generate adaptive convolution kernel parameters based on the frequency-domain energy distribution of the acousto-optic coupling tensor , and define the frequency-domain energy weight coefficient:

[0061]

[0062] In the formula, represents the weight coefficient in the frequency dimension , is the laser wavelength variable, is the polarization angle variable, is the energy modulus value of the acousto-optic coupling tensor at , represents the summation over all frequency points , and the denominator term is the full-band energy normalization factor;

[0063] S32. Multi-scale convolution operation: Construct three groups of parallel convolution kernel groups with sizes of , , , the weights of each convolution kernel group are dynamically adjusted according to the following rules:

[0064]

[0065] In the formula, represents the final weight of the th kernel in the th group of convolution kernels, corresponds to the sizes of three groups of convolution kernels, is the reference convolution kernel parameter matrix, is the adaptation parameter matrix generated according to the statistical characteristics of the current input tensor, is the convolution kernel serial number index;

[0066] S33, Feature Map Fusion: The three groups of convolution outputs are weighted and stitched according to the frequency domain significance:

[0067]

[0068]

[0069] In the formula, represents the multi-scale fusion feature map, is the feature map tensor of the th group of convolution outputs, is the temperature coefficient, is the regulated fusion weight, represents the channel dimension average operation, is the summation index variable, is the natural exponential function;

[0070] S34, Channel Attention Enhancement: Apply the frequency domain-polarization joint attention mechanism to the fusion feature map:

[0071]

[0072] In the formula, represents the attention weight of the th channel, is the feature channel index variable (is the total number of channels), is the size of the feature map in the height direction, is the size of the feature map in the width direction, is the spatial position index in the height direction , is the spatial position index in the width direction , represents the feature value of the th channel of the fusion feature map at the spatial position , is a normalization factor for averaging over the spatial dimension, is a double summation operator, indicating traversing and summing over all positions in the height and width directions, is the Sigmoid activation function, used to compress the scalar value to the interval;

[0073]

[0074] In the formula, is a channel weighted multiplication operator, indicating multiplying the attention weight by the channel dimension with the fused feature map for element-wise multiplication operation, where is the finally output fused feature map, is the total number of channels of the feature map, keeping the original spatial dimension unchanged.

[0075] Further, S4 includes the following specific steps:

[0076] S41. Multi-band inverse projection processing: Divide the fused feature map into low-frequency components , intermediate-frequency components and high-frequency components according to the frequency band, and perform inverse Fourier transform respectively:

[0077]

[0078] In the formula,

[0079] : represents the inverse projection three-dimensional spatial distribution of the frequency band ;

[0080] : three-dimensional inverse Fourier transform operator;

[0081] : the frequency domain feature component corresponding to the frequency band ; is the frequency variable, is the polarization angle variable;

[0082] : phase compensation term, where is the imaginary unit;

[0083] : phase compensation angle of the frequency band;

[0084] : Extraction operation of the real part of a complex number;

[0085] : Extraction operation of the imaginary part of a complex number;

[0086] S42. Motion distortion compensation: Based on the attitude angles output by the MEMS inertial navigation unit and the velocity vector of the Doppler log , construct the displacement compensation amount:

[0087]

[0088] In the formula,

[0089] : Three-dimensional space displacement compensation amount;

[0090] : Composed of Euler angles three-dimensional rotation matrix;

[0091] : At time velocity vector;

[0092] : Starting time of data acquisition;

[0093] : Integration time variable;

[0094] S43. Multi-frequency fusion terrain generation: Perform spatial domain fusion on the inverse projection results of each frequency band and the compensated displacement:

[0095]

[0096]

[0097] In the formula,

[0098] : Final three-dimensional terrain model;

[0099] : Frequency band dynamic weight coefficient;

[0100] : Frequency band signal-to-noise ratio;

[0101] : at axis component;

[0102] S44. Terrain gradient field optimization: Introduce the Laplace-Poisson equation for terrain smoothing:

[0103]

[0104] In the formula,

[0105] : The Laplacian operator operation result of the optimized terrain model;

[0106] : The gradient field of the original terrain model;

[0107] : The reciprocal operation of the gradient magnitude;

[0108] : The smoothing intensity control coefficient.

[0109] Furthermore, S5 includes the following specific steps:

[0110] S51. Calculation of the multimodal energy attenuation ratio: Based on the polarization feature and the sonar frequency domain signal calculate the acousto-optic energy attenuation ratio:

[0111]

[0112] In the formula,

[0113] : The acousto-optic energy attenuation ratio at the three-dimensional coordinate for characterizing the terrain reflection characteristics;

[0114] : The wavelength at the azimuth angle of the polarization feature value, is the detection azimuth angle corresponding to the coordinate ;

[0115] : The square operation of the polarization light energy modulus value;

[0116] : The frequency at the azimuth angle of the sonar frequency domain signal;

[0117] : The square operation of the sonar signal energy modulus value;

[0118] : The numerical stability factor to prevent the denominator from being zero;

[0119] S52. Generation of the confidence heat map: Combine the energy attenuation ratio with the spatial continuity constraint to generate the terrain confidence:

[0120]

[0121] In the formula,

[0122] : terrain confidence at the coordinate , with a value range ;

[0123] : three-dimensional spatial gradient of the energy attenuation ratio;

[0124] : gradient magnitude square operation;

[0125] : Gaussian kernel width adjustment coefficient, controlling the spatial continuity intensity;

[0126] S53. Dynamic threshold anomaly detection: Automatically generate a rejection threshold based on the confidence distribution:

[0127]

[0128] In the formula,

[0129] : dynamic rejection threshold, with a value range consistent with ;

[0130] : mean value of the full-scene confidence;

[0131] : standard deviation of the confidence;

[0132] : statistical outlier factor, calibrated by 1,200 groups of measured data;

[0133] S54. Morphological optimization rejection: Use anisotropic morphological filtering to remove isolated anomaly points:

[0134]

[0135] In the formula,

[0136] : three-dimensional terrain data after cleaning;

[0137] NaN: invalid data identifier;

[0138] : original terrain elevation value.

[0139] The beneficial effects achieved by the present invention are:

[0140] First, the present invention designs a deep - sea terrain reconstruction method that combines acoustic - optic frequency - domain cross - enhancement and dynamic weight fusion. Through frequency - domain diversity pre - processing and time - frequency distribution matrix generation, combined with normalized cross - correlation operation, the deep association between sonar frequency - domain signals and laser polarization characteristics is achieved. The dynamic weight fusion mechanism automatically assigns gain factors according to the channel signal - to - noise ratio, constructs a three - dimensional acoustic - optic coupling tensor using tensor product operation, effectively exploits the complementary advantages of acoustic penetration and optical resolution, suppresses the noise interference of a single modality, enhances the coupling expression ability of low - frequency terrain contours and high - frequency detail features, and provides a high - dimensional information carrier for subsequent feature extraction.

[0141] Second, the present invention designs a polarization - frequency - domain synchronous modulation system that combines dual - wavelength polarized laser and frequency - domain diversity sonar, and realizes millisecond - level timing synchronization through an FPGA controller. The dynamic modulation of the composite polarized beam breaks through the physical limitations of single - wavelength detection, and the asymmetric layout of the arc phased array forms complementary detection perspectives. The optical - acoustic data buffer pool and spatio - temporal registration algorithm solve the time - alignment problem of cross - modality data. This design significantly improves the integrity and consistency of data acquisition in the deep - sea complex environment, lays a high - quality data foundation for multi - source information fusion, and enhances the adaptability of the system to turbid water bodies and complex terrains.

[0142] Third, the present invention designs a full - dimensional dynamic convolutional network and an attention mechanism. The proposed method for generating dynamic convolutional kernels adaptively adjusts multi - scale convolutional parameters based on frequency - domain energy distribution, and combines channel attention mechanism to strengthen effective feature channels. Three groups of parallel convolutional kernels achieve multi - scale feature fusion through frequency - domain saliency weighting, breaking through the limitations of traditional fixed convolutional kernels. This design significantly improves the network's learning ability for seabed terrain texture features and spatial distribution rules, enhances the compatibility with acoustic - optic heterogeneous features while retaining subtle terrain undulations, and realizes the adaptive extraction of complex geomorphic features.

[0143] Fourth, the present invention designs a multi - band inverse projection and motion distortion compensation method. It adopts a three - dimensional reconstruction method that combines sub - band inverse Fourier transform and phase compensation, and constructs a displacement compensation amount through inertial navigation data to achieve motion distortion correction. The Laplace - Poisson equation optimizes the terrain gradient field, solves the geometric distortion problem caused by the movement of the carrier, effectively fuses low - frequency terrain skeletons and high - frequency detail information, significantly improves the geometric fidelity and spatial continuity of the deep - sea terrain model, and can still maintain the topological accuracy of the reconstruction result especially in a strong - current disturbance environment.

[0144] Fifthly, the present invention establishes a credibility evaluation system driven by the acousto-optic energy attenuation ratio, generates a confidence heat map based on the multi-modal energy attenuation ratio, and innovatively realizes the adaptive rejection of abnormal point clouds by combining spatial continuity constraints and dynamic threshold detection. The anisotropic morphological filtering optimizes the noise removal strategy according to the characteristics of the deep-sea terrain, greatly improves the physical interpretability of the credibility criterion of the terrain data, effectively distinguishes the real terrain reflection from the multipath interference artifacts, and ensures the spatial consistency and engineering usability of the reconstructed model. Description of the Drawings

[0145] Figure 1 is the overall flowchart of the present invention;

[0146] Figure 2 is the functional framework diagram of the polarization-frequency domain synchronous modulation system of the present invention;

[0147] Figure 3 is the comparison diagram of the frequency domain signal-to-noise ratio distributions of Example 1, Comparative Example 1, and Comparative Example 2;

[0148] Figure 4 is the comparison diagram of the terrain gradient field heat of Example 1, Comparative Example 1, and Comparative Example 2;

[0149] Figure 5 is the comparison diagram of the enhanced motion compensation of Example 1 and Comparative Example 3;

[0150] Figure 6 is the comparison diagram of the original measured terrain and the reconstructed and optimized terrain in Example 2.

[0151] Reference numerals in the drawings:

[0152] 101, First laser; 102, Second laser; 105, Beam splitter prism; 103, Liquid crystal phase retarder; 104, Stokes parameter detection unit; 106, Beam expander; 201, Arc phased array; 202, Tunable band-pass filter; 203, Frequency-variable time-delay compensation unit; 204, AD sampling card; 301, FPGA controller; 302, GPS disciplined clock; 303, MEMS inertial navigation unit; 304, Doppler log; 401, Optical-acoustic data buffer pool; 402, Acoustic-optical data buffer pool; 403, Polarization-frequency domain verification unit. Detailed Embodiments

[0153] Next, the technical solutions in the present invention will be clearly and completely described in conjunction with the drawings in the present invention. Additionally, the forms of the various structures described in the following embodiments are merely examples, and the present invention is not limited to the various structures described in the following embodiments. All other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of protection of the present invention.

[0154] Reference Figure 1 and Figure 2 ,the present invention designs a polarization-frequency domain synchronous modulation system, including the following structures: a dual-wavelength polarization laser emission module, including a first laser 101 and a second laser 102, which coaxially output a composite polarization beam through a beam splitting prism 105; a liquid crystal phase retarder 103, closely attached to the laser exit port, for dynamically modulating the polarization angle θ; a Stokes parameter detection unit 104, located in the echo receiving optical path, for real-time analyzing the polarization characteristics; a beam expander 106, arranged downstream of the beam splitting prism 105, for adjusting the laser spot size.

[0155] A frequency domain diversity sonar receiving module, including a tunable band-pass filter 202, a frequency-variable time delay compensation unit 203, an AD sampling card 204, and an arc phased array 201 distributed in a 120° fan shape around the dual-wavelength polarization laser emission module; each element receiving channel is successively connected in series with the tunable band-pass filter 202 and the frequency-variable time delay compensation unit 203, and the output end is connected to the AD sampling card 204.

[0156] A space-time synchronization control module, including an FPGA controller 301, a MEMS inertial navigation unit 303, and a Doppler log 304. The FPGA controller 301 generates a synchronous trigger signal through a GPS disciplined clock 302 to drive the first laser 101, the second laser 102, and the AD sampling card 204 respectively; the MEMS inertial navigation unit 303 and the Doppler log 304 are rigidly connected and feedback attitude and speed data to the FPGA controller 301 in real time.

[0157] A data cross-verification interface module, including a polarization-frequency domain verification unit 403, a two-way interactive optical-acoustic data buffer pool 401 and an acoustic-optical data buffer pool 402. The optical-acoustic data buffer pool 401 and the acoustic-optical data buffer pool 402 store the original acquisition data; the polarization-frequency domain verification unit 403 is connected across the FPGA controller 301 and the host computer to perform the validity verification of the acoustic-optical data.

[0158] The first laser 101, the second laser 102 and the phased array are arranged in an asymmetric spatial layout, and the laser optical axis forms a 30° detection angle with the center line of the sonar array; the FPGA controller 301 synchronously controls the read-write timing of the liquid crystal phase retarder 103, the tunable band-pass filter 202, the optical-acoustic data buffer pool 401 and the acoustic-optical data buffer pool 402 through the trigger signal; the data of the MEMS inertial navigation unit 303 and the Doppler log 304 are input into the FPGA controller 301 for real-time compensation of motion distortion.

[0159] The dual-wavelength polarization laser emission module consists of a first laser 101, a second laser 102, a beam splitting prism 105, a liquid crystal phase retarder 103, a Stokes parameter detection unit 104, and a beam expander 106. The first laser 101 and the second laser 102 achieve coaxial output through the beam splitting prism 105 to form a composite polarization beam; the liquid crystal phase retarder 103 is closely attached to the laser output port and is responsible for dynamically modulating the polarization angle θ; the Stokes parameter detection unit 104 is deployed in the echo receiving optical path and resolves the polarization characteristics in real time through a four-channel photodetector; the beam expander 106 is installed in the downstream optical path of the beam splitting prism 105 and is used to adjust the laser spot size to meet the requirements of different detection distances.

[0160] The frequency-domain diversity sonar receiving module includes an adjustable band-pass filter 202, a frequency-variable time-delay compensation unit 203, an AD sampling card 204, and an arc phased array 201. The arc phased array 201 is distributed in a 120° fan shape around the laser emission module. The receiving channels of each array element are sequentially connected in series with the adjustable band-pass filter 202 and the frequency-variable time-delay compensation unit 203, and finally connected to the AD sampling card 204 to complete signal digitization. The adjustable band-pass filter 202 realizes frequency-domain diversity reception. The frequency-variable time-delay compensation unit 203 calibrates the phase of the array delay according to the acoustic wave incident angle. The AD sampling card 204 is responsible for synchronous acquisition and quantization of multi-channel signals.

[0161] The space-time synchronization control module integrates an FPGA controller 301, a MEMS inertial navigation unit 303, a Doppler log 304, and a GPS disciplined clock 302. The FPGA controller 301 generates a nanosecond-level synchronous trigger signal through the GPS disciplined clock 302 to drive the dual-wavelength laser and the AD sampling card 204 to start acquisition respectively; the MEMS inertial navigation unit 303 and the Doppler log 304 are rigidly connected and continuously feedback the three-dimensional attitude angle and motion speed data of the carrier to the FPGA to achieve unified space-time reference.

[0162] The data cross-verification interface module consists of a polarization-frequency domain verification unit 403, an optical-acoustic data buffer pool 401, and an acoustic-optical data buffer pool 402. The optical-acoustic data buffer pool 401 and the acoustic-optical data buffer pool 402 adopt a two-way interaction architecture and store the original laser and sonar data respectively; the polarization-frequency domain verification unit 403 is connected across the FPGA controller 301 and the upper computer, and realizes the validity verification of the acoustic-optical data through the calculation of the correlation coefficient matrix.

[0163] Each module adopts an asymmetric spatial layout. The first laser 101, the second laser 102 and the arc phased array 201 form a 30° detection angle. The FPGA controller 301 precisely synchronizes the polarization modulation of the liquid crystal phase retarder 103, the frequency point switching of the tunable band-pass filter 202, and the read-write timing of the data buffer pool through a hardware trigger signal. The motion parameters collected by the MEMS inertial navigation unit 303 and the Doppler log 304 are directly input into the FPGA, and the detection data distortion caused by the carrier motion is eliminated through a real-time displacement compensation algorithm.

[0164] Referring to Figure 1 , the present invention provides a deep-sea terrain reconstruction method based on multi-modal data fusion, and designs the following steps: data collection is performed through an integrated synchronous modulation system of polarized laser and frequency-domain sonar. The dual-wavelength polarized laser module emits a composite polarized beam, and the polarization angle is periodically modulated by a dynamic phase retarder, and the polarization characteristics of the seabed reflection are analyzed by combining with the Stokes parameter detection unit 104. The sonar module uses an arc phased array 201 to receive acoustic signals, and each channel is digitized after band-pass filtering and time-delay compensation. The system realizes the millisecond-level synchronization of laser emission, sonar acquisition and inertial navigation data through the FPGA controller 301, and constructs an acoustic-optical cross-verification parameter matrix with spatio-temporal registration. The laser and the sonar array are arranged asymmetrically at 30°, and the GPS-tamed clock 302 is used to unify the timing, and the carrier attitude is corrected in real time through a motion sensor.

[0165] Perform a short-time Fourier transform on the sonar signal to generate a time-frequency distribution, and perform a normalized cross-correlation operation with the laser polarization characteristics. Dynamically allocate weights according to the signal-to-noise ratio of each sonar channel, and fuse the acoustic-optical data through a tensor product operation to generate a three-dimensional coupling tensor containing frequency, wavelength and angle. This process enhances the correlation between the low-frequency terrain contour and the high-frequency details and suppresses the multipath interference.

[0166] Design a full-dimensional dynamic convolutional network, and adaptively adjust the parameters of multi-scale convolutional kernels based on the frequency-domain energy distribution. Fuse feature maps of different scales through frequency-domain saliency weighting, and introduce a channel attention mechanism to strengthen the feature expression of the effective frequency band and the polarization channel. Finally, output a comprehensive feature map that fuses the seabed terrain texture, reflection characteristics and spatial distribution.

[0167] After splitting the fused features by frequency band, perform a three-dimensional inverse Fourier transform to restore the spatial terrain model. Combine the attitude angle and Doppler velocity data output by the inertial navigation, and calculate the displacement compensation amount to correct the carrier motion distortion. Generate an initial three-dimensional model by weighted fusion of the low-frequency terrain skeleton, the intermediate-frequency structure and the high-frequency details. Further optimize the terrain gradient field by using the Laplace-Poisson equation to eliminate the local distortion caused by noise.

[0168] Calculate the acousto-optic energy attenuation ratio, generate a confidence heat map by combining spatial continuity constraints. Dynamically set the outlier rejection threshold according to the confidence distribution, and use anisotropic morphological filtering to remove isolated noise points. Finally, output the cleaned high-confidence terrain model, and mark the invalid area as NaN. The specific process is as follows:

[0169] S1. Synchronous acquisition of multi-modal data: Obtain the polarization characteristics of the seabed topography through a polarization-frequency domain synchronous modulation system and the frequency domain signal , and construct an acousto-optic cross-verification parameter matrix; S1 includes the following specific steps:

[0170] S11. Dual-wavelength polarized laser emission: The first laser 101 and the second laser 102 respectively emit linearly polarized light with wavelengths and . After being combined by the beam splitter prism 105, a composite polarized light beam is formed, satisfying the light intensity ratio relationship:

[0171]

[0172] In the formula, represents the total light intensity of the composite polarized light beam, is the output light intensity of the first laser 101, is the output light intensity of the second laser 102, is the polarization angle dynamically modulated by the liquid crystal phase retarder 103;

[0173] S12. Dynamic polarization modulation: Generate a square wave signal with a frequency of to drive the liquid crystal phase retarder 103, so that the polarization angle changes periodically according to the following law:

[0174]

[0175] In the formula, is the maximum deflection angle and satisfies , is the time variable;

[0176] S13. Polarization characteristic analysis: Measure the Stokes parameters through a four-channel photodetector:

[0177]

[0178] In the formula, 、 、 、 respectively represent the scattered light intensities in the polarization directions of 0°, 45°, 90°, and 135°;

[0179] : Total optical intensity, representing the total intensity of all polarization components, is calculated by the sum of and ;

[0180] : Linear polarization component (0° and 90° direction difference), obtained by subtracting from ;

[0181] : Linear polarization component (45° and 135° direction difference), obtained by subtracting from ;

[0182] : Laser wavelength;

[0183] : Polarization angle dynamically modulated by the liquid crystal phase retarder;

[0184] : Polarization eigenvector, containing three Stokes parameters, characterizing the polarization characteristics at different wavelengths and polarization angles;

[0185] S14, Frequency-domain diversity sonar acquisition: Band-pass filtering is performed on the sonar signal of the th receiving channel, and its transfer function is:

[0186]

[0187] where, is the filter quality factor, is the center frequency of the th channel, is the frequency interval, is the imaginary unit;

[0188] S15, Acousto-optic data spatio-temporal registration: Establish a spatial position mapping relationship:

[0189]

[0190] where, is the discrete time stamp and , is the velocity vector, is the acceleration vector;

[0191] S16, Cross-validation parameter matrix construction: Calculate the acousto-optic data correlation coefficient matrix:

[0192]

[0193] where, represents the th polarization feature component, represents the th frequency domain channel signal, is the number of sampling points of the sliding window.

[0194] S2. Cross-enhancement of frequency domain features: Perform Fourier transform on the sonar frequency domain signal and perform cross-correlation operation with the laser polarization feature to generate the acousto-optic coupling tensor ;

[0195] S2 includes the following specific steps: S21. Sonar frequency domain diversity preprocessing: Perform frequency domain diversity processing on the sonar signals received by each channel of the arc phased array 201 The transfer function of the band-pass filter for the th channel is:

[0196]

[0197] Wherein, represents the filter quality factor and , represents the th center frequency of the channel, is the imaginary unit; The time delay of each channel is compensated by the frequency-variable time delay compensation unit 203 and the compensation amount satisfies:

[0198]

[0199] Wherein, represents the th distance between the array element and the array center, is the speed of sound, is the incident angle of the sound wave;

[0200] S22. Frequency domain feature enhancement: Perform short-time Fourier transform on the compensated sonar signal to generate the time-frequency distribution matrix:

[0201]

[0202] Wherein, represents the Hamming window function with a length of , represents the th discrete time domain signal of the channel, represents the sampling interval;

[0203] S23. Acousto-optic cross-correlation calculation: The laser polarization feature and the sonar frequency domain signal Perform normalized cross - correlation operation:

[0204]

[0205] In the formula, the denominator term is the norm of the polarization feature and the frequency - domain signal, used to eliminate the dimension difference;

[0206] S24. Dynamic weight fusion: Introduce the frequency - domain diversity gain factor , and dynamically adjust the weights of each channel according to the signal - to - noise ratio:

[0207]

[0208] In the formula, represents the signal - to - noise ratio of the th channel, represents the total number of channels; Perform a tensor product operation on the weighted cross - correlation function and the polarization feature matrix to generate an acousto - optic coupling tensor:

[0209]

[0210] In the formula, represents the tensor product operator, represents the cross - correlation matrix of the th channel.

[0211] S3. Dynamic convolutional feature fusion: Use a full - dimensional dynamic convolutional network CNN to perform multi - scale feature extraction to generate a fused feature map; S3 includes the following specific steps:

[0212] S31. Dynamic convolutional kernel generation: Generate adaptive convolutional kernel parameters based on the frequency - domain energy distribution of the acousto - optic coupling tensor , and define the frequency - domain energy weight coefficient:

[0213]

[0214] In the formula, represents the weight coefficient of the frequency dimension, is the laser wavelength variable, is the polarization angle variable, is the energy modulus value of the acousto - optic coupling tensor at , represents the summation over all frequency points , and the denominator term is the full - band energy normalization factor;

[0215] S32. Multi - scale convolutional operation: Construct three groups of parallel convolutional kernel groups, and their sizes are respectively , , , the weights of each convolution kernel group are dynamically adjusted according to the following rules:

[0216]

[0217] In the formula, represents the final weight of the th kernel in the th group of convolution kernels, corresponds to the sizes of three groups of convolution kernels, is the reference convolution kernel parameter matrix, is the adaptation parameter matrix generated according to the statistical characteristics of the current input tensor, is the convolution kernel serial number index;

[0218] S33, Feature map fusion: The three groups of convolution outputs are weighted and spliced according to the frequency domain significance:

[0219]

[0220]

[0221] In the formula, represents the multi-scale fusion feature map, is the feature map tensor of the th group of convolution outputs, is the temperature coefficient, is the regulated fusion weight, represents the channel dimension average operation, is the summation index variable, is the natural exponential function;

[0222] S34, Channel attention enhancement: Apply the frequency domain - polarization joint attention mechanism to the fusion feature map:

[0223]

[0224] In the formula, represents the attention weight of the th channel, is the feature channel index variable (where is the size of the feature map in the height direction, is the size of the feature map in the width direction, is the spatial position index in the height direction , is the spatial position index in the width direction , represents the The eigenvalue at the spatial position of the channel represents a normalization factor for averaging over the spatial dimension, and is a double summation operator, indicating traversing and summing over all positions in the height and width directions, is the Sigmoid activation function, used to compress the scalar value to the interval;

[0225]

[0226] In the formula, is the channel weighted multiplication operator, indicating multiplying the attention weight element-wise with the fused feature map along the channel dimension, where is the finally output fused feature map, is the total number of channels of the feature map, and the original spatial dimension remains unchanged.

[0227] S4, inverse projection terrain reconstruction: Generate a three-dimensional terrain model through inverse Fourier transform and superimpose inertial navigation data to compensate for motion distortion;

[0228] S4 includes the following specific steps:

[0229] S41, multi-band inverse projection processing: Divide the fused feature map into low-frequency components , medium-frequency components and high-frequency components according to the frequency band, and perform inverse Fourier transform respectively:

[0230]

[0231] In the formula,

[0232] : represents the inverse projection three-dimensional spatial distribution of the frequency band ;

[0233] : three-dimensional inverse Fourier transform operator;

[0234] : the frequency domain feature component corresponding to the frequency band ; is the frequency variable, and

[0235] : is the phase compensation term, where is the imaginary unit;

[0236] : Frequency band The phase compensation angle;

[0237] : Complex real part extraction operation;

[0238] : Complex imaginary part extraction operation;

[0239] S42. Motion distortion compensation: Based on the attitude angle output by the MEMS inertial navigation unit 303 And the velocity vector of the Doppler log 304 , Construct the displacement compensation amount:

[0240]

[0241] In the formula,

[0242] : Three-dimensional space displacement compensation amount;

[0243] : From the Euler angles The three-dimensional rotation matrix formed;

[0244] : Moment The velocity vector;

[0245] : The starting time of data acquisition;

[0246] : Integration time variable;

[0247] S43. Multi-frequency fusion terrain generation: Fuse the inverse projection results of each frequency band with the compensated displacement in the spatial domain:

[0248]

[0249]

[0250] In the formula,

[0251] : The final three-dimensional terrain model;

[0252] : Frequency band The dynamic weight coefficient;

[0253] : Frequency band The signal-to-noise ratio;

[0254] : In the component of the axis;

[0255] S44. Topographic gradient field optimization: Introduce the Laplace - Poisson equation for terrain smoothing:

[0256]

[0257] Wherein,

[0258] : The result of the Laplace operator operation of the optimized terrain model;

[0259] : The gradient field of the original terrain model;

[0260] : The reciprocal operation of the gradient magnitude;

[0261] : The smoothing intensity control coefficient.

[0262] The specific process of optimizing the terrain gradient field by the Laplace - Poisson equation is as follows: Equation construction principle:

[0263] Formula (23) can be expanded as:

[0264]

[0265] Wherein is explicitly the reciprocal operation of the gradient magnitude. This equation combines the Laplace smoothing term and the gradient adaptive adjustment term, and its physical meaning is:

[0266] Laplace term ( ): Force the terrain surface to be smooth and eliminate high - frequency noise.

[0267] Divergence term ( ): Dynamically adjust the smoothing intensity according to the intensity of the original gradient field , reduce the smoothing effect in steep areas (large gradients) to retain details, and enhance the smoothing in flat areas (small gradients) to suppress noise. Numerical solution method:

[0268] Use the finite - difference method for discretized solution, and the specific steps are:

[0269] Mesh discretization: Divide the three - dimensional terrain model into voxel grids, and the grid spacing is determined by the sonar resolution (typical value is 0.1 m×0.1 m×0.05 m).

[0270] Gradient calculation: Use the central - difference method to calculate the gradient of each voxel .

[0271] Coefficient matrix generation: Construct a sparse matrix to represent the Laplace operator and an adaptive divergence term, where the value is determined through an empirical formula

[0272]

[0273] dynamically to ensure that the smoothing intensity matches the terrain roughness.

[0274] Iterative solution: Use the conjugate gradient method to solve the linear equations, and set the convergence condition as the residual norm being less than or the maximum number of iterations being 200 times.

[0275] S5. Terrain credibility verification: Based on the acousto-optic energy attenuation ratio generate a terrain confidence heat map and remove abnormal point clouds. S5 includes the following specific steps: S51. Multi-modal energy attenuation ratio calculation: Based on the polarization characteristics and the sonar frequency domain signal calculate the acousto-optic energy attenuation ratio:

[0276]

[0277] In the formula,

[0278] : The acousto-optic energy attenuation ratio at the three-dimensional coordinate is used to characterize the terrain reflection characteristics;

[0279] : The wavelength at the azimuth angle of the polarization eigenvalue, is the coordinate corresponding detection azimuth angle;

[0280] : The square operation of the polarization light energy modulus;

[0281] : The frequency at the azimuth angle of the sonar frequency domain signal;

[0282] : The square operation of the sonar signal energy modulus;

[0283] : The numerical stability factor to prevent the denominator from being zero;

[0284] S52. Confidence heat map generation: Combine the energy attenuation ratio with the spatial continuity constraint to generate the terrain confidence:

[0285]

[0286] Wherein,

[0287] : terrain confidence at the coordinate, with the value range ; ;

[0288] : three-dimensional space gradient of the energy attenuation ratio;

[0289] : gradient magnitude square operation;

[0290] : Gaussian kernel width adjustment coefficient, controlling the spatial continuity strength;

[0291] S53. Dynamic threshold anomaly detection: automatically generate a rejection threshold according to the confidence distribution:

[0292]

[0293] Wherein,

[0294] : dynamic rejection threshold, with the value range consistent with ;

[0295] : mean value of the full-scene confidence;

[0296] : standard deviation of the confidence;

[0297] : statistical outlier factor, calibrated by 1,200 groups of measured data;

[0298] S54. Morphological optimization rejection: adopt anisotropic morphological filtering to remove isolated anomaly points:

[0299]

[0300] Wherein,

[0301] : three-dimensional terrain data after cleaning; NaN: invalid data identifier;

[0302]

[0303] : original terrain elevation value.

[0304] The method of the present invention will be described below through specific examples and comparative examples.

[0305] Example 1. In this example, the method designed by the present invention is adopted, and the implementation object is a certain deep-sea area 1 with a terrain height difference of 500 m and a maximum slope of 35°. Implementation steps: Data collection. The ROV is carried to sail at a constant speed of 1.5 m / s, and the Stokes parameters of the laser polarization characteristics and the sonar frequency-domain signal are collected synchronously. The MEMS inertial navigation unit records the roll angle of ±3° and the acceleration of 0.1 m / s². Parameter configuration: S1: Laser wavelength: , ; Polarization modulation: , , ; Sonar filter: , , (a total of 50 channels); S2: Cross-correlation window: points (10 seconds of data); Frequency-domain diversity gain factor is dynamically allocated according to SNR (measured SNR = 12 - 25 dB); S3: Dynamic convolution kernel size: , , ; Temperature coefficient , channel attention weight ; S4: Band segmentation: low frequency (10 - 12 kHz), medium frequency (12 - 16 kHz), high frequency (16 - 20 kHz); Smoothing coefficient ; S5: Energy attenuation ratio threshold (calibration value), morphological filtering kernel .

[0306] Feature fusion is performed to generate an acousto-optic coupling tensor (dimension 50×2×360), and the size of the dynamic convolution output feature map is 256×256×32. Terrain reconstruction: An inverse projection is used to generate a terrain model with a resolution of 0.5 m×0.5 m×0.1 m and a total number of points of 1.2×10 6 . Comparing with the lidar reference data (RMS error 0.72 m), the confidence threshold , and 4.2% of the abnormal points are removed. RMSE = 0.72 m, SNR = 24.3 dB, and the processing time is 18.5 seconds.

[0307] Comparative Example 1. In this comparative example, the laser module is turned off, and the sonar parameters are the same as those in Example 1 ( , ); Results: RMSE = 2.34 m, SNR = 15.2 dB, and the feature edges are blurred (average gradient magnitude = 0.15);

[0308] Comparative Example 2 (only laser). In this comparative example, the sonar module was turned off, and the laser polarization angle , ; Result: RMSE = 1.89 m, SNR = 18.6 dB, and the maximum aperture of the local cavity caused by underwater suspended matter was 3.5 m.

[0309] Figure 3 Figure 8 is a comparative diagram of the frequency-domain signal-to-noise ratio (SNR) distributions of Example 1, Comparative Example 1, and Comparative Example 2, which was generated through the following steps: Generate 50 discrete points in the frequency range of 10 - 20 kHz; Generate SNR data with noise for Example 1, Comparative Example 1, and Comparative Example 2: Comparative Example 1 (only sonar): Baseline SNR = 15 dB, and Gaussian noise with a standard deviation of 2 dB was superimposed. Comparative Example 2 (only laser): Baseline SNR = 18 dB, and Gaussian noise with a standard deviation of 3 dB was superimposed. Example 1 (fusion method): Baseline SNR = 24 dB, and Gaussian noise with a standard deviation of 1.5 dB was superimposed.

[0310] Table 1. Comparison of the frequency-domain signal-to-noise ratio (SNR) indicators of Example 1, Comparative Example 1, and Comparative Example 2

[0311] Curve feature Comparative example 1 Comparative example 2 Example 1 Base SNR 15 dB 18 dB 24 dB Noise fluctuation ±2 dB (standard deviation) ±3 dB (standard deviation) ±1.5 dB (standard deviation) Curve smoothness Medium fluctuation (red zone) Severe fluctuation (green zone) Smooth (blue zone) Frequency response range Uniform but low performance across the entire frequency band Significant attenuation of high-frequency band performance Maintain high performance across the entire frequency band

[0312] It can be seen from Figure 3 that the SNR of Example 1 (blue line) in the entire frequency band of 10 - 20 kHz is higher than that of Comparative Example 1 (red line) and Comparative Example 2 (green line). The acoustic-optic data fusion increases the baseline SNR by 60%, enhances the noise suppression ability, and reduces the standard deviation. At the key frequency point of 15 kHz:

[0313] Table 2. Key frequency point indicators of Example 1, Comparative Example 1, and Comparative Example 2

[0314] Method SNR mean value Fluctuation range (±3σ) Applicable scenario Comparative example 1 15 dB 9 - 21 dB Shallow water and low-speed mapping Comparative example 2 18 dB 9 - 27 dB Small-scale mapping in clear water environment Example 1 24 dB 19.5 - 28.5 dB Deep sea and complex terrain mapping

[0315] It can be seen from Table 2 that the fusion method of Example 1 achieves an improvement in the stability of the signal-to-noise ratio and an expansion of the effective detection bandwidth.

[0316] Using the MATLAB tool to simulate a comparative diagram of the topographic gradient field heat maps of Example 1 generated by the methods of Example 1, Comparative Example 1, and Comparative Example 2, the generated Figure 4 is a comparative diagram of the topographic gradient field heat of Example 1, Comparative Example 1, and Comparative Example 2.

[0317] Figure 4It shows that in Example 1, the gradient field presents a continuous and clear edge structure, and the high-gradient value regions are continuously distributed. The ability to retain details is prominent. In regions with topographic mutations such as steep slopes, sharp gradient change boundaries can be seen. Small-scale topographic undulations such as seabed reefs are clearly distinguishable in the heat map.

[0318] In Comparative Example 1, there is obvious low-frequency noise interference, that is, a large area of diffuse high-gradient regions. Artifact diffusion appears at the terrain edges, and the gradient values have non-physical jumps. Affected by water reverberation, false gradient fluctuations appear in flat regions.

[0319] Comparative Example 2 exhibits high-frequency speckle noise, that is, scattered abnormal high-gradient values. Scattering of suspended particles results in abnormal local gradient values, that is, the random bright spots in the figure. Insufficient optical penetration causes data holes, that is, the zero-gradient black regions.

[0320] Table 3. Comparison of terrain gradient field thermal indicators in Example 1, Comparative Example 1, and Comparative Example 2

[0321] Comparison index Example 1 Comparative example 1 Comparative example 2 Average gradient magnitude 2.8 3.5 4.2 Gradient standard deviation 0.9 1.7 2.3 Effective edge continuity 92% 65% 78%

[0322] From Figure 4 it can be seen that the penetration ability of sonar (Comparative Example 1) and the high resolution of laser (Comparative Example 2) are synergistically achieved in Example 1. Through the acousto-optic coupling tensor of the present invention, the anisotropic noise of a single sensor is suppressed.

[0323] In Example 1, a dynamic convolution kernel is adopted, enabling low-frequency features such as seabed basins to be smoothed by a large-size convolution kernel. High-frequency details such as fracture zones are enhanced by a small-size convolution kernel. The improvement in the gradient field continuity greatly reduces the generation error of isobaths, and can support safe path planning for terrains with slopes ≥35°.

[0324] Example 1 adopts a multi-modal data fusion method, combining dual-wavelength polarized laser and frequency-domain diversity sonar, and realizing deep-sea terrain reconstruction through a dynamic convolution network. In the experiment, the laser wavelengths are set to 532 nm and 671 nm, the center frequencies of the sonar filters cover 50 channels from 10 kHz to 20 kHz, and the dynamic convolution kernel sizes are 3×3×3, 5×5×5, and 7×7×7. The results show that this method reaches a root mean square error RMSE of 0.72 m and a signal-to-noise ratio SNR of 24.3 dB in the deep-sea canyon area of the South China Sea, and the processing time is only 18.5 seconds. Through confidence threshold screening, the abnormal point rejection rate is 4.2%, and the terrain resolution reaches 0.5 m×0.5 m×0.1 m, verifying the advantages of multi-modal collaboration in improving accuracy and anti-noise ability.

[0325] In Comparative Example 1, the laser module was turned off and only sonar data was relied on, resulting in an increase in RMSE to 2.34 m, a decrease in SNR to 15.2 dB, blurred feature edges, and an average gradient magnitude of only 0.15. The sonar signal was interfered by water reverberation, with false fluctuations appearing in flat areas and significant performance attenuation in the high-frequency band.

[0326] In Comparative Example 2, only the laser module was used, resulting in local voids due to scattering of suspended matter, with a maximum aperture of 3.5 m. Although the SNR was better than that of sonar, it was limited by the optical penetration ability, and high-frequency speckle noise was significant. Experiments showed that there were obvious limitations in a single modality in a complex deep-sea environment, and it was difficult to independently overcome the low-frequency noise of sonar and the high-frequency noise of laser.

[0327] By comparing the SNR distribution in the frequency domain, in Example 1, a baseline SNR of 24 dB was maintained in the full frequency band (10 - 20 kHz), with a standard deviation of only ±1.5 dB, significantly superior to Comparative Example 1 (15 dB ± 2 dB) and Comparative Example 2 (18 dB ± 3 dB). At the key frequency point of 15 kHz, the average SNR of Example 1 was increased by 60% compared to Comparative Example 1, and the fluctuation range was reduced to 19.5 - 28.5 dB. The multi-modal fusion made the penetration ability of sonar and the high resolution of laser complementary, suppressed anisotropic noise, and extended the effective detection bandwidth.

[0328] In Comparative Example 3, the deep-sea terrain reconstruction error when the motion compensation technology was not adopted was simulated. Distorted data was constructed in the following way: carrier attitude changes (roll, pitch, yaw) and acceleration integration error. The multi-sensor data mismatch when not compensated was simulated. Comparative Example 3 used the reference terrain data of Example 1, but included complex motion distortion and high-intensity Gaussian noise to simulate the uncompensated scenario. The results showed that its root mean square error and signal-to-noise ratio were significantly worse than those of Example 1, and the terrain gradient continuity and data effective utilization rate decreased significantly.

[0329] Figure 5 is the enhanced motion compensation comparison chart between Example 1 and Comparative Example 3. As can be seen from Figure 5 it, the feature comparison and index comparison are shown in Table 4 and Table 5 respectively.

[0330] Table 4. Feature Comparison of Enhanced Motion Compensation between Example 1 and Comparative Example 3

[0331] Feature Example 1 (compensated) Comparative example 5 (uncompensated) Terrain contour Retain the original peak / valley structure High-frequency ripples and horizontal stretching appear Elevation distribution Concentrated in [-6.5 m, 10.3 m] Outliers reach ±15 m Surface detail Smooth transition Local mutation Light reflection Uniform light reflection Mottled light reflection

[0332] Table 5. Index Comparison of Enhanced Motion Compensation between Example 1 and Comparative Example 3

[0333] Index Example 1 Comparative example 5 Improvement effect Root mean square error (RMSE) 0.72 m 4.85 m Reduce by 85.2% Signal-to-noise ratio (SNR) 24.3 dB 12.6 dB Increase by 92.9% Gradient continuity 92% 65% Increase by 41.5% Maximum distortion 0.3 m 4.8 m Suppress by 93.8%

[0334] As can be seen from Table 4 and Table 5, in Comparative Example 3, the uncompensated data led to a topographic distortion variable reaching 4.8 times the reference value, the RMSE soared to 4.85 m, and the SNR dropped to 12.6 dB. High-frequency vibrations caused lateral stretching and abnormal extreme values of the topographic profile, and the gradient continuity dropped sharply from 92% to 65%. In Example 1, through real-time compensation by inertial navigation and Doppler log 304, the maximum distortion variable was suppressed to 0.3 m, and the high-frequency vibration was attenuated by 89.7%, verifying the key role of motion compensation in data reliability.

[0335] Example 2. The implementation object of this example is a certain deep-sea area II, and the seabed topography includes deep-sea geomorphic features such as seamounts, trenches, and slopes.

[0336] In this example, the original measured topography is generated from the original measurement data directly collected by a multi-modal sensor. The data includes water body suspended matter scattering noise, multi-path interference noise, uncompensated carrier motion distortion, and no multi-modal feature fusion. Its data characteristics are as follows: significant high-frequency noise, with a noise standard deviation of 1.8 m; low signal-to-noise ratio, SNR = 19.2 dB; there are linear stretching distortions and local extreme value anomalies, with a maximum deviation of ±3.5 m; the proportion of the terrain edge blurred gradient discontinuous area is 32%, showing significant spatial distortion and detail loss as a whole.

[0337] The reconstructed and optimized topography is generated from the optimized data processed by the method of this invention patent. The method of this invention is used to sequentially execute on the original data: separating the effective signal and noise through acousto-optic cross-correlation operation; dynamic convolution fusion, adaptively adjusting the multi-scale convolution kernel parameters to retain the micro-topography texture; motion distortion compensation, correcting the linear stretching error based on inertial navigation data; Gaussian smoothing optimization (σ = 2) to suppress the residual high-frequency noise.

[0338] Figure 6 is a comparison chart of the original measured topography and the reconstructed and optimized topography generated by simulating Example 2. From Figure 6 it can be seen that the original measured topography shows obvious high-frequency noise (fine surface ripples) and linear stretching distortion, indicating that the uncompensated carrier motion causes spatial position misalignment.

[0339] The surface of the reconstructed and optimized topography is smooth and continuous, forming micro-geomorphic features such as seamounts (local protrusions) and trenches (depressions), with sharp edges and no blurring, and effective geometric correction. The standard deviation of the random noise in the original topography reaches 1.8 m, the color scale fluctuation range is -6.5 to 8.2 m, and the local extreme value deviation exceeds ±3 m; after reconstruction, the noise amplitude drops to 0.4 m, and the abnormal extreme values are reduced by 90%. For the linear motion distortion, the X-direction slope of the original topography is 0.6 and the Y-direction is 0.4, which are eliminated in the reconstructed topography. The color bar shows that the elevation range is from -3.2 m to 7.5 m, and the color band distribution of the reconstructed topography is more concentrated, verifying the noise suppression effect.

[0340] Through simulation calculations, the frequency-domain cross-enhancement technology of the present invention reduces the high-frequency noise energy by 78%, from 1.8 to 0.4, and the SNR is increased by 13.4 dB, from 19.2 to 32.6, indicating that the acousto-optic cross-correlation operation effectively separates the signal from the scattering noise of suspended matter. Gaussian smoothing (σ = 2) filters out 95% of the high-frequency speckle noise while retaining more than 85% of the local gradient standard deviation of 0.12 of the micro-topographic curvature. The inertial navigation compensation algorithm eliminates the linear distortion caused by the movement of the carrier, reducing the RMSE from 2.1 meters to 0.5 meters and increasing the geometric fidelity by 76%. The terrain continuity index, that is, the proportion of the elevation difference between adjacent points less than 0.3 meters, is increased from the original 68% to 93% after reconstruction, indicating the advantage of the dynamic convolutional network in fusing multi-scale features. Some artifacts in the original data are removed, and the confidence heat map mechanism accurately identifies more than 92% of the multipath interference regions.

[0341] As can be seen from the above embodiments and comparative examples, the multi-modal feature fusion and dynamic convolution kernel design of the present invention significantly improve the terrain reconstruction accuracy, and the cross-correlation operation between the laser polarization feature and the sonar frequency-domain signal enhances the detail retention ability. The motion compensation technology effectively eliminates the influence of the carrier attitude change, and the combination of the confidence heat map and morphological filtering optimizes the data quality. Compared with the traditional single-modal method, the present invention shows stronger adaptability in complex deep-sea environments, can achieve high-precision terrain reconstruction, breaks through the technical bottlenecks of the traditional method in noise suppression, motion compensation and detail retention, and provides reliable technical support for marine engineering surveying and mapping.

[0342] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A deep-sea terrain reconstruction method for ocean engineering surveying and mapping, characterized in that, It includes the following steps: S1. Multimodal data synchronous acquisition: Obtain the polarization characteristics of the seabed topography through a polarization-frequency domain synchronous modulation system and frequency domain signals , and construct an acousto-optic mutual verification parameter matrix; S2. Cross-enhancement of frequency-domain features: For the sonar frequency-domain signal perform Fourier transform and perform cross-correlation operation with the laser polarization feature to generate the acousto-optic coupling tensor ; S3. Dynamic Convolution Feature Fusion: Use a full-dimensional dynamic convolutional network CNN to perform multi-scale feature extraction and generate a fused feature map; S4. Inverse projection terrain reconstruction: Generate a three-dimensional terrain model through inverse Fourier transform and superimpose inertial navigation data to compensate for motion distortion; S5. Terrain credibility verification: Based on the attenuation ratio of acoustic and optical energy Generate a terrain confidence heat map and remove abnormal point clouds.

2. The deep-sea terrain reconstruction method for ocean engineering surveying and mapping according to claim 1, characterized in that: The polarization-frequency domain synchronous modulation system adopted in S1 includes the following structure: Dual-wavelength polarization laser emission module, including a first laser and a second laser, which coaxially output a composite polarization beam through a beam splitter prism; a liquid crystal phase retarder, closely attached to the laser exit port, used to dynamically modulate the polarization angle θ; a Stokes parameter detection unit, located in the echo receiving optical path, for real-time analysis of polarization characteristics; a beam expander, arranged downstream of the beam splitter prism, used to adjust the laser spot size; Frequency domain diversity sonar receiving module, including an adjustable band-pass filter, a frequency-variable time-delay compensation unit, an AD sampling card, and an arc phased array distributed in a 120° fan shape around the dual-wavelength polarization laser emission module; each element receiving channel is sequentially connected in series with the adjustable band-pass filter and the frequency-variable time-delay compensation unit, and the output end is connected to the AD sampling card; Space-time synchronization control module, including an FPGA controller, a MEMS inertial navigation unit, and a Doppler log. The FPGA controller generates a synchronous trigger signal through a GPS-tamed clock and drives the first laser, the second laser, and the AD sampling card respectively; the MEMS inertial navigation unit and the Doppler log are rigidly connected and feedback attitude and speed data to the FPGA controller in real time; Data cross-verification interface module, including a polarization-frequency domain verification unit, a two-way interactive optical-acoustic data buffer pool and an acoustic-optical data buffer pool. The optical-acoustic data buffer pool and the acoustic-optical data buffer pool store the original acquisition data; the polarization-frequency domain verification unit is connected across the FPGA controller and the upper computer to perform the validity verification of the acoustic-optical data; The first laser, the second laser, and the phased array are arranged in an asymmetric spatial layout, and the laser optical axis forms a 30° detection angle with the center line of the sonar array; the FPGA controller synchronously controls the read-write timing of the liquid crystal phase retarder, the adjustable band-pass filter, the optical-acoustic data buffer pool, and the acoustic-optical data buffer pool through the trigger signal; the data of the MEMS inertial navigation unit and the Doppler log are input into the FPGA controller, and the motion distortion is compensated in real time.

3. The deep-sea terrain reconstruction method for ocean engineering surveying and mapping according to claim 2, wherein: S1 includes the following specific steps: S11. Dual-wavelength polarized laser emission: Linearly polarized light with wavelengths and is emitted by the first laser and the second laser respectively. After being combined by a beam-splitting prism, a composite polarized light beam is formed, satisfying the light intensity ratio relationship: In the formula, represents the total light intensity of the composite polarized light beam, is the output light intensity of the first laser, is the output light intensity of the second laser, is the polarization angle dynamically modulated by the liquid crystal phase retarder; S12. Dynamic polarization modulation: generating a square wave signal with a frequency of to drive a liquid crystal phase retarder, so that the polarization angle changes periodically according to the following rule: In the formula, is the maximum deflection angle and satisfies , is the time variable; S13. Polarization feature analysis: Measure the Stokes parameters through a four-channel photodetector: Wherein, , , , respectively represent the scattered light intensities at polarization analysis directions of 0°, 45°, 90°, and 135°; : Total light intensity, representing the total intensity of all polarization components, is calculated by the sum of and ; : The linearly polarized component (the difference in the 0° and 90° directions), obtained by subtracting from ; : The linearly polarized component (difference in the 45° and 135° directions), obtained by subtracting from to obtain; : Laser wavelength; : Polarization angle dynamically modulated by a liquid crystal phase retarder; : The polarization eigenvector, which contains three Stokes parameters and characterizes the polarization properties at different wavelengths and polarization angles; S14. Frequency-domain diversity sonar acquisition: Perform band-pass filtering on the sonar signals of the th receiving channel, and its transfer function is: In the formula, is the filter quality factor, is the th channel center frequency, is the frequency interval, is the imaginary unit; S15. Acoustic-optical data space-time registration: Establish a spatial position mapping relationship: wherein, is a discrete timestamp and , is a velocity vector, is an acceleration vector; S16. Cross-verification parameter matrix construction: Calculate the correlation coefficient matrix of the acoustic-optical data: In the formula, represents the th polarization feature component, represents the th frequency domain channel signal, is the number of sampling points of the sliding window.

4. The deep-sea terrain reconstruction method for ocean engineering surveying and mapping according to claim 2, wherein: S2 includes the following specific steps: S21. Sonar frequency-domain diversity preprocessing: Perform frequency-domain diversity processing on the sonar signals received by each channel of the arc phased array. Perform frequency-domain diversity processing, where the transfer function of the band-pass filter for the th channel is: Wherein, represents the filter quality factor and , represents the center frequency of the th channel, and is the imaginary unit; the time delay of each channel is compensated by the frequency-variable time delay compensation unit, and the compensation amount satisfies: In the formula, represents the distance between the th array element and the center of the array, is the speed of sound, is the incident angle of the sound wave; S22. Frequency domain feature enhancement: Perform short-time Fourier transform on the compensated sonar signal to generate a time-frequency distribution matrix: In the formula, represents a Hamming window function with a length of ; represents the discrete time-domain signal of the -th channel; represents the sampling interval. S23. Acousto-optic cross-correlation calculation: The laser polarization characteristics are subjected to normalized cross-correlation operation with the sonar frequency-domain signal as follows: In the formula, the denominator term is the norm of the polarization feature and the frequency-domain signal, which is used to eliminate the dimensional difference; S24. Dynamic weight fusion: introducing a frequency-domain diversity gain factor , dynamically adjusting the weights of each channel according to the signal-to-noise ratio: wherein, represents the signal-to-noise ratio of the th channel, represents the total number of channels; perform a tensor product operation on the weighted cross-correlation function and the polarization characteristic matrix to generate an acousto-optic coupling tensor: In the formula, represents the tensor product operator, represents the cross-correlation matrix of the 5. The deep-sea terrain reconstruction method for ocean engineering surveying and mapping according to claim 2, characterized in that: S3 includes the following specific steps: S31. Dynamic convolution kernel generation: Based on the frequency-domain energy distribution of the acousto-optic coupling tensor to generate adaptive convolution kernel parameters, and define the frequency-domain energy weight coefficient: In the formula, represents the weight coefficient in the frequency dimension , is the laser wavelength variable is the polarization angle variable is the energy modulus value of the acousto-optic coupling tensor at , represents the summation over all frequency points , and the denominator term is the full-band energy normalization factor; S32. Multi-scale convolution operation: Construct three groups of parallel convolution kernels with sizes of , , respectively. The weights of each group of convolution kernels are dynamically adjusted according to the following rules: wherein, represents the final weight of the th kernel in the th group of convolution kernels, corresponding to the sizes of three groups of convolution kernels, is the reference convolution kernel parameter matrix, is the adaptation parameter matrix generated according to the statistical characteristics of the current input tensor, is the convolution kernel serial number index; S33. Feature map fusion: Weightedly splice the three sets of convolution outputs according to the frequency domain significance: In the formula, represents the multi-scale fusion feature map, is the feature map tensor output by the -th group of convolutional operations, is the temperature coefficient, is the regulated fusion weight, represents the channel dimension average operation, is the summation index variable, is the natural exponential function; S34. Channel attention enhancement: Apply a frequency domain-polarization joint attention mechanism to the fused feature map: Wherein, represents the attention weight of the th channel, is the feature channel index variable (where is the height dimension of the feature map, is the width dimension of the feature map, is the spatial position index in the height direction , is the spatial position index in the width direction , represents the feature value of the th channel of the fused feature map at the spatial position , represents the normalization factor for averaging over the spatial dimension, is the double summation operator, indicating traversing and summing over all positions in the height and width directions, is the Sigmoid activation function, used to compress the scalar value to the interval; In the formula, is the channel weighted multiplication operator, indicating that the attention weight is multiplied element-wise with the fused feature map along the channel dimension, where is the finally output fused feature map, is the total number of channels of the feature map, and the original spatial dimension remains unchanged.

6. The method for reconstructing deep-sea terrain in ocean engineering surveying and mapping according to claim 2, characterized in that: S4 includes the following specific steps: S41. Multi-band inverse projection processing: Divide the fused feature map into low-frequency components according to frequency bands , intermediate-frequency components and high-frequency components , and perform inverse Fourier transform on them respectively: In the formula, : Represents the frequency band Inverse projection three-dimensional spatial distribution; : Three-dimensional inverse Fourier transform operator; : corresponding frequency band of the frequency-domain characteristic component, where \(f\) is the frequency variable, and \(\theta\) is the polarization angle variable; : Phase compensation term, where is the imaginary unit; : Frequency band Phase compensation angle of; : Complex real part extraction operation; : Complex imaginary part extraction operation; S42. Motion distortion compensation: Based on the attitude angles output by the MEMS inertial navigation unit and the Doppler log velocity vector , construct the displacement compensation amount: In the formula, : Three-dimensional space displacement compensation amount; : A three-dimensional rotation matrix composed of Euler angles ; : At the moment of the velocity vector; : The starting time of data acquisition; : Integration time variable; S43. Multi-frequency fusion terrain generation: Perform spatial domain fusion on the inverse projection results of each frequency band and the compensated displacement: In the formula, : Final 3D terrain model; : Frequency band of the dynamic weight coefficient; : Frequency band Signal-to-noise ratio of : in the component of the axis; S44. Terrain gradient field optimization: Introduce the Laplace-Poisson equation for terrain smoothing: In the formula, : Laplacian operator operation result of the optimized terrain model; : Gradient field of the original terrain model; : Inverse operation of the gradient magnitude; : Smoothing intensity control coefficient.

7. The deep - sea terrain reconstruction method for ocean engineering surveying and mapping according to claim 2, wherein: S5 includes the following specific steps: S51. Multimodal energy attenuation ratio calculation: Based on polarization characteristics and sonar frequency-domain signal Calculate the acousto-optic energy attenuation ratio: wherein, : Acoustic-optical energy attenuation ratio at three-dimensional coordinates which is used to characterize terrain reflection characteristics; : Wavelength Polarization eigenvalue at azimuth , is the coordinate corresponding detection azimuth; : Polarized light energy modulus square operation; : Frequency Sonar frequency domain signal at the azimuth ; : Sonar signal energy modulus square operation; : Numerical stability factor to prevent the denominator from being zero; S52. Confidence heat map generation: Combine the energy decay ratio with the spatial continuity constraint to generate terrain confidence: wherein, : Coordinates Terrain confidence at, value range ; : Three-dimensional spatial gradient of the energy attenuation ratio; : Gradient magnitude square operation; : Gaussian kernel width adjustment coefficient, which controls the intensity of spatial continuity; S53. Dynamic threshold anomaly detection: Automatically generate a rejection threshold according to the confidence distribution: wherein, : The dynamic rejection threshold, whose value range is the same as that of ; : Mean confidence of all scenarios; : Standard deviation of confidence level; : Statistical outlier factor, calibrated by 1,200 groups of measured data; S54. Morphological optimization rejection: Use anisotropic morphological filtering to remove isolated outliers: wherein, : Three-dimensional terrain data after cleaning; NaN: Invalid data identifier; : Original terrain elevation value.

Images