A method for reconstructing deep-sea terrain in marine engineering surveying and mapping

Abstract

The present invention relates to the field of marine engineering surveying and mapping technology, and provides a method for reconstructing deep-sea terrain for marine engineering surveying and mapping, including synchronous acquisition of multimodal data, cross-enhancement of frequency domain features, dynamic convolution feature fusion, inverse projection terrain reconstruction, and terrain credibility verification. The present invention adopts acoustic-optical multimodal fusion technology, combining the advantages of acoustic penetration and optical high resolution to enhance the ability to capture complex terrain details; the dynamic convolution network adaptively adjusts the convolution kernel parameters through frequency domain energy to enhance the multi-scale feature fusion effect and improve the accuracy of terrain texture expression; the motion distortion compensation method uses inertial navigation data to correct the carrier attitude drift in real time, thereby ensuring the geometric fidelity of the reconstructed model. The present invention is applicable to complex deep-sea environments and provides high-precision terrain reconstruction method support for marine engineering surveying and mapping and marine resource exploration.

Description

Technical Field

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

[0002] Seabed topography reconstruction for marine engineering surveying and mapping refers to the process of high-precision restoration of the three-dimensional spatial structure of the seabed topography through multi-sensor data fusion. Seabed topography reconstruction for marine engineering surveying and mapping, especially deep-sea topography reconstruction, is an important technical means to support marine resource development, submarine pipeline laying, marine environmental monitoring, and underwater engineering safety. As deep-sea exploration extends to areas with complex terrain, higher requirements are placed on the accuracy, anti-interference capability, and real-time performance of deep-sea terrain modeling. Traditional methods mainly rely on single-mode detection of multi-beam sonar or side-scan sonar, supplemented by inertial navigation systems for motion compensation. However, in deep-sea turbid water environments, they face problems such as severe acoustic signal attenuation and insufficient optical detection penetration, resulting in blurred terrain details and significant multipath interference, making it difficult to meet the needs of high-precision engineering surveying and mapping.

[0003] In existing technologies, the physical limitations of single-modal detection lead to incomplete capture of complex terrain features and under-exploitation of the complementary advantages of acoustic and optical data. Cross-modal data synchronization accuracy is insufficient, and motion distortion compensation models fail to account for multi-band phase response differences. Feature fusion networks lack dynamic adaptability to acoustic and optical coupling features, making it difficult to balance detail preservation and noise suppression. Terrain credibility assessment lacks multi-physics field joint judgment criteria, and outlier filtering easily leads to loss of effective terrain information. Specifically, terrain reconstruction methods based on single sensors suffer from the inherent flaw of missing multidimensional information. While sonar systems have strong penetration, they lack resolution for micro-topography. While laser detection can obtain high-precision surface information, it is limited by water turbidity and detection depth. Furthermore, multi-sensor collaborative detection faces challenges in spatiotemporal registration. Temporal synchronization errors in different modal data can easily lead to feature misalignment, and geometric distortion caused by attitude drift of the moving carrier further reduces the topological accuracy of the reconstructed model. At the data processing level, traditional convolutional neural networks have limited ability to integrate heterogeneous acoustic and optical features. Fixed-scale convolution kernels struggle to adapt to the multi-scale representation of deep-sea terrain, resulting in loss of high-frequency detail and distortion of low-frequency contours. Furthermore, existing confidence assessment methods often rely on single-value thresholds and lack multimodal energy attenuation correlation analysis, making it difficult to effectively distinguish between true terrain reflections and suspended matter artifacts, resulting in insufficient accuracy in removing abnormal point clouds.

[0004] These problems seriously restrict the accuracy and reliability of terrain reconstruction of deep-sea engineering surveying and mapping equipment under complex sea conditions. It is of great significance to design a deep-sea terrain reconstruction method for marine engineering surveying and mapping to solve the above problems. Summary of the Invention

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

[0006] S1. Synchronous multi-modal data acquisition: Acquiring laser polarization characteristics through a polarization-frequency domain synchronous modulation system With frequency domain signal , construct the acoustic-optical mutual verification parameter matrix;

[0007] S2. Frequency domain feature cross enhancement: for sonar frequency domain signals Perform Fourier transform and compare with the laser polarization characteristics Perform cross-correlation operations to generate the acousto-optic coupling tensor ;

[0008] S3, dynamic convolution feature fusion: using full-dimensional dynamic convolutional network CNN to Perform multi-scale feature extraction and generate fusion feature maps;

[0009] S4, inverse projection terrain reconstruction: Generate a 3D terrain model through inverse Fourier transform and superimpose inertial navigation data to compensate for motion distortion;

[0010] S5. Terrain credibility verification: based on the acoustic and optical energy attenuation ratio Generate terrain confidence heatmap 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] The dual-wavelength polarized laser emission module includes a first laser and a second laser, which coaxially output a composite polarized beam through a beam splitter prism; a liquid crystal phase retarder, located close to the laser output port, is used to dynamically modulate the polarization angle θ; a Stokes parameter detection unit, located in the echo receiving optical path, analyzes the polarization characteristics in real time; and a beam expander, located downstream of the beam splitter prism, is used to adjust the laser spot size.

[0013] The frequency-domain diversity sonar receiving module includes an adjustable bandpass filter, a frequency-varying time delay compensation unit, an AD sampling card, and an arc-shaped phased array distributed in a 120-degree fan shape around the dual-wavelength polarized laser transmitting module. Each receiving channel of the arc-shaped phased array is connected in series with an adjustable bandpass filter and a frequency-varying time delay compensation unit, and the output end is connected to the AD sampling card.

[0014] The spatiotemporal synchronization control module includes an FPGA controller, a MEMS inertial navigation unit, and a Doppler odometer. The FPGA controller generates synchronization trigger signals using the GPS disciplined clock to drive the first and second lasers and the AD sampling card, respectively. The MEMS inertial navigation unit and the Doppler odometer are rigidly connected and provide real-time feedback of attitude and velocity data to the FPGA controller.

[0015] The data mutual verification interface module includes a polarization-frequency domain verification unit, a bidirectional interactive optical-acoustic data buffer pool, and an acoustic-optical data buffer pool, which store the original collected data. The polarization-frequency domain verification unit is connected between the FPGA controller and the host computer to perform acoustic-optical data validity verification.

[0016] The first laser, second laser and phased array are arranged in an asymmetric spatial arrangement, with the laser optical axis and the center line of the sonar array forming a 30° detection angle; the FPGA controller synchronously controls the read and write timing of the liquid crystal phase delay, adjustable bandpass filter, optical-acoustic data buffer and acoustic-optical data buffer through trigger signals; the data of the MEMS inertial navigation unit and the Doppler odometer are input into the FPGA controller, and motion distortion is compensated in real time.

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

[0018] S11, dual wavelength polarized laser emission: the first laser and the second laser emit wavelengths of and The linearly polarized light is combined by the beam splitter prism to form a composite polarized light beam, which satisfies the light intensity ratio relationship:

[0019]

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

[0021] S12, dynamic polarization modulation: the generated frequency is The square wave signal drives the liquid crystal phase retarder to make the polarization angle It changes periodically according to the following rules:

[0022]

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

[0024] S13. Polarization Characteristic Analysis: Stokes Parameters Measured by Four-Channel Photodetector:

[0025]

[0026] Where, 、 、 、 Respectively represent the scattered light intensity at 0°, 45°, 90°, and 135° analysis directions; : Total light intensity, which represents the total intensity of all polarization components, is given by and and calculation; : Linear polarization component (the difference between 0° and 90° directions), given by minus get; : Linear polarization component (difference between 45° and 135° directions), given by minus get; : laser wavelength; : Polarization angle dynamically modulated by liquid crystal phase retarder; : Polarization characteristic vector, including three Stokes parameters, characterizing the polarization characteristics at different wavelengths and polarization angles;

[0027] S14, frequency domain diversity sonar acquisition: The sonar signal of each receiving channel is band-pass filtered, and its transfer function is:

[0028]

[0029] Where, is the filter quality factor, For the Channel center frequency, is the frequency interval, is an imaginary unit;

[0030] S15. Spatiotemporal registration of acoustic and optical data: Establishing spatial position mapping relationship:

[0031]

[0032] Where, is a discrete timestamp and , is the velocity vector, is the acceleration vector;

[0033] S16. Construction of mutual verification parameter matrix: Calculation of the correlation coefficient matrix of acoustic and optical data:

[0034]

[0035] Where, Indicates the polarization characteristic components, Indicates the frequency domain channel signals, is the number of sliding window sampling points.

[0036] Furthermore, S2 includes the following specific steps:

[0037] S21. Sonar frequency domain diversity preprocessing: Sonar signals received by each channel of the arc phased array Perform frequency domain diversity processing, where The bandpass filter transfer function of the channel is:

[0038]

[0039] Where, represents the filter quality factor and , Indicates the The center frequency of each channel; the delay of each channel is compensated by the frequency-varying delay compensation unit Compensation is performed, and the compensation amount satisfies:

[0040]

[0041] Where, Indicates the The distance between each element and the center of the array, is the speed of sound, is the incident angle of the sound wave;

[0042] S22, frequency domain feature enhancement: Perform short-time Fourier transform on the compensated sonar signal to generate a time-frequency distribution matrix:

[0043]

[0044] Where, Indicates the length is The Hamming window function, Indicates the The discrete time domain signal of each channel, represents the sampling interval;

[0045] S23. Acousto-optic cross-correlation calculation: The laser polarization characteristics Sonar frequency domain signal Perform normalized cross-correlation:

[0046]

[0047] In the formula, the denominator is the polarization characteristic and the frequency domain signal norm, used to eliminate dimensional differences;

[0048] S24, Dynamic Weight Fusion: Introducing Frequency Domain Diversity Gain Factor , dynamically adjust the weight of each channel according to the signal-to-noise ratio:

[0049]

[0050] Where, Indicates the The signal-to-noise ratio of each channel, represents the total number of channels; the weighted cross-correlation function is tensor-producted with the polarization characteristic matrix to generate the acousto-optic coupling tensor:

[0051]

[0052] Where, represents the tensor product operator, Indicates the The cross-correlation matrix of the channels.

[0053] Furthermore, S3 includes the following specific steps:

[0054] S31. Dynamic convolution kernel generation: based on the acousto-optic coupling tensor The frequency domain energy distribution generates the adaptive convolution kernel parameters and defines the frequency domain energy weight coefficient:

[0055]

[0056] Where, Represents the frequency dimension The weight coefficient of is the laser wavelength variable, is the polarization angle variable, is the acousto-optic coupling tensor in The energy modulus at Indicates that for all frequencies The sum is calculated, and the denominator is the full-band energy normalization factor;

[0057] S32, multi-scale convolution operation: construct three parallel groups of convolution kernels, whose sizes are 、 、 , the weights of each convolution kernel group are dynamically adjusted according to the following rules:

[0058]

[0059] Where, Indicates the The first convolution kernel in the group The final weight of each core, Corresponding to three sets of convolution kernel sizes, is the benchmark 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;

[0060] S33, feature map fusion: The three sets of convolution outputs are weighted and spliced ​​according to the frequency domain significance:

[0061]

[0062]

[0063] Where, represents the multi-scale fusion feature map, For the The feature map tensor output by the group convolution, is the temperature coefficient, is the fusion weight of the regulation, represents the channel dimension average operation, To sum index variables, is the natural exponential function;

[0064] S34, Channel Attention Enhancement: Applying frequency domain-polarization joint attention mechanism on the fused feature map:

[0065]

[0066] Where, Indicates the The attention weight of each channel, is the feature channel index variable is the total number of channels), is the height dimension of the feature map, is the width dimension of the feature map, Index of spatial position in height direction , Spatial position index in the width direction , Represents the fusion feature map Channel location in space The eigenvalue at represents the normalization factor for averaging the spatial dimensions, It is a double summation operator, which means traversing and summing all positions in the height and width directions. Sigmoid activation function is used to compress scalar values ​​into interval;

[0067]

[0068] Where, is the channel weighted multiplication operator, which means that the attention weight By channel dimension and fusion feature map Perform element-wise multiplication, where is the final output fusion feature map, is the total number of feature map channels, Keep the original spatial dimensions unchanged.

[0069] Furthermore, S4 includes the following specific steps:

[0070] S41, multi-band back projection processing: fusion feature map Split into low-frequency components by frequency band , intermediate frequency components and high frequency components , respectively perform inverse Fourier transform:

[0071]

[0072] Where, : Indicates frequency band The inverse projection three-dimensional spatial distribution of : 3D inverse Fourier transform operator; : Corresponding frequency band The frequency domain characteristic components of is the frequency variable, is the polarization angle variable; : Phase compensation term, where is an imaginary unit; : Frequency band Phase compensation angle; : Complex real part extraction operation; : Complex imaginary part extraction operation;

[0073] S42, Motion distortion compensation: attitude angle based on MEMS inertial navigation unit output and Doppler log velocity vector , construct the displacement compensation:

[0074]

[0075] Where, : three-dimensional spatial displacement compensation; : By Euler angle The three-dimensional rotation matrix formed; :time The velocity vector of : the starting time of data collection; : integral time variable;

[0076] S43, Multi-frequency fusion terrain generation: Fusion of the back projection results of each frequency band with the compensation displacement in the spatial domain:

[0077]

[0078]

[0079] Where, : Final 3D terrain model; : Frequency band Dynamic weight coefficient of : Frequency band signal-to-noise ratio; : exist The weight of the axis;

[0080] S44. Terrain gradient field optimization: Introducing the Laplace-Poisson equation for terrain smoothing:

[0081]

[0082] Where, : Laplace operator calculation results of the optimized terrain model; : gradient field of the original terrain model; : The inverse operation of the gradient modulus; : Smoothing intensity control coefficient.

[0083] Furthermore, S5 includes the following specific steps:

[0084] S51. Calculation of multimodal energy attenuation ratio: based on polarization characteristics Sonar frequency domain signal Calculate the acousto-optic energy attenuation ratio:

[0085]

[0086] Where, : 3D coordinates The acoustic and optical energy attenuation ratio at is used to characterize the terrain reflection characteristics; :wavelength In azimuth The polarization eigenvalue at For coordinates The corresponding detection azimuth; : square operation of polarized light energy modulus; :frequency In azimuth Sonar frequency domain signal at ; : square operation of sonar signal energy modulus; : numerical stability factor to prevent the denominator from being zero;

[0087] S52. Confidence heat map generation: Combine the energy attenuation ratio with the spatial continuity constraint to generate terrain confidence:

[0088]

[0089] Where, :coordinate The terrain confidence at the location, the value range is ; : three-dimensional spatial gradient of energy attenuation ratio; : Gradient modulus square operation; : Gaussian kernel width adjustment coefficient, controlling the strength of spatial continuity;

[0090] S53, dynamic threshold anomaly detection: automatically generate rejection thresholds based on confidence distribution:

[0091]

[0092] Where, : Dynamic elimination threshold, the value range is consistent; : The mean confidence value of the entire scene; : confidence standard deviation; : Statistical outlier factors were calibrated using 1,200 sets of measured data;

[0093] S54, morphological optimization and elimination: Use anisotropic morphological filtering to remove isolated outliers:

[0094]

[0095] Where, : Cleaned 3D terrain data; NaN: invalid data identifier; : Original terrain elevation value.

[0096] The beneficial effects achieved by the present invention are:

[0097] First, the present invention designs a deep-sea terrain reconstruction method that combines acoustic-optical frequency-domain cross-enhancement with dynamic weight fusion. Through frequency-domain diversity preprocessing and time-frequency distribution matrix generation, combined with normalized cross-correlation operations, it achieves deep correlation between sonar frequency-domain signals and laser polarization characteristics. The dynamic weight fusion mechanism automatically allocates gain factors based on the channel signal-to-noise ratio and constructs a three-dimensional acoustic-optical coupling tensor using tensor product operations. This effectively exploits the complementary advantages of acoustic penetration and optical resolution, suppresses noise interference from a single mode, and enhances the coupled expression of low-frequency terrain contours and high-frequency detailed features, providing a high-dimensional information carrier for subsequent feature extraction.

[0098] Second, the present invention designs a polarization-frequency domain synchronous modulation system, combining dual-wavelength polarized lasers with frequency-domain diversity sonar, achieving millisecond-level timing synchronization through an FPGA controller. Dynamic modulation of the composite polarized beam overcomes the physical limitations of single-wavelength detection. The asymmetric layout of the arc-shaped phased array forms a complementary detection perspective. The optical-acoustic data buffer and spatiotemporal registration algorithm solve the challenge of temporal alignment of cross-modal data. This design significantly improves the integrity and consistency of data acquisition in complex deep-sea environments, lays a foundation for high-quality data for multi-source information fusion, and enhances the system's adaptability to turbid waters and complex terrain.

[0099] Third, the present invention designs a full-dimensional dynamic convolutional network and attention mechanism. The proposed dynamic convolution kernel generation method adaptively adjusts multi-scale convolution parameters based on frequency domain energy distribution, and combines the channel attention mechanism to strengthen effective feature channels. Three sets of parallel convolution kernels achieve multi-scale feature fusion through frequency domain significance weighting, breaking through the limitations of traditional fixed convolution kernels. This design significantly improves the network's ability to learn the texture characteristics and spatial distribution patterns of seabed terrain, while retaining subtle terrain undulations while enhancing compatibility with acoustic and optical heterogeneous features, and realizing adaptive extraction of complex landform features.

[0100] Fourth, the present invention devised a multi-band inverse projection and motion distortion compensation method, employing a three-dimensional reconstruction method combining frequency-band inverse Fourier transform with phase compensation. Motion distortion correction is achieved by constructing displacement compensation values ​​from inertial navigation data. The Laplace-Poisson equation optimizes the terrain gradient field, resolving the geometric distortion caused by carrier motion. This effectively integrates low-frequency terrain skeletons with high-frequency detail information, significantly improving the geometric fidelity and spatial continuity of deep-sea terrain models. This approach maintains the topological accuracy of the reconstruction results, particularly in environments with strong current disturbances.

[0101] Fifth, the present invention establishes a credibility assessment system driven by the acoustic-optical energy attenuation ratio. This system generates confidence heatmaps based on multimodal energy attenuation ratios and innovatively implements adaptive removal of anomalous point clouds by combining spatial continuity constraints with dynamic threshold detection. Anisotropic morphological filtering optimizes noise removal strategies tailored to the characteristics of deep-sea terrain, significantly improving the physical interpretability of the credibility criteria for terrain data. It effectively distinguishes true terrain reflections from multipath interference artifacts, ensuring the spatial consistency and engineering usability of the reconstructed model. BRIEF DESCRIPTION OF THE DRAWINGS

[0102] Figure 1 It is the overall flow chart of the present invention;

[0103] Figure 2 This is a functional framework diagram of the polarization-frequency domain synchronous modulation system of the present invention;

[0104] Figure 3 This is a comparison diagram of the frequency domain signal-to-noise ratio distribution of Example 1, Comparative Example 1 and Comparative Example 2;

[0105] Figure 4 This is a thermodynamic comparison diagram of the terrain gradient field of Example 1, Comparative Example 1 and Comparative Example 2;

[0106] Figure 5 A comparison diagram of enhanced motion compensation of Example 1 and Comparative Example 3;

[0107] Figure 6 This is a comparison diagram of the original measured terrain and the reconstructed optimized terrain in Example 2.

[0108] Numbers in the figure:

[0109] 101. First laser; 102. Second laser; 105. Beam splitter; 103. Liquid crystal phase retarder; 104. Stokes parameter detection unit; 106. Beam expander; 201. Arc phased array; 202. Adjustable bandpass filter; 203. Frequency-varying delay compensation unit; 204. AD sampling card; 301. FPGA controller; 302. GPS disciplined clock; 303. MEMS inertial navigation unit; 304. Doppler odometer; 401. Optical-acoustic data buffer; 402. Acoustic-optical data buffer; 403. Polarization-frequency domain verification unit. DETAILED DESCRIPTION

[0110] The technical solutions of the present invention will be described clearly and completely below in conjunction with the drawings in the present invention. In addition, the forms of the various structures described in the following embodiments are merely examples. The present invention is not limited to the various structures described in the following embodiments. All other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0111] Reference Figure 1 and Figure 2 The present invention designs a polarization-frequency domain synchronous modulation system, which includes the following structure: a dual-wavelength polarized laser emission module, including a first laser 101 and a second laser 102, which coaxially output a composite polarized light beam through a beam splitter prism 105; a liquid crystal phase retarder 103, which is close to the laser output port and is used to dynamically modulate the polarization angle θ; a Stokes parameter detection unit 104, which is located in the echo receiving optical path and analyzes the polarization characteristics in real time; a beam expander 106, which is arranged downstream of the beam splitter prism 105 and is used to adjust the laser spot size.

[0112] The frequency domain diversity sonar receiving module includes an adjustable bandpass filter 202, a frequency-varying 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 polarized laser transmitting module; each array element receiving channel of the arc phased array is connected in series with the adjustable bandpass filter 202 and the frequency-varying time delay compensation unit 203 in sequence, and the output end is connected to the AD sampling card 204.

[0113] The spatiotemporal synchronization control module includes an FPGA controller 301, a MEMS inertial navigation unit 303, and a Doppler odometer 304. The FPGA controller 301 generates a synchronization trigger signal through the 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 odometer 304 are rigidly connected and provide real-time feedback of attitude and velocity data to the FPGA controller 301.

[0114] The data mutual verification interface module includes a polarization-frequency domain verification unit 403, a bidirectional 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 collected data; the polarization-frequency domain verification unit 403 is connected between the FPGA controller 301 and the host computer to perform acoustic and optical data validity verification.

[0115] The first laser 101, the second laser 102 and the phased array are arranged in an asymmetric spatial arrangement, 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 and write timing of the liquid crystal phase delay 103, the adjustable bandpass filter 202, the optical-acoustic data buffer pool 401 and the acoustic-optical data buffer pool 402 through trigger signals; the data of the MEMS inertial navigation unit 303 and the Doppler odometer 304 are input into the FPGA controller 301, and motion distortion is compensated in real time.

[0116] The dual-wavelength polarized laser emission module consists of a first laser 101, a second laser 102, a beam splitter 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 are coaxially output through the beam splitter 105 to form a composite polarized beam. The liquid crystal phase retarder 103 is located close 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 uses a four-channel photodetector to analyze the polarization characteristics in real time. The beam expander 106 is installed in the optical path downstream of the beam splitter 105 to adjust the laser spot size to meet different detection distance requirements.

[0117] The frequency-domain diversity sonar receiver module includes an adjustable bandpass filter 202, a frequency-variable delay compensation unit 203, an AD sampling card 204, and an arc-shaped phased array 201. The arc-shaped phased array 201 is arranged in a 120-degree fan-shaped pattern surrounding the laser transmitter module. Each receiving channel of the arc-shaped phased array is connected in series with the adjustable bandpass filter 202 and the frequency-variable delay compensation unit 203, and finally connected to the AD sampling card 204 for signal digitization. The adjustable bandpass filter 202 implements frequency-domain diversity reception, the frequency-variable delay compensation unit 203 performs phase calibration of the array delay based on the acoustic wave's incident angle, and the AD sampling card 204 is responsible for the synchronous acquisition and quantization of multi-channel signals.

[0118] The spatiotemporal synchronization control module integrates an FPGA controller 301, a MEMS inertial navigation unit 303, a Doppler odometer 304, and a GPS disciplined clock 302. The FPGA controller 301 generates nanosecond synchronization trigger signals from the GPS disciplined clock 302, driving the dual-wavelength laser and the AD sampling card 204 to initiate data acquisition. The MEMS inertial navigation unit 303 and the Doppler odometer 304 are rigidly connected and continuously feed the FPGA with the carrier's three-dimensional attitude angle and velocity data, achieving a unified spatiotemporal reference.

[0119] The data mutual verification interface module consists of a polarization-frequency domain verification unit 403, an optical-acoustic data buffer 401, and an acoustic-optical data buffer 402. The optical-acoustic data buffer 401 and the acoustic-optical data buffer 402 utilize a bidirectional interactive architecture, storing raw laser and sonar data, respectively. The polarization-frequency domain verification unit 403, connected between the FPGA controller 301 and the host computer, verifies the validity of the acoustic-optical data by calculating a correlation coefficient matrix.

[0120] Each module utilizes an asymmetric spatial layout, with the first laser 101, the second laser 102, and the arc-shaped phased array 201 forming a 30° detection angle. The FPGA controller 301 uses hardware trigger signals to precisely synchronize the polarization modulation of the liquid crystal phase retarder 103, the frequency switching of the adjustable bandpass filter 202, and the read and write timing of the data buffer. Motion parameters collected by the MEMS inertial navigation unit 303 and the Doppler odometer 304 are directly input into the FPGA, and a real-time displacement compensation algorithm is used to eliminate detection data distortion caused by carrier motion.

[0121] Reference Figure 1 The present invention provides a deep-sea topography reconstruction method based on multimodal data fusion, and designs the following steps: data acquisition is performed through a synchronous modulation system integrating polarized laser and frequency domain sonar. The dual-wavelength polarized laser module emits a composite polarized light beam, and the polarization angle is periodically modulated by a dynamic phase retarder, and the polarization characteristics of the seabed reflection are analyzed in combination with the Stokes parameter detection unit 104. The sonar module uses an arc-shaped phased array 201 to receive acoustic wave signals, and each channel is digitized after bandpass filtering and time delay compensation. The system uses an FPGA controller 301 to achieve millisecond-level synchronization of laser emission, sonar acquisition and inertial navigation data, and constructs an acoustic-optical mutual verification parameter matrix for spatiotemporal registration. The laser and sonar arrays are arranged in an asymmetric manner at 30°, and the GPS disciplined clock 302 is used to unify the timing, and the carrier posture is corrected in real time through motion sensors.

[0122] The sonar signal is subjected to a short-time Fourier transform to generate a time-frequency distribution, which is then normalized cross-correlated with the laser polarization characteristics. Weights are dynamically assigned based on the signal-to-noise ratio of each sonar channel. The acoustic and optical data are fused via a tensor product operation to generate a three-dimensional coupling tensor containing frequency, wavelength, and angle. This process enhances the correlation between low-frequency terrain contours and high-frequency details, suppressing multipath interference.

[0123] A full-dimensional dynamic convolutional network is designed to adaptively adjust multi-scale convolution kernel parameters based on frequency domain energy distribution. Feature maps of different scales are fused using frequency domain saliency weighting, and a channel attention mechanism is introduced to enhance the representation of features in effective frequency bands and polarization channels. The final output is a comprehensive feature map that integrates seafloor topography texture, reflectivity, and spatial distribution.

[0124] The fused features are segmented by frequency band and then subjected to a 3D inverse Fourier transform to restore the spatial terrain model. The attitude angles and Doppler velocity data from the inertial navigation output are combined to calculate displacement compensation to correct for motion distortion. An initial 3D model is generated by weighted fusion of the low-frequency terrain skeleton, mid-frequency structure, and high-frequency details. The terrain gradient field is further optimized using the Laplace-Poisson equation to eliminate local distortion caused by noise.

[0125] The acoustic and optical energy attenuation ratio is calculated and combined with spatial continuity constraints to generate a confidence heat map. The outlier rejection threshold is dynamically set based on the confidence distribution, and anisotropic morphological filtering is used to remove isolated noise points. The final output is a cleaned, high-confidence terrain model, with invalid areas marked as NaN. The specific process is as follows:

[0126] S1. Synchronous multi-modal data acquisition: Acquiring laser polarization characteristics through a polarization-frequency domain synchronous modulation system With frequency domain signal , construct the acousto-optic mutual verification parameter matrix; S1 includes the following specific steps:

[0127] S11, dual-wavelength polarized laser emission: the first laser 101 and the second laser 102 emit wavelengths and The linearly polarized light is combined by the beam splitter prism 105 to form a composite polarized light beam, which satisfies the light intensity ratio relationship:

[0128]

[0129] Where, represents the total intensity of the composite polarized 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;

[0130] S12, dynamic polarization modulation: the generated frequency is The square wave signal drives the liquid crystal phase retarder 103, making the polarization angle It changes periodically according to the following rules:

[0131]

[0132] Where, is the maximum deflection angle and satisfies , is the time variable;

[0133] S13. Polarization Characteristic Analysis: Stokes Parameters Measured by Four-Channel Photodetector:

[0134]

[0135] Where, 、 、 、 Respectively represent the scattered light intensity at 0°, 45°, 90°, and 135° analysis directions; : Total light intensity, which represents the total intensity of all polarization components, is given by and and calculation; : Linear polarization component (the difference between 0° and 90° directions), given by minus get; : Linear polarization component (difference between 45° and 135° directions), given by minus get; : laser wavelength; : Polarization angle dynamically modulated by liquid crystal phase retarder; : Polarization characteristic vector, including three Stokes parameters, characterizing the polarization characteristics at different wavelengths and polarization angles;

[0136] S14, frequency domain diversity sonar acquisition: The sonar signal of each receiving channel is band-pass filtered, and its transfer function is:

[0137]

[0138] Where, is the filter quality factor, For the Channel center frequency, is the frequency interval, is an imaginary unit;

[0139] S15. Spatiotemporal registration of acoustic and optical data: Establishing spatial position mapping relationship:

[0140]

[0141] Where, is a discrete timestamp and , is the velocity vector, is the acceleration vector;

[0142] S16. Construction of mutual verification parameter matrix: Calculation of the correlation coefficient matrix of acoustic and optical data:

[0143]

[0144] Where, Indicates the polarization characteristic components, Indicates the frequency domain channel signals, is the number of sliding window sampling points.

[0145] S2. Frequency domain feature cross enhancement: for sonar frequency domain signals Perform Fourier transform and compare with the laser polarization characteristics Perform cross-correlation operations to generate the acousto-optic coupling tensor S2 includes the following specific steps: S21, sonar frequency domain diversity preprocessing: the sonar signal received by each channel of the arc phased array 201 Perform frequency domain diversity processing, where The bandpass filter transfer function of the channel is:

[0146]

[0147] Where, represents the filter quality factor and , Indicates the The center frequency of each channel, The delay of each channel is compensated by the frequency-varying delay compensation unit 203. Compensation is performed, and the compensation amount satisfies:

[0148]

[0149] Where, Indicates the The distance between each element and the center of the array, is the speed of sound, is the incident angle of the sound wave;

[0150] S22, frequency domain feature enhancement: Perform short-time Fourier transform on the compensated sonar signal to generate a time-frequency distribution matrix:

[0151]

[0152] Where, Indicates the length is The Hamming window function, Indicates the The discrete time domain signal of each channel, represents the sampling interval;

[0153] S23. Acousto-optic cross-correlation calculation: The laser polarization characteristics Sonar frequency domain signal Perform normalized cross-correlation:

[0154]

[0155] In the formula, the denominator is the polarization characteristic and the frequency domain signal norm, used to eliminate dimensional differences;

[0156] S24, Dynamic Weight Fusion: Introducing Frequency Domain Diversity Gain Factor , dynamically adjust the weight of each channel according to the signal-to-noise ratio:

[0157]

[0158] Where, Indicates the The signal-to-noise ratio of each channel, represents the total number of channels; the weighted cross-correlation function is tensor-producted with the polarization characteristic matrix to generate the acousto-optic coupling tensor:

[0159]

[0160] Where, represents the tensor product operator, Indicates the The cross-correlation matrix of the channels.

[0161] S3, dynamic convolution feature fusion: using full-dimensional dynamic convolutional network CNN to Perform multi-scale feature extraction and generate fusion feature maps; S3 includes the following specific steps: S31, dynamic convolution kernel generation: based on the acoustic-optical coupling tensor The frequency domain energy distribution generates the adaptive convolution kernel parameters and defines the frequency domain energy weight coefficient:

[0162]

[0163] Where, Represents the frequency dimension The weight coefficient of is the laser wavelength variable, is the polarization angle variable, is the acousto-optic coupling tensor in The energy modulus at Indicates that for all frequencies The sum is calculated, and the denominator is the full-band energy normalization factor;

[0164] S32, multi-scale convolution operation: construct three parallel groups of convolution kernels, whose sizes are 、 、 , the weights of each convolution kernel group are dynamically adjusted according to the following rules:

[0165]

[0166] Where, Indicates the The first convolution kernel in the group The final weight of each core, Corresponding to three sets of convolution kernel sizes, is the benchmark 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;

[0167] S33, feature map fusion: The three sets of convolution outputs are weighted and spliced ​​according to the frequency domain significance:

[0168]

[0169]

[0170] Where, represents the multi-scale fusion feature map, For the The feature map tensor output by the group convolution, is the temperature coefficient, is the fusion weight of the regulation, represents the channel dimension average operation, To sum index variables, is the natural exponential function;

[0171] S34, Channel Attention Enhancement: Applying frequency domain-polarization joint attention mechanism on the fused feature map:

[0172]

[0173] Where, Indicates the The attention weight of each channel, is the feature channel index variable is the total number of channels), is the height dimension of the feature map, is the width dimension of the feature map, Index of spatial position in height direction , Spatial position index in the width direction , Represents the fusion feature map Channel location in space The eigenvalue at represents the normalization factor for averaging the spatial dimensions, It is a double summation operator, which means traversing and summing all positions in the height and width directions. Sigmoid activation function is used to compress scalar values ​​into interval;

[0174]

[0175] Where, is the channel weighted multiplication operator, which means that the attention weight By channel dimension and fusion feature map Perform element-wise multiplication, where is the final output fusion feature map, is the total number of feature map channels, Keep the original spatial dimensions unchanged.

[0176] S4, reverse projection terrain reconstruction: generate a three-dimensional terrain model by inverse Fourier transform, and superimpose inertial navigation data to compensate for motion distortion; S4 includes the following specific steps: S41, multi-band reverse projection processing: fusion feature map Split into low-frequency components by frequency band , intermediate frequency components and high frequency components , respectively perform inverse Fourier transform:

[0177]

[0178] Where, : Indicates frequency band The inverse projection three-dimensional spatial distribution of : 3D inverse Fourier transform operator; : Corresponding frequency band The frequency domain characteristic components of is the frequency variable, is the polarization angle variable; : Phase compensation term, where is an imaginary unit; : Frequency band Phase compensation angle; : Complex real part extraction operation; : Complex imaginary part extraction operation;

[0179] S42, motion distortion compensation: based on the attitude angle output by the MEMS inertial navigation unit 303 and Doppler log 304 velocity vector , construct the displacement compensation:

[0180]

[0181] Where, : three-dimensional spatial displacement compensation; : By Euler angle The three-dimensional rotation matrix formed; :time The velocity vector of : the starting time of data collection; : integral time variable;

[0182] S43, Multi-frequency fusion terrain generation: Fusion of the back projection results of each frequency band with the compensation displacement in the spatial domain:

[0183]

[0184]

[0185] Where, : Final 3D terrain model; : Frequency band Dynamic weight coefficient of : Frequency band signal-to-noise ratio; : exist The axis's weight;

[0186] S44, Terrain Gradient Field Optimization: Introducing the Laplace-Poisson equation for terrain smoothing:

[0187]

[0188] Where, : Laplace operator calculation results of the optimized terrain model; : gradient field of the original terrain model; : The inverse operation of the gradient modulus; : Smoothing intensity control coefficient.

[0189] The specific process of optimizing the terrain gradient field using the Laplace-Poisson equation is as follows: Equation construction principle:

[0190] Formula (23) can be expanded as: ;in It is clearly defined as the inverse operation of the gradient modulus. This equation combines the Laplace smoothing term and the gradient adaptive adjustment term. Its physical meaning is: the Laplace term ( : Forces the terrain surface to be smooth and removes high-frequency noise. Divergence term ( ): According to the original gradient field The smoothing strength is dynamically adjusted by the intensity of the gradient. In steep areas (large gradients), the smoothing effect is reduced to preserve details, and in flat areas (small gradients), the smoothing effect is enhanced to suppress noise. The finite difference method is used for discretization solution. The specific steps are as follows:

[0191] 3D terrain model The voxel grid is divided into two parts, the grid spacing is determined by the sonar resolution (typical value is 0.1 m × 0.1 m × 0.05 m). The gradient of each voxel is calculated using the central difference method. . Construct a sparse matrix to represent the Laplacian operator and an adaptive divergence term, where The value is obtained through empirical formula Dynamic determination to ensure that the smoothing strength matches the terrain roughness. The conjugate gradient method is used to solve the linear equations, and the convergence condition is set to the residual norm less than Or the maximum number of iterations is 200.

[0192] S5. Terrain credibility verification: based on the acoustic and optical 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 polarization characteristics Sonar frequency domain signal Calculate the acousto-optic energy attenuation ratio:

[0193]

[0194] Where, : 3D coordinates The acoustic and optical energy attenuation ratio at is used to characterize the terrain reflection characteristics; :wavelength In azimuth The polarization eigenvalue at For coordinates The corresponding detection azimuth; : square operation of polarized light energy modulus; :frequency In azimuth Sonar frequency domain signal at ; : square operation of sonar signal energy modulus; : numerical stability factor to prevent the denominator from being zero;

[0195] S52. Confidence heat map generation: Combine the energy attenuation ratio with the spatial continuity constraint to generate terrain confidence:

[0196]

[0197] Where, :coordinate The terrain confidence at the location, the value range is ; : three-dimensional spatial gradient of energy attenuation ratio; : Gradient modulus square operation; : Gaussian kernel width adjustment coefficient, controlling the strength of spatial continuity;

[0198] S53, dynamic threshold anomaly detection: automatically generate rejection thresholds based on confidence distribution:

[0199]

[0200] Where, : Dynamic elimination threshold, the value range is consistent; : The mean confidence value of the entire scene; : confidence standard deviation; : Statistical outlier factors were calibrated using 1,200 sets of measured data;

[0201] S54, morphological optimization and elimination: Use anisotropic morphological filtering to remove isolated outliers:

[0202]

[0203] Where, : Cleaned 3D terrain data; NaN: invalid data identifier; : Original terrain elevation value.

[0204] The method of the present invention is described below through specific examples and comparative examples.

[0205] Example 1: This example adopts the method designed by the present invention. The implementation object is a deep sea area with a terrain height difference of 500 m and a maximum slope of 35°. The implementation steps are as follows: data acquisition, ROV sails at a constant speed of 1.5 m / s, and synchronously collects the Stokes parameters of the laser polarization characteristics. and sonar frequency domain signals. The MEMS inertial navigation unit records roll angle ±3° and acceleration 0.1 m / s. Parameter configuration: S1: Laser wavelength: 、 Polarization modulation: , , ; Sonar filter: , , (50 channels in total); S2: Cross-correlation window: point (10 seconds of data); frequency domain diversity gain factor Dynamic allocation according to SNR (measured SNR = 12~25 dB); S3: Dynamic convolution kernel size: 、 、 ; Temperature coefficient , channel attention weight ; S4: Frequency band split: low frequency (1012 kHz), medium frequency (1216 kHz), high frequency (16-20 kHz); smoothing coefficient ; S5: Energy decay ratio threshold (calibrated value), morphological filter kernel .

[0206] Feature fusion to generate the acoustic-optical coupling tensor (Dimension 50×2×360), dynamic convolution output feature map size 256×256×32. Terrain reconstruction, reverse projection generates 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 Compared with the LiDAR benchmark data (RMS error 0.72 m), the confidence threshold , 4.2% outliers were removed. RMSE = 0.72 m, SNR = 24.3 dB, processing time 18.5 seconds.

[0207] 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, feature edges are blurred (mean gradient modulus = 0.15);

[0208] Comparative Example 2 (Laser Only): In this example, the sonar module is turned off and the laser polarization angle , Results: RMSE = 1.89 m, SNR = 18.6 dB, and the maximum diameter of the local cavity caused by underwater suspended matter is 3.5 m.

[0209] Figure 3 This is a comparison of the frequency domain signal-to-noise ratio (SNR) distributions for Example 1, Comparative Example 1, and Comparative Example 2. These plots were generated using the following steps: 50 discrete points were generated in the frequency range of 10-20 kHz; noisy SNR data were generated for Example 1, Comparative Example 1, and Comparative Example 2: Comparative Example 1 (sonar only): baseline SNR = 15 dB, superimposed Gaussian noise with a standard deviation of 2 dB. Comparative Example 2 (laser only): baseline SNR = 18 dB, superimposed Gaussian noise with a standard deviation of 3 dB. Example 1 (fusion method): baseline SNR = 24 dB, superimposed Gaussian noise with a standard deviation of 1.5 dB.

[0210] Table 1. Comparison of frequency domain signal-to-noise ratio (SNR) indicators among Example 1, Comparative Example 1 and Comparative Example 2

[0211] Curve characteristics Comparative Example 1 Comparative Example 2 Example 1 Basic SNR 15 dB 18 dB 24 dB Noise fluctuations ±2 dB (standard deviation) ±3 dB (standard deviation) ±1.5 dB (standard deviation) Curve stability Moderate volatility (red zone) Severe volatility (green zone) Stable (blue zone) Frequency response range Uniform but poor performance across the entire frequency band The high-frequency performance is significantly degraded Maintain high performance across all frequency bands

[0212] from Figure 3 As can be seen from the figure, Example 1 (blue line) has a higher SNR than Comparative Example 1 (red line) and Comparative Example 2 (green line) across the entire 10-20 kHz frequency band. The fusion of acoustic and optical data improves the baseline SNR by 60%, enhancing noise suppression and reducing standard deviation. At the key frequency of 15 kHz:

[0213] Table 2, key frequency indexes of Example 1, Comparative Example 1 and Comparative Example 2

[0214] method Mean SNR Fluctuation range (±3σ) Applicable Scenarios Comparative Example 1 15 dB 9-21 dB Shallow water low-speed mapping Comparative Example 2 18 dB 9-27 dB Small-scale mapping of clear water environment Example 1 24 dB 19.5-28.5 dB Deep-sea complex terrain mapping

[0215] As can be seen from Table 2, the fusion method of Example 1 achieves the improvement of signal-to-noise ratio stability and the expansion of effective detection bandwidth. Figure 4 This is a thermodynamic comparison diagram of the terrain gradient field of Example 1, Comparative Example 1 and Comparative Example 2.

[0216] Figure 4 The results show that in Example 1, the gradient field exhibits a continuous and clear edge structure, with high gradient value regions coherently distributed. Detail preservation is outstanding, with sharp gradient change boundaries visible in areas with sudden changes in terrain, such as steep slopes. Small-scale terrain undulations, such as submarine reefs, are clearly discernible in the thermal map.

[0217] In Comparative Example 1, there is significant low-frequency noise interference, namely, large, diffuse areas of high gradient. Artifacts diffuse along terrain edges, with unnatural jumps in gradient values. Spurious gradient fluctuations appear in flat areas due to water reverberation.

[0218] Comparative Example 2 exhibits high-frequency speckle noise, characterized by scattered, abnormally high gradient values. Scattering from suspended particles leads to localized gradient anomalies, represented by the random bright spots in the image. Insufficient optical penetration results in data voids, represented by black areas with zero gradient.

[0219] Table 3. Comparison of thermal indices of terrain gradient field among Example 1, Comparative Example 1 and Comparative Example 2

[0220] Comparison indicators Example 1 Comparative Example 1 Comparative Example 2 Average gradient modulus 2.8 3.5 4.2 Gradient standard deviation 0.9 1.7 2.3 Effective edge continuity 92% 65% 78%

[0221] from Figure 4 As can be seen from the figure, the penetration ability of sonar (Comparative Example 1) and the high resolution of laser (Comparative Example 2) are synergistically achieved in Example 1. The anisotropic noise of a single sensor is suppressed by the acousto-optic coupling tensor of the present invention.

[0222] In Example 1, a dynamic convolution kernel is used to smooth low-frequency features such as submarine basins using a large convolution kernel. High-frequency details such as fault zones are enhanced using a small convolution kernel. This improved gradient field continuity significantly reduces the error in depth contour generation, enabling safe path planning on terrain with steep slopes of 35° or greater.

[0223] Example 1 adopts a multimodal data fusion method, combines dual-wavelength polarized laser and frequency domain diversity sonar, and realizes deep-sea terrain reconstruction through a dynamic convolutional network. In the experiment, the laser wavelength is set to 532nm and 671nm, the sonar filter center frequency covers 50 channels from 10kHz to 20kHz, and the dynamic convolution kernel size is 3×3×3, 5×5×5 and 7×7×7. The results show that this method achieves a root mean square error (RMSE) of 0.72m and a signal-to-noise ratio (SNR) of 24.3dB in the deep-sea canyon area of ​​the South China Sea, with a processing time of only 18.5 seconds. Through confidence threshold screening, the outlier rejection rate is 4.2%, and the terrain resolution reaches 0.5m×0.5m×0.1m, verifying the advantages of multimodal collaboration in improving accuracy and noise resistance.

[0224] In Comparative Example 1, disabling the laser module and relying solely on sonar data resulted in an increase in RMSE to 2.34m, a decrease in SNR to 15.2dB, blurred feature edges, and a mean gradient modulus of only 0.15. The sonar signal was affected by water reverberation, resulting in spurious fluctuations in flat areas and significant performance degradation in high-frequency bands.

[0225] Comparative Example 2, using only the laser module, resulted in localized voids due to scattering from suspended matter, resulting in a maximum aperture of 3.5m. While superior to sonar in SNR, this was limited by optical penetration and resulted in significant high-frequency speckle noise. This experiment demonstrated the significant limitations of a single modality in complex deep-sea environments, and it was difficult to overcome both the sonar's low-frequency noise and the laser's high-frequency noise independently.

[0226] Comparing the frequency domain signal-to-noise ratio (SNR) distribution, Example 1 maintains a 24dB baseline SNR across the entire frequency band (10-20kHz), with a standard deviation of only ±1.5dB, significantly outperforming Comparative Example 1 (15dB±2dB) and Comparative Example 2 (18dB±3dB). At the critical 15kHz frequency, the mean SNR of Example 1 is 60% higher than that of Comparative Example 1, with the fluctuation range narrowed to 19.5-28.5dB. Multimodal fusion complements sonar penetration with laser high resolution, suppressing anisotropic noise and expanding the effective detection bandwidth.

[0227] Comparative Example 3 simulates the deep-sea terrain reconstruction error without motion compensation. Distorted data is constructed using the following method: carrier attitude changes (roll, pitch, and yaw) and integrated acceleration errors. This simulates the multi-sensor data mismatch in the absence of compensation. Comparative Example 3 uses the baseline terrain data from Example 1, but incorporates compound motion distortion and high-intensity Gaussian noise to simulate the uncompensated scenario. The results show significantly worse root mean square error and signal-to-noise ratio than those of Example 1, with a significant decrease in terrain gradient continuity and data efficiency.

[0228] Figure 5 This is a comparison diagram of enhanced motion compensation between Example 1 and Comparative Example 3. Figure 5 As can be seen from the table, the feature comparison and index comparison are shown in Table 4 and Table 5 respectively.

[0229] Table 4. Comparison of enhanced motion compensation features between Example 1 and Comparative Example 3

[0230] feature Example 1 (Compensation) Comparative Example 5 (uncompensated) Terrain Contours Preserve original peak / valley structure High-frequency ripples and lateral stretching occur Elevation distribution Concentrated [-6.5 m, 10.3 m] Abnormal values ​​up to ±15 m Surface details Smooth transition Local mutation Light reflection Uniform reflection Mottled reflections

[0231] Table 5. Comparison of enhanced motion compensation indicators between Example 1 and Comparative Example 3

[0232] index Example 1 Comparative Example 5 Improved results Root mean square error (RMSE) 0.72 m 4.85 m Reduced by 85.2% Signal-to-noise ratio (SNR) 24.3 dB 12.6 dB Increased by 92.9% Gradient continuity 92% 65% Increased by 41.5% Maximum distortion 0.3 m 4.8 m Inhibit 93.8%

[0233] As can be seen from Tables 4 and 5, the uncompensated data in Example 3 resulted in terrain distortion 4.8 times the baseline value, with the RMSE soaring to 4.85m and the SNR dropping to 12.6dB. High-frequency vibrations caused lateral stretching and abnormal extremes in the terrain profile, and gradient continuity plummeted from 92% to 65%. In Example 1, through real-time compensation using inertial navigation and Doppler odometer 304, the maximum distortion was suppressed to 0.3m and high-frequency vibration was attenuated by 89.7%, demonstrating the critical role of motion compensation in data reliability.

[0234] Example 2: This example is implemented in a certain deep-sea area 2, where the seabed topography includes deep-sea landform features such as seamounts, trenches, and slopes.

[0235] In this embodiment, the original measured terrain is generated by the original measurement data directly collected by the multimodal sensor. The data contains water suspended matter scattering noise, multipath interference noise, uncompensated carrier motion distortion, and no multimodal feature fusion. The data characteristics are as follows: significant high-frequency noise, with a noise standard deviation of 1.8 meters; low signal-to-noise ratio, SNR = 19.2 dB; linear stretching distortion and local extreme value anomalies, with a maximum deviation of ±3.5 meters; the terrain edge fuzzy gradient discontinuity area accounts for 32%, and the overall image presents significant spatial distortion and detail loss.

[0236] The reconstructed optimized terrain is generated by optimizing the data processed by the patented method of the present invention. The original data is sequentially processed using the following methods: separation of effective signal and noise through acousto-optic cross-correlation operation; dynamic convolution fusion, adaptive adjustment of multi-scale convolution kernel parameters, and preservation of micro-topography texture; motion distortion compensation, correction of linear stretching error based on inertial navigation data; and Gaussian smoothing optimization (σ=2) to suppress residual high-frequency noise.

[0237] Figure 6 This is a comparison diagram of the original measured terrain and the reconstructed optimized terrain generated by the simulation embodiment 2. Figure 6 It can be seen that the original measured topography exhibits obvious high-frequency noise (fine surface ripples) and linear stretching distortion, indicating that the uncompensated carrier motion leads to spatial position misalignment.

[0238] The reconstructed and optimized terrain surface is smooth and continuous, with micro-relief features such as seamounts (localized protrusions) and trenches (depressions) showing sharp, unblurred edges and effective geometric correction. The standard deviation of random noise in the original terrain reached 1.8 meters, and the color scale fluctuated from -6.5 to 8.2 meters, with local extreme deviations exceeding ±3 meters. After reconstruction, the noise amplitude was reduced to 0.4 meters, reducing anomalous extremes by 90%. Linear motion distortion, with slopes of 0.6 in the original terrain's X direction and 0.4 in the Y direction, was eliminated in the reconstructed terrain. The color bar displays elevations ranging from -3.2 to 7.5 meters, and the reconstructed terrain's color bands are more concentrated, verifying the effectiveness of noise suppression.

[0239] Simulations show that the frequency-domain cross-enhancement technique of our invention reduces high-frequency noise energy by 78%, from 1.8 to 0.4, and improves the signal-to-noise ratio (SNR) by 13.4 dB, from 19.2 to 32.6. This demonstrates that the acousto-optic cross-correlation algorithm effectively separates the signal from suspended object scattering noise. Gaussian smoothing (σ = 2) removes 95% of high-frequency speckle noise while retaining >85% of the local gradient standard deviation of micro-topography curvature (0.12). An inertial navigation compensation algorithm eliminates linear distortion caused by vehicle motion, reducing the RMSE from 2.1 meters to 0.5 meters and improving geometric fidelity by 76%. The terrain continuity index (the proportion of adjacent points with elevation differences less than 0.3 meters) increases from 68% in the original data to 93% in the reconstructed data, demonstrating the advantages of the dynamic convolutional network in fusing multi-scale features. Artifacts in the original data are removed, and the confidence heatmap mechanism accurately identifies >92% of multipath interference areas.

[0240] It can be seen from the above embodiments and comparative examples that the multimodal feature fusion and dynamic convolution kernel design of the present invention significantly improve the accuracy of terrain reconstruction, and the cross-correlation operation of laser polarization characteristics and sonar frequency domain signals enhances the ability to retain details. Motion compensation technology effectively eliminates the influence of carrier posture changes, and the combination of confidence heat map and morphological filtering optimizes data quality. Compared with traditional single-modal methods, the present invention shows stronger adaptability in complex deep-sea environments, can achieve high-precision terrain reconstruction, break through the technical bottlenecks of traditional methods in noise suppression, motion compensation and detail retention, and provide reliable technical support for marine engineering surveying and mapping.

[0241] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for reconstructing deep-sea terrain for marine engineering surveying and mapping, characterized in that: It includes the following steps: S1. Synchronous multi-modal data acquisition: Acquiring laser polarization characteristics through a polarization-frequency domain synchronous modulation system Sonar frequency domain signal , construct the acousto-optic mutual verification parameter matrix; where: ； Where, : Laser polarization characteristics, including three Stokes parameters, characterize the polarization characteristics at different wavelengths and polarization angles; 、 、 、 Respectively represent the scattered light intensity at 0°, 45°, 90°, and 135° analysis directions; : Total light intensity, which represents the total intensity of all polarization components, is given by and and calculation; : linear polarization component, given by minus get; : linear polarization component, given by minus get; : laser wavelength; : Polarization angle dynamically modulated by liquid crystal phase retarder; ； Where, : Sonar frequency domain signal, representing the output signal of the frequency domain diversity sonar receiving module; : frequency variable of sonar frequency domain signal; : angle of incidence of sound waves; : Length is Hamming window function; :No. Discrete time domain signal of channels; : natural constant; :imaginary unit; : sampling interval; S2. Frequency domain feature cross enhancement: for sonar frequency domain signals Perform Fourier transform and compare with the laser polarization characteristics Perform cross-correlation operations to generate the acousto-optic coupling tensor ; S3, dynamic convolution feature fusion: using full-dimensional dynamic convolutional network CNN to Perform multi-scale feature extraction and generate fusion feature maps; S4, inverse projection terrain reconstruction: Generate a 3D terrain model through inverse Fourier transform and superimpose inertial navigation data to compensate for motion distortion; S5. Terrain credibility verification: based on the acoustic and optical energy attenuation ratio Generate terrain confidence heatmap and remove abnormal point clouds.

2. The method for reconstructing deep-sea terrain for marine engineering surveying and mapping according to claim 1, characterized in that: The polarization-frequency domain synchronous modulation system used in S1 includes the following structure: The dual-wavelength polarized laser emission module includes a first laser and a second laser, which coaxially output a composite polarized beam through a beam splitter prism; a liquid crystal phase retarder, located close to the output port of the composite polarized beam, is used to dynamically modulate the polarization angle θ; a Stokes parameter detection unit, located in the echo receiving optical path, analyzes the polarization characteristics in real time; and a beam expander, located downstream of the beam splitter prism, is used to adjust the laser spot size. The frequency domain diversity sonar receiving module includes an adjustable bandpass filter, a frequency-variant time delay compensation unit, an AD sampling card, and an arc-shaped phased array distributed in a 120-degree fan shape around the dual-wavelength polarized laser transmitting module. Each receiving channel of the arc-shaped phased array is connected in series with the adjustable bandpass filter and the frequency-variant time delay compensation unit, and the output end of the frequency-variant time delay compensation unit is connected to the AD sampling card. The spatiotemporal synchronization control module includes an FPGA controller, a MEMS inertial navigation unit, and a Doppler odometer. The FPGA controller generates synchronization trigger signals using the GPS disciplined clock to drive the first and second lasers and the AD sampling card, respectively. The MEMS inertial navigation unit and the Doppler odometer are rigidly connected and provide real-time feedback of attitude and velocity data to the FPGA controller. The data mutual verification interface module includes a polarization-frequency domain verification unit, a bidirectional interactive optical-acoustic data buffer pool, and an acoustic-optical data buffer pool, which store the original collected data. The polarization-frequency domain verification unit is connected between the FPGA controller and the host computer to perform acoustic-optical data validity verification. The first laser, second laser and arc-shaped phased array are arranged in an asymmetric spatial layout, with the laser optical axis and the center line of the sonar array forming a 30° detection angle; the FPGA controller synchronously controls the read and write timing of the liquid crystal phase delay, adjustable bandpass filter, optical-acoustic data buffer pool and acoustic-optical data buffer pool through trigger signals; the data of the MEMS inertial navigation unit and the Doppler odometer are input into the FPGA controller, and motion distortion is compensated in real time.

3. The method for reconstructing deep-sea terrain for marine engineering surveying and mapping according to claim 2, characterized in that: S1 includes the following specific steps: S11, dual wavelength polarized laser emission: the first laser and the second laser emit wavelengths of and The linearly polarized light is combined by the beam splitter prism to form a composite polarized light beam, which satisfies the light intensity ratio relationship: Where, represents the total intensity of the composite polarized beam, is the output intensity of the first laser, is the output intensity of the second laser, is the polarization angle dynamically modulated by the liquid crystal retarder; S12, dynamic polarization modulation: the generated frequency is The square wave signal drives the liquid crystal phase retarder to make the polarization angle It changes periodically according to the following rules: Where, is the maximum deflection angle and satisfies , is the time variable; S13. Polarization characteristics analysis: Laser polarization characteristics are measured using a four-channel photodetector: S14, frequency domain diversity sonar acquisition: The sonar signal of each receiving channel is band-pass filtered, and its transfer function is: Where, is the filter quality factor, For the Channel center frequency, is the frequency interval, is an imaginary unit; S15. Spatiotemporal registration of acoustic and optical data: Establishing spatial position mapping relationship: Where, is a discrete timestamp and , is the velocity vector, is the acceleration vector; S16. Construction of mutual verification parameter matrix: Calculation of the correlation coefficient matrix of acoustic and optical data: Where, Indicates the polarization characteristic components, Indicates the frequency domain channel signals, is the number of sliding window sampling points.

4. The method for reconstructing deep-sea terrain for marine engineering surveying and mapping according to claim 2, characterized in that: S2 includes the following specific steps: S21. Sonar frequency domain diversity preprocessing: Sonar signals received by each channel of the arc phased array Perform frequency domain diversity processing, where The bandpass filter transfer function of the channel is: Where, represents the filter quality factor and , Indicates the The center frequency of each channel, is an imaginary unit; the delay of each channel is compensated by the frequency-varying delay compensation unit Compensation is performed, and the compensation amount satisfies: Where, Indicates the The distance between each 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 sonar frequency domain signal: S23. Acousto-optic cross-correlation calculation: The laser polarization characteristics Sonar frequency domain signal Perform normalized cross-correlation: In the formula, the denominator is the polarization characteristic and the frequency domain signal norm, used to eliminate dimensional differences; S24, Dynamic Weight Fusion: Introducing Frequency Domain Diversity Gain Factor , dynamically adjust the weight of each channel according to the signal-to-noise ratio: Where, Indicates the The signal-to-noise ratio of each channel, represents the total number of channels; the weighted cross-correlation function is tensor-producted with the polarization characteristic matrix to generate the acousto-optic coupling tensor: Where, represents the tensor product operator, Indicates the The cross-correlation matrix of the channels.

5. The method for reconstructing deep-sea terrain for marine 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 acousto-optic coupling tensor The frequency domain energy distribution generates the adaptive convolution kernel parameters and defines the frequency domain energy weight coefficient: Where, represents the frequency domain energy weight coefficient, is the laser wavelength variable, is the polarization angle variable, is the acousto-optic coupling tensor in The energy modulus at Indicates that for all frequencies The sum is calculated, and the denominator is the full-band energy normalization factor; S32, multi-scale convolution operation: construct three parallel groups of convolution kernels, whose sizes are 、 、 , the weights of each convolution kernel group are dynamically adjusted according to the following rules: Where, Indicates the The first convolution kernel in the group The final weight of each core, Corresponding to three sets of convolution kernel sizes, is the benchmark 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: The three sets of convolution outputs are weighted and spliced ​​according to the frequency domain significance: Where, represents the multi-scale fusion feature map, For the The feature map tensor output by the group convolution, is the temperature coefficient, is the fusion weight of the regulation, represents the channel dimension average operation, To sum index variables, is the natural exponential function; S34, Channel Attention Enhancement: Applying frequency domain-polarization joint attention mechanism on the fused feature map: Where, Indicates the The attention weight of each channel, is the feature channel index variable is the total number of channels, is the height dimension of the feature map, is the width dimension of the feature map, Index of spatial position in height direction , Spatial position index in the width direction , Represents the fusion feature map Channel location in space The eigenvalue at represents the normalization factor for averaging the spatial dimensions, It is a double summation operator, which means traversing and summing all positions in the height and width directions. Sigmoid activation function is used to compress scalar values ​​into interval; Where, is the channel weighted multiplication operator, which means that the attention weight By channel dimension and fusion feature map Perform element-wise multiplication, where is the final output fusion feature map, is the total number of feature map channels, Keep the original spatial dimensions unchanged.

6. The method for reconstructing deep-sea terrain for marine engineering surveying and mapping according to claim 2, characterized in that: S4 includes the following specific steps: S41, multi-band back projection processing: fusion feature map Split into low-frequency components by frequency band , intermediate frequency components and high frequency components , respectively perform inverse Fourier transform: Where, : Indicates frequency band The inverse projection three-dimensional spatial distribution of : 3D inverse Fourier transform operator; : Corresponding frequency band The frequency domain characteristic components of is the frequency variable, is the polarization angle variable; : Phase compensation term, where is an imaginary unit; : Frequency band Phase compensation angle; : Complex real part extraction operation; : Complex imaginary part extraction operation; S42, Motion distortion compensation: attitude angle based on MEMS inertial navigation unit output and Doppler log velocity vector , construct the displacement compensation: Where, : three-dimensional spatial displacement compensation; : By Euler angle The three-dimensional rotation matrix formed; :time The velocity vector of : the starting time of data collection; : The time when data collection ends; : integral time variable; S43, Multi-frequency fusion terrain generation: Fusion of the back projection results of each frequency band with the compensation displacement in the spatial domain: Where, : Final 3D terrain model; : Frequency band Dynamic weight coefficient of : Frequency band signal-to-noise ratio; : exist The axis's weight; S44, Terrain Gradient Field Optimization: Introducing the Laplace-Poisson equation for terrain smoothing: Where, : Laplace operator calculation results of the optimized terrain model; : gradient field of the original terrain model; : The inverse operation of the gradient modulus; : Smoothing intensity control coefficient.

7. The method for reconstructing deep-sea terrain for marine engineering surveying and mapping according to claim 2, characterized in that: S5 includes the following specific steps: S51. Calculation of multimodal energy attenuation ratio: based on polarization characteristics Sonar frequency domain signal Calculate the acousto-optic energy attenuation ratio: Where, : 3D coordinates The acoustic and optical energy attenuation ratio at is used to characterize the terrain reflection characteristics; :wavelength In azimuth The polarization eigenvalue at For coordinates The corresponding detection azimuth; : square operation of polarized light energy modulus; :frequency In azimuth Sonar frequency domain signal at ; : square operation of sonar signal energy modulus; : numerical stability factor to prevent the denominator from being zero; S52. Confidence heat map generation: Combine the energy attenuation ratio with the spatial continuity constraint to generate terrain confidence: Where, :coordinate The terrain confidence at the location, the value range is ; : three-dimensional spatial gradient of energy attenuation ratio; : Gradient modulus square operation; : Gaussian kernel width adjustment coefficient, controlling the strength of spatial continuity; S53, dynamic threshold anomaly detection: automatically generate rejection thresholds based on confidence distribution: Where, : Dynamic elimination threshold, the value range is consistent; : The mean confidence value of the entire scene; : confidence standard deviation; : Statistical outlier factors were calibrated using 1,200 sets of measured data; S54, morphological optimization and elimination: Use anisotropic morphological filtering to remove isolated outliers: Where, : Cleaned 3D terrain data; NaN: invalid data identifier; : Original terrain elevation value.

Images