Coral reef substrate classification method based on lidar and side-scan sonar data fusion

Abstract

This invention discloses a method for classifying coral reef substrate based on the fusion of lidar and side-scan sonar data, mainly including the following steps: (1) By analyzing the structure of the photon-counting lidar scanning system, a precise calculation model for the coordinates of underwater depth sounding points of UAV-borne lidar is constructed; (2) Through fine processing steps such as seabed line tracking and joint correction of radiation distortion, a geocoded side-scan sonar image is constructed; (3) The backscattering intensity of lidar and side-scan sonar is fused to construct a seabed topographic image; (4) The standard wavelet BP neural network is improved; (5) Based on the fused image, coral reef species samples are extracted, and feature parameters with high self-cohesion and resolution are selected for sample training. The sample training results are used to classify the coral reef substrate in the fused image. Through the above steps, coral reef substrate classification based on the fusion of lidar and side-scan sonar data can be realized, providing technical support for identifying coral reef types.

Description

Technical Field

[0001] This invention relates to a method for classifying coral reef substrate based on the fusion of lidar and side-scan sonar data, and particularly to a method for calculating the coordinates of underwater depth sounding points using a lidar based on a photon counting mechanism using underwater light tracking, as well as a fusion technology for lidar and side-scan sonar data. Background Technology

[0002] Coral reefs are among the marine ecosystems with the highest biodiversity and primary productivity, playing a vital role in the health and sustainable development of human society and the marine environment. Therefore, the investigation, monitoring, and protection of coral reefs have become urgent and important tasks. Compared with underwater manual photography or surveys, UAV-borne lidar measurement of coral reefs offers high efficiency, high precision, and high coverage. However, due to water absorption of laser light, the measurement depth is limited, generally ranging from a few meters to tens of meters. Shipborne side-scan sonar, on the other hand, features long detection range and high-density scanning, but its positioning accuracy is not high. By leveraging the complementary advantages of both methods, high-precision, high-density images of coral reef substrates at great depths can be constructed. Based on these images, an improved wavelet BP neural network can be used to classify various coral reef substrate types.

[0003] In summary, although UAV-borne lidar and shipborne side-scan sonar have certain advantages in detecting coral reefs, current technologies still have the following shortcomings:

[0004] (1) The UAV-borne lidar is easily affected by wind, which can cause the attitude sensor to jitter too much, resulting in a large attitude error and ultimately affecting the mapping accuracy of the lidar.

[0005] (2) Due to the limitations of its own measurement mechanism, lidar has not been able to effectively detect coral reefs at great depths.

[0006] (3) Since the shipborne side-scan sonar towed fish is softly connected to the support point on the ship, the accuracy of the towed fish's position calculation will be affected and needs to be improved.

[0007] (4) The mixed substrate of coral reefs has not been effectively distinguished, and an effective method for decomposing mixed pixels needs to be found. Summary of the Invention

[0008] The technical problem this invention aims to solve is the low efficiency and accuracy of current side-scan sonar for detecting coral reef bottoms. Firstly, by analyzing the structure of a photon-counting lidar scanning system, the geometric relationship between the incident angle and azimuth angle of the reflected laser beam on the water surface and the normal vector of the reflector is established. Furthermore, a precise calculation model for the underwater depth sounding point coordinates of an unmanned aerial vehicle (UAV) lidar is constructed using a constant-speed-of-light ray tracking algorithm within the underwater layer. Secondly, through fine processing steps such as seabed line tracking, combined radiation distortion correction, slant range correction, towed fish position estimation, and echo sampling point coordinate calculation, a geocoded side-scan sonar image is constructed. Thirdly, it utilizes… Z The backscattering intensities of lidar and side-scan sonar are normalized and converted into grayscale values. Simultaneously, the two data sets are fused using a near-homogeneous point rule. A fused image is then constructed through geocoding and image resampling. Next, based on a standard wavelet BP neural network, improvements are made to the network learning rate, momentum factor, and initial values ​​to enhance the training rate and fitting accuracy. Finally, based on the high-precision, high-resolution fused image, various coral reef species samples are extracted from corresponding locations in the fused image using geographic coordinates captured by a grab or underwater camera. These samples are then trained using statistical feature parameters with high self-cohesion and resolution. The training results are used to classify the coral reef substrate in the fused image.

[0009] To achieve the above objectives, the present invention provides a method for classifying coral reef substrate based on the fusion of lidar and side-scan sonar data, comprising the following steps:

[0010] (1) By analyzing the structure of the photon counting lidar scanning system, a laser scanning reference coordinate system and its transition coordinate system are established. Then, the geometric relationship between the incident angle and azimuth angle of the laser reflected light on the water surface and the normal vector of the reflector is established to obtain the three-dimensional coordinates of the laser incident point on the water surface. Furthermore, by constructing an underwater light speed profile and an underwater constant light speed ray tracking algorithm, a precise calculation model for the underwater depth sounding point coordinates of the UAV-borne lidar is constructed.

[0011] (2) Through fine processing steps such as seabed line tracking, radiation distortion joint correction, slant range correction, towed position estimation, and echo sampling point coordinate calculation, a side-scan sonar image with geocoding is constructed.

[0012] (3) Favorable Z The score normalizes the backscattering intensity of lidar and side-scan sonar and converts it into grayscale values. At the same time, it fuses the two types of data through the near-homologous point rule, and then constructs a fused image through geocoding and image resampling.

[0013] (4) Based on the standard wavelet BP neural network, the training rate and fitting accuracy of the wavelet BP neural network are improved by improving the network learning rate, momentum factor and initial value.

[0014] (5) Based on the high-precision and high-resolution fused image, various coral reef species samples are extracted from the corresponding positions of the fused image according to the geographic coordinates captured by the grab or underwater camera. Then, some statistical feature parameters with high self-cohesion and resolution are selected for sample training. The fused image is used to classify the coral reef substrate.

[0015] In one embodiment of the present invention, the geometric relationship between the incident angle and azimuth angle of the laser reflected light on the water surface and the normal vector of the reflecting mirror is established. Furthermore, a precise calculation model for the coordinates of underwater depth sounding points of an unmanned aerial vehicle (UAV) lidar is constructed using a constant-speed-of-light ray tracing algorithm within the underwater layer. This mainly includes the following steps:

[0016] (1) Analyze the structure of the photon counting lidar scanning system;

[0017] (2) Establishment of the laser scanning reference coordinate system and its transition coordinate system;

[0018] (3) Establishment of a model relating the incident angle and azimuth angle of the reflected light on the water surface to the normal vector of the mirror in the laser scanning reference coordinate system;

[0019] (4) Calculate the three-dimensional coordinates of the laser incident point on the water surface;

[0020] (5) Construct an underwater light speed profile;

[0021] (6) Establish a constant light speed ray tracing model in the underwater layer;

[0022] (7) Construct an accurate calculation model of the coordinates of underwater sounding points in the laser scanning reference coordinate system.

[0023] In one embodiment of the present invention, a geocoded side-scan sonar image is constructed through fine processing steps such as seabed line tracking, combined radiation distortion correction, slant range correction, towed fish position estimation, and echo sampling point coordinate calculation. This mainly includes the following steps:

[0024] (1) Tracking the seabed line;

[0025] (2) Combined correction of radiation distortion;

[0026] (3) Slope distance correction;

[0027] (4) Calculation of the position of the dragfish;

[0028] (5) Calculation of coordinates of side-scan sonar echo sampling points;

[0029] (6) Construct side-scan sonar images with geocoding.

[0030] In one embodiment of the present invention, it is advantageous to ZThe scores normalize the backscattering intensities of lidar and side-scan sonar and convert them into grayscale values. Simultaneously, the two types of data are fused using the near-corresponding point rule. Finally, a fused image is constructed through geocoding and image resampling. The main steps include:

[0031] (1) Favorable Z The fraction normalizes the backscattering intensity of lidar and side-scan sonar;

[0032] (2) Convert the normalized data of both into grayscale values;

[0033] (3) Establish rules for near-identical points;

[0034] (4) Use the near-corresponding point rule to fuse the two types of data;

[0035] (5) Re-geocode the merged data;

[0036] (6) Image resampling;

[0037] (7) Construct a fusion image of lidar and side-scan sonar.

[0038] In one embodiment of the present invention, based on a standard wavelet BP neural network, the training rate and fitting accuracy of the wavelet BP neural network are improved by modifying the network learning rate, momentum factor, and initial values. This mainly includes the following steps:

[0039] (1) Establish the basic framework of wavelet BP neural network, namely, one input layer, one hidden layer and one output layer, with the wavelet function used as the transfer function in the hidden layer;

[0040] (2) Improvement of network learning rate;

[0041] (3) Improvement of momentum factor;

[0042] (4) Considering the relationship between the initial values ​​of network weights, neuron transfer functions, and learning samples, construct an optimal initial weight model.

[0043] In one embodiment of the present invention, image samples of various coral reef types are extracted based on the fused image, and the feature values ​​of the image samples are statistically calculated. An improved wavelet BP neural network is used to classify the coral reef substrate in the fused image, mainly including the following steps:

[0044] (1) Based on the fused images, extract image samples of various types of coral reefs according to the actual shooting coordinates of the grab or camera;

[0045] (2) Select some statistical feature parameters with high self-cohesion and resolution (such as standard deviation, entropy, third moment, kurtosis, skewness, etc.) as image sample features;

[0046] (3) Select 70% and 30% of the samples respectively as training and verification, and perform image sample training and verification;

[0047] (4) Use the sample training results to classify the coral reef substrate in the fused images.

[0048] In one embodiment of the present invention, the constant light speed ray tracing algorithm in the underwater layer is as follows:

[0049] 1) The laser beam undergoes the process of... N A water column composed of layers, where light travels at the constant speed of light within each layer, according to Snell's law:

[0050] (10)

[0051] 2) Assume the thickness of the water column is... Then the beam is in the layer i Horizontal displacement within y i and transmission time t i for:

[0052] (11)

[0053] (12)

[0054] According to equations (11) and (12), the horizontal distance and propagation time of the light beam through the entire water column are respectively:

[0055] (13)

[0056] (14)

[0057] Assuming the beam of light does not pass through the entire water column, but... Z r The beam disappears at this point, and at this time the horizontal displacement of the beam in that layer is... The vertical displacement is The time it takes for the light beam to travel through the entire water column is... t all The time spent on this layer is t r The optical path length experienced by the beam in that layer is... for:

[0058] (15)

[0059] (16)

[0060] (17)

[0061] Therefore, the total horizontal and vertical displacements of the light beam in the water are respectively:

[0062] (18)

[0063] (19).

[0064] The technical problems to be solved by this invention mainly include the following aspects:

[0065] (1) Analyze the structure of the single-photon lidar scanning system and establish a geometric relationship model between the laser reflected ray and the normal vector of the mirror;

[0066] (2) Construct a coordinate calculation model of the laser radar incident point on the water surface in the laser scanning reference coordinate system;

[0067] (3) A constant light speed ray tracing model for underwater layers is proposed;

[0068] (4) Establish a coordinate calculation model for underwater laser depth sounding points in the laser scanning reference coordinate system;

[0069] (5) Construct high-resolution side-scan sonar images;

[0070] (6) Fusion processing of lidar and side-scan sonar data;

[0071] (7) Improve the standard wavelet BP neural network;

[0072] (8) Classify the coral reef substrate in the fused image.

[0073] The beneficial effects of the present invention through the above technical solution are:

[0074] (1) By studying the scanning structure of photon counting lidar, a geometric relationship model between the laser reflected ray and the normal vector of the mirror was constructed, which can accurately calculate the coordinates of the laser incident point on the water surface;

[0075] (2) A constant light speed ray tracing model based on the light speed profile is proposed, and a coordinate calculation model of the underwater laser sounding point in the laser scanning reference coordinate system is established, which can greatly improve the measurement accuracy of the underwater sounding point.

[0076] (3) Favorable Z The score is used to fuse data from lidar and side-scan sonar using different mechanisms;

[0077] (4) Improve the accuracy of coral reef substrate classification by improving the traditional wavelet BP neural network. Attached Figure Description

[0078] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0079] Figure 1 This is a schematic diagram of the elliptical scanning system of the photon counting lidar of the present invention;

[0080] Figure 2 These are the two Cartesian coordinate systems of the normal direction vector of the reflector in this invention;

[0081] Figure 3 This refers to the geometric angle of the reflected light rays in the sensor coordinate system according to the present invention;

[0082] Figure 4 This is a schematic diagram of the geometric structure of the emitted laser calculated from the change of normal in this invention;

[0083] Figure 5 This is a schematic diagram of the laser incident point on the sea surface according to the present invention;

[0084] Figure 6 This is a schematic diagram of the intralayer constant light speed ray tracing of the present invention;

[0085] Figure 7 This is a schematic diagram of the underwater optical path of the lidar of the present invention;

[0086] Figure 8 This is a schematic diagram of the seabed tracking method of the present invention;

[0087] Figure 9 This invention relates to the amplitude threshold method for seabed tracking;

[0088] Figure 10 This invention relates to the window slope method for seabed tracking;

[0089] Figure 11 This invention relates to the effect of the tow height on the beam angle.

[0090] Figure 12 This invention relates the beam angle to the height of the towed fish and the propagation distance.

[0091] Figure 13 This relates to the positional relationship between the echo, the seabed line, and the towed fish in this invention.

[0092] Figure 14 This is an example of the slant range correction for side-scan sonar waterfall images according to the present invention;

[0093] Figure 15This is a schematic diagram of the towed fish coordinate calculation of the present invention;

[0094] Figure 16 This is a schematic diagram of the echo position calculation of the present invention;

[0095] Figure 17 This is a schematic diagram of the waterfall image geocoding of the present invention;

[0096] Figure 18 This invention relates to the image resampling and scanning filling method;

[0097] Figure 19 This invention relates to an improved wavelet BP neural network;

[0098] Figure 20 This is a flowchart of the fused image coral reef substrate classification process of the present invention;

[0099] Figure 21 This is an overall flowchart of the coral reef substrate classification method based on the fusion of lidar and side-scan sonar data of the present invention. Detailed Implementation

[0100] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below with reference to specific illustrations.

[0101] First, this invention relates to the following technical terms:

[0102] Photon counting lidar

[0103] Photon-counting lidar, also known as single-photon lidar, is a type of lidar with high sensitivity and high temporal resolution. It uses a single-photon detector—a photodetector capable of detecting weak echo signals down to the level of a single photon—as the photoelectric conversion device. Combined with high-precision time-correlated single-photon counting (TCSPC) technology, it can achieve high-precision detection of weak signals, making it suitable for scenarios with limited echo intensity, such as long-distance targets and low-reflectivity targets.

[44] .

[0104] Inertial navigation system

[0105] An inertial navigation system (INS) is an autonomous navigation system that does not rely on external information and does not radiate energy to the outside.

[45] Its working environment includes not only the air and ground, but also underwater. The basic working principle of inertial navigation is based on Newton's laws of motion. By measuring the acceleration of the carrier in the inertial reference frame, integrating it over time, and transforming it into the navigation coordinate system, information such as velocity, yaw angle, and position in the navigation coordinate system can be obtained.

[0106] Ray tracing

[0107] Ray tracing is a method based on underwater light speed profiles for calculating the coordinates of underwater laser depth sounding points (projection points) within a lidar scanning reference coordinate system.

[46] Assuming that photons in each water column move at a constant speed, the incident angle and refraction angle of the light at the interface of each water column are calculated according to Snell's law. The travel time and horizontal displacement of the light in each water column are then calculated until the light disappears at some point on the interface or within the layer.

[0108] Side-scan sonar

[0109] A side-scan sonar system consists of several parts, including a workstation, display unit, winch, tow cable, towed fish, and GPS receiver. It also requires auxiliary equipment such as pressure sensors, compasses, and attitude sensors to provide relevant motion parameters. To accurately obtain the towed fish's real-time position, an underwater positioning system for the towed fish is also necessary.

[47] .

[0110] This invention relates to a coral reef substrate classification method based on the fusion of lidar and side-scan sonar data, which mainly includes the following steps, see below. Figure 21 :

[0111] (1) By analyzing the structure of the photon counting lidar scanning system, a laser scanning reference coordinate system and its transition coordinate system are established. Then, the geometric relationship between the incident angle and azimuth angle of the laser reflected light on the water surface and the normal vector of the reflector is established to obtain the three-dimensional coordinates of the laser incident point on the water surface. Furthermore, by constructing an underwater light speed profile and an underwater constant light speed ray tracking algorithm, a precise calculation model for the underwater depth sounding point coordinates of the UAV-borne lidar is constructed.

[0112] (2) Through fine processing steps such as seabed line tracking, radiation distortion joint correction, slant range correction, towed position estimation, and echo sampling point coordinate calculation, a side-scan sonar image with geocoding is constructed.

[0113] (3) Favorable Z The score normalizes the backscattering intensity of lidar and side-scan sonar and converts it into grayscale values. At the same time, it fuses the two types of data through the near-homologous point rule, and then constructs a fused image through geocoding and image resampling.

[0114] (4) Based on the standard wavelet BP neural network, the training rate and fitting accuracy of the wavelet BP neural network are improved by improving the network learning rate, momentum factor and initial value.

[0115] (5) Based on the high-precision and high-resolution fused image, various coral reef species samples are extracted from the corresponding positions of the fused image according to the geographic coordinates captured by the grab or underwater camera. Then, some statistical feature parameters with high self-cohesion and resolution are selected for sample training. The fused image is used to classify the coral reef substrate.

[0116] See Figures 1 to 20 As shown, the specific embodiments of the present invention will now be described in detail below:

[0117] (1) Overall technical solution

[0118] First, by analyzing the structure of the photon-counting lidar scanning system, the geometric relationship between the incident angle and azimuth angle of the reflected laser light on the water surface and the normal vector of the reflector was established. Furthermore, a precise calculation model for the underwater depth sounding point coordinates of the UAV-borne lidar was constructed using a constant-speed-of-light ray tracking algorithm within the underwater layer. Second, through fine processing steps such as seabed line tracking, combined radiation distortion correction, slant range correction, towed fish position estimation, and echo sampling point coordinate calculation, a geocoded side-scan sonar image was constructed. Third, the backscattering intensity of the lidar and side-scan sonar was fused using Z-scores. Finally, based on the standard wavelet BP neural network, the training rate and fitting accuracy of the wavelet BP neural network were improved by modifying the network learning rate, momentum factor, and initial values. Using the improved wavelet BP neural network and extracted seabed samples, the fused image was used to classify the coral reef seabed.

[0119] (2) Fine processing of marine lidar data

[0120] 1) Structure of the photon counting lidar elliptical scanning system

[0121] The marine lidar described in this article is a conventional elliptical scanning structure. Figure 1 As shown, a prism that can rotate around a rotation axis is used as a reflector to control the direction of the emitted laser beam. The emitted laser is reflected by the prism and points towards the sea surface. The angle between the normal direction of the prism and the rotation axis is 7.5°. When the prism surface rotates around the rotation axis, the laser traces a path on the sea surface with an incident angle of approximately 15°. Since the incident angle is not always equal to 15° during one scan (depending on the normal direction), the final laser point trajectory on the sea surface when the aircraft is hovering is approximately elliptical and oval. Therefore, this scanning structure is also called an oval scanning structure.

[0122] 2) LiDAR scanning reference coordinate system

[0123] Definition of the lidar scanning reference coordinate system: with the center point of the reflector as the origin. O , X s The axis points in the negative direction of the emitted laser.Y s The axis points in the direction of flight. Z s shaft and X s , Y s Construct a right-handed coordinate system with the axis pointing vertically upwards. The incident laser and the motor axis are in the same plane. X s Z s (surface), laser incident horizontally (along) X s (In the negative direction of the axis), the incident point of the laser beam on the mirror is the center of the mirror. For ease of understanding, as... Figure 2 As shown, the original X s Y s Z s Coordinate system around Y s A new coordinate system is obtained by rotating the axis counterclockwise by 45°. ,at this time The shaft and motor rotate in the same direction. The normal to the reflecting mirror is at... X s Z s Projection of the surface and Z s Angle between axes ,exist Y s Z s Projection of the surface and Z s Angle between axes .

[0124] like Figure 3 As shown, the reflected light is at X s Z s , Y s Z s Projection of a plane and Z The included angles of the axes are respectively x , y Its nadir is Because the laser travels along... X s Incident along the negative direction of the axis, the normal lies in Y s Zs Projection of the surface and Z s Angle between axes Equivalent to reflected laser at Y s Z s Projection of the surface and Z s Angle between axes y (Because the incident laser line, the mirror normal, and the reflected laser line are coplanar, and the incident laser line is perpendicular to the mirror normal) Y s Z s According to the theorem that if one plane passes through a perpendicular line to another plane, the two planes are orthogonal, therefore, in Y s Z s On a plane, the angle of rotation of the normal is synchronized with the angle of rotation of the reflected ray (i.e., the normal rotates). Angle, the reflected light also rotates. (Angle). And in X s Z s On a plane, when the mirror rotates (i.e., the normal is rotated)... When the angle is 2, the reflected ray rotates 2. Angle. When the angle of the normal changes, the included angle... And so it changed, from Easy to solve x And then calculate the nadir angle of the beam. and azimuth Therefore, the change in the angle of the normal is the key.

[0125] 3) Direction vector of the mirror normal

[0126] exist Figure 2 In the middle, the normal of the mirror is in Normal vector of the coordinate system ( ):

[0127] (1)

[0128] Then by going around Rotating the coordinate axes 45° clockwise will yield the result. X s Y s Z s The normal vector of the mirror in the coordinate system ( ):

[0129] (2)

[0130] 4) The relevant angles of the reflected light rays in the laser scanning reference coordinate system

[0131] Depend on Figure 2 , Figure 4 According to geometric relationships, Therefore:

[0132] (3)

[0133] Depend on Figure 2 According to geometric relationships, ,and ,so:

[0134] (4)

[0135] Depend on Figure 3 The geometric relationship yields the nadir angle. and azimuth for:

[0136] (5)

[0137] (6)

[0138] 5) Coordinates of water surface light spots and footprints in the lidar scanning reference coordinate system

[0139] like Figure 5 As shown, if the sea surface is a plane, the laser incident point on the sea surface is... P 1. The center of the laser is represented as S The slant distance of the laser beam in the air is L 1. Azimuth angle is The measured height of the center of the reflector is H Then the laser incident point on the sea surface P The position coordinates of 1 are:

[0140] (7)

[0141] (8)

[0142] (9)

[0143] (3) Calculation of the coordinates of underwater light spot footprints in the lidar scanning reference coordinate system

[0144] 1) Ray tracing algorithm based on the assumption of constant light speed within the water layer

[0145] Because the temperature, salinity, and density of each water mass differ along the vertical direction, the speed of light traveling within each mass also varies. Furthermore, light refracts at the interfaces between different water masses. Therefore, it is necessary to use an ocean light speed profiler to obtain the vertical sequence of water depth and light speed values, and then precisely track the light rays to obtain high-precision underwater light spot footprint coordinates.

[0146] Assuming the laser beam undergoes a process of... N A water column composed of layers, within which light travels at the constant speed of light. Figure 6 According to Snell's Law, we have:

[0147] (10)

[0148] like Figure 6 As shown, let the thickness of the water column be... Then the beam is in the layer i Horizontal displacement within y i and transmission time t i for:

[0149] (11)

[0150] (12)

[0151] According to equations (11) and (12), the horizontal distance and propagation time of the light beam through the entire water column are respectively:

[0152] (13)

[0153] (14)

[0154] Assuming the beam of light does not pass through the entire water column, but... Z r The beam disappears at this point, and at this time the horizontal displacement of the beam in that layer is... y r The vertical displacement is z r The time it takes for the light beam to travel through the entire water column is... t all The time spent on this layer is t r The optical path length experienced by the beam in that layer is... S r for:

[0155] (15)

[0156] (16)

[0157] (17)

[0158] Therefore, the total horizontal and vertical displacements of the light beam in the water are respectively:

[0159] (18)

[0160] (19)

[0161] 2) Coordinates of underwater laser footprints in the laser scanning reference coordinate system

[0162] like Figure 7 As shown, the laser has an incident angle of... On the water surface P The light is incident at point 1, and the angle of refraction is... 0, passing through different water layers in sequence P 2. P 3. Finally reach the point P Light energy disappeared in 4 places. The distance from the projection of the underwater lidar depth sounding point onto the water surface to the origin of the coordinate system. L s Then its value is:

[0163] (20)

[0164] Then the lidar underwater depth measurement point x w , y w The coordinate values ​​are as follows:

[0165] (twenty one)

[0166] (twenty two)

[0167] (twenty three)

[0168] 3) Coordinates of the underwater depth sounding points of the lidar in the WGS84 rectangular coordinate system

[0169] (twenty four)

[0170] and( X GPS , Y GPS , Z GPS )for:

[0171] (25)

[0172] In equations (24) and (25), ( () represents the coordinates of the underwater depth sounding point of the lidar in the WGS84 rectangular coordinate system; The coordinates of the center of the GPS antenna on the hovercraft in the WGS84 Cartesian coordinate system; This is the rotation matrix for transforming the body coordinate system to the local navigation coordinate system; It includes two parts: the eccentricity difference between the center of the laser scanning reference coordinate system and the center of the IMU body coordinate system, and the eccentricity difference between the center of the GPS antenna and the center of the IMU body coordinate system. The offset angle of the laser scanning reference coordinate system relative to the IMU body coordinate system.

[0173] (4) Fine processing of side-scan sonar data

[0174] 1) Submarine line tracking

[0175] When there are few suspended particles in the water, a clear boundary, known as the seabed line, can be observed between the water column and the seabed echo image in a waterfall image. This line is formed by the strong echo line of the first seabed echo in each row. The process of determining the position of the first seabed echo in each row is usually called seabed tracking. Figure 8 As shown. Generally, the seabed lines on the left and right sides are symmetrical about the transmission line. The first echo from the seabed always comes from directly below the towed fish. The lateral distance between the seabed line and the transmission line is the height of the towed fish above the seabed. Figure 8 middle, N 0 represents the position of the transmission line. N b To indicate the location of the seabed line, the horizontal dimension of a single pixel in the sonar image is [missing information]. The height of the trawler above the seabed can be expressed as:

[0176] (26)

[0177] The seabed line is the starting line for the lateral gain of side-scan sonar and also the baseline for slant range correction. Accurate extraction of the seabed line is fundamental to subsequent processing of waterfall images. Here, we summarize commonly used methods for seabed tracking, including amplitude thresholding, window slope methods, and manual intervention.

[0178] ① Amplitude threshold method

[0179] Before the first seabed echo returns, the hydrophone of the side-scan sonar can only detect weak noise signals. After the first seabed echo arrives, the signal received by the hydrophone will undergo a step change. Therefore, by setting an appropriate threshold T ( Figure 9 ), find the first intensity in the echo sequence according to the receiving time. The echo is considered to be the first echo on the seabed surface, and the distance from this echo to the towed fish is the height of the towed fish to the seabed.

[0180] This method is simple and fast, but when the threshold is not chosen properly, it is easily affected by strong echoes formed by objects in the water. Therefore, the threshold needs to be continuously adjusted according to the actual situation during data processing.

[0181] ① Window slope method

[0182] Generally, the seabed surface changes relatively gently, so it can be assumed that the height at which fish are dragged to the seabed is the same or similar within a few adjacent pings. This can be addressed by setting the window length. d The mean of each column of echoes within the statistical window b j The derivative of the mean curve is obtained using the finite difference method, and the calculation formula is shown in equation (27). The position with the largest slope on the curve is the position of the seabed line (e.g., Figure 10 ).

[0183] (27)

[0184] ③ Artificial intervention method

[0185] Generally, both of the above methods can correctly extract the seabed line from sonar images. However, the marine environment is complex. When there are a large number of suspended objects in the water, the strong echoes of the suspended objects and the seabed echoes are mixed together, making it difficult to distinguish the true location of the seabed line. In order to better eliminate the influence of water columns, the seabed line in complex areas can be manually selected based on the trend of the seabed line in front and behind.

[0186] In addition, window-value filtering can be used to eliminate the influence of random noise on the extracted seabed lines, and a symmetric method can be used to improve the stability of the extracted seabed lines.

[0187] 2) Combined correction of radiation distortion

[0188] Radiation distortion has three main components: 1) different artificial gain; 2) grayscale intensity variations in images at the seabed are related to the beam angle, which is influenced by the beam pattern, with significant side effects; 3) grayscale intensity variations at the far end of the image are related to distance, and are related to absorption and spread losses during propagation, as well as overgain. Radiation distortion correction is mainly aimed at obtaining high-quality images with uniform grayscale variations that ensure accurate detection and identification of targets.

[0189] ① Distance-related radiation distortion correction

[0190] Taking the starboard side as an example, suppose at a certain Ping ( n The data sampling sequence in ) is N sAnd the grayscale value sequence of this sequence is Assume that the position of the seabed line obtained by seabed tracking is... b ( n ), N min The width of the image area corresponding to the maximum height of the dragged fish within the strip can be calculated using equation (28):

[0191] (28)

[0192] The correction factor for each Ping section position is calculated according to equation (29). δ :

[0193] (29)

[0194] in N P The number of Ping smoothing operations along the flight path. N P It should satisfy the requirement of eliminating the effects of local changes while ensuring that it does not cross two different seabed sediment zones. This is obtained through equation (29). δ ( i The sequence is not smooth enough; therefore, the following moving average method is used to smooth it. δ ( i ):

[0195] (30)

[0196] In equation (30), p ( i )=max(0, i - l ); q ( i )=min( N min -1, i + l ); l = N min / 50. Calculate the correction factor according to equation (31):

[0197] (31)

[0198] In summary, the correction formula for Ping cross-sectional echo data can be obtained as follows:

[0199] (32)

[0200] The aforementioned correction effect is closely related to the accuracy of seabed tracking. This method eliminates strong grayscale variations at the far end of the image, effectively improving the visual effect of the image, and ensuring clearer edges between image categories after processing.

[0201] ① Beam mode-related radiation distortion correction

[0202] The variation in gray intensity at the seabed location in side-scan sonar images is related to the beam angle, which can be further divided into two aspects: the influence of the towed fish height and the beam angle. Figure 11 The ratio of the transducer's height above the seabed to the beam's position on the seabed is shown. This ratio determines the beam pattern of each Ping echo and the position of each echo within the Ping scan line.

[0203] Figure 12 The relationship between beam angle and propagation distance and towed fish height is shown. The beam angle changes rapidly over a distance directly below the towed fish on the seabed. Once the propagation distance exceeds a certain multiple of the towed fish height, the influence of the propagation distance on the beam angle gradually weakens; moreover, the influence of the beam angle increases with the towed fish height.

[0204] Taking the seabed point and a point at a short distance from the seabed point as reference points, the echo intensity within the cross section is recalculated using equation (33). In order to prevent the influence of seabed changes, several Pings along the track direction need to be taken for equalization processing.

[0205] (33)

[0206] The distance from the reference point to the seabed point can be taken as 1 / 10 of the working range; L This is the ping number; + and - represent the port and starboard sides, respectively. b ( n ) represents the location of the seabed point; the Ping cross-section grayscale is corrected as shown in equation (34): (34)

[0207] Then, the correction coefficients at each point on the Ping section can be calculated using equation (35):

[0208] (35)

[0209] Finally, corrections for strong echoes near the seabed directly below the towed fish in the side-scan sonar image are performed based on equation (36):

[0210] (36)

[0211] The combined artificial gain elimination algorithm, the separation distance correlation and beam pattern correction algorithm are the joint correction methods for radiation distortion.

[0212] 3) Slope distance correction

[0213] Due to the influence of slant range recording, side-scan sonar images exhibit lateral tilt geometric distortion. The presence of water columns also causes the target directly below the towed fish to be separated into two sides. Therefore, after lateral equalization, slant range correction is still required for the side-scan sonar images. Constrained by the measurement mechanism, the side-scan sonar transducer cannot distinguish the direction of each echo, but can only obtain the slant range from the towed fish to each echo. Under the existing conditions, slant range correction can only be performed by introducing the following assumptions:

[0214] (1) The first seabed echo came from directly below the transducer;

[0215] (2) The seabed surface is gentle and approximately flat, and the vertical distance from the target to the towed fish is equal to the height of the towed fish to the seabed;

[0216] (3) Ignore the change in sound speed and assume that sound waves propagate in a straight line in seawater.

[0217] Based on the above assumptions, the horizontal distance of each echo can be calculated by utilizing the triangular relationship between the towed fish, the seabed, and the echo. Figure 13 For a ping echo in the waterfall image, N The total number of echo samples per side. n h The slant distance pixel width from the first echo from the seabed to the transmission line. n i The slant distance pixel width from the current echo to the transmission line is: The horizontal distance pixel width from this echo to the first echo on the seabed is: (37)

[0218] Side-scan sonar images after slant range correction, such as Figure 14 As shown, the image removes the water column area and reduces the influence of tilted geometric distortion on the target shape.

[0219] 4) Calculation of the position of the dragfish

[0220] The towed position is generally calculated approximated by the length of the towline. Towlines typically possess good strength and flexibility, and during towing, they significantly reduce the impact of changes in the ship's attitude on the towed fish. Therefore, when the ship is traveling at a constant speed in a straight line, the towed fish can be considered to be only subject to forward traction, and the horizontal distance from the towed fish to the towing point is calculated as follows: Figure 15 Derivation of geometric relationships in a middle triangle.

[0221] (38)

[0222] Due to the weight of the tow cable, the hypotenuse of the triangle is often taken as the length of the tow cable. L The horizontal distance from the towing point to the towing point is calculated according to formula (38), which is 0.9 times the towing point. hThe height of the winch above the water surface. f d The distance from the towed fish to the water surface can be obtained by a pressure sensor. Finally, based on the alignment of the towed fish with the ship's hull, the geographical coordinates of the towed fish are determined by the towing point (Equation (39)).

[0223] (39)

[0224] In the above formula, These are the coordinates of the towing point in the ship's coordinate system. These are the coordinates of the drag point in the geographic coordinate system. A To measure the ship's azimuth.

[0225] 5) Calculation of echo sampling point coordinates

[0226] After processing, the center pixel of each row of the waterfall image theoretically corresponds to the position directly below the towed fish. Based on the position of the towed fish, the geographic coordinates of the center pixel of each row can be determined. Each row of pixels is perpendicular to the current course of the towed fish, and the width from each pixel to the center is a horizontal distance. Based on these relationships, the position of each echo in geographic coordinates can be calculated.

[0227] like Figure 16 As shown, in a Cartesian coordinate system, the geographic coordinates of the projection point directly below the towed fish during the Ping echo measurement are: P 0 ( X 0, Y 0), the single-side scan amplitude of the side-scan sonar is R The sampling rate for each channel is N The sailing azimuth is α Since each ping echo is perpendicular to the direction of navigation, the azimuth angle of the port channel echo is... The azimuth angle of the echo from the starboard channel is... , P i For a certain channel's first i One echo, then P i geographic coordinates ( X i , Y i )for:

[0228] (40)

[0229] If we want to consider the effect of movement posture on the towing of the fish, the above needs to be corrected accordingly:

[0230] (41)

[0231] (5) Data fusion processing of lidar and side-scan sonar

[0232] 1) Beneficial Z Normalize the scores

[0233] Because the backscattering intensity of lidar and side-scan sonar is affected by their respective emission mechanisms, the ranges of their obtained backscattering intensity are different. Therefore, it is possible to... Z The fractions are normalized to their backscatter intensity before being converted to grayscale values.

[0234] Z Fractions are a common mathematical and statistical method used to generalize parameters from multiple different ranges of variation to a common range of variation. For any set of scattering intensity sequences... B (mean is) The standard deviation is ),but B Any scattering intensity value b of Z The score is: (42)

[0235] 2) Conversion between Z-score and grayscale value

[0236] Before mapping the seabed topography, it is necessary to... Z The formula for converting fractions to grayscale values ​​is: (43)

[0237] in Z For the backscattering intensity Z Fraction, Z min The smallest score in the sequence. Z max This is the largest score in the sequence. I This is the grayscale value after linear quantization.

[0238] 3) Merging of lidar and side-scan sonar point clouds near the same point

[0239] After obtaining the data of lidar sounding points and multibeam sonar sounding points, since their coordinates have been returned to the WGS84 spatial rectangular coordinate system, their sounding points form a full-coverage strip measurement. When two sounding points of different types meet the conditions of equation (44), it is considered that these two points are approximately the same point. Then the coordinates of the sounding points need to be reassigned, as shown in equation (45). Similarly, the gray value also needs to be reassigned, as shown in equation (46).

[0240] (44)

[0241] (45)

[0242] (46)

[0243] 4) Fusion image geocoding

[0244] After linearly quantizing the scattering intensity data, it is necessary to calculate the specific pixel location of each sampling point in the image, assuming a pixel resolution of . res The formula for calculating the position of the sampling point in the image is: (47)

[0245] in( X i , Y i ) are respectively the first i The pixel position of sampling point ( ) in the image. x i , y i ) are respectively the first i The geographical coordinates of sampling point No. 1, ( x min , y min Then, it represents the minimum value of the geographic coordinates of the overall sampling points.

[0246] 5) Image resampling

[0247] Image resampling primarily addresses the gap problem caused by the inconsistency between the longitudinal and transverse sampling rates of side-scan sonar echoes (e.g., Figure 17 To address the imaging characteristics of lidar and multibeam sonar images, a scan-fill-based image resampling method is presented here.

[0248] The basic principle of the scan-fill method is that for any closed region, each row of pixels in the region is scanned sequentially from top to bottom using horizontal scan lines. A series of intersection points generated by each scan line and the boundary are calculated. These intersection points are sorted according to the horizontal axis. The sorted intersection points are then taken out in pairs as the left and right boundary points. All pixels within the left and right boundary points are marked as fill points. When the entire region has been scanned, the region filling is completed.

[0249] like Figure 18 As shown, A 1. A 2 represents two adjacent echoes on the same scan line. B 3. B 4 represents two adjacent points on a nearby scan line. A 1. A 2. B 3. B4 form a closed connected region. Using the scan-fill method, all pixels inside the region can be marked. The pixel value of each pixel can be obtained by the inverse distance weighting method, as shown in the formula.

[0250] (48)

[0251] (6) Improved wavelet BP neural network

[0252] The standard wavelet backpropagation (BP) neural network is based on the standard BP neural network and consists of an input layer, hidden layers, and an output layer. Unlike the standard BP neural network, the hidden layers use wavelets or scaling functions instead of the sigmoid function. Typically, the activation function used in the hidden layers is the Morlet Gaussian function. The three-layer BP wavelet neural network structure...

[48] like Figure 19 As shown.

[0253] Wavelet transform uses a series of vibrational functions with different frequencies as window functions to analyze signals. The basic principle of wavelet transform is to fit the signal by constructing a series of special wavelet basis functions.

[49] Wavelet functions possess excellent local time-frequency analysis capabilities and can also perform multi-resolution analysis of non-stationary signals through scaling and translation operations. If the mother wavelet function is defined as... It represents the square of the set of real numbers. R Let represent the set of real numbers. Then the subwavelet function...

[50] Defined as the following formula: (49)

[0254] The activation functions used in the hidden layer and the output layer are the Morlet Gaussian function and the Sigmoid function, respectively, as shown in equations (50) and (51).

[0255] (50)

[0256] (51)

[0257] BP wavelet neural network model

[50] It can be represented as: (52)

[0258] In equation (52), For the input layer p The first sample i The input value of each node, w kj and w ji The sum represents the connection weights between the output layer and the hidden layer, and between the hidden layer and the input layer. b 1 j For the hidden layerj The threshold of each node, b 2 k For the output layer k The threshold of each node, a j and b j Hidden layers are respectively j The scaling factor and translation factor of each node.

[0259] The objective function for network training is the energy function.

[51] As shown in the following formula:

[0260] (53)

[0261] In equation (53), It is the output layer. p The first mode k One expected output value, It is the output layer. p The first mode k The actual network output value.

[0262] In standard wavelet backpropagation (BP) neural networks, a fixed learning rate and randomized initial values ​​for connection weights and thresholds can lead to slow convergence and jitter at steep error surfaces, requiring improvement. The learning rate setting is particularly crucial for the entire network's learning process. An excessively large learning rate can cause network oscillations and instability; a too small learning rate results in slow convergence and prolonged training time. Therefore, during network training, the learning rate needs to be adaptively adjusted to better correct the network's error performance surface. Based on the local error surface, the learning rate can be adjusted accordingly. When the network error gradually approaches the convergence accuracy by decreasing, it indicates the correction direction is correct, and the step size needs to be increased, with the learning rate correspondingly increasing; conversely, the step size needs to be decreased, and the learning rate correspondingly decreasing. Generally, the learning rate increases or decreases at a fixed rate, i.e., as long as... E ( k )Compare E ( k-1 The method of increasing (or decreasing) the learning rate based on its size (or size) without comparing the degree of difference between them is flawed. A fixed increase or decrease in the learning rate will affect network stability, thus slowing down network convergence. Therefore, it is necessary to construct a [structure / mechanism] that... E ( k- 1) / E ( k The learning rate function changes accordingly, and follows E ( k- 1) / E ( kThe learning rate is updated continuously according to the changes in the number of students, using the following formula:

[0263] (54)

[0264] in, ， , , And momentum factor The update formula is:

[0265] (55)

[0266] The convergence and convergence speed of a network learning process are closely related to the optimal initialization of wavelet BP neural network parameters. Generally, using random numbers to obtain excellent initial weights is not guaranteed. Obtaining optimal initial weights requires considering the relationship between the initial weights, the neuron transfer function, and the learning samples. Let the number of hidden nodes in a three-layer wavelet BP neural network be... J There are , and the number of input layer nodes is I One, first... w ji Initialization involves the following steps:

[0267] ① First, generate a uniformly distributed random number in the interval [-1, 1], and use this as... w ji The initial value, and used w ji0 express;

[0268] ②Then w ji0 Normalize by row: (56)

[0269] ③ Then multiply by a factor C This factor is related to the transfer function and the number of input layer neurons. I Number of hidden layer neural units J Related, C Option 2 is acceptable;

[0270] (57)

[0271] ④ Finally, consider the relationship with the learning samples. Let the first sample in the input layer be... i The maximum and minimum values ​​of the input samples for each neuron are respectively x imax , x imin Then we have:

[0272] (58)

[0273] After initializing the weights between the hidden layer and the input layer, initialize the threshold between the hidden layer and the input layer. b 1 j The process involves the following steps:

[0274] ① First, generate a uniformly distributed random number in the interval [-1, 1], and use this as... b 1 j The initial value, using b 1 j0 express;

[0275] ② Multiply by another factor C This factor is related to the transfer function and the number of input layer neurons. I Number of hidden layer neural units J Related, C The value of is the same as that in equation (57);

[0276] (59)

[0277] ① Finally, the learning samples and w ji Then we have: (60)

[0278] Scaling factor for wavelet functions a j With translation factor b j Initial value optimization, let the time domain center and radius of the mother wavelet be respectively... θ , According to wavelet theory, the concentrated region of the wavelet scaling system in the time domain is... To ensure that the wavelet scaling system covers the entire input vector, the scaling and translation parameters must be initialized according to the following formula:

[0279] (61)

[0280] From equation (61), we can obtain:

[0281] (62)

[0282] For the initialization of connection weights and thresholds between the hidden layer and the output layer, since the output layer generally uses linear neurons, the initial values ​​of the connection weights and thresholds between the two layers can be uniformly distributed random numbers generated in the interval [-1,1].

[0283] In a standard wavelet BP neural network, the determination of the number of hidden layer neurons is highly arbitrary. The choice of the number of hidden layer neurons is closely related to the network's convergence speed and the improvement of error accuracy. If the number of hidden layer neurons is too small, the network's learning capacity is limited, and it cannot store all the patterns contained in the training samples. If the number of hidden layer neurons is too large, it will increase the network's learning time cost and may even store non-regular content from the samples, thus reducing generalization ability. Therefore, it is necessary to discuss the setting of the number of hidden layer neurons. The value of the number of hidden layer neurons can be referenced in the following formula.

[52] :

[0284] (63)

[0285] In equation (63), J Indicates the number of hidden layer neural units. n Indicates the number of neural units in the input layer. m Indicates the number of neurons in the output layer. It represents a constant between 1 and 10.

[0286] (7) Fusion image background classification

[0287] Coral reef substrate classification is based on fused images from lidar and side-scan sonar. Therefore, the first step is to calculate the grayscale value and geographic coordinates of each pixel within the survey area. Then, after a series of preprocessing steps including noise reduction, resampling, and grayscale equalization, a high-precision, high-resolution fused image is generated. Next, based on the geographic coordinates captured by the grab or underwater camera, samples of various coral reef species are extracted from the corresponding locations in the fused image. Sample training is then performed using statistical feature parameters with high self-cohesion and resolution. The training results are then used to classify the coral reef substrate from the fused image. The specific classification process is detailed in [link to documentation]. Figure 20 .

[0288] The substrate classifier uses an improved three-layer wavelet BP neural network, which includes an input layer, a hidden layer, and an output layer. The number of units in the input layer is determined based on the number of samples, and the number of neural units in the hidden layer is determined according to the empirical formula (63). 70% of the samples are used for training, and 30% are used for checking. The results of the training samples are used as the internal compliance evaluation, and the results of the checking samples are used as the external compliance evaluation.

[0289] The training accuracy of coral reef substrate samples can be evaluated using a confusion matrix derived from remote sensing image classification accuracy. Classification accuracy refers to the proportion of samples that can be identified. The confusion matrix is ​​defined as follows: (64)

[0290] In equation (64), m ij For the first i The sample type of the class was classified intoj The total number of samples in the class, n The confusion matrix represents the number of categories. Each column represents field reference verification information, and the value of each column equals the number of field sample types corresponding to the corresponding type in the sample training results. Each row of the confusion matrix represents the classification information of the sample type, and the value of each row equals the number of sample types in the corresponding type of real samples. A larger value on the main diagonal of the confusion matrix indicates higher classification accuracy, and vice versa.

[0291] Therefore, the present invention solves the following technical problems:

[0292] (1) The structure of the photon counting lidar scanning system was analyzed, and a geometric relationship model between the laser reflected ray and the normal vector of the mirror was established;

[0293] (2) A coordinate calculation model of the laser radar incident point on the water surface in the laser scanning reference coordinate system was constructed;

[0294] (3) A constant light speed ray tracing model for underwater layers was proposed;

[0295] (4) A coordinate calculation model for underwater lidar footprints in the laser scanning reference coordinate system was established;

[0296] (5) Reset the coordinates of the underwater sounding points to the WGS84 spatial rectangular coordinate system.

[0297] In addition, the technical features of the present invention are as follows:

[0298] (1) The structure of the single-photon lidar scanning system was analyzed, and a geometric relationship model between the incident angle and azimuth angle of the laser reflected light and the normal vector of the reflector was established;

[0299] (2) A constant light speed ray tracking model for underwater layers was proposed, which accurately tracked underwater laser footprints and improved the coordinate accuracy of underwater laser depth sounding points.

[0300] Based on the above description, the beneficial effects of the present invention are:

[0301] (1) By studying the scanning structure of photon counting lidar, a geometric relationship model between the laser reflected ray and the normal vector of the mirror was constructed, which can accurately calculate the coordinates of the laser incident point on the water surface;

[0302] (2) A constant light speed ray tracing model based on the light speed profile is proposed, and a coordinate calculation model of the underwater laser sounding point in the laser scanning reference coordinate system is established, which can greatly improve the measurement accuracy of the underwater sounding point.

[0303] (3) Favorable Z The score is used to fuse data from lidar and side-scan sonar using different mechanisms;

[0304] (4) Improve the accuracy of coral reef substrate classification by improving the traditional wavelet BP neural network.

Claims

1. A method for classifying coral reef substrate based on the fusion of lidar and side-scan sonar data, characterized in that, Includes the following steps: (1) By analyzing the structure of the photon counting lidar scanning system, a laser scanning reference coordinate system and its transition coordinate system are established. Then, the geometric relationship between the incident angle and azimuth angle of the laser reflected light on the water surface and the normal vector of the reflector is established to obtain the three-dimensional coordinates of the laser incident point on the water surface. Furthermore, by constructing an underwater light speed profile and an underwater constant light speed ray tracking algorithm, a precise calculation model for the underwater depth sounding point coordinates of the UAV-borne lidar is constructed. (2) Through fine processing steps such as seabed line tracking, radiation distortion joint correction, slant range correction, towed position estimation, and echo sampling point coordinate calculation, a side-scan sonar image with geocoding is constructed. (3) The backscattering intensity of lidar and side-scan sonar is normalized and converted into gray value using Z-score. At the same time, the two types of data are fused by the near-homologous point rule. Then, the fused image is constructed by geocoding and image resampling. (4) Based on the standard wavelet BP neural network, the training rate and fitting accuracy of the wavelet BP neural network are improved by improving the network learning rate, momentum factor and initial value. (5) Based on the high-precision and high-resolution fused image, various coral reef species samples are extracted from the corresponding positions of the fused image according to the geographic coordinates captured by the grab or underwater camera. Then, some statistical feature parameters with high self-cohesion and resolution are selected for sample training. The fused image is used to classify the coral reef substrate.

2. The coral reef substrate classification method based on lidar and side-scan sonar data fusion as described in claim 1, characterized in that, Step (4) includes the following steps: 1) Establish the basic architecture of the wavelet BP neural network, namely, one input layer, one hidden layer and one output layer, with the hidden layer using a wavelet function as the transfer function; 2) Improvements in network learning rate; 3) Improvement of momentum factor; 4) Considering the relationship between the initial values ​​of network weights, neuron transfer functions, and learning samples, construct an optimal initial weight model.

3. The coral reef substrate classification method based on lidar and side-scan sonar data fusion as described in claim 1, characterized in that, Step (5) includes the following steps: 1) Based on the fused images, extract image samples of various types of coral reefs according to the actual shooting coordinates of the grab or camera; 2) Select some statistical feature parameters with high self-cohesion and resolution as image sample features; 3) Select 70% and 30% of the samples respectively as training and verification, and perform image sample training and verification; 4) Use the sample training results to classify the coral reef substrate in the fused images.

Images