Deep-sea mining sediment plume turbidity prediction method based on digital image processing

Abstract

The invention provides a deep-sea mining sediment plume turbidity prediction method based on digital image processing, and establishes a set of deep-sea mining sediment plume turbidity prediction method based on the digital image processing technology by utilizing the digital image processing technology based on different reflectivity principles of seawater and deep-sea mining sediment plume to light. According to the algorithm, deep-sea mining sediment plume video images are subjected to a series of processing through program calculation, such as deep-sea mining sediment plume image graying, image enhancement and image threshold segmentation, and finally deep-sea mining sediment plume image feature values are extracted. The feature value of the deep-sea mining sediment plume image can be calibrated with in-situ observation data to realize inversion of the real turbidity of the deep-sea mining sediment plume.

Description

Technical Field

[0001] The invention relates to the field of marine engineering geology technology, and in particular to a deep-sea mining sediment plume turbidity prediction method based on digital image processing. Background Art

[0002] In recent years, with the popularization of high-definition camera systems and seabed observation networks carried by ROVs, environmental monitoring technology based on computer vision has begun to be applied in the deep sea. Existing studies mostly use traditional methods such as image grayscale histogram statistics or RGB color space conversion to estimate by establishing an empirical relationship between pixel intensity and turbidity. However, the complex optical environment of the deep sea causes problems such as color cast attenuation (typical attenuation rate: about 0.2-0.5dB per meter attenuation in the red light band) and interference from suspended particle scattering, making it difficult to accurately model the nonlinear relationship between traditional image features and turbidity parameters. For example, Kawamura et al. (2021) showed that when the suspended matter concentration exceeds 200 mg / L, the prediction error of the traditional HSV color model can reach ±40NTU. Therefore, it is urgent to develop new visual feature extraction methods and adaptive prediction models. Summary of the invention

[0003] In order to make up for the shortcomings of the prior art, the present invention provides a deep-sea mining sediment plume turbidity prediction method based on digital image processing. Through the deep-sea mining high-definition video image data obtained by long-term in-situ observation in the deep sea, the target deep-sea mining sediment plume can be observed intuitively. However, the turbidity of the deep-sea mining sediment plume cannot be quantitatively obtained by the naked eye alone. Digital image processing technology can provide a solution for the post-processing of the collected underwater video images.

[0004] The present invention is implemented by the following technical solution: a method for predicting turbidity of deep-sea mining sediment plume based on digital image processing, specifically comprising the following steps: Step S1: Feature extraction: multimodal feature fusion First, the size of the deep-sea mining video image is 1080*1920. The high-resolution deep-sea mining video image is processed and the video is framed. Then, the image is processed every 90 frames, including: calculating the mean and standard deviation of the three channels of the RGB image, as well as the mean and variance after grayscale conversion, and calculating the entropy of the image and the contrast of the image. Finally, the video frame image is converted to the HSV space, and the mean values ​​of the H, S, and V channels of the image in the HSV space are calculated. Finally, the Gaussian mixture model is used to detect the image of the sediment plume and calculate the plume area ratio, with a total of 14 features. The RBR multi-parameter water quality meter observes the turbidity of the water body. After calculating the eigenvalues ​​of the image, it is necessary to determine the observation area of ​​the turbidity probe of the RBR multi-parameter water quality meter for in-situ observation, that is, area I. Then, according to the coordinates of area I on the image: 100, 100, a square with a side length of 200, where the image is 1920 long and 1080 wide, with the upper left corner as the origin and rightward and downward as positive; extract the 14 features corresponding to the eigenvalues ​​of the area; Step S2: Model construction: Adaptive deep learning framework Step S2-1, data preprocessing: In order to increase the generalization ability of the model and improve the richness of the data, The principle and function of using data enhancement are as follows: improving the model's adaptability to real scenarios through diversified training data; adding Gaussian noise (σ=0.05) to the original sonar signal to simulate the interference of deep-sea suspended particles; The principle and function of using regularization to regularize data are: to prevent overfitting by constraining model complexity, add Dropout ratio = 0.3 after the convolution layer, and apply L2 regularization λ = 0.001; The principle and function of using expand dimension to increase data dimension are: enhancing the feature expression ability by reconstructing the input data, segmenting the one-dimensional signal into multiple time windows, forming a "time-segment" two-dimensional input, enhancing the ability to extract temporal features, and enhancing the model's ability to capture local features related to turbidity by multi-dimensionally deconstructing and reorganizing the spatial features of image data; Step S2-2, model structure selection: Based on the grid search method, the optimal learning rate of 0.001, batch size of 32 and number of iterations of 200 are determined, and an early stopping mechanism is introduced: patience value = 10 rounds to prevent training redundancy; a 6-layer convolutional neural network is used, and each layer of the convolutional neural network is followed by batch normalization (Batch Normalization), L2 regularization λ = 0.001 and Dropout ratio = 0.3 are used to suppress overfitting; and a maximum pooling layer Maxpooling1D (2), the activation function is ReLU, and the output layer is a linear regression unit; Step S2-3: Model performance evaluation The performance of the prediction model is evaluated using three indicators: determination coefficient R2, mean square error MSE and mean absolute error MAE; the calculation formula is as follows:

[0005]

[0006]

[0007] Refers to the actual turbidity, the actual measured turbidity of the i-th water sample; Refers to the model predicted turbidity, which is the turbidity predicted by the model based on the image features of the i-th water sample; Refers to the average value of the true turbidity.

[0008] As a preferred solution, in step S1, the frame rate of the deep-sea mining video is 30, and the observation interval of the RBR turbidity meter is 3s.

[0009] As a preferred solution, the calculation method of the 14 features in step S1 is: Step S1-1, calculate the mean values ​​of the three channels of the RGB image: R_mean, G_mean, B_mean, and the standard deviations: R_std, G_std, B_std; The formula for calculating the average value of the three channels of the RGB image is as follows:

[0010]

[0011]

[0012] The standard deviation calculation formula for the RGB three channels is as follows:

[0013]

[0014]

[0015] Step S1-2, Gray mean and Gray variance after grayscale conversion: To calculate the grayscale mean and variance, first convert the RGB image into a grayscale image, and then calculate the grayscale mean and variance: First, convert the RGB image to a grayscale image. The calculation formula is as follows:

[0016] The grayscale mean calculation formula is as follows:

[0017] Grayscale variance reflects the degree of dispersion of the grayscale value of each pixel and the average grayscale value. It is the main metric for measuring the amount of information in an image. The larger the variance, the greater the amount of information in the image. The variance calculation formula is as follows:

[0018] Step S1-3, the average values ​​of the H hue, S saturation, and V brightness channels: H_mean, S_mean, V_mean, calculated as follows:

[0019] Step S1-4, contrast: calculated based on the gray level co-occurrence matrix GLCM to characterize the local gray level difference; Luminance, L:

[0020] Using the standard deviation method RMS Contrast, the contrast is measured by the standard deviation of brightness:

[0021] Step S1-5, image entropy: reflects information complexity, and the calculation formula is

[0022] Where p(x) is the pixel intensity distribution probability; Step S1-6: Use Gaussian mixture model GMM to segment the plume area and calculate the plume area ratio: Ratio = number of plume pixels / total number of pixels.

[0023] Due to the adoption of the above technical solution, the present invention has the following beneficial effects compared with the prior art: The present invention is based on the principle of different reflectivity of seawater and deep-sea mining sediment plumes to light, and utilizes digital image processing technology to establish a deep-sea mining sediment plume turbidity prediction method based on digital image processing technology. The algorithm performs a series of processing on deep-sea mining sediment plume video images through program calculation, such as: deep-sea mining sediment plume image grayscale, image enhancement, and image threshold segmentation, and finally extracts the deep-sea mining sediment plume image characteristic value. The deep-sea mining sediment plume image characteristic value can be calibrated with in-situ observation data to realize the inversion of the real turbidity of the deep-sea mining sediment plume.

[0024] By comparing the fitting effects of different models on deep-sea mining sediment plume images, the algorithm was optimized, and finally the best image prediction method for deep-sea mining sediment plume turbidity was established. It can be applied to the on-site prediction of deep-sea mining sediment plume turbidity. This method attempts to apply digital image processing technology to the measurement of deep-sea mining sediment plume turbidity. Its measurement accuracy only depends on the shooting accuracy of the measuring device and the improvement of the subsequent algorithm. Compared with the traditional deep-sea mining sediment plume turbidity measurement method, it has the advantages of intuitive, continuous observation and simple operation. Therefore, it has broad application prospects in marine surveys.

[0025] The present invention applies digital image processing technology to the measurement of deep-sea mining plume turbidity. The measurement accuracy of this method only depends on the shooting accuracy of the measuring device and the improvement of the subsequent algorithm. It overcomes the principle defects of traditional deep-sea mining plume turbidity measurement technology, such as being easily affected by the particle size of suspended sediments in the ocean and the particle size distribution and flocculation of deep-sea mining plumes. It also has the advantages of intuitiveness, continuous observation and simple operation. While ensuring the measurement accuracy, the present invention significantly reduces the cost and operation complexity of deep-sea turbidity monitoring, and provides innovative technical means for environmental protection of deep-sea resource development. Therefore, it has broad application prospects in marine surveys. This method promotes digital image processing technology to the field of deep-sea mining plume turbidity, which also reflects the intersection of multiple disciplines and has good research reference significance.

[0026] 1. Multimodal feature fusion: Combining RGB, HSV, texture and statistical features to comprehensively characterize the optical properties of the plume and break through the limitations of a single color space.

[0027] 2. Adaptive deep learning framework: Through the nonlinear modeling capability of CNN, it solves the problem of feature-turbidity nonlinear mapping in complex deep-sea optical environments.

[0028] 3. Low-cost, high-resolution monitoring: Relying only on video images and algorithm optimization, it replaces traditional high-cost sensor arrays, and the monitoring range can be extended to kilometer-level working surfaces.

[0029] 4. Real-time and continuity: Supports online video stream processing, achieves turbidity update in seconds, and meets the dynamic environmental monitoring needs of deep-sea mining.

[0030] Additional aspects and advantages of the present invention will become apparent from the following description or may be learned by practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] The above and / or additional aspects and advantages of the present invention will become apparent and easily understood from the description of the embodiments in conjunction with the following drawings, in which: Figure 1 This is a schematic diagram of the results of Gaussian mixture model (GMM) segmentation of plume areas; Figure 2 Build a schematic for the model; Figure 3 Schematic diagram of the convolutional neural network structure; Figure 4 The figure shows the performance evaluation results (coefficient of determination R2, mean square error MSE and mean absolute error MAE). DETAILED DESCRIPTION

[0032] In order to more clearly understand the above-mentioned purpose, features and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that the embodiments of the present application and the features in the embodiments can be combined with each other without conflict.

[0033] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the protection scope of the present invention is not limited to the specific embodiments disclosed below.

[0034] Combine the following Figures 1 to 3 The deep-sea mining sediment plume turbidity prediction method based on digital image processing according to an embodiment of the present invention is specifically described.

[0035] The present invention proposes a deep-sea mining sediment plume turbidity prediction method based on digital image processing, which specifically includes the following steps: Step S1: Feature extraction: multimodal feature fusion First, the size of the deep-sea mining video image is 1080*1920. Based on digital image processing technology, the present invention processes the acquired high-resolution deep-sea mining video image, performs frame processing on the video, and then performs image processing every 90 frames (the frame rate of the in-situ observation video is 30, and the observation interval of the RBR turbidity meter is 3s), including: calculating the mean and standard deviation of the three channels of the RGB image, as well as the mean and variance after grayscale, and calculating the entropy of the image, and the contrast of the image, and finally converting the video frame image into the HSV space, calculating the average values ​​of the H, S, and V channels of the image in the HSV space, and finally using the Gaussian mixture model to perform image detection of sediment plumes and calculate the plume area ratio, a total of 14 features; The calculation method of the 14 features is Step S1-1, calculate the mean values ​​of the three channels of the RGB image: R_mean, G_mean, B_mean, and the standard deviations: R_std, G_std, B_std; Calculate the three-channel mean and standard deviation of the RGB image of the deep-sea mining plume recorded by a high-definition camera at a depth of 5700 meters in the sea.

[0036] The formula for calculating the average value of the three channels of the RGB image is as follows:

[0037] The standard deviation calculation formula of the RGB three channels is as follows: Standard deviation is a statistic that measures the discrete degree of pixel value distribution in an image. It describes the variation of pixel values ​​in an image and can be used to evaluate the contrast and detail richness of an image. The larger the standard deviation, the wider the distribution of pixel values ​​in the image and the stronger the contrast.

[0038]

[0039] Step S1-2, Gray mean and variance after grayscale conversion: To calculate the gray mean and variance, first convert the RGB image to a grayscale image, and then calculate the gray mean and variance: First, convert the RGB image to a grayscale image. The calculation formula is as follows:

[0040] The grayscale mean calculation formula is as follows:

[0041] Grayscale variance reflects the degree of dispersion of the grayscale value of each pixel and the average grayscale value. It is the main metric for measuring the amount of information in an image. The larger the variance, the greater the amount of information in the image. The variance calculation formula is as follows:

[0042] Step S1-3, the average values ​​of H (hue), S (saturation), and V (brightness) channels: H_mean, S_mean, V_mean, calculated as follows:

[0043] Step S1-4, Contrast: calculated based on the gray level co-occurrence matrix (GLCM) to characterize the local gray level difference; Grayscale brightness (Luminance, L)

[0044] Using the standard deviation method (RMS Contrast), contrast can be measured by the standard deviation of brightness:

[0045] Step S1-5, Entropy of the image: reflects the complexity of information, and the calculation formula is

[0046] Where p(x) is the pixel intensity distribution probability; Step S1-6: Use Gaussian mixture model (GMM) to segment the plume area and calculate the plume area ratio: Ratio = number of plume pixels / total number of pixels.

[0047] When conducting in-situ observations of deep-sea mining, there are not only high-resolution seabed video images, but also RBR multi-parameter water quality meter observations (which can observe the turbidity of the water body). Therefore, after calculating the eigenvalues ​​of the image, it is necessary to determine the observation area of ​​the turbidity probe of the RBR multi-parameter water quality meter for in-situ observation, that is, area I, and then according to the coordinates of area I on the image: 100, 100, a square with a side length of 200, where the image is 1920 long and 1080 wide, with the upper left corner as the origin, and rightward and downward as positive; extract the 14 features corresponding to the eigenvalues ​​of the area, see Table 1; Table 1: Feature selection for training and prediction values, training value, validation value, prediction value Variables datapoints Training phase (0.7) Turbidity(NTU) 15160*0.7 Mean (RGB) 3*15160*0.7 Standard deviation 3*15160*0.7 average_gray, 15160*0.7 entropy 15160*0.7 Contrast 15160*0.7 variance_gray 15160*0.7 Mean(HSV) 3*15160*0.7 ratio 15160*0.7 Testing phase (0.3) Turbidity (NTU) 15160*0.3 Mean (RGB) 3*15160*0.3 Standard deviation 3*15160*0.3 average_gray, 15160*0.3 entropy 15160*0.3 Contrast 15160*0.3 variance_gray 15160*0.3 Mean (HSV) 3*15160*0.3 ratio 15160*0.3 validation phase (1) Mean (RGB) 3*15160 Standard deviation 3*15160 average_gray, 15160 entropy 15160 Contrast 15160 variance_gray 15160 Mean (HSV) 3*15160 ratio 15160 Turbidity (NTU) 15160 predicted value output Step S2: Model construction: Adaptive deep learning framework After extracting the above data, the image feature value is used as the independent variable, and the turbidity value obtained by in-situ observation is used as the dependent variable. There is a certain correlation between the two. We try to use convolutional neural networks to simulate the relationship between the two, establish a model for predicting water turbidity from image feature values, and use R2 to evaluate the performance of the model.

[0048] Step S2-1, data preprocessing: In order to increase the generalization ability of the model and improve the richness of the data, The principle and function of using data enhancement for data enhancement are: improving the adaptability of the model to real scenes through diversified training data; adding Gaussian noise (σ=0.05) to the original sonar signal to simulate the interference of deep-sea suspended particles; its core function is not to simply increase the amount of data, but to simulate the signal interference caused by suspended particles and other factors in the real deep-sea environment to "train" the model. The ultimate goal of this is to improve the adaptability and prediction (or analysis) accuracy of the model in the actual, complex and noisy deep-sea operating environment. By letting the model "rehearse" in advance how to deal with signals with interference, its reliability and robustness in real application scenarios can be significantly improved, thereby obtaining more reliable analysis results on deep-sea suspended particles.

[0049] The principle and function of using regularization to regularize data are: preventing overfitting by constraining model complexity, adding a Dropout ratio of 0.3 after the convolutional layer, and applying L2 regularization λ=0.001; the ultimate goal is to build a model with stronger generalization ability, that is, when applied to new and real (deep sea) images or sonar data, it can give more accurate and reliable predictions or analysis results.

[0050] The principle and function of using expand dimension to increase data dimension are: enhancing the feature expression ability by reconstructing the input data, segmenting the one-dimensional signal into multiple time windows, forming a "time-segment" two-dimensional input, enhancing the ability to extract temporal features, and enhancing the model's ability to capture local features related to turbidity by multi-dimensionally deconstructing and reorganizing the spatial features of image data; Step S2-2, model structure selection: In order to build a model that can accurately and automatically estimate the turbidity (turbidity value, in NTU) of a water body by analyzing its image alone, we chose convolutional neural network (CNN) as the core technology. CNN is particularly good at automatically learning and extracting visual features relevant to the target task from images. In terms of water turbidity, these features may include the color depth of the water, transparency, the pattern of light scattering, the presence of visible suspended matter or particulate matter, and so on.

[0051] However, the effectiveness of a CNN model depends largely on its "structural" design - such as how many layers of processing units (convolutional layers, pooling layers) it contains, how many "filters" (convolution kernels) are used in each layer to detect different visual patterns, and the size of these "filters". Different structural combinations will affect the effectiveness of the model in capturing turbidity-related information in the image. A too simple structure may not capture the complex visual clues that distinguish subtle turbidity differences; while an overly complex structure may require more data and computing resources, and is prone to "remembering" specific details of the training images (including irrelevant noise), resulting in poor prediction of new, unseen water images (this is called overfitting).

[0052] Therefore, it is crucial to find an "optimal" network structure that can best learn the mapping relationship between the visual features of water images and the actual turbidity values. To this end, we adopted two key strategies: Grid Search for Structure Optimization: The goal is to systematically explore different CNN structural possibilities to find the configuration that is most suitable for extracting turbidity-related information from water images. A series of candidate structural parameters are predefined (for example, try 2, 3, or 4 convolutional layers; try 16, 32, or 64 convolution kernels per layer; try 3x3 or 5x5 convolution kernel size, etc.). Grid search automatically tries all (or partially selected) combinations of these parameters like permutations and combinations. For each structural combination, we train the model with the same batch of water image data containing known turbidity values ​​and evaluate its accuracy in predicting turbidity (for example, R², MSE, MAE). By comparing the model performance under different structures, grid search helps us find the CNN structural configuration that can most accurately infer turbidity values ​​from image visual clues. This avoids blind guessing or relying purely on experience, making the structural selection process more scientific and data-driven.

[0053] Early Stopping for Generalization: The purpose is to prevent the model from overfitting to specific images in the training data (even irrelevant noise or lighting conditions in the images) during training, and to ensure that the model learns universal turbidity judgment rules that can be applied to new water images. During model training, we not only use training images to adjust model parameters, but also periodically use a batch of validation images that the model has never "seen" to test its current turbidity prediction performance. We will continue to monitor the prediction error of the model on the validation images. Once we find that the prediction error of the model on the validation images is no longer decreasing, or even starts to increase (even though its error on the training images may still be decreasing), we stop training immediately. This means that the model has learned enough general turbidity discriminant features, and continuing training is likely to start learning details that are unique to the training data and are not conducive to generalization. Through early stopping, we can obtain a model with stronger generalization ability, which not only performs well on trained images, but more importantly, can also give relatively reliable and stable turbidity prediction results when encountering new water images taken in different environments in the future, which is crucial for the practical application of the model (such as field water quality monitoring).

[0054] Based on the grid search method, the optimal learning rate of 0.001, batch size of 32 and number of iterations of 200 were determined, and an early stopping mechanism was introduced: patience value = 10 rounds to prevent training redundancy; finally, a 6-layer convolutional neural network was used, and each layer of the convolutional neural network was followed by batch normalization (Batch Normalization), L2 regularization λ = 0.001 and Dropout ratio = 0.3 were used to suppress overfitting; and a maximum pooling layer Maxpooling1D (2), ReLU was used as the activation function, and the output layer was a linear regression unit; Step S2-3: Model performance evaluation Using the coefficient of determination R 2 , mean square error MSE and mean absolute error MAE are used to evaluate the performance of the prediction model; the specific meanings and calculation formulas of the three indicators are as follows: Coefficient of determination (R²): measures the ability of the model to explain the variability of the data (target value > 0.85), indicating to what extent the model explains the actual turbidity changes, focusing on the goodness of fit and explanatory power of the model. It shows to what extent the image features can capture the law of turbidity changes. If R² is very low, it means that the image features themselves may not be sufficient to predict turbidity well, or the model structure is not properly selected.

[0055] Mean Squared Error (MSE): quantifies the average of the squared errors between the model's predicted values ​​and the true values. It is the error between the actual turbidity and the predicted turbidity of each water sample, and the square of the error is calculated to calculate the average of these squared errors. The unit of the mean squared error (MSE) is the square of the turbidity unit (NTU) 2 Since the calculation includes the square of the error between the true turbidity value and the predicted turbidity value, it is particularly sensitive to large turbidity prediction errors. One or several extreme prediction errors (such as very turbid water with very low predicted turbidity) will significantly increase the MSE value.

[0056] Mean absolute error (MAE): quantifies the average absolute error between the model's predicted turbidity value and the true turbidity value. Calculate the absolute value of the error between the true turbidity value and the predicted turbidity value, and find the average. The unit of mean absolute error (MAE) is the same as the turbidity unit (NTU). If MAE=5 NTU, it means that the turbidity value predicted by the model differs from the actual measured turbidity value by 5 NTU (it may be 5 NTU higher or 5 NTU lower). MAE gives the same weight to all sizes of turbidity prediction differences, does not amplify the impact of extreme errors, and better reflects the general average prediction accuracy of the model. The smaller the MAE, the smaller the average deviation of the turbidity value predicted by the model from the true value.

[0057]

[0058] Refers to the actual turbidity, the actual measured turbidity of the i-th water sample; Refers to the model predicted turbidity, which is the turbidity predicted by the model based on the image features of the i-th water sample; Refers to the average value of the true turbidity.

[0059] In the description of the present invention, the term "plurality" refers to two or more than two. Unless otherwise clearly defined, the orientation or positional relationship indicated by the terms "upper" and "lower" is based on the orientation or positional relationship shown in the drawings. It is only for the convenience of describing the present invention and simplifying the description, and does not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation. Therefore, it cannot be understood as a limitation of the present invention; the terms "connection", "installation", "fixation", etc. should be understood in a broad sense. For example, "connection" can be a fixed connection, a detachable connection, or an integral connection; it can be directly connected or indirectly connected through an intermediate medium. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0060] In the description of this specification, the description of the terms "one embodiment", "some embodiments", "specific embodiments", etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0061] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for predicting turbidity of deep-sea mining sediment plume based on digital image processing, characterized in that , specifically including the following steps: Step S1, feature extraction: multimodal feature fusion; First, the size of the deep-sea mining video image is 1080*1920. The high-resolution deep-sea mining video image is processed and the video is framed. Then, the image is processed every 90 frames, including: calculating the mean and standard deviation of the three channels of the RGB image, as well as the mean and variance after grayscale conversion, and calculating the entropy of the image and the contrast of the image. Finally, the video frame image is converted to the HSV space, and the mean values ​​of the H, S, and V channels of the image in the HSV space are calculated. Finally, the Gaussian mixture model is used to detect the image of the sediment plume and calculate the plume area ratio, with a total of 14 features. The RBR multi-parameter water quality meter observes the turbidity of the water body. After calculating the eigenvalues ​​of the image, it is necessary to determine the observation area of ​​the turbidity probe of the RBR multi-parameter water quality meter for in-situ observation, that is, area I. Then, according to the coordinates of area I on the image: 100, 100, a square with a side length of 200, where the image is 1920 long and 1080 wide, with the upper left corner as the origin and rightward and downward as positive; extract the 14 features corresponding to the eigenvalues ​​of the area; Step S2, model construction: adaptive deep learning framework; Step S2-1, data preprocessing: In order to increase the generalization ability of the model and improve the richness of the data, The principle and function of using data enhancement are as follows: improving the model's adaptability to real scenarios through diversified training data; adding Gaussian noise σ=0.05 to the original sonar signal to simulate the interference of deep-sea suspended particles; The principle and function of using regularization to regularize data are: to prevent overfitting by constraining model complexity, add Dropout ratio = 0.3 after the convolution layer, and apply L2 regularization λ = 0.001; The principle and function of using expand dimension to increase data dimension are: enhancing feature expression ability by reconstructing input data, segmenting one-dimensional signal into multiple time windows, forming "time-segment" two-dimensional input, enhancing time series feature extraction ability, and enhancing the model's ability to capture turbidity-related local features by multi-dimensional deconstruction and reorganization of spatial features of image data; Step S2-2, model structure selection: Based on the grid search method, the optimal learning rate of 0.001, batch size of 32 and number of iterations of 200 are determined, and an early stopping mechanism is introduced: patience value = 10 rounds to prevent training redundancy; a 6-layer convolutional neural network is used, and each layer of the convolutional neural network is followed by batch normalization. L2 regularization λ = 0.001 and Dropout ratio = 0.3 are used to suppress overfitting; and a maximum pooling layer Maxpooling1D (2), the activation function is ReLU, and the output layer is a linear regression unit; Step S2-3, model performance evaluation: Using the coefficient of determination R 2 , mean square error MSE and mean absolute error MAE are used to evaluate the performance of the prediction model; the calculation formula is as follows: ， ， ， Refers to the actual turbidity, the actual measured turbidity of the i-th water sample; Refers to the model predicted turbidity predicted turbidity, which is the turbidity predicted by the model based on the image features of the i-th water sample; Refers to the average value of the true turbidity.

2. A deep sea mining sediment plume turbidity prediction method based on digital image processing according to claim 1, characterized in that ,In the step S1, the frame rate of the deep sea mining video is 30, and the observation interval of the RBR turbidity meter is 3s.

3. A deep sea mining sediment plume turbidity prediction method based on digital image processing according to claim 1, characterized in that ,The calculation method of the 14 features in step S1 is: Step S1-1, calculate the mean values ​​of the three channels of the RGB image: R_mean, G_mean, B_mean, and the standard deviations: R_std, G_std, B_std; The formula for calculating the average value of the three channels of the RGB image is as follows: ， ， ， The standard deviation calculation formula for the RGB three channels is as follows: ， ， ， Step S1-2, Gray mean and Gray variance after grayscale conversion: To calculate the grayscale mean and variance, first convert the RGB image into a grayscale image, and then calculate the grayscale mean and variance: First, convert the RGB image to a grayscale image. The calculation formula is as follows: ， The grayscale mean calculation formula is as follows: ， Grayscale variance reflects the degree of dispersion between the grayscale value of each pixel and the average grayscale value. It is the main metric for measuring the amount of image information. The larger the variance, the greater the amount of image information. The variance calculation formula is as follows: ， Step S1-3, the average values ​​of the H hue, S saturation, and V brightness channels: H_mean, S_mean, V_mean, calculated as follows: ， Step S1-4, contrast: calculated based on the gray level co-occurrence matrix GLCM to characterize the local gray level difference; Luminance, L: ， Using the standard deviation method RMS Contrast, the contrast is measured by the standard deviation of brightness: ， ， Step S1-5, Entropy of the image: reflects the complexity of information, and the calculation formula is: ， Where p(x) is the pixel intensity distribution probability; Step S1-6: Use Gaussian mixture model GMM to segment the plume area and calculate the plume area ratio: Ratio = number of plume pixels / total number of pixels.

Images