Seabed sediment classification method based on multilevel comparative learning and uncertainty measurement

Abstract

The invention discloses a seabed sediment classification method based on multilevel comparative learning and uncertainty measurement. The method comprises the following specific steps: dividing target backscattering intensity data to obtain a training set containing marked samples and unmarked samples and a test set containing marked samples; constructing a feature extraction model, and obtaining a multi-level fusion feature based on the training set; obtaining optimized feature representation based on multi-level comparative learning; screening based on the unmarked samples in the training set and the uncertainty measurement to obtain a target pseudo-tag sample; and carrying out joint learning based on the labeled samples in the training set and the target pseudo-label samples. According to the method, a high-quality pseudo label, namely a target pseudo label sample, is screened out through a pseudo label screening strategy based on uncertainty measurement, and the quality of the target pseudo label sample and the adaptability of model learning are ensured; and supervised and semi-supervised joint learning is carried out through the marked samples and the target pseudo-tag samples in the training set, unmarked samples are fully utilized, and the substrate classification performance is improved.

Description

Technical Field

[0001] The present invention relates to the field of seabed sediment classification, and in particular to a seabed sediment classification method based on multi-level contrast learning and uncertainty measurement. Background Art

[0002] The accurate classification of seabed sediments is of great significance in marine scientific research and practical applications. Seabed sediment information is not only the basis for marine geological research, but also the core data for assessing marine resources, protecting the marine ecological environment, predicting marine disasters, and supporting seabed exploration and engineering construction.

[0003] Seabed classification usually relies on sonar data, but high-quality bottom classification models require a large amount of labeled training data. However, the data annotation process not only requires the participation of domain experts, but also has high time and economic costs. In addition, marine surveys are limited by harsh environments and remote geographical locations, resulting in extremely limited labeled samples.

[0004] Therefore, it is particularly necessary to develop semi-supervised learning methods that can combine large amounts of unlabeled data. Summary of the invention

[0005] The present invention provides a seabed sediment classification method based on multi-level contrastive learning and uncertainty measurement. The method uses a student network to take target pseudo-label samples as guidance and combines labeled samples in a training set to perform supervised and semi-supervised joint learning, making full use of unlabeled samples and gradually improving the sediment classification performance.

[0006] The technical solution adopted by the present invention is:

[0007] A method for seafloor sediment classification based on multi-level contrastive learning and uncertainty measurement includes the following specific steps:

[0008] S1. Obtain target backscatter intensity data;

[0009] S2, dividing the target backscattering intensity data into a training set including labeled samples and unlabeled samples and a test set including labeled samples; wherein the labeled samples are target backscattering intensity data with known seabed sediment types, and the unlabeled samples are target backscattering intensity data with unknown seabed sediment types;

[0010] S3, build a feature extraction model, and obtain multi-level fusion features based on the training set;

[0011] S4, embedding the labeled samples and unlabeled samples in the training set into the same feature space, and obtaining optimized feature representation based on multi-level contrastive learning;

[0012] S5, obtaining target pseudo-label samples based on unlabeled samples and uncertainty measures in the training set;

[0013] S6. Perform joint learning based on the labeled samples in the training set and the target pseudo-label samples.

[0014] Further, S1 includes the following specific steps:

[0015] S11, acquiring seabed multi-beam data, and preprocessing the seabed multi-beam data to obtain target seabed multi-beam data;

[0016] S12, extracting backscatter intensity data from the target seabed multi-beam data;

[0017] S13, performing minimum-maximum normalization processing on the extracted backscattering intensity data to obtain target backscattering intensity data;

[0018] Among them, the minimum-maximum normalization processing is as follows:

[0019]

[0020] In formula (1), I min and I max are the minimum and maximum values ​​in the backscattering intensity data, respectively.

[0021] Furthermore, the seafloor multi-beam data preprocessing includes the following specific steps:

[0022] 1) Using mean filtering method or median filtering method to eliminate random noise and obtain the first seafloor multi-beam data;

[0023] 2) using a statistical method to detect and remove abnormal seabed multi-beam data from the first seabed multi-beam data to obtain second seabed multi-beam data;

[0024] 3) Convert the second seafloor multi-beam data from the original sensor coordinate system into a unified geographic coordinate system to obtain the target seafloor multi-beam data.

[0025] Furthermore, S2 also includes:

[0026] 1) Perform weak enhancement on the labeled samples in the training set to obtain the target labeled samples;

[0027] 2) Perform strong enhancement settings on the unlabeled samples in the training set to obtain the target unlabeled samples.

[0028] Further, S3 includes the following specific steps:

[0029] S31, build feature extraction model deeplabv3+;

[0030] S32, using the training set as input features, obtain the multi-level features of the feature extraction model deeplabv3+ (F 1 ,F 2 ,...,F L );

[0031] S33, performing feature fusion on the multi-level features to obtain multi-level fusion features, thereby capturing local detail features and global spatial distribution features of the seabed sediment;

[0032] Among them, the acquisition of multi-level fusion features includes the following specific steps:

[0033] Assume that the characteristics of the teacher network are represented as The characteristics of the student network are expressed as Where L represents the number of layers;

[0034] Through feature projection, we get the multi-level fusion feature vector z of the teacher network and the student network. T and z S :

[0035] z T =g(F T )(2)

[0036] z S =g(F S )(3)

[0037] In equations (2) to (3), g(·) represents feature projection. A multi-layer perceptron is used to upsample features from different levels to match the resolution, thereby obtaining multi-level fusion feature vectors of the teacher network and the student network.

[0038] Furthermore, S4 includes the following specific steps:

[0039] S41, the multi-level fusion feature vector z obtained by the teacher network T The feature quality is filtered by the predicted confidence to obtain a high-confidence multi-level fusion feature vector

[0040]

[0041] In formula (4), is the feature set of seabed sediment category C, is the feature extracted from the i-th unlabeled sample, and τ is the confidence threshold;

[0042] S42, after predicting q(·), the multi-level fusion feature vector with high confidence The multi-level fusion feature vector z obtained by the student network SGroup them according to the predicted categories to achieve feature alignment within the seabed sediment category;

[0043] S43. Calculate the contrast loss based on the feature vector obtained after feature quality filtering and category grouping.

[0044] Further, S43 includes the following specific steps:

[0045] 1) Based on a specific category of attention module S c,θ , calculate the weight value w∈[0,1] of each feature, the calculation formula is as follows:

[0046]

[0047] In formula (5), Prediction vector for category C The weight of Pc is the number of features of category C; S c,θ (z S ) is the feature weight generated by the category attention module;

[0048] 2) Based on cosine similarity, the multi-level fusion feature vector z of the student network is calculated S and the multi-level fusion feature vector z of the teacher network T The similarity is calculated as follows:

[0049]

[0050] In formula (6), C(z S ,z T ) is the similarity between the multi-level fusion feature vector of the student network and the multi-level fusion feature vector of the teacher network; <·,·> represents the inner product operation, || || 2 Indicates L 2 norm;

[0051] 3) Based on weight and The distance between the multi-level fusion feature vector of the student network and the multi-level fusion feature vector of the teacher network is weighted:

[0052]

[0053] In formula (7), is the distance between the multi-level fusion feature vector of the student network and the multi-level fusion feature vector of the teacher network;

[0054] 4) Calculate the contrast loss. The specific calculation formula is as follows:

[0055]

[0056] In formula (8), L contrastive is the contrast loss; C is the total number of seabed sediment categories; is the number of features of the student network in the seabed sediment category C; is the number of features of the teacher network in the seabed sediment category C; is the distance between the characteristics of the student network and the teacher network.

[0057] Further, S5 includes the following specific steps:

[0058] S51, the teacher network performs forward inference on the unlabeled samples in the training set to obtain the initial pseudo labels, including the following steps:

[0059] 3) Input the unlabeled samples in the training set into the teacher network;

[0060] 4) The prediction p of the unlabeled sample is converted into a probability distribution P:

[0061] P = Softmax(p) (9)

[0062] In formula (9), P∈R N×C represents the predicted probability distribution of N unlabeled samples, and C is the total number of seabed sediment categories;

[0063] 3) Obtain the maximum predicted probability P of each unlabeled sample from the probability distribution P max And the corresponding pseudo label Y get the initial pseudo label:

[0064] P max ,Y=max(P)(10)

[0065] In formula (10), P max ∈R N is the maximum predicted probability for each unlabeled sample, Y∈R N is the corresponding pseudo label;

[0066] S52, screening the initial pseudo labels based on the uncertainty measure to obtain target pseudo label samples;

[0067] 1) Use the entropy function H(·) to measure the uncertainty of each initial pseudo-label prediction:

[0068]

[0069] In formula (11), H∈R N Represents the entropy value of each initial pseudo-label. A low entropy value indicates that the model has high confidence in the prediction of the initial pseudo-label, and a high entropy value indicates that the model has high uncertainty in the prediction of the initial pseudo-label.

[0070] 2) According to the uncertainty H of the initial pseudo-label, dynamically adjust the threshold of the initial pseudo-label screening:

[0071] T dyn =α·(1-H) (12)

[0072] In formula (12), T dyn ∈R N is the dynamic threshold, and α is a hyperparameter used to control the adjustment amplitude of the threshold;

[0073] 3) Dynamically calculate category weight W c , the calculation formula is as follows:

[0074]

[0075] In formula (13), count c represents the number of initial pseudo labels for category c, ∈ is a smoothing term to prevent the denominator from being 0, W c ∈R C The weight for each category;

[0076] 4) For each initial pseudo-label, its adjusted probability is defined as:

[0077]

[0078] In formula (14), P adj ∈R N is the adjusted probability, W Y is the category weight corresponding to the pseudo label;

[0079] 5) Based on the adjusted probability P adj and dynamic threshold T dyn , filter out high-quality pseudo labels to obtain target pseudo label samples:

[0080] m=P adj >T dyn (15)

[0081] In formula (15), m∈{0,1} N The filter mask.

[0082] Further, S6 includes the following specific steps:

[0083] S61, inputting the labeled samples and target pseudo-label samples in the training set into the student network;

[0084] S62. The student network optimizes the performance of the feature extraction model through joint training of supervised learning and semi-supervised learning:

[0085] 1) Using labeled samples x l Carry out basic ability training of the classifier, let H(y1 ,y 2 ) are two probability distributions y containing N pixel categories 1 ,y 2 The weighted cross entropy loss between the lists of is calculated as follows:

[0086]

[0087] In formula (16), C is the total number of seabed sediment types; N is y 1 The number of pixels in ; α c is the weight set by category; β n The weights are set per pixel.

[0088] 2) Calculate the supervision loss L sup , the calculation formula is as follows:

[0089]

[0090] In formula (17), is the labeled data x l A weakly enhanced version of .

[0091] 3) Use the target pseudo-label samples to expand the model's learning ability for unlabeled samples. For each unlabeled sample x u , obtain high-quality pseudo labels through pseudo label screening strategy based on uncertainty measurement Calculate the unlabeled sample x by cross entropy u The pseudo label loss L pseudo , the calculation formula is as follows:

[0092]

[0093] In formula (18), is x u A strong enhanced version of u The number of augmentations applied.

[0094] 4) Direct entropy minimization is applied to the student network for unlabeled samples x u The predicted class distribution, as a regularization loss L ent , and its calculation formula is as follows:

[0095]

[0096] In formula (19), C is the number of categories; N is the number of pixels; A is the u The number of enhancements applied.

[0097] 4) Calculate the total loss function, the calculation formula is as follows:

[0098] L=λ sup L sup +λ pseudo L pseudo +λ ent L ent +λ contrastive L contrastive (20)

[0099] In formula (20), λ is the weight factor, which dynamically adjusts the influence of each part of the loss function;

[0100] Furthermore, S62 also includes:

[0101] After the student network training is completed, the updated parameter status is synchronized to the teacher network;

[0102] In each training step, the teacher network f ζ Updated via:

[0103] ζ=τζ+(1-τ)θ(21)

[0104] In formula (21), ζ is the weight of the teacher network; τ is the decay rate; θ is the student network f θ The weight of

[0105] Among them, the teacher network weight ζ is the exponential moving average of the student network weight θ, and the decay rate is τ∈[0,1].

[0106] Compared with the prior art, the present invention has the following beneficial effects:

[0107] 1) Input the target backscatter intensity data as input features into the constructed feature extraction model to obtain multi-level features, and perform feature fusion on the multi-level features to obtain multi-level fusion features, which are used to capture the local detail features and global spatial distribution features of the seabed sediment;

[0108] 2) Adopt the pseudo-label screening strategy based on uncertainty measurement to screen high-quality pseudo-labels to obtain target pseudo-label samples and remove unreliable pseudo-label samples to ensure the quality of pseudo-label samples and the adaptability of model learning;

[0109] 3) Through the student network, the target pseudo-label samples are used as guidance, combined with the labeled samples in the training set, to conduct supervised and semi-supervised joint learning, make full use of unlabeled samples, and gradually improve the bottom classification performance

[0110] 4) The test set retains some labeled samples for independent verification of model performance, ensuring that the labeled samples cover different categories of sediment characteristics and avoiding classification bias caused by excessive sample imbalance. BRIEF DESCRIPTION OF THE DRAWINGS

[0111] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments:

[0112] Figure 1 It is a schematic diagram of the process of the present invention;

[0113] Figure 2 It is a schematic diagram of the optimization process based on multi-level contrast learning in the present invention;

[0114] Figure 3 It is a schematic diagram of the pseudo-label screening process based on uncertainty measurement in the present invention. DETAILED DESCRIPTION

[0115] Example

[0116] like Figures 1 to 3 As shown, a seafloor sediment classification method based on contrastive learning and uncertainty measurement includes the following specific steps:

[0117] S1. Obtain target backscatter intensity data;

[0118] Specifically, S11, acquiring seabed multi-beam data, and preprocessing the seabed multi-beam data to obtain target seabed multi-beam data;

[0119] S12, extracting backscatter intensity data from the target seabed multi-beam data;

[0120] S13, performing minimum-maximum normalization processing on the extracted backscattering intensity data to obtain target backscattering intensity data;

[0121] Among them, the seabed multi-beam data preprocessing specifically includes the following steps:

[0122] 1) Using mean filtering method or median filtering method to eliminate random noise and obtain the first seafloor multi-beam data;

[0123] 2) using a statistical method to detect and remove abnormal seabed multi-beam data from the first seabed multi-beam data to obtain second seabed multi-beam data;

[0124] 3) converting the second seafloor multi-beam data from the original sensor coordinate system into a unified geographic coordinate system to obtain target seafloor multi-beam data;

[0125] Note: The seabed multi-beam data comes from the multi-beam sonar system, which acquires seabed backscatter signals by emitting and receiving high-frequency sound waves; the backscatter intensity reflects the characteristics of the interaction between sound waves and the seabed surface, and is an important basis for distinguishing different bottom types. Since the multi-beam sonar system may be affected by factors such as environmental noise and instrument deviation, the seabed multi-beam data needs to be preprocessed: first, the mean filter or median filter method is used to eliminate random noise to obtain the first seabed multi-beam data; then, a statistical method is used to detect and eliminate the abnormal first seabed multi-beam data to obtain the second seabed multi-beam data; the second seabed multi-beam data is converted from the original sensor coordinate system to a unified geographic coordinate system to obtain the target seabed multi-beam data, and the coordinate system conversion ensures spatial consistency; the backscatter intensity data is extracted from the target seabed multi-beam data as the main input feature. The backscatter intensity is determined by the interaction between the sound waves and the seabed surface, and different bottom types present unique intensity distribution patterns;

[0126] To ensure data consistency, the extracted backscattering intensity data is normalized to obtain target backscattering intensity data. This embodiment uses minimum-maximum normalization, and the formula is as follows:

[0127]

[0128] In formula (1), I min and I max are the minimum and maximum values ​​in the backscattering intensity data, respectively; the data is mapped to the interval [0, 1] to reduce the dimensional difference and avoid the subsequent model training being affected by the backscattering intensity data range being too large or too small;

[0129] S2, dividing the target backscattering intensity data into a training set including labeled samples and unlabeled samples and a test set including labeled samples; wherein the labeled samples are target backscattering intensity data with known seabed sediment types, and the unlabeled samples are target backscattering intensity data with unknown seabed sediment types;

[0130] Note: The training set and the test set are divided in proportion. The labeled samples in the training set are used for supervised learning, and the unlabeled samples in the training set are used for pseudo-label sample generation and semi-supervised learning. In this embodiment, two different enhancement settings are used: weak enhancement settings are performed for the labeled samples in the training set, and strong enhancement settings are performed for the unlabeled samples in the training set. The specific settings are shown in the following table:

[0131] Enhancement Methods Weak Enhancement Strong Enhancement Classmix 0.2 0.8 Flip 0.5 0.5 Contrast Adjustment 0.2 0.8

[0132] The test set retains some labeled samples for independent verification of model performance, ensuring that the labeled samples cover different categories of sediment characteristics and avoiding classification bias caused by excessive sample imbalance.

[0133] S3, build a feature extraction model, and obtain multi-level fusion features based on the training set;

[0134] Specifically, S31, construct the feature extraction model deeplabv3+;

[0135] S32, using the training set as input features, obtain the multi-level features of the feature extraction model deeplabv3+ (F 1 ,F 2 ,...,F L );

[0136] S33, performing feature fusion on the multi-level features to obtain multi-level fusion features, thereby capturing local detail features and global spatial distribution features of the seabed sediment;

[0137] Among them, the acquisition of multi-level fusion features includes the following specific steps:

[0138] Assume that the characteristics of the teacher network are represented as The characteristics of the student network are expressed as Where L represents the number of layers;

[0139] Through feature projection, we get the multi-level fusion feature vector z of the teacher network and the student network. T and z S :

[0140] z T =g(F T )(2)

[0141] z S =g(F S )(3)

[0142] In formulas (2) to (3), g(·) represents feature projection. A multi-layer perceptron is used to upsample features from different levels to match the resolution, thus achieving multi-level fusion feature vector acquisition of the teacher network and the student network.

[0143] S4, embedding the labeled samples and unlabeled samples in the training set into the same feature space, and obtaining optimized feature representation based on multi-level contrastive learning;

[0144] Specifically, S41, the multi-level fusion feature vector z obtained by the teacher network T The feature quality is filtered by the predicted confidence to obtain a high-confidence multi-level fusion feature vector

[0145]

[0146] In formula (4), is the feature set of seabed sediment category C, is the feature extracted from the i-th unlabeled sample, and τ is the confidence threshold;

[0147] S42, after predicting q(·), the multi-level fusion feature vector with high confidence The multi-level fusion feature vector z obtained by the student network S Group them according to the predicted categories to achieve feature alignment within the seabed sediment category;

[0148] S43, based on the feature vector obtained after feature quality filtering and category grouping, calculating the contrast loss, specifically including the following steps:

[0149] 1) Based on a specific category of attention module S c,θ , calculate the weight value w∈[0,1] of each feature, the calculation formula is as follows:

[0150]

[0151] In formula (5), Prediction vector for category C The weight of Pc is the number of features of category C; S c,θ (z S ) is the feature weight generated by the category attention module;

[0152] 2) Based on cosine similarity, the multi-level fusion feature vector z of the student network is calculated S and the multi-level fusion feature vector z of the teacher network T The similarity is calculated as follows:

[0153]

[0154] In formula (6), C(z S ,z T ) is the similarity between the multi-level fusion feature vector of the student network and the multi-level fusion feature vector of the teacher network; <·,·> represents the inner product operation, || || 2 Indicates L 2 norm;

[0155] 3) Based on weight and The distance between the multi-level fusion feature vector of the student network and the multi-level fusion feature vector of the teacher network is weighted:

[0156]

[0157] In formula (7), is the distance between the multi-level fusion feature vector of the student network and the multi-level fusion feature vector of the teacher network;

[0158] 4) Calculate the contrast loss. The specific calculation formula is as follows:

[0159]

[0160] In formula (8), L contrastive is the contrast loss; C is the total number of seabed sediment categories; is the number of features of the student network in the seabed sediment category C; is the number of features of the teacher network in the seabed sediment category C; is the distance between the characteristics of the student network and the teacher network;

[0161] Note: In order to ensure the efficiency of contrastive learning, it is necessary to perform feature quality filtering on the extracted multi-level features of the seabed sediments; in order to confirm the importance of the feature vector to the contrastive learning task, it is necessary to calculate the specific weight of each feature, which is used as the weight factor of the contrastive loss function.

[0162] S5, obtaining target pseudo-label samples based on unlabeled samples and uncertainty measures in the training set;

[0163] Specifically, S51, the teacher network performs forward inference on the unlabeled samples in the training set to obtain initial pseudo labels, including the following steps:

[0164] 5) Input the unlabeled samples in the training set into the teacher network;

[0165] 6) The prediction p of the unlabeled sample is converted into a probability distribution P:

[0166] P = Softmax(p) (9)

[0167] In formula (9), P∈R N×C represents the predicted probability distribution of N unlabeled samples, and C is the total number of seabed sediment categories;

[0168] 3) Obtain the maximum predicted probability P of each unlabeled sample from the probability distribution P max And the corresponding pseudo label Y get the initial pseudo label:

[0169] P max ,Y=max(P)(10)

[0170] In formula (10), P max ∈R N is the maximum predicted probability for each unlabeled sample, Y∈R N is the corresponding pseudo label;

[0171] S52, screening the initial pseudo labels based on the uncertainty measure to obtain target pseudo label samples;

[0172] 1) Use the entropy function H(·) to measure the uncertainty of each initial pseudo-label prediction:

[0173]

[0174] In formula (11), H∈R N Represents the entropy value of each initial pseudo-label. A low entropy value indicates that the model has high confidence in the prediction of the initial pseudo-label, and a high entropy value indicates that the model has high uncertainty in the prediction of the initial pseudo-label.

[0175] 2) According to the uncertainty H of the initial pseudo-label, dynamically adjust the threshold of the initial pseudo-label screening:

[0176] T dyn =α·(1-H)(12)

[0177] In formula (12), T dyn ∈R N is the dynamic threshold, and α is a hyperparameter used to control the adjustment amplitude of the threshold;

[0178] 3) Dynamically calculate category weight W c , the calculation formula is as follows:

[0179]

[0180] In formula (13), count c represents the number of initial pseudo labels for category c, ∈ is a smoothing term to prevent the denominator from being 0, W c ∈R C The weight for each category;

[0181] 4) For each initial pseudo-label, its adjusted probability is defined as:

[0182]

[0183] In formula (14), P adj ∈R N is the adjusted probability, W Y is the category weight corresponding to the pseudo label;

[0184] 5) Based on the adjusted probability P adj and dynamic threshold T dyn , filter out high-quality pseudo labels to obtain target pseudo label samples:

[0185] m=P adj >T dyn (15)

[0186] In formula (15), m∈{0,1} N is the filter mask;

[0187] Note: For each initial pseudo-label, the entropy value is calculated based on the probability distribution predicted by the model; since different categories may be uneven in the pseudo-label distribution, in order to balance the category influence during the training process, the category weight is dynamically adjusted according to the pseudo-label distribution of unlabeled samples to avoid a few categories being ignored. In this embodiment, the category weight W is dynamically calculated. c ; High-confidence samples are retained as training pseudo-labels, and high-uncertainty samples are removed to avoid noise interference with model learning; and the screening threshold is adjusted through each round of training to adapt to the learning state of the model at different stages;

[0188] S6, joint learning of supervision and semi-supervision based on labeled samples in the training set and target pseudo-label samples;

[0189] Specifically, S61, input the labeled samples and target pseudo-label samples in the training set into the student network

[0190] S62. The student network optimizes the performance of the feature extraction model through joint training of supervised learning and semi-supervised learning:

[0191] 1) Using labeled samples x l Carry out basic ability training of the classifier, let H(y 1 ,y 2 ) are two probability distributions y containing N pixel categories 1 ,y 2 The weighted cross entropy loss between the lists of is calculated as follows:

[0192]

[0193] In formula (16), C is the total number of seabed sediment types; N is y 1 The number of pixels in ; α c is the weight set by category; β n The weights are set per pixel.

[0194] 2) Calculate the supervision loss L sup , the calculation formula is as follows:

[0195]

[0196] In formula (17), is the labeled data x l A weakly enhanced version of .

[0197] 3) Use the target pseudo-label samples to expand the model's learning ability for unlabeled samples. For each unlabeled sample x u, obtain high-quality pseudo labels through pseudo label screening strategy based on uncertainty measurement Calculate the unlabeled sample x by cross entropy u The pseudo label loss L pseudo , the calculation formula is as follows:

[0198]

[0199] In formula (18), is x u A strong enhanced version of u The number of augmentations applied.

[0200] 4) Direct entropy minimization is applied to the student network for unlabeled samples x u The predicted class distribution, as a regularization loss L ent , and its calculation formula is as follows:

[0201]

[0202] In formula (19), C is the number of categories; N is the number of pixels; A is the u The number of enhancements applied.

[0203] 4) Calculate the total loss function, the calculation formula is as follows:

[0204] L=λ sup L sup +λ pseudo L pseudo +λ ent L ent +λ contrastive L contrastive (20)

[0205] In formula (20), λ is the weight factor, which dynamically adjusts the influence of each part of the loss function;

[0206] 5) After the student network training is completed, the updated parameter state is synchronized to the teacher network; the teacher model can provide more accurate and robust predictions, so in each training step, the teacher network f ζ Instead of optimizing via gradient descent, it is updated in the following way:

[0207] ζ=τζ+(1-τ)θ(21)

[0208] In formula (21), ζ is the weight of the teacher network; τ is the decay rate; θ is the student network f θ The weight of

[0209] Among them, the teacher network weight ζ is the exponential moving average of the student network weight θ, with a decay rate of τ∈[0,1];

[0210] Note: Labeled samples provide clear supervision signals for the student network, and unlabeled samples participate in training by generating pseudo labels, making full use of the potential of massive unlabeled samples; the student network combines high-quality pseudo labels with labeled samples and adopts a joint learning strategy to gradually improve the model performance. At the same time, the feedback mechanism of the teacher-student network architecture is used to optimize the entire training process.

[0211] In this embodiment, a certain coast is taken as the study area of ​​the embodiment, and the study area includes five types of seabed substrates: sediment, gravel, sand, clay, and bedrock.

[0212] The present invention performs coordinate system conversion and normalization preprocessing on seabed multi-beam data to ensure spatial consistency and data consistency; divides target backscatter intensity data into a training set and a test set; constructs a feature extraction model, uses the training set as input features, obtains multi-level features and performs feature fusion to capture local detail features and global spatial distribution features of seabed sediments; embeds labeled samples and unlabeled samples into the same feature space, and optimizes feature representation based on multi-level contrast learning; infers unlabeled samples through a teacher network to generate initial pseudo-labels; a pseudo-label screening strategy based on uncertainty measurement screens out high-quality pseudo-labels, i.e., target pseudo-label samples, and removes unreliable pseudo-labels to ensure the quality of target pseudo-label samples and the adaptability of model learning; uses a student network to take target pseudo-label samples as guidance and combines labeled samples in a training set to perform supervised and semi-supervised joint learning, fully utilizes unlabeled samples, and gradually improves sediment classification performance.

[0213] The embodiments described above are merely descriptions of preferred implementation modes of the present invention and are not intended to limit the scope of the present invention. Without departing from the principles and essence of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the scope of protection determined by the claims of the present invention.

Claims

1. A method for seafloor sediment classification based on multi-level contrastive learning and uncertainty measurement, characterized in that: The specific steps include: S1. Obtain target backscatter intensity data; S2, dividing the target backscattering intensity data into a training set including labeled samples and unlabeled samples and a test set including labeled samples; wherein the labeled samples are target backscattering intensity data with known seabed sediment types, and the unlabeled samples are target backscattering intensity data with unknown seabed sediment types; S3, build a feature extraction model, and obtain multi-level fusion features based on the training set; S4, embedding the labeled samples and unlabeled samples in the training set into the same feature space, and obtaining optimized feature representation based on multi-level contrastive learning; S5, obtaining target pseudo-label samples based on unlabeled samples and uncertainty measures in the training set; S6. Perform joint learning based on the labeled samples in the training set and the target pseudo-label samples.

2. A method for classifying seabed sediments based on multi-level contrastive learning and uncertainty measurement according to claim 1, characterized in that: S1 includes the following specific steps: S11, acquiring seabed multi-beam data, and preprocessing the seabed multi-beam data to obtain target seabed multi-beam data; S12, extracting backscatter intensity data from the target seabed multi-beam data; S13, performing minimum-maximum normalization processing on the extracted backscattering intensity data to obtain target backscattering intensity data; Among them, the minimum-maximum normalization processing is as follows: In formula (1), I min and I max are the minimum and maximum values ​​in the backscattering intensity data, respectively.

3. A method for classifying seabed sediments based on multi-level contrast learning and uncertainty measurement according to claim 2, characterized in that: The seafloor multi-beam data preprocessing comprises the following specific steps: 1) Using mean filtering method or median filtering method to eliminate random noise and obtain the first seafloor multi-beam data; 2) using a statistical method to detect and remove abnormal seabed multi-beam data from the first seabed multi-beam data to obtain second seabed multi-beam data; 3) Convert the second seafloor multi-beam data from the original sensor coordinate system into a unified geographic coordinate system to obtain the target seafloor multi-beam data.

4. A method for classifying seabed sediments based on multi-level contrastive learning and uncertainty measurement according to claim 1, characterized in that: S2 also includes: 1) Perform weak enhancement on the labeled samples in the training set to obtain the target labeled samples; 2) Perform strong enhancement settings on the unlabeled samples in the training set to obtain the target unlabeled samples.

5. A method for classifying seabed sediments based on multi-level contrastive learning and uncertainty measurement according to claim 1, characterized in that: S3 includes the following specific steps: S31, build feature extraction model deeplabv3+; S32, using the training set as input features, obtain the multi-level features (F1, F2, ..., F L ); S33, performing feature fusion on the multi-level features to obtain multi-level fusion features, thereby capturing local detail features and global spatial distribution features of the seabed sediment; Among them, the acquisition of multi-level fusion features includes the following specific steps: Assume that the characteristics of the teacher network are represented as The characteristics of the student network are expressed as Where L represents the number of layers; Through feature projection, we get the multi-level fusion feature vector z of the teacher network and the student network. T and z S : z T =g(F T )(2) z S =g(F S )(3) In equations (2) to (3), g(·) represents feature projection. A multi-layer perceptron is used to upsample features from different levels to match the resolution, thereby obtaining multi-level fusion feature vectors of the teacher network and the student network.

6. A method for seabed sediment classification based on multi-level contrast learning and uncertainty measurement according to claim 1, characterized in that: S4 includes the following specific steps: S41, the multi-level fusion feature vector z obtained by the teacher network T The feature quality is filtered by the predicted confidence to obtain a high-confidence multi-level fusion feature vector In formula (4), is the feature set of seabed sediment category C, is the feature extracted from the i-th unlabeled sample, and τ is the confidence threshold; S42, after predicting q(·), the multi-level fusion feature vector with high confidence The multi-level fusion feature vector z obtained by the student network S Group them according to the predicted categories to achieve feature alignment within the seabed sediment category; S43. Calculate the contrast loss based on the feature vector obtained after feature quality filtering and category grouping.

7. A method for classifying seabed sediments based on multi-level contrastive learning and uncertainty measurement according to claim 6, characterized in that: S43 includes the following specific steps: 1) Based on a specific category of attention module S c,θ , calculate the weight value w∈[0,1] of each feature, the calculation formula is as follows: In formula (5), Prediction vector for category C The weight of Pc is the number of features of category C; S c,θ (z S ) is the feature weight generated by the category attention module; 2) Based on cosine similarity, the multi-level fusion feature vector z of the student network is calculated S and the multi-level fusion feature vector z of the teacher network T The similarity is calculated as follows: In formula (6), C(z S ,z T ) is the similarity between the multi-level fusion feature vector of the student network and the multi-level fusion feature vector of the teacher network; <·,·> represents the inner product operation, || ||2 represents the L2 norm; 3) Based on weight and The distance between the multi-level fusion feature vector of the student network and the multi-level fusion feature vector of the teacher network is weighted: In formula (7), is the distance between the multi-level fusion feature vector of the student network and the multi-level fusion feature vector of the teacher network; 4) Calculate the contrast loss. The specific calculation formula is as follows: In formula (8), L contrastive is the contrast loss; C is the total number of seabed sediment categories; is the number of features of the student network in the seabed sediment category C; is the number of features of the teacher network in the seabed sediment category C; is the distance between the characteristics of the student network and the teacher network.

8. A method for seabed sediment classification based on multi-level contrast learning and uncertainty measurement according to claim 1, characterized in that: S5 includes the following specific steps: S51, the teacher network performs forward inference on the unlabeled samples in the training set to obtain the initial pseudo labels, including the following steps: 1) Input the unlabeled samples in the training set into the teacher network; 2) The prediction p of the unlabeled sample is converted into a probability distribution P: P = Softmax(p) (9) In formula (9), P∈R N×C represents the predicted probability distribution of N unlabeled samples, and C is the total number of seabed sediment categories; 3) Obtain the maximum predicted probability P of each unlabeled sample from the probability distribution P max And the corresponding pseudo label Y get the initial pseudo label: P max ,Y=max(P)(10) In formula (10), P max ∈R N is the maximum predicted probability for each unlabeled sample, Y∈R N is the corresponding pseudo label; S52, screening the initial pseudo labels based on the uncertainty measure to obtain target pseudo label samples; 1) Use the entropy function H(·) to measure the uncertainty of each initial pseudo-label prediction: In formula (11), H∈R N Represents the entropy value of each initial pseudo-label. A low entropy value indicates that the model has high confidence in the prediction of the initial pseudo-label, and a high entropy value indicates that the model has high uncertainty in the prediction of the initial pseudo-label. 2) According to the uncertainty H of the initial pseudo-label, dynamically adjust the threshold of the initial pseudo-label screening: T dyn =α·(1-H)(12) In formula (12), T dyn ∈R N is the dynamic threshold, and α is a hyperparameter used to control the adjustment amplitude of the threshold; 3) Dynamically calculate category weight W c , the calculation formula is as follows: In formula (13), count c represents the number of initial pseudo labels for category c, ∈ is a smoothing term to prevent the denominator from being 0, W c ∈R C The weight for each category; 4) For each initial pseudo-label, its adjusted probability is defined as: In formula (14), P adj ∈R N is the adjusted probability, W Y is the category weight corresponding to the pseudo label; 5) Based on the adjusted probability P adj and dynamic threshold T dyn , filter out high-quality pseudo labels to obtain target pseudo label samples: m=P adj >T dyn (15) In formula (15), m∈{0,1} N is the filter mask.

9. A method for seabed sediment classification based on multi-level contrast learning and uncertainty measurement according to claim 1, characterized in that: S6 includes the following specific steps: S61, inputting the labeled samples and target pseudo-label samples in the training set into the student network; S62. The student network optimizes the performance of the feature extraction model through joint training of supervised learning and semi-supervised learning: 1) Using labeled samples x l To train the basic capabilities of the classifier, let H(y1,y2) be the weighted cross entropy loss between two lists containing N pixel category probability distributions y1 and y2, and its calculation formula is as follows: In formula (16), C is the total number of seabed sediment categories; N is the number of pixels in y1; α c is the weight set by category; β n The weights are set per pixel. 2) Calculate the supervision loss L sup , the calculation formula is as follows: In formula (17), is the labeled data x l A weakly enhanced version of . 3) Use the target pseudo-label samples to expand the model's learning ability for unlabeled samples. For each unlabeled sample x u , obtain high-quality pseudo labels through pseudo label screening strategy based on uncertainty measurement Calculate the unlabeled sample x by cross entropy u The pseudo label loss L pseudo , the calculation formula is as follows: In formula (18), is x u A strongly enhanced version of the conduct; A is the sample x u The number of augmentations applied. 4) Direct entropy minimization is applied to the student network for unlabeled samples x u The predicted class distribution, as a regularization loss L ent , and its calculation formula is as follows: In formula (19), C is the number of categories; N is the number of pixels; A is the u The number of enhancements applied. 4) Calculate the total loss function, the calculation formula is as follows: L=λ sup L sup +λ pseudo L pseudo +λ ent L ent +λ contrastive L contrastive (20) In formula (20), λ is the weight factor, which dynamically adjusts the influence of each part of the loss function.

10. A method for seabed sediment classification based on multi-level contrast learning and uncertainty measurement according to claim 9, characterized in that: The S62 also includes: After the student network training is completed, the updated parameter status is synchronized to the teacher network; In each training step, the teacher network f ζ Updated via: ζ=τζ+(1-τ)θ(21) In formula (21), ζ is the weight of the teacher network; τ is the decay rate; θ is the student network f θ The weight of Among them, the teacher network weight ζ is the exponential moving average of the student network weight θ, and the decay rate is τ∈[0,1].

Images