Method and system for inverting ocean vertical temperature structure from offshore bottom-mounted in situ observations

Abstract

The application discloses a method and system for in-situ observation of offshore seabed to inverse marine vertical temperature structure, through constructing a machine learning model of physical mechanism, performing machine learning and physical information constraint-based marine vertical temperature profile inversion, realizing high-precision and high-efficiency mapping from in-situ observation data of a single point of seabed to a temperature profile of full water depth, not only improving modeling precision and robustness, but also expanding the application range of existing inversion technology, which can be directly applied to intelligent monitoring of offshore marine environment, digital management of marine ranch, underwater operation support and marine scientific research and other fields, and has important scientific research and engineering application value. The application also solves the technical bottlenecks of applicability, real-time and accuracy of the method for mapping marine vertical temperature from sea surface observation in offshore waters, and is superior to existing statistical regression technology in terms of inversion accuracy, application range, model flexibility and engineering application convenience, and has important application prospect and popularization value.

Description

Technical Field

[0001] This invention belongs to the field of intelligent monitoring and inversion technology of marine environment, specifically involving a method and system for inverting the vertical temperature structure of the ocean through in-situ observation of the nearshore seabed. Background Technology

[0002] As the ocean plays an increasingly important role in climate regulation, resource development, and disaster early warning, acquiring high-precision three-dimensional ocean temperature structure information has become a key technology in ocean observation and forecasting. However, current ocean internal observation data is limited by spatial coverage and temporal continuity, and current observation methods cannot meet the demand for high-resolution temperature profiles.

[0003] To alleviate the scarcity of ocean internal data, existing technologies have proposed inversion schemes that map the vertical temperature structure of the ocean using statistical regression methods based on sea surface observations. These methods construct empirical modes based on historical ocean temperature profiles and establish the correspondence between sea surface observation variables (such as sea surface temperature and sea level height) and the internal structure of the water column through regression analysis. The basic idea is to project currently available satellite-obtainable sea surface information onto historical empirical space, estimate modal coefficients representing different vertical structures through regression, and then calculate the vertical structure of the target variable, thereby reconstructing the temperature profile. This type of method has a clear structure, high computational efficiency, and is practically applicable in areas with good observation conditions and relatively stable ocean conditions, especially in open deep-sea areas far from nearshore waters.

[0004] However, existing statistical regression methods are generally based on the assumption of linear mapping, making it difficult to effectively characterize nonlinear changes in ocean dynamic processes. In structurally complex regions such as fronts, eddies, upwellings, or cold water tongues beneath sea ice, these methods often struggle to accurately reconstruct the true vertical temperature profile. Furthermore, these methods heavily rely on historical empirical modes, resulting in limited generalization ability; in regions or periods with significant changes in the distribution of observed samples, prediction accuracy drops markedly. Simultaneously, traditional statistical models lack adaptive updating capabilities, failing to flexibly absorb and utilize ever-increasing new observational data, thus limiting their applicability and scalability in dynamically changing contexts. In recent years, although a few studies have explored the use of purely deep learning models (such as fully connected neural networks and recurrent neural networks) for temperature profile inversion in open ocean regions, their applications are mainly limited to the relatively stable and homogeneous open ocean. More importantly, these methods rely entirely on data-driven models in their model architecture and training objectives, without introducing any physical constraints, leading to outputs that may violate fundamental ocean physics laws (such as hydrostatic equilibrium and mass conservation), producing physically unrealistic solutions. In key sea areas with complex ocean dynamic processes and strong multi-scale interactions, such as nearshore areas, straits, and continental slopes, this purely data-driven model, lacking physical guidance, is easily misled by observation noise and data bias, and its robustness and reliability remain questionable.

[0005] Therefore, although statistical regression methods have certain applicability in marine areas with ideal observation conditions and stable structures, their modeling capabilities and adaptability have gradually become key bottlenecks restricting their further development in more complex and variable marine environments. To meet the demand for high-precision, high-resolution vertical temperature structure inversion, it is urgent to develop a novel inversion method with nonlinear modeling capabilities, physical process compliance, strong generalization performance, and self-learning characteristics, in order to improve its application effectiveness and reliability in highly dynamic and high-dimensional marine environments.

[0006] On the other hand, most current mainstream vertical temperature retrieval methods rely on sea surface remote sensing data, such as sea surface temperature, sea surface height anomalies, and wind fields. These satellite-based methods are well-suited for use in open ocean areas such as the high seas, but they have the following significant shortcomings in nearshore waters:

[0007] First, nearshore remote sensing data is often hampered by factors such as coastline reflection, complex topographical obstruction, and frequent cloud cover, resulting in information gaps, significant noise, and poor quality, leading to a substantial decrease in the accuracy of inversion results in key areas. Furthermore, current mainstream inversion methods generally fail to fully integrate high-quality, continuous measured data provided by fixed nearshore stations (such as seabed temperature sensors and moored observation systems), resulting in underutilization of this important information source. Technically, existing methods largely rely on simplified statistical models or empirical function expansions, which have inherent limitations in spatial and vertical resolution, making it difficult to effectively capture common high-gradient structures in nearshore waters such as fronts and strata. Therefore, their ability to characterize fine dynamic processes and support local analysis is limited. It should also be noted that satellite observations typically provide data at the daily average scale, whose temporal resolution is insufficient to reflect diurnal variations or high-frequency dynamic processes at shorter timescales in nearshore waters, further restricting the revelation of the fine evolution mechanisms of the nearshore environment.

[0008] In summary, existing statistical regression methods and remote sensing-based sea surface inversion techniques have certain limitations in applicability and capability. There is an urgent need to introduce new methods that are more adaptable, have nonlinear expressive power, and data fusion capabilities to improve the accuracy and application breadth of inversion of the vertical temperature structure of nearshore and complex ocean regions. Summary of the Invention

[0009] To overcome the shortcomings of existing technologies and achieve more accurate, continuous, and high-resolution reconstruction of the vertical temperature structure of ocean water, this invention adopts the following technical solution:

[0010] The method for inverting the vertical temperature structure of the ocean through in-situ observation of the nearshore seabed includes the following steps:

[0011] Step 1: Obtain the in-situ temperature time series observed at the seabed timescale of the nearshore observation station;

[0012] Step 2: Construct a neural network model. Extract historical long-term seabed temperature data from the in-situ temperature time series and location information of nearshore observation stations and the information of matched neighboring observation points. Divide the data into time periods and establish a nonlinear mapping relationship between seabed temperature and vertical temperature profile for the observation stations to obtain the optimal weight for each time period. Use a composite loss function that combines mean square error and temperature gradient physical constraints to train the neural network model and obtain the optimized neural network model.

[0013] Step 3: Obtain the time series and location information of the in-situ seabed temperature of the nearshore observation station to be predicted, as well as the information of the matching neighboring observation points. Then, use the optimized neural network model to predict the seabed vertical temperature profile model of the nearshore observation station to be predicted.

[0014] Furthermore, the composite loss function in step 2 includes a data loss term and a physical loss term. The data loss term calculates the mean square error between the temperature prediction value and the actual temperature value of the neural network model, while the physical loss term constructs a penalty term to penalize temperature distributions that do not conform to physical laws.

[0015] Furthermore, the physical loss term is the average of the product of the temperature gradient value predicted by the neural network model and the temperature diffusion coefficient under the gradient in the vertical direction, calculated at the depth level.

[0016] This invention establishes a coupled modeling mechanism between input features and profile structure, and expresses the vertical structure in the form of layer-by-layer temperature values ​​at the output end. The modeling mechanism of "fixed-point input-fixed-layer output" enhances the model's ability to express local hydrological features and improves its generalization performance in areas with strong spatial heterogeneity.

[0017] Further, step 2 includes the following steps:

[0018] Step 2.1: Collect high-resolution ocean reanalysis temperature profile data of the nearshore area, and combine the location of the observation station with the matching of neighboring grid points to extract historical long-term temperature data and construct the first effective dataset for training the neural network model;

[0019] Step 2.2: Perform standardization preprocessing on the first valid dataset, including data denoising, outlier removal, missing measurement handling, physical quantity unit unification, and data format standardization. Then, classify the processed data according to a certain time period to construct the second valid dataset. This step also supports statistical analysis and data visualization for data distribution and quality checks.

[0020] Step 2.3: Divide the second valid dataset into training and testing sets, input them into the neural network model, train and test them according to time periods, and perform dynamic hyperparameter optimization using the selected optimizer (Adam) and composite loss function to finally obtain a dynamic temperature profile inversion model with generalization ability. This step supports cross-validation, early stopping strategy, and training process monitoring.

[0021] Furthermore, the neural network model in step 2 includes an input layer, a first hidden layer, a second hidden layer, a third hidden layer, and an output layer connected in sequence;

[0022] The input layer acquires features of the temperature time series, including seabed temperature, station coordinates, observation depth, and time information of the observation station.

[0023] The first hidden layer and the second hidden layer respectively perform batch normalization and activation operations on the acquired features. The normalization operation is used to alleviate the internal covariate offset problem, and the activation operation introduces nonlinear mapping and uses regularization terms to suppress overfitting.

[0024] The third hidden layer performs activation operations on the acquired features to further compress the feature dimensions and extract key features to support subsequent temperature profile inversion tasks, and uses regularization terms to suppress overfitting.

[0025] The output layer employs linear activation to output predicted temperature values ​​at each depth layer of the seabed at the observation station, thereby constructing a vertical temperature profile.

[0026] Furthermore, the regularization term is the addition of the squared L2 norm of the weight vector as a penalty term to the loss function.

[0027] Furthermore, in step 1, the high-frequency temperature time series on the original time scale of the seabed at the nearshore observation station is obtained, and quality control and daily average processing are performed, including the following steps:

[0028] Step 1.1: Perform preliminary quality control on the temperature data at the original time resolution, including removing outliers, filling in short-term missing values, and filtering noise to ensure the continuity and accuracy of the data;

[0029] Step 1.2: Calculate the daily average temperature profile of the seabed at the observation station; average the temperature profiles of all sampling time points within a certain natural day to calculate the daily average temperature profile and obtain the time series of the daily average temperature of the seawater at the bottom of the observation station.

[0030] Step 1.3: Use the daily average temperature time series as the effective temperature input data series for the observation station.

[0031] This invention presents a method for vertical temperature structure inversion using seabed observation data as input. It is the first to propose using in-situ real-time seabed observation data as the main input variable, replacing the traditional inversion path that relies on sea surface remote sensing data. This method overcomes the application limitations of existing technologies in near-shore obscured areas and low-precision environments, and has significant practicality and innovation.

[0032] Furthermore, in step 1, the information of the neighboring observation points is long-term continuous in-situ seabed observation data of the nearshore marine ranch area. The marine ranch is a natural ecological aquaculture system built in a specific nearshore sea area. It continuously monitors marine environmental parameters to ensure the safety of aquaculture. The seabed observation data is marine bottom environmental parameter observation data obtained through a fixed long-term in-situ seabed observation system. By making full use of the information of the neighboring observation points, local high-precision vertical temperature inversion can be achieved, making up for the deficiency of insufficient vertical characterization capability of on-site observation.

[0033] A system for in-situ observation and inversion of the vertical temperature structure of the ocean from the nearshore seabed includes an input module, a data training module, and an inversion result output module.

[0034] The input module acquires the in-situ observation temperature time series on the seabed time scale of the nearshore observation station, and extracts historical long-term seabed temperature data by combining the location information of the nearshore observation station and the information of the matched neighboring observation points.

[0035] The data training module divides the historical seabed long-term temperature data into time periods, establishes a nonlinear mapping relationship between seabed temperature and vertical temperature profile for the observation station, obtains the optimal weight for each time period, and uses a composite loss function that combines mean square error and temperature gradient physical constraints to train the neural network model, thereby obtaining an optimized neural network model.

[0036] The inversion result output module obtains the seabed in-situ observation temperature time series and location information of the nearshore observation station to be predicted, as well as the information of the matching neighboring observation points. Through the optimized neural network model, it predicts the seabed vertical temperature profile model of the nearshore observation station to be predicted.

[0037] Furthermore, the data training module employs a multilayer perceptron, which includes an input unit, a first hidden unit, a second hidden unit, a third hidden unit, and an output unit.

[0038] The input unit acquires features of the temperature time series, including seabed temperature at the observation station, station coordinates, observation depth, and time information.

[0039] The first and second hidden units respectively perform batch normalization and activation operations on the acquired features. The normalization operation is used to alleviate the internal covariate offset problem, and the activation operation introduces nonlinear mapping and uses regularization terms to suppress overfitting.

[0040] The third hidden unit performs an activation operation on the acquired features to further compress the feature dimensions and extract key features to support subsequent temperature profile inversion tasks, and uses a regularization term to suppress overfitting.

[0041] The output unit uses linear activation to output the predicted temperature values ​​at each depth layer of the seabed at the observation station, in order to construct a vertical temperature profile.

[0042] The advantages and beneficial effects of this invention are as follows:

[0043] This invention overcomes the limitations of linear assumptions by constructing a deep multilayer perceptron neural network, enabling the learning of complex nonlinear mappings between seabed observation data and complete vertical temperature profiles, thus enhancing the reconstruction capability of complex marine dynamic structures. This invention fully utilizes seabed observation data as input, making it particularly suitable for nearshore areas where traditional satellite remote sensing data has significant errors, greatly improving inversion accuracy and compensating for the shortcomings of existing technologies in these regions. The invention employs an end-to-end neural network training mode, improving the model's generalization ability and adaptive performance, allowing the model to effectively adapt to changes in different sea areas, seasons, and dynamic processes, and supporting continuous updates and optimization, making it suitable for long-term operation and practical engineering applications. This invention reduces reliance on complex empirical model construction, simplifies the inversion process, and improves inversion accuracy. This invention improves efficiency and automation, reducing the consumption of human and computational resources. Through a multi-level output structure design, it achieves high-resolution reconstruction of vertical temperature profiles, effectively enhancing the ability to analyze fine-scale ocean processes. The invention introduces a temperature diffusion equation into the loss function, overcoming the limitations of traditional data-driven network models and ensuring the physical consistency of the retrieved temperature profiles. The marine ranch seabed observation data used in this invention has high-frequency acquisition characteristics. When the original time-series data is directly used as model input, it can effectively support the generation of high-temporal-resolution temperature profile retrieval results. The high-frequency sampling characteristics overcome the inherent time-level resolution limitations of traditional satellite and other remote sensing methods, potentially enabling the refined capture and reproduction of hourly or even minute-level changes in vertical temperature structure. Attached Figure Description

[0044] Figure 1 This is a diagram illustrating the system architecture and execution process in an embodiment of the present invention.

[0045] Figure 2 This is a flowchart of the method in an embodiment of the present invention.

[0046] Figure 3This is a map showing the distribution of ocean temperature field around the observation station in this embodiment of the invention.

[0047] Figure 4 This is a comparison diagram of the training process of the mean square error and mean absolute error models corresponding to the observation stations in this embodiment of the invention.

[0048] Figure 5 This is a comparison diagram of the original temperature profile and the inverted temperature profile in an embodiment of the present invention.

[0049] Figure 6 This is a vertical cross-sectional comparison diagram of the original average temperature and the inverted average temperature based on time in an embodiment of the present invention. Detailed Implementation

[0050] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0051] To overcome the limitations of existing statistical regression methods and remote sensing inversion techniques in nearshore areas and complex marine environments, such as weak nonlinear modeling capabilities, poor generalization ability, limited resolution, and insufficient utilization of nearshore observation data, this invention proposes a system for inverting the vertical temperature structure of the ocean through in-situ observation of the nearshore seabed. The vertical temperature structure refers to the distribution characteristics of seawater temperature along the vertical direction (water depth), and is a key parameter affecting marine physical and ecological processes. This invention does not rely on sea surface remote sensing data and is particularly suitable for nearshore areas where sea surface remote sensing accuracy is insufficient or missing, but continuous in-situ seabed observation data is available, such as nearshore marine ranching areas, nearshore marine environmental monitoring stations, and areas with offshore wind power and oil platforms. These stations or areas often deploy in-situ marine environmental monitoring systems on the ocean floor, enabling long-term, real-time, and stable acquisition of continuous observation time series of key ocean floor elements, such as depth, temperature, salinity, and current fields. The ocean internal temperature inversion system of the present invention has the advantages of high accuracy, high efficiency and flexible deployment. It is suitable for a variety of marine environmental application scenarios such as nearshore marine environmental monitoring and early warning, digital monitoring and management of marine ranch environment, and underwater operation support for engineering facilities such as offshore wind power and oil platforms, which require high dynamics and high resolution.

[0052] This invention utilizes historical reanalysis temperature profile data to train a Multilayer Perceptron (MLP) model. By constructing a nonlinear mapping model and leveraging the neural network of the MLP model, it establishes a nonlinear mapping relationship between seabed observation data (such as seabed temperature and water depth) and the temperature structure of the entire water column (such as the temperature structure of a complete water column), thereby improving the ability to represent complex ocean processes. A multilayer perceptron is a feedforward artificial neural network consisting of at least three layers (input layer, hidden layer, and output layer), which approximates complex functions through nonlinear activation functions. This invention enhances the model's adaptability to dynamic environmental changes by introducing an incremental training mechanism; during the inference phase, only the observed seabed temperature data needs to be input to quickly output a high-resolution vertical temperature profile.

[0053] The system of this invention has advantages such as strong nonlinear expression capability, good generalization performance, and high inference efficiency, such as... Figure 1 As shown, it specifically includes a data input module, a data training module, and an inversion result output module.

[0054] Data Input Module: Acquires continuous in-situ seabed observation data from the target site, including parameters such as seawater temperature, depth, and latitude and longitude. The raw observation data is standardized and normalized to construct feature vectors for model input. By fully integrating nearshore station observation information, and effectively utilizing long-term continuous in-situ high-precision seabed observation data from nearshore marine ranching areas, local high-precision vertical temperature inversion is achieved, compensating for the shortcomings of in-situ observations in vertical characterization. Marine ranching refers to natural ecological aquaculture systems constructed in specific sea areas using modern engineering technology, requiring continuous monitoring of marine environmental parameters to ensure aquaculture safety. Seabed observation data refers to marine bottom environmental parameter observation data acquired through a fixed long-term in-situ seabed observation system (marine environmental element sensor array).

[0055] Data Training Module: Based on a pre-trained MLP neural network model, a nonlinear mapping relationship is established between seabed observation data and vertical temperature profiles at target sites. Through a neural network model trained on historical multi-source profile data, accurate predictions of temperature structures under different regions, seasons, and dynamic processes are achieved, adapting to constantly changing observation environments and data structures, and improving the model's generalization ability and adaptability. Based on the neural network model, the expressive power limitations of traditional empirical function expansion methods are overcome, enhancing the ability to characterize the vertical structure of the ocean. This makes it suitable for refined ocean process research and nearshore environmental monitoring, improving vertical resolution and structure reconstruction capabilities.

[0056] The inversion result output module takes the seabed temperature time series of the target site as model input and outputs the temperature information of the corresponding depth layer at each time point of the target site, realizing high-precision inversion of the complete vertical temperature profile. This MLP neural network model can be customized for training and deployment according to the characteristics of different sites, and has good generalization ability and real-time adaptability.

[0057] like Figure 2 As shown, this invention also proposes a method for inverting the vertical temperature structure of the ocean based on in-situ observation of the nearshore seabed using the above-mentioned system, comprising the following steps:

[0058] Step 1: Obtain the in-situ temperature time series of the seabed at the nearshore observation station on the seabed time scale.

[0059] Specifically, the bottom temperature observation data of the target site is collected and preprocessed. Long-term in-situ seabed temperature data is usually collected continuously at a minute-level time resolution. High-frequency temperature data of the site is obtained to reflect short-term changes in seawater temperature. Based on the elimination of periodic changes in tidal frequency and the fact that daily periodic changes can generally meet the needs of marine environmental monitoring and forecasting, the raw minute-scale high-frequency temperature time series data needs to be quality controlled and processed with daily average values. The specific steps are as follows:

[0060] Step 1.1: Perform preliminary quality control on the raw minute-resolution temperature data, including removing outliers, filling in short-term missing values, and noise filtering, to ensure the continuity and accuracy of the data.

[0061] Step 1.2: Calculate the daily average temperature profile; for a given natural day, average the temperature profiles at all sampling time points to calculate the daily average temperature profile:

[0062]

[0063] in, This represents the temperature value at the i-th sampling time, and N represents the number of valid samples within that day. This indicates the average temperature.

[0064] Step 1.3: The generated daily average bottom seawater temperature time series is the effective temperature input data series for the target station.

[0065] This invention presents a method for vertical temperature structure inversion using seabed observation data as input. It is the first to propose using in-situ real-time seabed observation data as the main input variable, replacing the traditional inversion path that relies on sea surface remote sensing data. This method overcomes the application limitations of existing technologies in near-shore obscured areas and low-precision environments, and has significant practicality and innovation.

[0066] Step 2: Construct a neural network model. Extract historical long-term seabed temperature data from the in-situ temperature time series and location information of nearshore observation stations and the information of matched neighboring observation points. Divide the data into time periods and establish a nonlinear mapping relationship between seabed temperature and vertical temperature profile for the observation stations to obtain the optimal weights for each time period. Use a composite loss function that combines mean square error and temperature gradient physical constraints to train the neural network model and obtain the optimized neural network model.

[0067] Specifically, based on the MLP neural network model, using historical ocean reanalysis datasets, the model is trained month by month to construct a nonlinear mapping relationship between bottom seawater temperature and vertical temperature profile at the target site, thereby obtaining the optimal weights of the model for each month. The mean square error and temperature diffusion equation are used as loss functions for data and physical information, and the Adam optimizer is used to perform multiple rounds of dynamic training iterations on the MLP neural network model to collaboratively optimize the model parameters, obtain the dynamically optimal model weight parameters, and finally form the optimized MLP model for the target site.

[0068] This invention is based on the nonlinear profile inversion model structure and training mechanism of MLP, and designs a multilayer perceptron neural network framework suitable for temperature profile inversion. It uses deep neural networks to learn the nonlinear mapping relationship between seabed point observation data and the temperature distribution of the entire water body. This framework supports the direct output of multi-level, depth-by-depth vertical temperature structure, which significantly improves the inversion accuracy and model adaptability.

[0069] Specifically, based on historical ocean reanalysis data, an MLP neural network model is trained and iteratively optimized for the target station to construct a nonlinear mapping relationship between bottom seawater temperature and the vertical temperature profile of the station, including the following steps:

[0070] Step 2.1: Collect high-resolution ocean reanalysis temperature profile data of the nearshore area, and combine the location of the target observation point with the matching of neighboring grid points to extract historical long-term temperature data, constructing the first effective dataset for training the model. This step supports automatic access to a specified ocean reanalysis database, such as the GLORYS (Global Ocean Reanalysis and Simulation) ocean reanalysis data product obtained from the Copernicus Marine Environment Monitoring Service (CMEMS), and implements coordinate matching and data extraction functions. The ocean reanalysis data is a spatiotemporally continuous ocean environment dataset generated by fusing numerical models and observational data, used to reproduce the basic state of the historical ocean and support scientific research and applications.

[0071] Step 2.2: Perform standardization preprocessing on the first valid dataset, including data denoising, outlier removal, missing value handling, unit standardization of physical quantities, and data format normalization. Then, categorize the processed data by month to construct the second valid dataset. This step also supports statistical analysis and data visualization for data distribution and quality checks.

[0072] Step 2.3: Input the completed training dataset into the Multilayer Perceptron (MLP) model. Divide the dataset into training and test sets according to a set ratio (e.g., 8:2), train the model by month, and perform dynamic hyperparameter optimization using the selected optimizer (Adam) and loss function. The final output is a dynamic temperature profile inversion model with generalization ability. This step supports cross-validation, early stopping strategy, and training process monitoring.

[0073] The neural network structure of a multilayer perceptron (MLP) model includes an input layer, a first hidden layer, a second hidden layer, a third hidden layer, and an output layer. The specific execution process is as follows:

[0074] Input layer: Acquires input features, specifically including seabed temperature, station latitude, longitude, observation depth, and time information such as year, month, and day. All input features are standardized before input to improve the stability and convergence speed of model training.

[0075] First hidden layer (first layer after input): The input layer is followed by a fully connected hidden layer containing 128 neurons. This layer first performs batch normalization to adjust the feature distribution and alleviate the internal covariate shift problem; then, ReLU (Rectified Linear Unit) is used as the activation function to introduce non-linear mapping capability. All parameters in this layer incorporate L2 regularization terms to suppress overfitting. L2 regularization refers to adding a penalty term equal to the squared L2 norm of the weight vector to the loss function. The regularized loss function is:

[0076]

[0077] in, Represents the original loss function. The regularization coefficient controls the strength of regularization. It is the weight vector in the model.

[0078] The second hidden layer is a fully connected layer containing 64 neurons. It also undergoes batch normalization before input and uses the ReLU activation function for non-linear transformation. This layer also incorporates an L2 regularization term to improve the model's generalization ability and stability.

[0079] The third hidden layer is a fully connected layer containing 32 neurons. This layer uses the ReLU activation function but does not perform batch normalization. Its purpose is to further compress the feature dimension and extract key features to support subsequent temperature profile inversion tasks. This layer also uses L2 regularization constraints.

[0080] Output Layer: The output layer is a linearly activated fully connected layer with the number of neurons equal to the set standard depth layer number, which is 102 layers in this embodiment. This layer directly outputs the predicted temperature value at each depth layer, forming a complete vertical temperature profile. The output results are denormalized after the model calculation is completed to restore the true physical quantity (unit: degrees Celsius).

[0081] The training and optimization of the Multilayer Perceptron (MLP) model employs a composite loss function combining Mean Squared Error (MSE) and temperature gradient physical constraints. The Adam algorithm is used as the optimizer, with an initial learning rate of 0.001. During model training, a typical supervised learning workflow is employed, involving training, validation, and test set splitting on historical profile data (8:2). The training batch size is set to 32 based on memory and sample size, and the training epochs are typically 100. Early stopping is used to control overfitting risk. The complete loss function definition is as follows:

[0082]

[0083] in, The data loss term representing the model is calculated using MSE to determine the error between the model's predicted values ​​and the actual values:

[0084]

[0085] in, This represents the temperature value predicted by the model. This represents the target ground truth value for model training, and N represents the number of training samples.

[0086] in, The physical loss term representing the model is used to enhance the model's generalization ability and structural rationality by penalizing temperature distributions that do not conform to physical laws.

[0087]

[0088] in, This represents the temperature gradient value predicted by the model. Represents the temperature diffusivity. This represents the vertical depth layer number.

[0089] This invention establishes a coupled modeling mechanism between input features and profile structure. At the input end, a feature vector containing multiple elements such as seabed temperature and station location (latitude and longitude) is constructed, and at the output end, the vertical structure is expressed as layer-by-layer temperature values. This "fixed-point input - fixed-layer output" modeling mechanism enhances the model's ability to express local hydrological characteristics and improves its generalization performance in areas with strong spatial heterogeneity.

[0090] In this model, by constructing and training a neural network that incorporates physical information, the temperature diffusion equation is introduced not only as prior knowledge but also as a mandatory physical constraint. By constructing it as a physical residual term in the loss function, the network is guided to simultaneously fit observed data and follow physical laws during training, thereby overcoming the limitations of purely data-driven approaches and ensuring that the retrieved temperature profile is physically reliable.

[0091] In another embodiment, regarding the model architecture, in addition to multilayer perceptrons, structures such as convolutional neural networks, recurrent neural networks, long short-term memory networks, or Transformers can also be used to construct a nonlinear mapping relationship between seabed observation data and vertical temperature profiles, thereby enhancing the model's ability to extract complex features and its generalization performance.

[0092] Secondly, at the system design level, a multi-input multi-output architecture can be adopted to integrate multi-source observation variables, or an ensemble learning and model fusion strategy can be introduced to further improve the stability and accuracy of the inversion results.

[0093] Furthermore, in response to the heterogeneity of observation data from different sea areas, methods such as transfer learning and federated learning can serve as effective training alternatives, enabling the effective application and knowledge sharing of models in cross-regional scenarios while ensuring data security.

[0094] In addition, alternative feature engineering techniques such as autoencoder dimensionality reduction and principal component analysis can be used in the data preprocessing stage to optimize the quality of input data and improve the learning efficiency and performance of the model.

[0095] It should be noted that although this invention is constructed and verified based on seabed observation data from marine ranches, its methodological framework has good portability. When applied to other sites or scenarios, it only needs to be modeled and adapted to the specific observation system and environmental characteristics.

[0096] Step 3: Obtain the time series and location information of the in-situ seabed temperature of the nearshore observation station to be predicted, as well as the information of the matching neighboring observation points. Then, use the optimized neural network model to predict the seabed vertical temperature profile model of the nearshore observation station to be predicted.

[0097] Specifically, the time series of seawater temperature observations at the bottom of the target station and basic station information are input into the optimized MLP model for that station. The model then calculates and outputs the vertical temperature profile of the target station. This invention, as a model application for observation station deployment, proposes a model scheme that allows for customized training and deployment for specific observation stations, mooring areas, and other practical application environments. It is suitable for scenarios such as long-term continuous operation, edge-end inference, and nearshore observation grids. This mechanism supports efficient model operation in resource-constrained environments and has good engineering feasibility and promotional value.

[0098] To verify the feasibility of this invention, the seabed temperature data for August 2022 obtained by a seabed observation system at a marine ranching demonstration area (location: 120.9167°E, 37.8333°N, water depth: 17m) was used as input, and the ocean reanalysis dataset was used as model training and testing data. The test results for this site are presented, such as... Figure 3 , Figure 4 As shown, the mean squared error loss function used in the training process significantly outperforms the mean absolute error loss used in training. Figure 5 , Figure 6 As shown, the temperature profile inverted by the neural network model that introduces physical information in this invention can well reflect the original temperature profile, and the average deviation between the vertical profile of the time-based inverted average temperature and the vertical profile of the original average temperature is only 0.011℃.

[0099] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for inverting the vertical temperature structure of the ocean through in-situ observation of the nearshore seabed, characterized in that... Includes the following steps: Step 1: Obtain the in-situ temperature time series on the seabed time scale of the nearshore observation station. The nearshore observation station includes, but is not limited to, one or more of the following: nearshore ranching areas, nearshore environmental monitoring stations, offshore wind power and oil platforms. Step 2: Construct a neural network model. Historical long-term seabed temperature data is extracted from in-situ temperature time series and location information of nearshore observation stations, along with information from matched neighboring observation points. This data is divided into time periods, and a nonlinear mapping relationship between seabed temperature and vertical temperature profiles is established for each observation station to obtain the optimal weights for each time period. A composite loss function combining mean square error and physical constraints on temperature gradients is used to train the neural network model, resulting in an optimized model. The composite loss function includes a data loss term and a physical loss term. The data loss term calculates the mean square error between the temperature predictions and actual temperatures of the neural network model. The physical loss term penalizes temperature distributions that do not conform to physical laws by constructing a penalty term. The physical loss term is the square of the product of the temperature gradient value predicted by the neural network model and the temperature diffusion coefficient under the vertical gradient, calculated at the depth level. Step 3: Obtain the time series and location information of the in-situ seabed temperature of the nearshore observation station to be predicted, as well as the information of the matching neighboring observation points. Then, use the optimized neural network model to predict the seabed vertical temperature profile model of the nearshore observation station to be predicted.

2. The method for inverting the vertical temperature structure of the ocean through in-situ observation of the nearshore seabed according to claim 1, characterized in that: Step 2 includes the following steps: Step 2.1: Collect high-resolution ocean reanalysis temperature profile data of the nearshore area, and combine the location of the observation station with the matching of neighboring grid points to extract historical long-term temperature data and construct the first effective dataset for training the neural network model; Step 2.2: Perform standardization preprocessing on the first valid dataset, and classify the processed data according to a certain time period to construct the second valid dataset; Step 2.3: Divide the second effective dataset into a training set and a test set, input the neural network model, train and test it according to time periods, and perform dynamic hyperparameter optimization with the selected optimizer and composite loss function to finally obtain the dynamic temperature profile inversion model.

3. The method for inverting the vertical temperature structure of the ocean through in-situ observation of the nearshore seabed according to claim 1, characterized in that: The neural network model in step 2 includes an input layer, a first hidden layer, a second hidden layer, a third hidden layer, and an output layer connected in sequence. The input layer acquires features of the temperature time series, including seabed temperature, station coordinates, observation depth, and time information of the observation station. The first hidden layer and the second hidden layer respectively perform batch normalization and activation operations on the acquired features, and use regularization terms to suppress overfitting; The third hidden layer performs activation operations on the acquired features to further compress the feature dimensions and extract key features, and uses regularization terms to suppress overfitting. The output layer employs linear activation to output predicted temperature values ​​at each depth layer of the seabed at the observation station, thereby constructing a vertical temperature profile.

4. The method for inverting the vertical temperature structure of the ocean through in-situ observation of the nearshore seabed according to claim 3, characterized in that: The regularization term is the addition of the L2 norm squared weight vector as a penalty term to the loss function.

5. The method for inverting the vertical temperature structure of the ocean through in-situ observation of the nearshore seabed according to claim 1, characterized in that: Step 1 involves acquiring high-frequency temperature time series on the original timescale of the seabed at nearshore observation stations, and performing quality control and daily average processing, including the following steps: Step 1.1: Perform preliminary quality control on the temperature data at the original time resolution; Step 1.2: Calculate the daily average temperature profile of the seabed at the observation station; average the temperature profiles of all sampling time points within a certain natural day to calculate the daily average temperature profile and obtain the time series of the daily average temperature of the seawater at the bottom of the observation station. Step 1.3: Use the daily average temperature time series as the effective temperature input data series for the observation station.

6. The method for inverting the vertical temperature structure of the ocean through in-situ observation of the nearshore seabed according to claim 1, characterized in that: In step 1, the neighboring observation point information is long-term continuous in-situ seabed observation data of the nearshore marine ranch area. The marine ranch is a natural ecological aquaculture system built in a specific nearshore sea area, which continuously monitors marine environmental parameters. The seabed observation data is marine bottom environmental parameter observation data obtained through a fixed long-term in-situ seabed observation system.

7. A system for inverting the vertical temperature structure of the ocean through in-situ observation of the nearshore seabed, comprising an input module, a data training module, and an inversion result output module, characterized in that: The method for retrieving ocean vertical temperature structure by in-situ observation of nearshore seabed as described in claim 1; The input module acquires the in-situ observation temperature time series on the seabed time scale of the nearshore observation station, and extracts historical long-term seabed temperature data by combining the location information of the nearshore observation station and the information of the matched neighboring observation points. The data training module divides the historical seabed long-term temperature data into time periods, establishes a nonlinear mapping relationship between seabed temperature and vertical temperature profile for the observation station, obtains the optimal weight for each time period, and uses a composite loss function that combines mean square error and temperature gradient physical constraints to train the neural network model, thereby obtaining an optimized neural network model. The inversion result output module obtains the seabed in-situ observation temperature time series and location information of the nearshore observation station to be predicted, as well as the information of the matching neighboring observation points. Through the optimized neural network model, it predicts the seabed vertical temperature profile model of the nearshore observation station to be predicted.

8. The system for in-situ observation and inversion of ocean vertical temperature structure from the nearshore seabed according to claim 7, characterized in that: The data training module employs a multilayer perceptron, which includes an input unit, a first hidden unit, a second hidden unit, a third hidden unit, and an output unit. The input unit acquires features of the temperature time series, including seabed temperature at the observation station, station coordinates, observation depth, and time information. The first and second hidden units respectively perform batch normalization and activation operations on the acquired features, and use regularization terms to suppress overfitting. The third hidden unit performs an activation operation on the acquired features to further compress the feature dimensions and extract key features, and uses a regularization term to suppress overfitting. The output unit uses linear activation to output the predicted temperature values ​​at each depth layer of the seabed at the observation station, in order to construct a vertical temperature profile.

Images