A method for retrieving ocean salinity vertical profile by matching brightness temperature and buoy in space and time

Abstract

The present application relates to a kind of methods for matching the space-time of brightness temperature-buoy and the vertical profile of ocean salinity, which is matched by obtaining and satellite brightness temperature and buoy salinity profile data, construct space-time matching scatter data set, respectively standardize and save parameters to feature and label. Improved Transformer model is constructed, the input embedding layer projects the feature into high-dimensional vector, the encoder encodes the feature, and the layered output head is predicted by three parallel sub-networks respectively to predict the middle and deep layer salinity, and spliced into complete profile. The normalized feature to be inverted is input into the trained model, and the predicted value is de-normalized to obtain the vertical salinity profile with physical units. The present application realizes the inversion of ocean vertical salinity structure from surface signal by time cycle coding and layered output head design, and solves the inverse problem of radiation transfer by deep learning.

Description

Technical Field

[0001] This invention relates to the field of marine remote sensing and marine environment numerical inversion technology, specifically to a method for inverting marine salinity vertical profiles using a scatter plot method with brightness temperature-buoy spatiotemporal matching. Background Technology

[0002] Ocean salinity is a key state parameter describing the physical and chemical properties of seawater, and it is fundamental to processes such as ocean circulation, water mass distribution, stratification stability, global mass and energy transport, and air-sea interactions. Therefore, obtaining high-precision, high-spatiotemporal resolution vertical profiles of ocean salinity is crucial for marine scientific research, climate prediction, fisheries resource management, and environmental protection.

[0003] Currently, obtaining large-scale ocean salinity information mainly relies on two technical approaches: one is direct measurement based on on-site observation, and the other is indirect inversion based on satellite remote sensing.

[0004] In terms of in-situ observation, the global network of automated profiling buoys can provide high-quality temperature and salinity profile data at discrete points from the sea surface to a depth of approximately 2,000 meters. However, its observations are discontinuous and sparse in spatiotemporal distribution, making it difficult to directly generate continuous salinity field products that meet the needs of regional or global dynamic process research.

[0005] In satellite remote sensing, L-band passive microwave radiometers (such as the ESA's SMOS, NASA's Aquarius / SAC-D, and SMAP satellites) can effectively retrieve sea surface salinity by measuring sea surface brightness temperature. This technology enables near real-time and quasi-continuous observation of sea surface salinity over a wide area, greatly compensating for the spatial coverage limitations of on-site observations. However, this type of satellite remote sensing technology has limitations: its observation signals mainly come from the sea surface and a very thin layer below it, and it cannot directly penetrate the water body to detect salinity information in the subsurface and deeper layers.

[0006] To overcome the aforementioned limitations, existing technologies attempt to combine sea surface remote sensing information with sparse field profile data through data assimilation, statistical models, or machine learning methods to infer the salinity structure of the subsurface and even deeper layers. However, these methods generally suffer from the following drawbacks:

[0007] 1. Most methods rely on spatiotemporally gridding (interpolating or reanalyzing) the raw, irregularly distributed satellite brightness temperature and buoy profile data to generate regular grid data for modeling. This preprocessing smooths out local, instantaneous, fine-scale dynamic features, introduces interpolation errors, and causes the model to learn distorted "mean field" information rather than true "point-to-point" correlations.

[0008] 2. Traditional statistical methods (such as empirical orthogonal function regression and multiple linear regression) or early shallow neural network models typically only establish inversion relationships for a single depth layer (especially the sea surface layer), and cannot invert the entire vertical profile in one go and in a coordinated manner. Even when attempting to model by depth, the interlayer physical connections and continuity constraints of salinity in the vertical direction are ignored.

[0009] 3. Some methods use monthly or climatological average data as training sets, which means that the model learns statistical relationships in a climatological sense, but cannot capture and invert instantaneous salinity changes on daily to weekly time scales driven by synoptic or mesoscale processes, resulting in insufficient accuracy of the model in reflecting the real instantaneous state of the ocean.

[0010] 4. When using neural networks for multi-depth joint inversion, if an equal mean squared error loss is adopted, the model will overemphasize the deep layers because the deep salinity changes gradually and is relatively easy to predict. This will cause the optimization direction of the entire model to be biased towards fitting the deep layer mean, thereby ignoring the surface salinity details that change more dramatically and are more sensitive to dynamic processes, resulting in the phenomenon of "mean degradation".

[0011] Therefore, there is an urgent need in this field for a novel salinity profile inversion method that can directly utilize raw, unmeshed matching data, uniformly model salinity vertical profiles, maintain consistency in instantaneous matching scales, and embed prior physical knowledge of ocean vertical changes into the model structure and training strategy, in order to overcome the above-mentioned shortcomings of existing technologies and improve the accuracy and physical rationality of inverting full-profile salinity from sea surface remote sensing information. Summary of the Invention

[0012] This invention provides a method for inverting ocean salinity vertical profiles using a scatter plot with brightness temperature-buoy spatiotemporal matching. The purpose is to solve the problems of insufficient accuracy and lack of physical rationality in ocean salinity profile inversion caused by traditional methods, which rely on gridded data, cannot collaboratively invert complete vertical profiles, ignore instantaneous matching relationships, and lack physical guidance.

[0013] To achieve the above objectives, the first aspect of the present invention provides a method for retrieving vertical ocean salinity profiles by scatter point inversion using brightness temperature-buoy spatiotemporal matching, comprising the following steps:

[0014] Satellite brightness temperature data and on-site buoy salinity profile data are acquired and spatiotemporally matched to construct a spatiotemporally matched scatter dataset containing input feature vectors and label vectors. The input feature vectors include time, space and physical features, and the label vectors include salinity values ​​corresponding to multiple target depths.

[0015] The input feature vector and label vector in the spatiotemporal matching scatter dataset are standardized to obtain standardized input feature vector and standardized label vector, and the corresponding standardization parameters are saved.

[0016] Construct and train an improved Transformer neural network model, which includes:

[0017] An input embedding layer is used to receive the standardized input feature vector and project it into a high-dimensional embedding vector.

[0018] A Transformer encoder is used to perform feature encoding on the high-dimensional embedding vector;

[0019] The layered output header includes three independent sub-headers that are set in parallel and correspond to the surface, middle and deep salinity predictions, respectively. These sub-headers are used to predict the normalized salinity values ​​of each depth layer based on the output of the Transformer encoder and then splice them together to form a complete normalized salinity vertical profile prediction vector.

[0020] The input feature vector, which is standardized in the same way and corresponds to the data to be inverted, is input into the trained improved Transformer neural network model to obtain the standardized salinity vertical profile prediction value.

[0021] By using the saved standardized parameters corresponding to the label vector, the obtained standardized salinity vertical profile prediction values ​​are inversely normalized to obtain salinity vertical profile inversion results with physical units.

[0022] Furthermore, methods for spatiotemporal matching include:

[0023] For each on-site buoy profile, based on its buoy profile center time and geographical location, find the closest matching time in the satellite observation time set that satisfies data validity.

[0024] The matching time is determined by the following formula:

[0025]

[0026] in, For matching time; Center time; For satellite observation time sets, These represent longitude and latitude, respectively. Represents the set of satellite observation times One of the candidate times; Indicates time Geographical location Satellite brightness temperature observation vector at the location; It is a verification function used to determine whether the brightness temperature observation vector has no missing data and passes the quality control flag;

[0027] If no satisfactory time is found within the preset time window of the buoy profile center time, If valid satellite observations are available, the buoy profile sample should be discarded.

[0028] Furthermore, the input embedding layer performs the following operations to project the normalized input feature vector into a high-dimensional embedding vector:

[0029] Dimension is The input feature vector is regarded as being composed of A sequence consisting of feature patches, wherein For batch size, Input feature dimension;

[0030] Each feature patch is projected into a high-dimensional embedding space through a one-dimensional convolutional layer. The kernel size of the one-dimensional convolutional layer is 1, and the number of output channels is [missing information]. Thus, the initial embedding tensor is obtained. ;

[0031] Generate a learnable positional encoding parameter that matches the size of the initial embedding tensor;

[0032] The learnable position encoding parameters are added element-wise to the initial embedding tensor to obtain a high-dimensional embedding vector containing order information.

[0033] Furthermore, the Transformer encoder performs the following operations to perform feature encoding on the high-dimensional embedding vector:

[0034] The high-dimensional embedding vector is input into an encoder consisting of stacked L layers of coding blocks, where L is an integer greater than 1;

[0035] Each layer of coded blocks performs the following operations in sequence:

[0036] The input tensor is layer normalized, processed by a multi-head self-attention mechanism, and the processed result is discarded through a random path operation and added to the input tensor residual to obtain the first intermediate feature.

[0037] The first intermediate feature is normalized by a layer and processed by a feedforward network. The processed result is discarded through a random path and added to the residual of the first intermediate feature to obtain the output feature of the coding block.

[0038] The feedforward network adopts a convolutional feedforward network structure, using one-dimensional convolutional layers to replace traditional fully connected layers to achieve feature transformation.

[0039] Furthermore, the layered output header performs the following operations to predict the normalized salinity values ​​for each depth layer based on the output of the Transformer encoder, and then concatenates them to form a complete normalized salinity vertical profile prediction vector:

[0040] The feature tensor output by the Transformer encoder is flattened to obtain a high-dimensional feature vector.

[0041] The high-dimensional feature vector is simultaneously input into the heads of three parallel and independent multilayer perceptrons:

[0042] The first subheading is the surface head, which is used to predict the salinity values ​​of the first set of depth layers most affected by surface dynamics.

[0043] The second subhead is the middle head, which is used to predict the salinity values ​​of the second set of depth layers near the thermocline;

[0044] The third subhead is the deep head, which is used to predict the salinity values ​​of the relatively stable third group of depth layers.

[0045] The prediction vectors output from the first, second, and third sub-headers are concatenated along the depth dimension to form a complete standardized salinity vertical profile prediction vector.

[0046] Furthermore, methods for training improved Transformer neural network models include:

[0047] The spatiotemporal matching scatter dataset is divided into a training set and a validation set in chronological order.

[0048] The model parameters are iteratively optimized using the training set, and the model performance is monitored on the validation set.

[0049] The goal of model optimization is to minimize the depth-weighted mean squared error loss function, which is calculated using the following formula:

[0050]

[0051] in, The loss function; For batch size, The total number of layers at the target depth. Indicates the index of the depth layer. and The first The sample at the th Standardized true salinity values ​​and standardized predicted salinity values ​​at each depth layer For the preset first The weights for each depth layer are set according to the vertical variation characteristics of ocean salinity, with the surface depth weight being higher than the deep depth weight.

[0052] The AdamW optimizer is used to update the model parameters, and a cosine annealing learning rate scheduler is used to dynamically adjust the learning rate for each iteration.

[0053] During training, a copy of the model parameters that minimizes the loss on the validation set is continuously saved.

[0054] If the loss on the validation set does not decrease within a set number of consecutive training cycles, an early stopping mechanism is triggered to terminate the training process.

[0055] Furthermore, methods for inverse normalization of the obtained standardized salinity vertical profile predictions include:

[0056] Load the label normalization parameter file saved during the training phase;

[0057] Using the scaling parameters stored in the label standardization parameter file, the predicted value of the standardized salinity vertical profile is inversely normalized.

[0058] The inverse normalization calculation is applied simultaneously to the standardized predicted values ​​output by the model and the standardized true values ​​corresponding to the test set, in order to obtain the predicted and true values ​​of the salinity vertical profile with real physical units.

[0059] Furthermore, after obtaining the salinity vertical profile inversion results in physical units, the method also includes a step of evaluating the model inversion performance, the evaluation step including:

[0060] Instantaneous profile performance analysis: Directly compare the predicted salinity value of each scatter sample in the inversion results with the corresponding true salinity value, and calculate the mean absolute error, root mean square error and coefficient of determination for each target depth layer respectively;

[0061] Monthly-scale aggregate evaluation: Grouping samples using the month labels attached to each sample when constructing the dataset, calculating the absolute error between the true average salinity and the predicted average salinity within each month group and at each depth layer, and using this error as the model's performance metric on a macro-monthly scale.

[0062] To achieve the above objectives, a second aspect of the present invention provides an electronic device including a memory and a processor, the memory being used to store a program that supports the processor in executing the brightness temperature-buoy spatiotemporal matching scatter inversion method for vertical ocean salinity profiling, and the processor being configured to execute the program stored in the memory.

[0063] To achieve the above objectives, a third aspect of the present invention provides a computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, performs the steps of the brightness temperature-buoy spatiotemporal matching scatter inversion method for vertical ocean salinity profile.

[0064] The beneficial effects of this invention are:

[0065] Compared with existing technologies, this invention provides a method for retrieving ocean salinity vertical profiles by using a scatter plot of brightness temperature-buoy spatiotemporal matching. This method directly utilizes raw, non-gridized instantaneous matching scatter plot data of the "satellite brightness temperature-buoy profile" as the training basis, avoiding the information smoothing and errors introduced by gridded interpolation. By constructing an improved Transformer model with embedded physical priors, its encoder captures complex correlations between features. Furthermore, it employs a "layered output head" that conforms to the vertical structure characteristics of the ocean, decomposing the single, complex full-profile regression task into three physically meaningful parallel sub-tasks—surface, mid-depth, and deep—for collaborative inversion, achieving unified modeling of the salinity vertical profile. Further, by designing a "depth-weighted loss function," higher weights are assigned to the drastically changing surface layer during training, forcing the model to focus on learning surface details and effectively suppressing the "mean degradation" tendency caused by the gradual changes in the deeper layers. Finally, this method significantly improves the accuracy and physical plausibility of retrieving multi-depth salinity profiles from sea surface brightness temperature while maintaining consistency in instantaneous matching scale. Attached Figure Description

[0066] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below.

[0067] Figure 1 This is a flowchart of a method for retrieving vertical ocean salinity profiles by scatter point inversion using brightness temperature-buoy spatiotemporal matching, as disclosed in an embodiment of the present invention.

[0068] Figure 2 This is a schematic diagram of the physical principle of microwave brightness temperature inversion of ocean vertical salinity structure based on layered medium radiative transmission disclosed in an embodiment of the present invention.

[0069] Figure 3 This is an example diagram of the average brightness temperature level distribution in May 2017, as disclosed in an embodiment of the present invention.

[0070] Figure 4 This is a model inversion plotting diagram of underwater salinity levels at multiple depths in May 2017, obtained according to an embodiment of the present invention. Figure 4(a)-(e) in the table are the salinity prediction statistics for (a) 5 meters (mean 33.61, standard deviation 0.23), (b) 20 meters (33.69, 0.18), (c) 50 meters (34.01, 0.19), (d) 100 meters (34.473, 0.065), and (e) 200 meters (34.5452, 0.0060), respectively.

[0071] Figure 5 This is a predicted effect diagram of a vertical cross-section in May 2017, as disclosed in an embodiment of the present invention. Figure 5 (a)-(d) are salinity profiles for longitudes of 111.00°, 113.00°, 115.00°, and 117.00°, respectively. Detailed Implementation

[0072] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0073] According to embodiments of the present invention, it should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the following methods, in some cases the steps shown or described may be executed in a different order than that shown here.

[0074] like Figure 1 As shown, this invention provides a method for retrieving vertical ocean salinity profiles using a scatter plot method based on brightness temperature-buoy spatiotemporal matching, comprising the following steps:

[0075] Step S100: Obtain satellite brightness temperature data and on-site buoy salinity profile data, and perform spatiotemporal matching to construct a spatiotemporal matching scatter dataset containing input feature vectors and label vectors. The input feature vectors include time, space and physical features, and the label vectors include salinity values ​​corresponding to multiple target depths.

[0076] This step avoids the information smoothing and error introduction caused by the gridding interpolation in traditional methods, and directly uses the original observations that are successfully "paired" between the satellite and the buoy at a specific time and space point.

[0077] First, raw data from two sources is acquired. One is point-by-point brightness temperature data from L-band passive microwave remote sensing satellites (such as the European Space Agency's Soil Moisture and Ocean Salinity Satellite, SMOS). The method of this invention is based on the principle of microwave radiative transfer. According to the radiative transfer equation, the brightness temperature observed by the satellite ( The dielectric constant of the sea surface is a function of sea surface emissivity, physical temperature, and atmospheric contributions, directly containing information about the sea surface dielectric constant, which is a function of salinity and temperature. Therefore, It is closely related to sea surface salinity (SSS). Brightness temperature is a measure of the brightness of microwave radiation emitted from the Earth's surface or sea surface as received by a satellite, measured in Kelvin (K). Its value is related to various physical parameters such as sea surface salinity and temperature. Although the penetration depth of L-band microwaves is only on the centimeter level, the salinity structure of deep seawater is not independent of the surface state, but is tightly coupled through vertical mixing processes and seasonal stratification, and is controlled by seasonal solar radiation and air-sea flux exchange. Secondly, there is in-situ profiling data from a global network of automatic profiling buoys (such as Argo), which provides high-precision measurements of temperature, salinity, pressure, etc., at discrete depths ranging from the sea surface to a depth of approximately 200 meters.

[0078] Secondly, feature engineering and label construction are performed. For each successfully matched "satellite-buoy" data pair, the input features and output labels of the model need to be constructed.

[0079] The input features are a 10-dimensional vector, specifically including: four time-cyclic encoded features, which encode the month and day of the year of the observation using sine and cosine functions to characterize the seasonal physical driving forces (such as solar radiation cycles) that drive ocean vertical stratification and mixing; two spatial features, namely the longitude (lon) and latitude (lat) of the observation point; and four physical features, namely the horizontal polarization brightness temperature, vertical polarization brightness temperature, and the real and imaginary parts of the third and fourth Stokes parameters of the brightness temperature observed by the satellite. This polarization information provides physical constraints: H / V polarization ( It is primarily sensitive to sea surface roughness (wind) and surface dielectric constant (salinity / temperature), providing key information on total energy and surface state; while HV cross-polarization ( Sensitive to anisotropic sea surface roughness, this helps decouple the effects of wind and waves. It is also sensitive to refractive index changes caused by subsurface density gradients (salinity gradients), providing indirect information about deep structure. These parameters collectively characterize the microwave radiation properties of the sea surface and their multi-layered correlation with ocean conditions.

[0080] The output label is a 9-dimensional vector corresponding to the measured salinity values ​​extracted from the Argo buoy profile at 9 preset target depth layers (e.g., 5 m, 10 m, 20 m, 30 m, 50 m, 75 m, 100 m, 150 m, 200 m), in Practical Salinity Units (PSU).

[0081] This invention employs an adaptive time-matching (or multi-day rollback) strategy. The method includes:

[0082] Step S101: Set the center time for each Argo buoy profile. and its geographical location In the satellite observation time set In the process, find a matching time that meets both the data quality control requirements (such as no missing values ​​and qualified flag bits) and is closest in time. Its mathematical expression is:

[0083]

[0084] in, For matching time; Center time; For satellite observation time sets, These represent longitude and latitude, respectively. Represents the set of satellite observation times One of the candidate times; Indicates time Geographical location Satellite brightness temperature observation vector at the location; It is a verification function used to determine whether the brightness temperature observation vector has no missing data and passes the quality control flag.

[0085] Step S102: If no valid satellite matching data is found within a preset time window (e.g., ±2 days) before or after the buoy observation time, the buoy profile sample will be discarded to ensure the quality of the training data and the immediacy of the matching.

[0086] Finally, to facilitate subsequent analysis and evaluation at different time scales, metadata tags are attached to each of the final "matching scatter" samples, including the year and month in which the observations occurred.

[0087] Step S200: Standardize the input feature vector and label vector in the spatiotemporal matching scatter dataset to obtain standardized input feature vector and standardized label vector, and save the corresponding standardization parameters.

[0088] Step S200 preprocesses the constructed original dataset to eliminate the influence caused by differences in units and numerical ranges between different features and labels, making the model training process more stable and efficient. This step includes two main parts: dataset partitioning and the calculation and storage of standardized parameters. The processing principle is to strictly avoid "time leakage," that is, to ensure that future data information used for model evaluation is not perceived by the model during the training phase.

[0089] First, the dataset is partitioned. Considering the significant temporal correlation and seasonal variations in ocean parameters, to prevent the model from obtaining false high-performance evaluation results by "sneaking" into future data, this step employs a chronological partitioning strategy. For example, the entire dataset can be divided by year, with all matching scatter plots for the entire year of 2017 reserved as a separate test set; while the data from 2010 to 2016 and 2018 to 2020 are combined as the training set. This partitioning method simulates the business scenario in real-world applications where historical data is used to train the model and predict future periods, ensuring the scientific validity and reliability of the evaluation results.

[0090] Next, standardization is performed, specifically using a min-max scaling method. This linearly transforms the original data to a specified range (e.g., [0, 1] or [-1, 1]), ensuring that the values ​​of different features are on the same order of magnitude. Here, independent scaling calculations are required for both the input features (10 dimensions) and the output labels (9-dimensional salinity values).

[0091] It is important to note that all statistical parameters used for standardization (i.e., the minimum and maximum values ​​for each feature or label dimension) must be calculated solely from the training set data. This is because the test set is unknown during model training, and calculating parameters using the entire dataset, including the test set, would lead to information leakage, giving the model an unfair advantage during testing and overestimating its actual generalization ability.

[0092] Next, save the standardized parameters. Save the minimum and maximum values ​​corresponding to each feature dimension and each layer depth salinity label calculated based on the training set to two separate JSON format configuration files, for example, named features_scaling_params.json and target_scaling_params.json.

[0093] JSON is a lightweight data-interchange format that is easy for both humans and machines to read and parse. Saving these parameters is crucial for two reasons: firstly, during model training, these parameters are needed to perform real-time standardization transformations on the training and validation sets; secondly, during model deployment, when salinity inversion is required for new, unseen satellite brightness temperature data, the input features of this new data must be standardized using the exact same parameters as during training, and the model's output predictions must be inversely normalized to obtain salinity values ​​with the correct physical units (PSUs). Therefore, saving and reusing these standardized parameters is key to ensuring consistency and comparability of results throughout the entire methodology.

[0094] Step S300: Construct and train an improved Transformer neural network model, which includes:

[0095] Step S301: Input embedding layer, used to receive the standardized input feature vector and project it into a high-dimensional embedding vector;

[0096] This step treats each input feature as a "token" with potential information and maps it to a continuous vector space.

[0097] Specifically, this step begins with a "patch embedding" operation. The model's input is a model with dimension [missing information]. The feature matrix of , where This represents the batch size, i.e., the number of samples processed simultaneously. At this point, the 10-dimensional input features (temporal, spatial, and physical features) constructed in step S100 are considered. These 10 feature values ​​are viewed as a sequence of 10 "feature patches". This sequence is processed by a one-dimensional convolutional layer (nn.Conv1d) with a kernel size of 1. This convolutional layer acts as a fully connected projection: it independently maps each individual one-dimensional feature "patch" to a higher-dimensional embedding space, such as a 200-dimensional vector. After this operation, the input is transformed into a vector with a dimension of... The tensor, i.e. 200 is referred to as the "embedded dimension" ( This high-dimensional tensor contains richer and more abstract information than the original features.

[0098] Subsequently, this step introduces "learnable positional encoding." In the standard Transformer model, positional encoding is used to inject the model with the order or positional information of each element in the sequence, because the self-attention mechanism itself does not have the ability to perceive order. In this invention, the order of features (such as temporal encoding, spatial coordinates, and the arrangement of different polarization brightness temperatures) may contain important information.

[0099] Therefore, a learnable parametric tensor is created that matches the size of the aforementioned embedding tensor. This parameter is optimized during model training, just like other weights. Finally, this learnable positional encoding parameter is added element-wise to the initial embedding tensor obtained from the "patch embedding". This addition operation ensures that the information at each feature position in the final high-dimensional embedding vector incorporates its specific order encoding, enabling the model to distinguish and utilize the meaning represented by different feature arrangements.

[0100] Step S302: Transformer encoder, used to perform feature encoding on the high-dimensional embedding vector;

[0101] This invention employs an improved Transformer encoder architecture, which efficiently accomplishes this task through a multi-layered stacked "block" structure and the introduction of specific regularization and enhancement designs.

[0102] This encoder consists of L identical code blocks stacked sequentially, where L is a configurable hyperparameter, for example, it can be set to 8. Each code block is processed twice in sequence, each time following the pattern of "normalization-body computation-residual connection", and random depth regularization is introduced.

[0103] First, the initial processing sub-block focuses on capturing global correlations between features. It performs layer normalization on the input tensor X. Layer normalization is a technique for stabilizing neural network training; it normalizes the data across all feature dimensions of a sample, keeping the data distribution stable.

[0104] Next, the normalized data is fed into a multi-head self-attention mechanism. This mechanism allows the model to simultaneously focus on information from multiple different representation subspaces (i.e., multiple "heads") across all positions in the input sequence, thereby computing a new representation for each position based on the global context. This attention output undergoes a stochastic depth operation, a regularization technique that randomly "drops out" (i.e., temporarily skips) entire sub-blocks during training, helping to prevent overfitting.

[0105] Finally, the output of this sub-block is directly added to the initial input X via a residual connection. Residual connections effectively mitigate the vanishing gradient problem in deep networks, ensuring that information flows efficiently even in deeper layers of the network.

[0106] Subsequently, the data flows into the second processing sub-block, whose main function is to perform feature transformation and nonlinear enhancement. Similarly, the output of the first sub-block is first normalized. Then, the data enters a feedforward network. This invention innovates by using a Convolutional Feedforward Network (ConvFFN) structure instead of the traditional fully connected feedforward network. Specifically, a one-dimensional convolutional layer (nn.Conv1d) is used to replace the first fully connected layer in the traditional MLP. One-dimensional convolution can better capture the local correlation between adjacent feature patches, thereby enhancing the model's ability to model the local structure of the input feature sequence. The output of this ConvFFN also undergoes random depth operations and is residually added to the input of the second sub-block. Through these L layers of encoding blocks, the input signal is continuously refined and transformed, and finally the encoder outputs a depth-encoded feature tensor, providing a feature representation for the subsequent hierarchical prediction head.

[0107] Step S303: Layered output header, which includes three independent sub-headers set in parallel and corresponding to surface, middle and deep salinity prediction respectively, used to predict the normalized salinity value of each depth layer according to the output of the Transformer encoder, and splice them to form a complete normalized salinity vertical profile prediction vector.

[0108] This step transforms the globally synthesized features learned by the Transformer encoder into specific predictions of the vertical profile of ocean salinity. Unlike traditional single output layers, this step employs a "layered output head" structure. This structure, based on the physical characteristics of vertical ocean salinity variation (variable at the surface, transitional in the middle, and stable in the deep), decomposes the complex multi-depth regression task into three sub-tasks with more clearly defined physical meanings and varying degrees of complexity. These sub-tasks are then completed collaboratively by three parallel multilayer perceptron sub-heads.

[0109] First, the features output by the encoder need to be reshaped. The Transformer encoder ultimately outputs a feature with one dimension. A high-dimensional tensor (where B is the batch size). To input this tensor into a subsequent fully connected network for prediction, it needs to be "flattened" into a one-dimensional high-dimensional feature vector. Specifically, this involves first performing layer normalization on the tensor to stabilize the data distribution, and then flattening its shape from... Transform into (Right now A vector of length 2000 aggregates the comprehensive information of all input features after deep encoding.

[0110] Next, the flattened high-dimensional feature vectors are simultaneously input into three parallel, independent multilayer perceptron heads. A multilayer perceptron is a feedforward artificial neural network consisting of a series of fully connected layers and nonlinear activation functions, capable of learning complex nonlinear mappings between inputs and outputs.

[0111] Surface Head: This head is responsible for predicting the surface seawater salinity, which is most significantly affected by dynamic processes such as air-sea exchange and wind-wave mixing, corresponding to target depths of 5 meters, 10 meters, and 20 meters. Due to its complex variations, this head is designed as the deepest network structure, such as a 4-layer MLP (network structure example: 2000→512→256→128→3), possessing the most parameters to learn subtle and varied patterns in the surface layer.

[0112] The middle head: This head is responsible for predicting salinity near the thermocline (the water layer with the largest vertical temperature gradient), corresponding to depths of 30 meters, 50 meters, 75 meters, and 100 meters. Salinity changes in this layer are transitional. Therefore, the head has moderate complexity, for example, a 3-layer MLP (network structure example: 2000→256→128→4).

[0113] Deep head: This head is responsible for predicting the salinity of deep seawater at greater depths (150 meters and 200 meters) that are less affected by transient surface disturbances and are relatively stable. Since the changes are relatively gradual, the simplest network structure is used, such as a 3-layer MLP (network structure example: 2000→128→64→2).

[0114] Finally, the outputs are concatenated. The three sub-headers will each output with dimensions of... , and The predicted vectors are then concatenated along the "depth" dimension (i.e., the feature dimension) using operations such as torch.cat to obtain a complete vector. The 3D prediction vector has nine values ​​that correspond to standardized salinity predictions at nine target depths, ranging from 5 meters to 200 meters. This hierarchical output header design essentially constructs an inversion function. The model is structured to correspond to the physical stratification of the ocean: the surface head (0-30 meters) is designed to be the deepest to capture high-frequency, variable signals influenced by atmospheric forcing (wind, evaporation, precipitation); the middle head (30-100 meters) has moderate complexity to capture the characteristics of the transition layer driven by seasonal solar radiation and exhibiting significant thermocline / halogenation fluctuations; and the deep head (150-200 meters) has the simplest structure to capture the low-frequency characteristics of the relatively stable background field, which is less affected by transient surface disturbances. This allows the model to decode and invert high-precision vertical salinity profiles from the limited surface microwave signal (TB) and the time signal (t) containing periodic driving forces.

[0115] like Figure 2 (The diagram illustrates the physical principle of microwave brightness temperature inversion of ocean vertical salinity structure based on layered medium radiative transfer.) The total brightness temperature observed by satellite is the result of attenuated superposition of radiative contributions from the sea surface and deeper layers. The principle of inverting deep salinity is based on the influence of the established seawater stratification model on the variation of L-band brightness temperature. This scheme utilizes the complementarity between information from different polarization channels to physically partially remove interference such as sea surface roughness, and uses the temporal characteristics of the mixing layer depth variation as a constraint, thereby inverting the vertical salinity structure of the ocean interior from limited surface microwave signals.

[0116] In step S300, after constructing the model, it needs to be trained to learn the mapping relationship from input features to salinity profiles. The training method of this invention aims to optimize model performance and prevent overfitting.

[0117] First, perform standardized data partitioning and iterative optimization. Divide the spatiotemporal matching scatter dataset constructed in step S100 into training and validation sets in chronological order. For example, reserve data from a specific year (e.g., 2017) as the validation set, and use the remaining years as the training set. The training process employs mini-batch gradient descent iteratively, with a batch size of 32 or 64. The model is computed on a graphics processing unit (GPU) to accelerate training.

[0118] Secondly, depth-weighted mean squared error is used as the loss function, and the formula for calculating this loss function is as follows:

[0119]

[0120] in, The loss function; For batch size, The total number of layers at the target depth. This represents the index of the depth layer (corresponding to 9 target depths). and The first The sample at the th Standardized true salinity values ​​and standardized predicted salinity values ​​at each depth layer For the preset first The weights for each depth layer are set based on prior physical knowledge of the vertical variation of ocean salinity, with higher weights for surface layers and lower weights for deeper layers. For example, they could be set to [3.0, 2.5, 2.0, 1.5, 1.5, 1.2, 1.0, 0.8, 0.6]. This design forces the model to focus more on the drastically changing and more difficult-to-fit surface salinity signals during training, thus effectively suppressing the model's tendency to simply tend towards the predicted mean (i.e., "mean degradation") due to the gradual change in deep salinity.

[0121] Next, the optimization algorithm and learning rate scheduling strategy are configured. The AdamW optimizer is used to update the model's learnable parameters (such as weights and biases). Simultaneously, a cosine annealing learning rate scheduler is used, which dynamically decays the learning rate from its initial value to near zero according to a cosine function curve during training. This smooth decay helps the model converge more stably to a better solution.

[0122] Finally, a model saving and early stopping mechanism is implemented to ensure the acquisition of the optimal model. During training, a copy of the model parameters at which the loss on the validation set is lowest is continuously saved; if the loss on the validation set does not decrease within a set number of consecutive training epochs, the early stopping mechanism is triggered to terminate the training process. This prevents the model from overfitting on the training set, which would lead to a decline in generalization performance. Through the above training process, a salinity profile inversion model that performs optimally on the validation set and has strong generalization ability is finally obtained.

[0123] Step S400: Input the input feature vector corresponding to the data to be inverted, which has been standardized in the same way, into the trained improved Transformer neural network model to obtain the standardized salinity vertical profile prediction value.

[0124] The input feature vector here specifically refers to satellite brightness temperature observation data for which the corresponding salinity profile is not yet known and needs to be predicted. These data usually come from new satellite transit records, and their format is exactly the same as the input features used when constructing the training set in step S100, that is, each sample also needs to contain 10-dimensional features: 4 time cyclic codes, 2 spatial coordinates (longitude and latitude), and 4 physical brightness temperature features.

[0125] The key point is that when standardizing the features of this new data, the standardized parameter file saved from the training phase must be strictly used. This is because the model learns and trains on the standardized data distribution, and only by mapping the new data to the same numerical range as the training data can the model make correct predictions.

[0126] The standardized new data feature vector is fed as input into the loaded, trained, best-in-class model. The model then propagates forward according to its inherent computational graph: the data passes sequentially through the input embedding layer, the Transformer encoder, and finally reaches the hierarchical output header. The layers within the model work together to encode, transform, and map the input features. After a series of nonlinear transformations, the model directly generates a prediction vector containing nine values ​​at the output layer.

[0127] Ultimately, the output is one or more standardized salinity vertical profile predictions. Each output vector is a 9-dimensional array, with each dimension corresponding to a salinity prediction for 9 preset target depths (e.g., 5 meters to 200 meters).

[0128] It should be clarified that the predicted value obtained at this time is still a standardized and scaled value, and its physical meaning and dimensions have not yet been recovered. This output is an intermediate result.

[0129] Step S500: Using the saved standardized parameters corresponding to the label vector, the obtained standardized salinity vertical profile prediction values ​​are inversely normalized to obtain salinity vertical profile inversion results with physical units.

[0130] This step involves restoring the dimensionless, standardized salinity predictions output by the model to ocean salinity values ​​with clear physical meaning and actual dimensions, thereby obtaining a final salinity vertical profile inversion product that can be directly used for scientific research and operational applications.

[0131] First, the standardized parameters saved in the previous step S200 are loaded. Next, inverse normalization is performed. Using the label-standardized parameters loaded in the previous step, the standardized salinity vertical profile predictions output by the model in step S400 are inversely transformed. Specifically, this involves applying mathematical operations that are the reverse of the normalization process during training, mapping the predicted values ​​for each depth layer from their standardized range (e.g., [0,1]) back to the original true salinity value range. This process transforms the abstract model output into salinity values ​​measured in Practical Salinity Units (PSUs) with a clear oceanographic meaning.

[0132] Furthermore, it is worth noting that during the model evaluation phase, this denormalization process needs to be applied to two sets of data simultaneously: the model's predicted output on the test set data and the corresponding true salinity label values ​​on the test set (these true labels have also undergone the same standardization process before being input into the model). Only by converting both the predicted and true values ​​back to their original physical units can fair and accurate error calculation and performance evaluation (such as calculating the mean absolute error MAE) be performed on the same benchmark, thereby objectively measuring the model's inversion capability.

[0133] Finally, after processing in step S500, a vertical salinity profile inversion result with real physical units (PSU) is obtained from the surface to deep layers (e.g., 5 meters to 200 meters). This result exists in the form of a 9-dimensional salinity vector corresponding to each matched scatter point, forming the basis data for further drawing horizontal distribution maps, vertical profile maps, or performing regional statistical analysis.

[0134] Preferably, after completing model training and obtaining salinity profile inversion results, a systematic, multi-scale scientific evaluation of the model's performance is required to fully verify its effectiveness and practicality. The evaluation method of this invention includes two analysis modes: one is instantaneous profile performance analysis, and the other is monthly-scale aggregated evaluation.

[0135] For instantaneous profile performance analysis, this method directly evaluates the model's prediction accuracy at each independent matched scatter point. Specifically, the predicted salinity values ​​with physical units obtained after denormalization in step S500 are compared sample-by-sample and depth-by-depth with the corresponding actual Argo observed salinity values ​​in the test set. Based on this, the mean absolute error (MAE), root mean square error (RMSE), and R0 are calculated for each of the nine depth layers (5m, 10m, ...). 2 coefficient.

[0136] During the evaluation, not only were the individual indicators of the nine depth layers reported, but the depths were also divided into three groups based on the vertical structure of the ocean: surface (0-30 meters), middle layer (30-100 meters), and deep layer (150-200 meters). The average performance of each group was calculated to determine the model’s performance in different physical stratification regions.

[0137] For the monthly-scale aggregated assessment, the "month" label attached to each sample in the preceding steps is used to group all predicted and actual observation results of the entire test set by month. Within each monthly group, for each of the nine target depth layers, the arithmetic mean of the true salinity of all samples for that month is calculated. and the predicted salinity arithmetic mean For each depth layer, calculate the absolute error between the two months' averages using the following formula:

[0138]

[0139] This value ( This represents the monthly average bias of the model's inversion results at that month and depth, expressed in PSU. By analyzing the monthly scale errors at different months and depths, the accuracy of the model's products in characterizing the salinity seasonal cycle and vertical climatological features can be assessed.

[0140] It should be noted that in this embodiment, the 2017 data corresponding to the test set has been removed from the training set to ensure the independence of the evaluation. Taking the test set from March to June 2017 as an example, the performance of the monthly-scale aggregate evaluation is shown in Table 1. This table shows the MAE of each depth layer in different months and its average value from March to June, which intuitively reflects the inversion stability and accuracy of the model during seasonal changes.

[0141] Table 1: Monthly-scale aggregate performance evaluation (2017 test set, March-June, unit PSU)

[0142]

[0143] As shown in Table 1, the model performs differently at different depths and in different months. For example, the MAE is generally low and stable in deep layers (e.g., 200 meters), while the error is large in some months in the middle layers (e.g., 50 meters, 75 meters). This reflects the seasonal variation of the vertical structure of salinity and the challenges of model inversion.

[0144] By combining the two evaluation methods described above, we can not only precisely control the model's ability to recreate the instantaneous ocean state, but also confirm the consistency between the data products it generates and the macroscopic observation and analysis results from the application level, thereby comprehensively demonstrating the reliability and practical value of the method of this invention.

[0145] Preferably, before generating a continuous and intuitive two-dimensional horizontal distribution map or vertical profile map from the salinity values ​​obtained by model inversion and discretely distributed at various points in space, the process includes preprocessing and visualization of the raw satellite brightness temperature data, specifically including:

[0146] Raw data reading and filtering: The program reads raw satellite observation data files (NetCDF format) in batches within a specified time range. For each file, it extracts the required brightness temperature channel data, time information, and corresponding longitude and latitude coordinates.

[0147] After reading and preliminary filtering through the above steps, the spatial distribution data of brightness temperature over a large area within this time period can be obtained. An example of its overall distribution is shown below. Figure 3 As shown. Figure 3 The average spatial distribution pattern of the horizontal polarization brightness temperature in this region in May 2017 is shown (average horizontal polarization = 85.8K, standard deviation = 6.5K), which intuitively reflects the macroscopic characteristics and spatial heterogeneity of the brightness temperature field.

[0148] Data cleaning and preprocessing: To ensure the quality and representativeness of the data, a series of preprocessing operations were implemented. First, a pre-prepared ocean mask file (.mat format) was applied, and specific mask ranges were selected according to requirements (such as deep-sea areas beyond 40km, 100km, and 200km from the coastline) to remove invalid land and nearshore observation points. Second, data points with invalid brightness temperature values ​​(e.g., less than or equal to 0) were filtered out.

[0149] Data deduplication and downsampling: At least three monthly nc files are selected to generate the plotting dataset. Given that overlapping satellite orbits may lead to multiple observations at the same location, and that the large volume of original data may affect plotting efficiency, this invention employs data deduplication and downsampling strategies. The deduplication operation identifies and removes redundant observation points that are identical in spatiotemporal location; the downsampling operation, based on a preset rule that if two adjacent rows have the same brightness temperature value, the next row is discarded, thus sparsifying the data.

[0150] Scatter plotting: Save the discrete latitude and longitude coordinates and corresponding brightness temperature values ​​obtained after the above processing as an intermediate data file (CSV format). Then, use a plotting tool to read the file and directly plot each data point on the map according to its latitude and longitude coordinates. The color or size of the point indicates the brightness temperature value.

[0151] Based on this, the salinity values ​​obtained from model inversion and discretely distributed at various points in space are further generated into a continuous and intuitive two-dimensional horizontal distribution map to show the spatial pattern of salinity at a certain depth layer.

[0152] When plotting the salinity level distribution map, the plotting dataset used was discrete observation points re-read from the SMOS NetCDF data and processed using the preprocessing steps described above. First, the latitude and longitude coordinates of each observation point were extracted from the preprocessed data. Then, their corresponding features (temporal, spatial, and physical features) were input into a trained improved Transformer neural network model to obtain the salinity prediction values ​​for each point at nine target depths. A specific depth layer was selected for visualization, and the latitude and longitude coordinates of each observation point and the salinity prediction values ​​for that depth layer were used as the plotting data.

[0153] Since the data consists of discrete points, it needs to be interpolated onto a regular latitude and longitude grid to generate a continuous image. This invention employs a hybrid interpolation method: sequentially attempting cubic interpolation, linear interpolation, and nearest neighbor interpolation to ensure that all grid points are effectively filled, avoiding blank areas. After generating the grid data, Gaussian smoothing can be selectively applied to improve the image's appearance.

[0154] like Figure 4 As shown, there are a total of 5 key depth layers, among which, Figure 4(a)-(e) in the figure represent the salinity prediction statistics at (a) 5 meters (mean 33.61, standard deviation 0.23), (b) 20 meters (33.69, 0.18), (c) 50 meters (34.01, 0.19), (d) 100 meters (34.473, 0.065), and (e) 200 meters (34.5452, 0.0060), respectively. Due to the large differences in salinity variation at different depths (large variation at the surface and small variation at the depth), fixing the color mark range will result in insufficient contrast in the deep images. Therefore, the color scale range strategy is automatically selected based on the salinity standard deviation and depth value of the current depth layer: for deep layers (such as depth ≥ 75 meters or standard deviation < 0.01), the fixed_range strategy is adopted, which uses a fixed interval with a narrow range centered on the salinity mean and displays more decimal places to highlight small changes; for the surface layer, a wide range strategy based on the standard deviation or percentile is automatically selected to fully show its dynamic range of change.

[0155] Figure 5 For multi-profile salinity comparison (longitude), among which Figure 5 In the diagram, (a)-(d) are salinity profiles for longitudes of 111.00° (N=5060 data points), 113.00° (N=6155 data points), 115.00° (N=7766 data points), and 117.00° (N=7678 data points), respectively. Figure 5 As shown, to draw a vertical profile, the first step is to determine the profile type and its location. Then, from all inverted scatter points, points within a certain tolerance range near the profile line are selected (e.g., longitude between 114.5°E and 115.5°E). The horizontal location information (e.g., latitude) of these selected scatter points is used as the X-axis, depth as the Y-axis, and the corresponding multi-depth salinity prediction values ​​as the Z-axis (data values). Since the distribution of these points in both the horizontal and vertical directions is irregular, a grid interpolation algorithm is used to interpolate them onto a regular (distance / latitude, depth) two-dimensional grid. Finally, a filled contour map is used to draw this two-dimensional grid data, clearly showing the contour distribution of salinity on the vertical profile. To conform the image to oceanographic drawing conventions (depth increases downwards), the Y-axis needs to be reversed. The final generated vertical profile visually reveals key oceanographic features such as the vertical stratification structure of salinity and the location of the thermocline.

[0156] In summary, by introducing time-cyclic coding (simulating seasonal solar radiation and heat flux that drive ocean stratification) and designing a stratified output header, this application is actually using a deep learning model to solve this complex inverse problem of radiative transfer, thereby achieving the goal of "seeing through" the vertical structure inside the ocean from the surface signal.

[0157] To make the objectives, technical solutions, and advantages of this invention clearer, the following is combined with... Figure 5 The present invention will be described in further detail below.

[0158] Example: Salinity profile inversion in a certain area (SCS)

[0159] Data preparation (corresponding to steps S100 and S200): In this embodiment, SMOS brightness temperature and Argo profile data of a certain region (longitude 109–119°E, latitude 5–23°N) are selected, and matching scatter plots are constructed according to step S100. Input features are 10-dimensional (4 temporal cyclic encodings + 2 spatial encodings + 4 physical encodings), and output labels are 9-dimensional (5m–200m). Training / testing split: 2010–2016 and 2018–2020 are used for training, and 2017 is used for testing. Standardized parameters are calculated based on the training set and saved to features_scaling_params.json and target_scaling_params.json.

[0160] Model Construction and Training (corresponding to step S300): The improved Transformer structure described in S300 is adopted. Specific parameters: embedding dimension is 200, encoder layer number is 8, and multi-head attention head number is 8. Hierarchical output heads: surface layer (3), middle layer (4), deep layer (2). The loss function is depth-weighted MSE, and the weight vector... Training parameters: batch size 32, learning rate 0.0005, optimizer AdamW, learning rate scheduling strategy using cosine annealing restart, and patience value for early stopping mechanism set to 25. The model was trained on cuda:0, and the model with the lowest validation loss was saved as best_model.pth.

[0161] Evaluation (corresponding to steps S400 and S500): Perform the two evaluations described in step S500 for the test year (2017):

[0162] Instantaneous assessment: Calculate the depth MAE of all instantaneous scatter points in 2017 and summarize the performance by dividing them into three groups: "surface", "middle" and "deep".

[0163] Monthly-scale assessment: Monthly-scale aggregation is performed on a monthly basis. Example: In July, samples with Month=7 are selected, and the mean of the actual and predicted 5m depths is calculated. , ,have to : .

[0164] Applications and Visualization: Retrain the production model using full-time data. Perform profile inversion on new, unmatched data (e.g., July 2021).

[0165] Horizontal distribution map: drawn according to the aforementioned process. The visualization module (paint_improved.py) automatically detects depths < 0.01 at 200m, triggering the fixed_range strategy and using the mean. The color swatch range and 4 decimal places precision.

[0166] Vertical profile: Draw according to the above procedure. Select the 20°N profile, and use griddata interpolation to draw a salinity distribution map at longitudes of 110°E–118°E and depths of 0–200m, and reverse the Y-axis.

[0167] According to another aspect of the embodiments of this application, an electronic device is also provided, including a processor and a memory, wherein the processor is configured to implement the steps of the method when executing a computer program stored in the memory.

[0168] In the above embodiments of the present invention, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0169] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For instance, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0170] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0171] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0172] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for retrieving vertical ocean salinity profiles using a scatter plot method with brightness temperature-buoy spatiotemporal matching, characterized in that, Includes the following steps: Satellite brightness temperature data and on-site buoy salinity profile data are acquired and spatiotemporally matched to construct a spatiotemporally matched scatter dataset containing input feature vectors and label vectors. The input feature vectors include time, space and physical features, and the label vectors include salinity values ​​corresponding to multiple target depths. The input feature vector and label vector in the spatiotemporal matching scatter dataset are standardized to obtain standardized input feature vector and standardized label vector, and the corresponding standardization parameters are saved. Construct and train an improved Transformer neural network model, which includes: An input embedding layer is used to receive the standardized input feature vector and project it into a high-dimensional embedding vector. A Transformer encoder is used to perform feature encoding on the high-dimensional embedding vector; The layered output header includes three independent sub-headers that are set in parallel and correspond to the surface, middle and deep salinity predictions, respectively. These sub-headers are used to predict the normalized salinity values ​​of each depth layer based on the output of the Transformer encoder and then splice them together to form a complete normalized salinity vertical profile prediction vector. The input feature vector, which is standardized in the same way and corresponds to the data to be inverted, is input into the trained improved Transformer neural network model to obtain the standardized salinity vertical profile prediction value. By using the saved standardized parameters corresponding to the label vector, the obtained standardized salinity vertical profile prediction values ​​are inversely normalized to obtain salinity vertical profile inversion results with physical units.

2. The method for retrieving vertical ocean salinity profiles by scatter point inversion using brightness temperature-buoy spatiotemporal matching as described in claim 1, characterized in that, Methods for spatiotemporal matching include: For each on-site buoy profile, based on its buoy profile center time and geographical location, find the closest matching time in the satellite observation time set that satisfies data validity. The matching time is determined by the following formula: in, For matching time; Center time; For the set of satellite observation times, These represent longitude and latitude, respectively. Represents the set of satellite observation times One of the candidate times; Indicates time Geographical location Satellite brightness temperature observation vector at the location; It is a verification function used to determine whether the brightness temperature observation vector has no missing data and passes the quality control flag; If no satisfactory time is found within the preset time window of the buoy profile center time, If valid satellite observations are available, the buoy profile sample should be discarded.

3. The method for retrieving vertical ocean salinity profiles by scatter point inversion using brightness temperature-buoy spatiotemporal matching as described in claim 1, characterized in that, The input embedding layer performs the following operations to project the standardized input feature vector into a high-dimensional embedding vector: Dimension is The input feature vector is regarded as being composed of A sequence consisting of feature patches, wherein For batch size, Input feature dimension; Each feature patch is projected into a high-dimensional embedding space through a one-dimensional convolutional layer. The kernel size of the one-dimensional convolutional layer is 1, and the number of output channels is [missing information]. Thus, the initial embedding tensor is obtained. ; Generate a learnable positional encoding parameter that matches the size of the initial embedding tensor; The learnable position encoding parameters are added element-wise to the initial embedding tensor to obtain a high-dimensional embedding vector containing order information.

4. The method for retrieving vertical ocean salinity profiles by scatter point inversion using brightness temperature-buoy spatiotemporal matching as described in claim 1, characterized in that, The Transformer encoder performs the following operations to encode features in the high-dimensional embedding vector: The high-dimensional embedding vector is input into an encoder consisting of stacked L layers of coding blocks, where L is an integer greater than 1; Each layer of coded blocks performs the following operations in sequence: The input tensor is layer normalized, processed by a multi-head self-attention mechanism, and the processed result is discarded through a random path operation and added to the input tensor residual to obtain the first intermediate feature. The first intermediate feature is normalized by a layer and processed by a feedforward network. The processed result is discarded through a random path and added to the residual of the first intermediate feature to obtain the output feature of the coding block. The feedforward network adopts a convolutional feedforward network structure, using one-dimensional convolutional layers to replace traditional fully connected layers to achieve feature transformation.

5. The method for retrieving vertical ocean salinity profiles by scatter point inversion using brightness temperature-buoy spatiotemporal matching as described in claim 1, characterized in that, The layered output header performs the following operations to predict the normalized salinity values ​​for each depth layer based on the output of the Transformer encoder, and then concatenates them to form a complete normalized salinity vertical profile prediction vector: The feature tensor output by the Transformer encoder is flattened to obtain a high-dimensional feature vector. The high-dimensional feature vector is simultaneously input into the heads of three parallel and independent multilayer perceptrons: The first subheading is the surface head, which is used to predict the salinity values ​​of the first set of depth layers most affected by surface dynamics. The second subhead is the middle head, which is used to predict the salinity values ​​of the second set of depth layers near the thermocline; The third subhead is the deep head, which is used to predict the salinity values ​​of the relatively stable third group of depth layers. The prediction vectors output from the first, second, and third sub-headers are concatenated along the depth dimension to form a complete standardized salinity vertical profile prediction vector.

6. The method for retrieving vertical ocean salinity profiles by scatter point inversion using brightness temperature-buoy spatiotemporal matching as described in claim 1, characterized in that, Methods for training improved Transformer neural network models include: The spatiotemporal matching scatter dataset is divided into a training set and a validation set in chronological order. The model parameters are iteratively optimized using the training set, and the model performance is monitored on the validation set. The goal of model optimization is to minimize the depth-weighted mean squared error loss function, which is calculated using the following formula: in, The loss function; For batch size, The total number of layers at the target depth. Indicates the index of the depth layer. and The first The sample at the th Standardized true salinity values ​​and standardized predicted salinity values ​​at each depth layer For the preset first The weights for each depth layer are set according to the vertical variation characteristics of ocean salinity, with the surface depth weight being higher than the deep depth weight. The AdamW optimizer is used to update the model parameters, and a cosine annealing learning rate scheduler is used to dynamically adjust the learning rate for each iteration. During training, a copy of the model parameters that minimizes the loss on the validation set is continuously saved. If the loss on the validation set does not decrease within a set number of consecutive training cycles, an early stopping mechanism is triggered to terminate the training process.

7. The method for retrieving vertical ocean salinity profiles by scatter point inversion using brightness temperature-buoy spatiotemporal matching as described in claim 6, characterized in that, Methods for inverse normalization of the obtained standardized salinity vertical profile predictions include: Load the label normalization parameter file saved during the training phase; Using the scaling parameters stored in the label standardization parameter file, the predicted value of the standardized salinity vertical profile is inversely normalized. The inverse normalization calculation is applied simultaneously to the standardized predicted values ​​output by the model and the standardized true values ​​corresponding to the test set, in order to obtain the predicted and true values ​​of the salinity vertical profile with real physical units.

8. The method for retrieving vertical ocean salinity profiles by scatter point inversion using brightness temperature-buoy spatiotemporal matching as described in claim 1, characterized in that, After obtaining the salinity vertical profile inversion results in physical units, the process also includes a step of evaluating the model inversion performance, the evaluation step including: Instantaneous profile performance analysis: Directly compare the predicted salinity value of each scatter sample in the inversion results with the corresponding true salinity value, and calculate the mean absolute error, root mean square error and coefficient of determination for each target depth layer respectively; Monthly-scale aggregated evaluation: Grouping samples using the month labels attached to each sample when constructing the dataset, calculating the absolute error between the true average salinity and the predicted average salinity within each month group and at each depth layer, and using this error as the model's performance metric on a macro-monthly scale.

9. An electronic device, comprising a memory and a processor, characterized in that, The memory is used to store programs that support the processor in executing any of the brightness temperature-buoy spatiotemporal matching scatter inversion ocean salinity vertical profile methods according to claims 1-8, and the processor is configured to execute the programs stored in the memory.

10. A computer-readable storage medium storing a computer program thereon, characterized in that, When the computer program is run by the processor, it executes the steps of the brightness temperature-buoy spatiotemporal matching scatter inversion method for vertical ocean salinity profiles as described in any of claims 1-8.

Images