A satellite sea surface salinity bias correction method based on deep learning
Through deep learning, the MLP and LSTM models are constructed, and the satellite sea surface salinity data is processed, which solves the problem of insufficient correction accuracy of satellite sea surface salinity data in the existing technology, especially in strong wind conditions, and achieves higher accuracy and more comprehensive consideration of influencing factors.
Patent Information
- Application Number
- CN202311806491.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-12-26
- Publication Date
- 2024-11-12
- Estimated Expiration
- 2043-12-26
AI Technical Summary
The existing satellite sea surface salinity data correction methods have large deviations in special environments such as strong winds, and have not fully considered factors such as sea conditions and atmospheric stability, resulting in insufficient accuracy.
Using a deep learning-based method, a multi-layer perceptron (MLP) and long-term memory network (LSTM) model was constructed. By acquiring and preprocessing satellite sea surface salinity-related data sets, a positive and negative salinity deviation classification model and deviation correction were constructed. Model, perform data interpolation matching and training to improve data accuracy.
The accuracy of satellite sea surface salinity data is improved, especially in special environments such as strong wind conditions, and influencing factors such as sea conditions and atmospheric stability are fully considered.
Smart Images

Figure CN118013346B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of sea surface salinity deviation correction, and in particular to a satellite sea surface salinity deviation correction based on deep learning. Background Art
[0002] Ocean salinity is one of the three most basic elements in ocean dynamics, and plays an important role in observing global atmospheric and oceanic processes such as ocean circulation, mesoscale ocean phenomena, and air-sea interactions. In the past few decades, the measured data of seawater salinity mostly came from Argo drifting buoys, CTD, anchored buoys, and some cruise observations. For example, in order to understand global climate change, atmospheric and ocean scientists proposed a plan to establish a "Global Ocean Observation Network (Array for Real-time Geostrophic Oceanography)" in 1998, namely the Argo plan. However, the average deployment interval of Argo buoys around the world is currently about 300km, and the buoys complete a profile measurement in 10-14 days. Therefore, the measured salinity data is very accurate, but the temporal and spatial resolution is insufficient in effectively depicting the fine temporal and spatial changes of ocean salinity. It was not until 2009 and 2011 that the European Soil Moisture and Ocean Salinity Satellite (SMOS) and the American Aquarius / SAC-D were launched successively, and in 2015 NASA's Soil Moisture Active Passive Probe (SMAP) was launched, that the problem of lack of sea surface salinity data was initially solved.
[0003] Remote sensing satellite sea surface salinity products usually include different levels of data processing and processing levels, which provide different resolutions, processing degrees and data content. Raw is the original format data received by the satellite, and L0 is the source format data obtained after Raw formatting. L1a is the corrected visibility obtained after unit conversion and calibration of L0; L1b is the brightness temperature in units of instrument snapshots after image reconstruction of L1a; L1c is the brightness temperature in units of orbits on the ISEA grid after geolocation reorganization. Level L2 iteratively inverts the L1c brightness temperature to obtain the orbital-level sea surface salinity field and corresponding uncertainty based on different inversion algorithms on the ISEA network. Level L3 obtains gridded sea surface salinity fields of different temporal and spatial resolutions and corresponding uncertainties after spatiotemporal reorganization of the L2 sea surface salinity.
[0004] At present, for L2 sea surface salinity data, predecessors have carried out a lot of work on salinity satellite brightness temperature reconstruction, error distribution characteristics, bias correction, etc. The main errors in the L2 stage include forward model errors and auxiliary data errors. The largest error source comes from the sea surface roughness effect caused by wind and waves. The most important factor affecting sea surface roughness is wind speed. Someone has developed a new SSS inversion algorithm that can significantly improve the coverage of global SMOS SSS, so that sea surface salinity inversion can be achieved in some key areas where there was no or little salinity before. The Barcelona Expert Center team (Barcelona Expert Center, BEC) used this method to generate a three-year time series of high-latitude SMOS SSS. With the development of machine learning, researchers began to use machine learning algorithms to correct satellite salinity bias. Mu Ziyao et al. used a generalized regression neural network (GRNN) to correct the SMOS salinity in the South China Sea and applied it to the assimilation system in coastal areas. The input of GRNN is SMOS SSS, longitude, latitude, time, SST and altimeter sea level anomaly (SLA). Existing methods have greatly improved the correction and accuracy of satellite sea surface salinity data, but there are still many shortcomings. Traditional algorithm models still have large deviations after deviation correction in special environments, such as strong winds, and do not fully consider influencing factors such as sea conditions and atmospheric stability. Machine learning method research mostly uses traditional machine learning methods and only considers a small number of sea surface parameters to correct satellite sea surface salinity data. Therefore, the correction and accuracy of satellite sea surface salinity data based on existing methods still need to be improved. Summary of the invention
[0005] The present invention aims to solve at least one of the technical problems existing in the prior art. To this end, the present invention proposes a satellite sea surface salinity deviation correction based on deep learning, which can correct the satellite sea surface salinity deviation and improve the accuracy of the data.
[0006] In a first aspect, an embodiment of the present invention provides a method for correcting a satellite sea surface salinity deviation based on deep learning, and the method for correcting a satellite sea surface salinity deviation based on deep learning includes:
[0007] Obtain satellite sea surface salinity related datasets;
[0008] Preprocessing the data in the satellite sea surface salinity related data set and interpolating and matching them to the same temporal and spatial resolution to obtain a processed satellite sea surface salinity related data set;
[0009] Based on the multi-layer perceptron, a classification model of salinity positive and negative deviations was constructed;
[0010] Based on the processed satellite sea surface salinity related data set, a first training data set is constructed, and the salinity positive and negative deviation classification model is trained using the first training data set to obtain a trained salinity positive and negative deviation classification model, so as to obtain the salinity data positive and negative deviation using the trained salinity positive and negative deviation classification model;
[0011] Based on the long short-term memory network, a deviation correction model is constructed;
[0012] Based on the positive and negative deviations of the salinity data, a second training data set is constructed, and the deviation correction model is trained using the second training data set to obtain a trained deviation correction model, so that the satellite sea surface salinity deviation can be corrected using the trained deviation correction model.
[0013] Compared with the prior art, the first aspect of the present invention has the following beneficial effects:
[0014] The method constructs a salinity positive and negative deviation classification model, pre-processes the data in the satellite sea surface salinity related data set and interpolates and matches them to the same spatiotemporal resolution, constructs a first training data set based on the processed satellite sea surface salinity related data set, and uses the first training data set to train the salinity positive and negative deviation classification model, which can improve the accuracy of the salinity positive and negative deviation classification model; uses the trained salinity positive and negative deviation classification model to obtain the salinity data positive and negative deviation, constructs a second training data set based on the salinity data positive and negative deviation, and uses the second training data set to train the deviation correction model, and the deviation correction model obtained by training the second training data set constructed based on the salinity data positive and negative deviation can correct the satellite sea surface salinity deviation and improve the accuracy of the data.
[0015] According to some embodiments of the present invention, the satellite sea surface salinity-related data set includes sea surface salinity, sea surface temperature, sea surface height, sea surface flow field, sea surface meridional wind speed, sea surface zonal wind speed, sea surface meridional wind stress and sea surface zonal wind stress.
[0016] According to some embodiments of the present invention, preprocessing the data in the satellite sea surface salinity related dataset and interpolating and matching them to the same spatiotemporal resolution to obtain a processed satellite sea surface salinity related dataset includes:
[0017] Dividing the sea surface salinity data in the satellite sea surface salinity related data set into a plurality of first data sets according to a preset time period;
[0018] Averaging the salinity values in each of the first data sets to obtain an average salinity value corresponding to each preset time period;
[0019] Constructing a second data set according to the average salinity value corresponding to each of the preset time periods;
[0020] Other data in the satellite sea surface salinity related data set are interpolated and matched to the second data set of the same temporal and spatial resolution to obtain a processed satellite sea surface salinity related data set.
[0021] According to some embodiments of the present invention, other data in the satellite sea surface salinity related dataset are interpolated and matched to the second dataset of the same spatiotemporal resolution in the following manner:
[0022]
[0023] Among them, (x, y) represents the coordinates of the data point to be estimated, Q 11 , Q 21 , Q 12 and Q 22 represents the four grid points closest to the data point to be estimated, x1 and x2 represent the horizontal coordinates of the four grid points, y1 and y2 represent the vertical coordinates of the four grid points, f(x, y) represents the result after interpolation, and f(Q 11 )、f(Q 21 )、f(Q 12 ) and f(Q 22 ) indicates Q 11 , Q 21 , Q 12 and Q 22 The corresponding interpolated data.
[0024] According to some embodiments of the present invention, the step of using the first training data set to train the salinity positive and negative deviation classification model to obtain a trained salinity positive and negative deviation classification model includes:
[0025] Normalizing the first training data set to obtain normalized first training data;
[0026] The normalized first training data is input into the salinity positive and negative deviation classification model, and the training is performed using a first loss function to obtain a trained salinity positive and negative deviation classification model.
[0027] According to some embodiments of the present invention, the first training data set is normalized in the following manner:
[0028]
[0029] Among them, X′ j represents the normalized set of j sea surface elements, X j represents the jth sea surface element dataset in the first training dataset, represents the minimum value of the jth sea surface feature dataset, Indicates the maximum value of the jth sea surface feature dataset.
[0030] According to some embodiments of the present invention, the first loss function is a cross entropy loss function.
[0031] According to some embodiments of the present invention, the step of training the deviation correction model using the second training data set to obtain a trained deviation correction model includes:
[0032] Normalizing the second training data set to obtain normalized second training data;
[0033] The normalized second training data is input into the deviation correction model, and training is performed using a second loss function to obtain a trained deviation correction model.
[0034] According to some embodiments of the present invention, the second loss function is a mean square error loss function.
[0035] According to some embodiments of the present invention, the step of using the second loss function for training to obtain a trained deviation correction model includes:
[0036] Using the second loss function to train the deviation correction model to obtain a first deviation correction model;
[0037] K-fold cross validation is used to perform a generalization evaluation on the first bias correction model, and the parameters of the first bias correction model are modified to obtain a trained bias correction model. BRIEF DESCRIPTION OF THE DRAWINGS
[0038] The above and / or additional aspects and advantages of the present invention will become apparent and easily understood from the description of the embodiments in conjunction with the following drawings, in which:
[0039] Figure 1 It is a flow chart of a method for correcting satellite sea surface salinity deviation based on deep learning according to an embodiment of the present invention;
[0040] Figure 2 is a schematic diagram of data interpolation according to an embodiment of the present invention;
[0041] Figure 3 It is a structural schematic diagram of a salinity positive and negative deviation classification model based on MLP according to an embodiment of the present invention;
[0042] Figure 4 It is a structural diagram of a deviation correction model based on LSTM according to an embodiment of the present invention. DETAILED DESCRIPTION
[0043] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.
[0044] In the description of the present invention, if there is a description of first, second, etc., it is only for the purpose of distinguishing technical features, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features or implicitly indicating the order of the indicated technical features.
[0045] In the description of the present invention, it should be understood that descriptions involving orientation, such as orientation or positional relationship indicated as up, down, etc., are based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as a limitation on the present invention.
[0046] In the description of the present invention, it should be noted that, unless otherwise clearly defined, terms such as setting, installing, connecting, etc. should be understood in a broad sense, and technicians in the relevant technical field can reasonably determine the specific meanings of the above terms in the present invention in combination with the specific content of the technical solution.
[0047] At present, for L2 sea surface salinity data, predecessors have carried out a lot of work on salinity satellite brightness temperature reconstruction, error distribution characteristics, bias correction, etc. The main errors in the L2 stage include forward model errors and auxiliary data errors. The largest error source comes from the sea surface roughness effect caused by wind and waves. The most important factor affecting sea surface roughness is wind speed. Someone has developed a new SSS inversion algorithm that can significantly improve the coverage of global SMOS SSS, so that sea surface salinity inversion can be achieved in some key areas where there was no or little salinity before. The Barcelona Expert Center team (Barcelona Expert Center, BEC) used this method to generate a three-year time series of high-latitude SMOS SSS. With the development of machine learning, researchers began to use machine learning algorithms to correct satellite salinity bias. Mu Ziyao et al. used a generalized regression neural network (GRNN) to correct the SMOS salinity in the South China Sea and applied it to the assimilation system in coastal areas. The input of GRNN is SMOS SSS, longitude, latitude, time, SST and altimeter sea level anomaly (SLA). Existing methods have greatly improved the correction and accuracy of satellite sea surface salinity data, but there are still many shortcomings. Traditional algorithm models still have large deviations after deviation correction in special environments, such as strong winds, and do not fully consider influencing factors such as sea conditions and atmospheric stability. Machine learning method research mostly uses traditional machine learning methods and only considers a small number of sea surface parameters to correct satellite sea surface salinity data. Therefore, the correction and accuracy of satellite sea surface salinity data based on existing methods still need to be improved.
[0048] To solve the above problems, the present invention constructs a salinity positive and negative deviation classification model, pre-processes the data in the satellite sea surface salinity related data set and interpolates and matches them to the same time and space resolution, constructs a first training data set based on the processed satellite sea surface salinity related data set, and uses the first training data set to train the salinity positive and negative deviation classification model, which can improve the accuracy of the salinity positive and negative deviation classification model; uses the trained salinity positive and negative deviation classification model to obtain the salinity data positive and negative deviation, constructs a second training data set based on the salinity data positive and negative deviation, and uses the second training data set to train a deviation correction model, and the deviation correction model obtained by training the second training data set constructed based on the salinity data positive and negative deviation can correct the satellite sea surface salinity deviation and improve the accuracy of the data.
[0049] Reference Figure 1The embodiment of the present invention provides a method for correcting satellite sea surface salinity deviation based on deep learning. The method for correcting satellite sea surface salinity deviation based on deep learning includes but is not limited to steps S100 to S600, wherein:
[0050] Step S100, obtaining a satellite sea surface salinity related data set;
[0051] Step S200, preprocessing the data in the satellite sea surface salinity related data set and interpolating and matching them to the same spatiotemporal resolution to obtain a processed satellite sea surface salinity related data set;
[0052] Step S300: constructing a salinity positive and negative deviation classification model based on a multi-layer perceptron;
[0053] Step S400, constructing a first training data set based on the processed satellite sea surface salinity related data set, and using the first training data set to train a salinity positive and negative deviation classification model to obtain a trained salinity positive and negative deviation classification model, so as to use the trained salinity positive and negative deviation classification model to obtain salinity data positive and negative deviations;
[0054] Step S500: constructing a deviation correction model based on the long short-term memory network;
[0055] Step S600: construct a second training data set based on the positive and negative deviations of the salinity data, and use the second training data set to train a deviation correction model to obtain a trained deviation correction model, so as to use the trained deviation correction model to correct the satellite sea surface salinity deviation.
[0056] In this embodiment, in order to improve the accuracy of the salinity positive and negative deviation classification model, this embodiment obtains a satellite sea surface salinity related data set, pre-processes the data in the satellite sea surface salinity related data set and interpolates and matches it to the same spatiotemporal resolution to obtain a processed satellite sea surface salinity related data set, constructs a salinity positive and negative deviation classification model based on a multi-layer perceptron, constructs a first training data set based on the processed satellite sea surface salinity related data set, and uses the first training data set to train the salinity positive and negative deviation classification model to obtain a trained salinity positive and negative deviation classification model, so that the trained salinity positive and negative deviation classification model can be used to obtain the salinity data positive and negative deviation; in order to correct the satellite sea surface salinity deviation and improve the accuracy of the data, this embodiment constructs a deviation correction model based on a long short-term memory network, constructs a second training data set based on the salinity data positive and negative deviation, and uses the second training data set to train the deviation correction model to obtain a trained deviation correction model, so that the trained deviation correction model can be used to correct the satellite sea surface salinity deviation.
[0057] In some embodiments, the satellite sea surface salinity related data set includes sea surface salinity, sea surface temperature, sea surface height, sea surface current field, sea surface meridional wind speed, sea surface zonal wind speed, sea surface meridional wind stress and sea surface zonal wind stress.
[0058] In this embodiment, by comprehensively considering the sea surface salinity, sea surface temperature, sea surface height, sea surface flow field, sea surface meridional wind speed, sea surface zonal wind speed, sea surface meridional wind stress and sea surface zonal wind stress, it is possible to fully consider the influencing factors such as sea conditions and atmospheric stability, and provide a substantial data basis for subsequent model training.
[0059] It should be noted that this embodiment includes sea surface salinity, sea surface temperature, sea surface height, sea surface flow field, sea surface meridional wind speed, sea surface zonal wind speed, sea surface meridional wind stress and sea surface zonal wind stress, but is not limited to these data and can be changed according to actual conditions.
[0060] In some embodiments, the data in the satellite sea surface salinity related dataset is preprocessed and interpolated to match the same spatiotemporal resolution to obtain a processed satellite sea surface salinity related dataset, including:
[0061] Dividing the sea surface salinity data in the satellite sea surface salinity related data set into a plurality of first data sets according to a preset time period;
[0062] Averaging the salinity values in each first data set to obtain an average salinity value corresponding to each preset time period;
[0063] Constructing a second data set according to the average salinity value corresponding to each preset time period;
[0064] Other data in the satellite sea surface salinity related data set are interpolated and matched to the second data set with the same temporal and spatial resolution to obtain a processed satellite sea surface salinity related data set.
[0065] In this embodiment, all data are interpolated and matched to the same spatiotemporal resolution, which can provide a good data foundation for subsequent model training.
[0066] In some embodiments, other data in the satellite sea surface salinity related dataset are interpolated and matched to the second dataset with the same spatiotemporal resolution in the following manner:
[0067]
[0068] Among them, (x, y) represents the coordinates of the data point to be estimated, Q 11 , Q 21 , Q 12 and Q 22represents the four grid points closest to the data point to be estimated, x1 and x2 represent the horizontal coordinates of the four grid points, y1 and y2 represent the vertical coordinates of the four grid points, f(x,y) represents the result after interpolation, and f(Q 11 )、f(Q 21 )、f(Q 12 ) and f(Q 22 ) indicates Q 11 , Q 21 , Q 12 and Q 22 The corresponding interpolated data.
[0069] In some embodiments, the first training data set is used to train the salinity positive and negative deviation classification model to obtain a trained salinity positive and negative deviation classification model, including:
[0070] Normalizing the first training data set to obtain normalized first training data;
[0071] The normalized first training data is input into the salinity positive and negative deviation classification model, and the first loss function is used for training to obtain a trained salinity positive and negative deviation classification model.
[0072] In this embodiment, since the first training data set contains sea surface salinity, sea surface temperature, sea surface height, sea surface flow field, sea surface meridional wind speed, sea surface zonal wind speed, sea surface meridional wind stress and sea surface zonal wind stress, the normalized first training data is used to train the salinity positive and negative deviation classification model, which can improve the accuracy of the salinity positive and negative deviation classification model in predicting the positive and negative deviations of salinity data.
[0073] In some embodiments, the first training data set is normalized as follows:
[0074]
[0075] Among them, X′ j represents the normalized set of j sea surface elements, X j represents the jth sea surface feature dataset in the first training dataset, represents the minimum value of the jth sea surface feature dataset, Indicates the maximum value of the jth sea surface feature dataset.
[0076] In some embodiments, the first loss function is a cross entropy loss function.
[0077] In this embodiment, the cross entropy loss function is used to train the salinity positive and negative deviation classification model, so that a trained salinity positive and negative deviation classification model with high accuracy can be obtained.
[0078] In some embodiments, using the second training data set to train the deviation correction model to obtain a trained deviation correction model includes:
[0079] Normalizing the second training data set to obtain normalized second training data;
[0080] The normalized second training data is input into the deviation correction model, and the second loss function is used for training to obtain a trained deviation correction model.
[0081] In this embodiment, since the second training data set contains sea surface salinity, sea surface temperature, sea surface height, sea surface flow field, sea surface meridional wind speed, sea surface latitudinal wind speed, sea surface meridional wind stress, sea surface latitudinal wind stress and positive and negative deviations of salinity data, the normalized second training data is used to train the deviation correction model, which can improve the accuracy of the deviation correction model in correcting the satellite sea surface salinity deviation.
[0082] In some embodiments, the second loss function is a mean square error loss function.
[0083] In this embodiment, the mean square error loss function is used to train the deviation correction model, so that a trained deviation correction model with high accuracy can be obtained.
[0084] In some embodiments, the second loss function is used for training to obtain a trained deviation correction model, including:
[0085] The deviation correction model is trained using the second loss function to obtain a first deviation correction model;
[0086] K-fold cross validation is used to perform generalization evaluation on the first bias correction model, and the parameters of the first bias correction model are modified to obtain a trained bias correction model.
[0087] In this embodiment, K-fold cross validation is used to perform generalization evaluation on the first deviation correction model, and the parameters of the first deviation correction model are modified, which can further improve the accuracy of the deviation correction model in correcting the satellite sea surface salinity deviation.
[0088] To facilitate understanding by those skilled in the art, a set of best embodiments is provided below:
[0089] In view of the problem that the accuracy of the existing satellite sea surface salinity L2 level product deviation correction method needs to be improved, this embodiment proposes a satellite sea surface salinity L2 level product deviation correction method based on deep learning fusion MLP-LSTM model. The method of this embodiment uses SMOS satellite L2 level salinity products, measured Argo data and historical marine environmental variable data, and predicts the positive and negative deviations between the L2 level product sea surface salinity and the measured data by using a multilayer perceptron (Multilayer Perceptron, MLP). By fusing the positive and negative deviations of the sea surface salinity L2 level product with the marine element data, a deviation correction model based on the long short-term memory network (Long Short-Term Memory, LSTM) is trained to correct the error of the SMOS satellite sea surface salinity L2 level product through the model.
[0090] Step 1: Obtain satellite sea surface salinity related datasets. The satellite sea surface salinity related datasets include satellite salinity L2 level products, measured Argo data, and related sea surface parameter product data, including:
[0091] The SMOS satellite sea surface salinity L2 level along-track data for n years were downloaded from the official website of the European Space Agency (ESA), and the Argo sea surface salinity data for n years were downloaded from the international Argo website. Other sea surface data include: sea surface temperature (SST) data are from the daily best interpolated V2 SST data of the National Oceanic and Atmospheric Administration (NOAA) of the United States, with a spatial resolution of 0.25°×0.25°; sea surface height data are from the SLA data with a daily resolution of 0.25°×0.25° provided by the Copernicus Marine and Environment Monitoring Service (CMEMS); sea surface current field data are from the Ocean Surface Current Analyses Realtime (OSCAR) website, which provides global surface current products with a resolution of 1 / 3° for about 5 days; sea surface wind speed and wind stress components are from the Advanced SCATterometer (ASCAT) daily data with a resolution of 0.25°×0.25° provided by Ifremer / cersat.
[0092] Step 2: preprocess the data obtained in step 1 and interpolate and match them to the same spatiotemporal resolution. The specific steps are as follows:
[0093] Step 2.1, SMOS satellite salinity L2 level along-track data is scanned by the satellite along the track, and partial regional data is transmitted back at regular intervals. Therefore, the data must first be aggregated into daily sea surface salinity data by day. The SMOS satellite salinity L2 level product data set for a certain day (i.e., the first data set) D is as follows:
[0094]
[0095] Among them, d(i,j) represents the jth sea surface element data at the i-th time point, N represents the total number of times the satellite returns data on that day, that is, the sample data, i=[1,...,N] represents the time from the 1st time point to the Nth time point on that day; j=[1,...,4] represents the data time, longitude, latitude and salinity value, and these four attributes correspond to the 4 columns of the matrix D. The average salinity of the day is calculated according to the following formula:
[0096]
[0097] The average salinity of the day is calculated according to the above method, and the obtained daily average salinity data is combined with time, longitude and latitude to form a new data set (i.e., the second data set), i.e., the daily average data of the SMOS satellite sea surface salinity L2 product.
[0098] Step 2.2: Since the L2-level satellite sea surface salinity product is scattered data, the gridded data of sea surface temperature, sea level, sea surface flow field, sea surface wind speed component and wind stress component should be interpolated to the L2-level satellite sea surface salinity field. Specifically:
[0099] Reference Figure 2 , determine the four nearest grid points Q11, Q12, Q21, Q22 around the data point P (x, y) to be estimated. First, interpolate Q11 and Q21, Q12 and Q22 in the x direction, and then interpolate the results in the y direction. The interpolation formula is as follows:
[0100]
[0101] Among them, (x, y) represents the coordinates of the data point to be estimated, Q 11 , Q 21 , Q 12 and Q 22 represents the four grid points closest to the data point to be estimated, x1 and x2 represent the horizontal coordinates of the four grid points, y1 and y2 represent the vertical coordinates of the four grid points, f(x,y) represents the result after interpolation, and f(Q 11 )、f(Q 21 )、f(Q 12 ) and f(Q 22 ) indicates Q 11 , Q21 , Q 12 and Q 22 The corresponding interpolated data.
[0102] This interpolation formula combines the interpolation in the x-direction and the y-direction. After expansion, the interpolation is first performed in the x-direction, and then the result is interpolated in the y-direction. These two steps are combined into one formula, which is directly substituted into the formula to obtain the result after interpolation in both directions. Repeat this process for each variable to obtain the estimated value of each variable at the corresponding salinity point, so that the sea surface temperature, sea level height, sea surface flow field, sea surface wind speed component and wind stress component have the same temporal and spatial resolution as the SMOS L2 sea surface salinity.
[0103] Step 3: Build a salinity positive and negative deviation classification model based on MLP. The specific steps are as follows:
[0104] Step 3.1, refer to Figure 3 , construct a positive and negative deviation classification model of salinity based on MLP to predict whether the salinity value of SMOS L2 salinity product is positive or negative compared with the measured data. Assuming that n feature variables and m samples are input, the input layer (Input layer) X∈R m×n , construct a hidden layer with h neurons, then the weight and bias of the hidden layer can be expressed as W h ∈R n×h , b h ∈R 1×h , if there are q output label values, the weight and bias parameters of the output layer are W o ∈R h×q , b O ∈R 1×q , the output of the hidden layer and the output of the output layer are expressed as follows:
[0105] H=XW h +b h
[0106] O=HW o +b O
[0107] Step 3.2, build three hidden layers. The first hidden layer is set to have h1 neurons, the second hidden layer has h2 neurons, and the third hidden layer has h3 neurons. Each layer is connected according to the above formula. Each hidden layer is followed by a ReLU activation layer, the formula is as follows:
[0108] ReLU(input) = max(0, input)
[0109] Adding an activation layer introduces nonlinear transformation to improve the expressiveness of the model. Finally, the output layer uses the sigmoid activation function, and the formula is as follows:
[0110]
[0111] The sigmoid function limits the output layer to between 0 and 1, thereby representing the category probability and realizing the binary classification of the model.
[0112] Step 4: Construct a data set to train the MLP-based salinity positive and negative deviation classification model. The specific steps are as follows:
[0113] Step 4.1, construct the MLP classification model data set (i.e., the first training data set) X. The aggregation method of the data set X is as follows:
[0114]
[0115] Among them, each sample in the data set X is denoted as x i,j It represents the jth sea surface element data of the i-th sample, and the longitude and latitude of all samples are consistent; i = [1, N] represents the 1st to Nth samples, one sample is a row in the matrix X, each row contains 9 elements, j = [1, 9] represents the 1st to 9th elements, corresponding to the 9 columns of the matrix X. The first 8 columns are the 8 sea surface factors of the input data: SMOS L2 sea surface salinity, sea surface temperature, sea surface height, sea surface flow field, sea surface meridional wind speed, sea surface zonal wind speed, sea surface meridional wind stress, sea surface zonal wind stress, and the last column is the positive and negative deviation of the label data for the salinity data, that is, the difference between the SMOS L2 salinity data and the measured data, the positive value is the positive deviation, and the negative value is the negative deviation. The paired data set X is divided into the training set X according to the ratio of 8:1:1. a , validation set X b With the test set X c The specific division method is: select the first 0.8×N rows of samples as the training set X a The remaining 0.1×N rows of samples are used as the validation set X b , 0.1×N rows of samples as the test set X c , there is no intersection among the three.
[0116] Step 4.2, input data normalization, select MinMax normalization method, normalize the sea surface temperature, sea surface height, sea surface flow field, sea surface wind speed (longitudinal, latitudinal), sea surface wind stress (longitudinal, latitudinal) and sea surface salinity data in the input data set X of a certain day to the interval [0,1]. The normalization formula expression is as follows:
[0117]
[0118] Among them, the input data set X of a certain day represents the input data set of a certain day aggregated in step 4.1 except for the positive and negative deviations of the ninth column of salinity, X j Represents an input sea surface element in the dataset, where X1 is sea surface salinity, X2 is sea surface temperature, X3 is sea surface height, X4 is sea surface current field, X5 is sea surface meridional wind speed, X6 is sea surface zonal wind speed, X7 is sea surface meridional wind stress, and X8 is sea surface zonal wind stress; is the jth sea surface feature dataset X j The minimum value of is the jth sea surface feature dataset X j The maximum value of each element set after normalization according to the formula is recorded as X′ j ; and Will be retained. During model validation and testing, the normalization of a batch of input data in the validation set and test set is also adopted and The calculation is performed because the statistical information contained in the training set is considered to be a known condition, while the statistical information of the test set and the validation set is an unknown condition. The purpose is to simulate the reasoning effect of the model on the unknown data set based on the existing information in most cases.
[0119] In step 4.3, cross entropy loss is selected as the loss function of the MLP-based salinity positive and negative deviation classification model. The loss function formula is as follows:
[0120]
[0121] Among them, y i represents the label of sample i, that is, the positive and negative deviations corresponding to the salinity value, the positive class is 1, the negative class is 0, and p i It represents the probability that sample i is predicted to be a positive class.
[0122] Step 4.4: Put the training set into the MLP-based salinity positive and negative deviation classification model for training, and fit the hidden layer weights and deviations W h , b h , the loss function is the cross entropy loss function, the number of initial iterations is epoch, epoch is a positive integer, the initial learning rate is lr, the initial batch size is batch size, and the optimizer is Adam.
[0123] Step 4.5, save the trained salinity positive and negative deviation classification model and model output parameters.
[0124] Step 5: Use the trained salinity positive and negative deviation classification model to obtain the positive and negative deviations of salinity data and construct an LSTM model data set (i.e., the second training data set). The specific steps are as follows:
[0125] Step 5.1, combined with the positive and negative deviations of the output salinity of the MLP in step 4.5, construct the LSTM model dataset X1. The aggregation method of the dataset X1 is as follows:
[0126]
[0127] Among them, each sample in the data set X1 is recorded as x i,j It represents the jth sea surface element data of the ith sample, and the longitude and latitude of all samples are consistent; i=[1, N] represents the 1st to Nth samples, one sample is a row in the matrix X1, each row contains 10 elements, j=[1, 10] represents the 1st to 10th elements, corresponding to the 10 columns of the matrix X1. The first 8 columns are the 8 sea surface factors of the input data: SMOS L2 level sea surface salinity, sea surface temperature, sea surface height, sea surface flow field, sea surface meridional wind speed, sea surface zonal wind speed, sea surface meridional wind stress, sea surface zonal wind stress, the ninth column is the positive and negative deviation of the salinity data, and the last column is the label: measured salinity data. Divide the paired data set X1 into training set X1 in a ratio of 8:1:1 a , validation set X1 b With the test set X1 c The specific division method is: select the first 0.8×N rows of samples as the training set X1 a The remaining 0.1×N rows of samples are used as the validation set X1 b , 0.1×N rows of samples as the test set X1 c , there is no intersection among the three.
[0128] Step 6: Build a deviation correction model based on LSTM. The specific steps are as follows:
[0129] Step 6.1, refer to Figure 4 , build a LSTM-based deviation correction model to revise the deviation of SMOS L2 sea surface salinity product data to reduce the error with the measured data. The LSTM-based deviation correction model is divided into input layer, LSTM layer, fully connected layer and output layer. Input layer: input n variables; LSTM layer: each LSTM unit contains a forget gate, an input gate and an output gate, and dropout is added between every two LSTM units; fully connected layer: mainly for learning the nonlinear combination of features and adjusting the salinity prediction value output by the previous layer according to the calculated deviation; output layer: processes the adjustment of the salinity prediction value by the fully connected layer to produce the final prediction result, that is, the salinity data after deviation correction.
[0130] Step 6.2, use the ReLU activation layer in the LSTM layer and the fully connected layer. The formula is as follows:
[0131] ReLU(input) = max(0, input)
[0132] The purpose of this activation function is to increase the nonlinearity of the model and the sparsity of the network.
[0133] Step 7: Use the constructed data set to train the LSTM-based deviation correction model. The specific steps are as follows:
[0134] Step 7.1, input data normalization, select MinMax normalization method, normalize the sea surface temperature, sea surface height, sea surface flow field, sea surface wind speed (longitudinal and latitudinal), sea surface wind stress (longitudinal and latitudinal) and sea surface salinity data in the input data set X1 of a certain day to the interval [0,1]. The normalization formula is as follows:
[0135]
[0136] Among them, the input data set X1 of a certain day represents the input data set of a certain day aggregated in step 5.1 except for the positive and negative deviations of the ninth column of salinity, X1 j Represents an input sea surface element in the dataset, where X11 is sea surface salinity, X12 is sea surface temperature, X13 is sea surface height, X14 is sea surface flow field, X15 is sea surface meridional wind speed, X16 is sea surface zonal wind speed, X17 is sea surface meridional wind stress, X18 is sea surface zonal wind stress, X1 10 is the measured salinity data; is the jth sea surface feature dataset X1 j The minimum value of is the jth sea surface feature dataset X1 j The maximum value of each element set after normalization according to the formula is recorded as X′1 j ; and Will be retained. During model validation and testing, the normalization of a batch of input data in the validation set and test set is also adopted and The calculation is performed because the statistical information contained in the training set is considered to be a known condition, while the statistical information of the test set and the validation set is an unknown condition. The purpose is to simulate the reasoning effect of the model on the unknown data set based on the existing information in most cases.
[0137] Step 7.2, select mean square error MSE as the loss function of the deviation correction model to measure the difference between the corrected salinity and the measured salinity. The formula is as follows:
[0138]
[0139] Among them, Y i represents the label of sample i, i.e. the measured salinity data, Y i ′ represents the value of sample i after model correction.
[0140] Step 7.3, put the training set into the LSTM bias correction model for training, the loss function is the mean square error MSE, the number of iterations is initialized to epoch, epoch is a positive integer, the initial learning rate is lr, the initial batch size is batchsize, and the optimizer is Adam.
[0141] Step 7.4, perform generalization evaluation of the model through K-fold cross-validation and modify model parameters. Divide the data set in step 5.1 into two parts according to a ratio of 7:3, one part for cross-validation and the other part for the final test set. Further divide the cross-validation data set into k subsets of equal size, use the collection of k-1 subsets as training data to train the model, and use the remaining subset as test data to evaluate model performance. At the end of each cycle, record the performance of the model on the test data to obtain the overall performance of the model, analyze the performance of the model on different subsets, and evaluate its generalization ability and stability. After completing all cross-validation cycles, use the previously retained test set to perform a final evaluation of the model. Compare the cross-validation results with the results on the final test set to verify the consistency of the model performance, and adjust the model parameters such as epoch and lr according to the cross-validation results.
[0142] Step 7.5, save the trained bias correction model and model output parameters.
[0143] Step 8: Correct the satellite sea surface salinity L2 product deviation (i.e., satellite sea surface salinity deviation) through the trained deviation correction model. The specific steps are as follows:
[0144] Step 8.1, use the root mean square error RMSE and mean absolute error MAE to evaluate the accuracy of the bias-corrected SMOS L2 salinity data output by the model in step 7.4. The formula is as follows:
[0145]
[0146]
[0147] Among them, Y i represents the label of sample i, i.e. the measured salinity data, Y i ′ represents the value of the model after correction of sample i. The RMSE and MAE of the output salinity data and the measured results are calculated to evaluate the model effect.
[0148] This embodiment uses existing public data sets and deep learning frameworks to build, train and predict models. By obtaining existing public data sets from the official website of ocean data products for model training and testing, the model is built based on deep learning MLP and LSTM, which is low-cost and easy to implement. This embodiment provides an alternative method for the deviation correction of satellite L2 salinity data, which can supplement the existing methods. Using deep learning methods, a deep learning model for super-resolution reconstruction of sea surface salinity is constructed. By using nearly ten years of ocean element data to train the model, and then directly using the trained model to achieve high-precision super-resolution reconstruction of sea surface salinity. The model constructed based on MLP and LSTM in this embodiment can be used as a set of general paradigms, applied to the deviation correction tasks of different salinity satellites or other ocean elements. By preprocessing other element data and model training according to the method, the corresponding ocean element can also be corrected to improve data accuracy.
[0149] The above is a specific description of the preferred implementation of the embodiments of the present application, but the embodiments of the present application are not limited to the above-mentioned implementation methods. Technical personnel familiar with the field can also make various equivalent modifications or substitutions without violating the spirit of the embodiments of the present application. These equivalent modifications or substitutions are all included in the scope defined by the claims of the embodiments of the present application.
Claims
1. A satellite sea surface salinity deviation correction method based on deep learning, characterized in that: The satellite sea surface salinity deviation correction method based on deep learning includes: Obtain satellite sea surface salinity related datasets; Preprocessing the data in the satellite sea surface salinity related data set and interpolating and matching them to the same temporal and spatial resolution to obtain a processed satellite sea surface salinity related data set; specifically including: Dividing the sea surface salinity data in the satellite sea surface salinity related data set into a plurality of first data sets according to a preset time period; Averaging the salinity values in each of the first data sets to obtain an average salinity value corresponding to each preset time period; Constructing a second data set according to the average salinity value corresponding to each of the preset time periods; Interpolate and match other data in the satellite sea surface salinity related data set to the second data set of the same temporal and spatial resolution to obtain a processed satellite sea surface salinity related data set; wherein: Interpolate and match other data in the satellite sea surface salinity related dataset to the second dataset of the same spatiotemporal resolution in the following manner: in, represents the coordinates of the data point to be estimated, , , and represents the four grid points closest to the data point to be estimated, , represents the horizontal coordinates of the four grid points, , represents the ordinates of the four grid points, represents the result after interpolation, , , and express , , and The corresponding interpolation data; Based on the multi-layer perceptron, a classification model of salinity positive and negative deviations was constructed; Based on the processed satellite sea surface salinity related data set, a first training data set is constructed, and the salinity positive and negative deviation classification model is trained using the first training data set to obtain a trained salinity positive and negative deviation classification model, so as to obtain the positive and negative deviations of salinity data using the trained salinity positive and negative deviation classification model; Based on the long short-term memory network, a deviation correction model is constructed; Based on the positive and negative deviations of the salinity data, a second training data set is constructed, and the deviation correction model is trained using the second training data set to obtain a trained deviation correction model, so that the satellite sea surface salinity deviation can be corrected using the trained deviation correction model.
2. The satellite sea surface salinity deviation correction method based on deep learning according to claim 1, characterized in that: The satellite sea surface salinity-related data set includes sea surface salinity, sea surface temperature, sea surface height, sea surface flow field, sea surface meridional wind speed, sea surface zonal wind speed, sea surface meridional wind stress and sea surface zonal wind stress.
3. The satellite sea surface salinity deviation correction method based on deep learning according to claim 1, characterized in that: The method of using the first training data set to train the salinity positive and negative deviation classification model to obtain a trained salinity positive and negative deviation classification model includes: Normalizing the first training data set to obtain normalized first training data; The normalized first training data is input into the salinity positive and negative deviation classification model, and the training is performed using a first loss function to obtain a trained salinity positive and negative deviation classification model.
4. The satellite sea surface salinity deviation correction method based on deep learning according to claim 3 is characterized in that: The first training data set is normalized in the following manner: in, Represents the normalized A collection of sea surface elements, Indicates the first training data set Sea surface feature datasets, Indicates The minimum value of the sea surface feature dataset, Indicates The maximum value of sea surface feature datasets.
5. The satellite sea surface salinity deviation correction method based on deep learning according to claim 3 is characterized in that: The first loss function is a cross entropy loss function.
6. The satellite sea surface salinity deviation correction method based on deep learning according to claim 1, characterized in that: The step of using the second training data set to train the deviation correction model to obtain a trained deviation correction model includes: Normalizing the second training data set to obtain normalized second training data; The normalized second training data is input into the deviation correction model, and training is performed using a second loss function to obtain a trained deviation correction model.
7. The satellite sea surface salinity deviation correction method based on deep learning according to claim 6 is characterized in that: The second loss function is a mean square error loss function.
8. The method for correcting satellite sea surface salinity deviation based on deep learning according to claim 6, characterized in that: The second loss function is used for training to obtain a trained deviation correction model, including: Using the second loss function to train the deviation correction model to obtain a first deviation correction model; K-fold cross validation is used to perform a generalization evaluation on the first bias correction model, and the parameters of the first bias correction model are modified to obtain a trained bias correction model.
