A method for retrieving nearshore sea surface temperature from microwave radiometer based on neural network
By constructing a radial basis function neural network model and using spatiotemporal matching of brightness temperature data from a multi-frequency microwave radiometer and measured sea surface temperature auxiliary data, the problem of microwave radiometers being unable to accurately retrieve sea surface temperature in nearshore areas in traditional methods was solved, achieving rapid and accurate nearshore sea surface temperature retrieval.
Patent Information
- Application Number
- CN202210223805.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-03-07
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2042-03-07
AI Technical Summary
Traditional microwave radiometers are affected by land pollution in nearshore areas and cannot accurately infer nearshore sea surface temperature. Existing methods cannot quickly and accurately obtain nearshore sea surface temperature.
A neural network-based method for retrieving nearshore sea surface temperature (SST) using a multi-frequency microwave radiometer is constructed. The model is trained by using a radial basis function neural network model and spatiotemporal matching of brightness temperature data from the multi-frequency microwave radiometer and measured SST auxiliary data to retrieve nearshore SST.
It achieves rapid and accurate nearshore sea surface temperature inversion, avoiding the complexity of physical algorithms and the shortcomings of traditional methods, and improving the accuracy and speed of nearshore sea surface temperature acquisition.
Smart Images

Figure CN114970663B_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of passive microwave remote sensing, and in particular relates to a nearshore sea surface temperature inversion method of a microwave radiometer based on a neural network model. Background Art
[0002] Sea surface temperature is one of the main factors affecting global climate change and long-term weather processes, and plays an important role in the global water cycle and the global surface energy balance. As human activities continue to expand into the ocean area, the acquisition of sea surface temperature products with high spatial resolution, high temporal resolution and high global coverage has attracted more and more attention. Traditional sea surface temperature measurement methods mainly include ship, shore-based and offshore buoy monitoring, which can directly obtain real-time sea surface temperature measurement data. However, due to factors such as observation cost, labor cost and weather conditions, these traditional observation methods cannot meet the requirements of spatial distribution continuity and temporal distribution continuity, and it is difficult to obtain global sea surface temperature products.
[0003] The development of remote sensing technology has made it possible to obtain sea surface temperatures with high temporal resolution and high spatial resolution on a global scale. The payloads of remote sensing satellites for measuring sea surface temperature mainly include infrared radiometers and microwave radiometers, both of which can provide sea surface temperatures with global coverage and high spatial and temporal resolution. However, infrared radiometers are affected by water vapor and clouds, which can easily cause data coverage holes. In contrast, microwave radiometers can penetrate thinner clouds and can observe sea surface temperatures uninterruptedly around the clock. However, microwave radiometers also have their own problems in obtaining global sea temperatures, because microwave radiometers mainly rely on low-frequency channels (such as 6GHz and 10GHz) to invert sea surface temperature, which has lower spatial resolution than high-frequency channels, that is, the ground coverage of each observation point is larger. This causes the microwave radiometer brightness temperature to be easily contaminated by land or sea ice when observing nearshore or near sea ice, making it impossible to obtain the correct sea temperature. The conventional practice is to use a land-expanding mask to remove the brightness temperature data within 50-200km of the land or sea ice, and only invert the sea temperature in the open sea. This results in no sea temperature inversion results for large areas of sea near the coast or near sea ice. The nearshore sea area is the main area for human production and life, and the acquisition of nearshore sea temperature is particularly important for the marine ecological environment, coastal tourism, and marine aquaculture. Therefore, it is urgent to solve the problem that the brightness temperature of the microwave radiometer is contaminated by land in the nearshore area and cannot be correctly inverted for the nearshore sea temperature. Summary of the invention
[0004] The traditional multi-frequency microwave radiometer sea temperature inversion method is only applicable to open sea areas, and it is difficult to physically model the nearshore scene, so it is impossible to quickly and accurately obtain the nearshore sea temperature based on the multi-frequency microwave radiometer. In view of the above problems, the present invention provides a multi-frequency microwave radiometer nearshore sea temperature inversion method based on a neural network. By constructing a neural network model suitable for nearshore sea temperature inversion of microwave radiometer, the sea temperature within 0 to 200 km offshore can be quickly and accurately obtained, which can effectively avoid the complexity of physical algorithms and the shortcomings of traditional methods.
[0005] In order to achieve the above object, the present invention provides a multi-frequency microwave radiometer nearshore sea temperature inversion method based on a neural network model. The method comprises:
[0006] Step (1) reading the brightness temperature observation data of the multi-frequency microwave radiometer, and selecting the brightness temperature data with a land percentage greater than 0% and less than 50% and no sea ice and no rainfall;
[0007] Step (2) uses the brightness temperature data selected in step (1) to perform spatiotemporal matching with the measured sea temperature auxiliary data to obtain a matching data set of multi-frequency brightness temperature and sea temperature data in the nearshore area;
[0008] Step (3) initializing the radial basis neural network model, using the matching data set obtained in step (2) as input data for model training;
[0009] Step (4) uses grid search to determine the key parameters of the model, obtains a radial basis function neural network model suitable for nearshore microwave radiometer sea temperature inversion, and based on the model, uses multi-frequency microwave radiometer brightness temperature observation data to invert the nearshore sea temperature.
[0010] Furthermore, the method also includes: step (5) using new microwave radiometer brightness temperature data to verify the nearshore sea temperature inversion accuracy of the radial basis neural network model obtained in step (4).
[0011] Furthermore, the measured sea temperature auxiliary data selects the iQuam measured sea temperature dataset.
[0012] Furthermore, the method for performing spatiotemporal matching in step (2) specifically includes: determining the size of the space window and the time window, screening out all matching points in the space window, and calculating the following formula:
[0013]
[0014] a=Lat1-Lat2
[0015] b=Long1-Long2
[0016] Among them, S represents the distance between the longitude and latitude of the observation points of the two data sources, Lat1 and Long1 are the longitude and latitude corresponding to the observation point of the brightness temperature data; Lat2 and Long2 are the longitude and latitude corresponding to the observation point of the measured sea temperature auxiliary data; R is the radius of the earth;
[0017] Then, these matching points are screened by time window, and the calculation formula is as follows:
[0018] T = Scantime - obser_time
[0019] Among them, Scantime is the observation time corresponding to the brightness temperature data, and obser_time is the observation time corresponding to the measured sea temperature auxiliary data.
[0020] Furthermore, the spatial window is selected as 25 km, the temporal window is selected as 30 min, and all matching points with S less than 25 km are screened out. Then, the observation locations with T less than 30 min are screened out from these matching points as the matching data set.
[0021] Furthermore, the step (3) specifically includes, according to the characteristics of the nearshore sea temperature inversion of the multi-frequency microwave radiometer, assuming that the input vector is a d-dimensional vector The output is a real value, so the initialization radial basis neural network model is constructed as:
[0022] Where q is the number of neurons in the hidden layer; c i is the center corresponding to the hidden layer neuron i, which can be obtained by; ω i is the weight corresponding to the hidden neuron i, ρ(x, c i ) is the radial basis function, defined as the square of the sample x and the center c i The Euclidean distance between .
[0023] Furthermore, the ρ(x, c i ) is the Gaussian radial basis function, formula: β i is the expansion factor related to the variance;
[0024] The loss function using least squares is:
[0025]
[0026] y i represents the label value corresponding to the training set, that is, the true value, and P represents the number of training sets;
[0027] There are three parameters to solve the model: c i , β i And ω i .
[0028] Furthermore, the step (4) specifically includes: using a grid search algorithm to arrange and combine all possible parameter values, calculating the results of all combinations to generate a grid, and finding the parameter combination with the highest accuracy. This parameter combination is the optimal model parameter suitable for multi-frequency microwave radiometer nearshore sea temperature inversion.
[0029] Furthermore, the optimal model parameters are: 6.925GHz-H, 6.925GHz-V, 10.7GHz-H, 10.7GHz-V, 18.7GHz-H, 18.7GHz-V, 23.8GHz-V, 37.0GHz-H, 37.0GHz-V, Lat_of_Observation_Point, Long_of_Observation_Point and Scan_Time.
[0030] Furthermore, the step (5) specifically includes: verifying the statistical information of the nearshore sea temperature bias Bias, the inverted nearshore sea temperature root mean square error RMSE and the correlation coefficient R output by the optimal radial basis neural network model in step (4), wherein the nearshore sea temperature accuracy RMSE is:
[0031]
[0032] The average deviation Bias is:
[0033]
[0034] The correlation coefficient R is:
[0035]
[0036] N is the number of samples, A i represents the inversion value, B i Represented as field measurement value or true value, represents the mean of N A, Represents the mean of N Bs.
[0037] The beneficial effects of the present invention are as follows:
[0038] 1. Use the land percentage mark to accurately select the brightness temperature polluted by land. Based on this brightness temperature, the sea temperature inversion model suitable for coastal waters is trained, which has the advantages of fast inversion speed and high accuracy.
[0039] 2. The present invention applies radial basis function neural network to the research of the method of inversion of nearshore sea temperature by multi-frequency microwave radiometer. Compared with the traditional method of inversion of sea temperature by multi-frequency microwave radiometer, the inversion method fills the gap of sea temperature inversion by microwave radiometer in nearshore waters.
[0040] 3. The multi-frequency microwave radiometer nearshore sea temperature inversion method based on the radial basis neural network model proposed in the present invention can effectively avoid the problems of complex and time-consuming calculations of physical algorithms and the lack of models for nearshore scenes. BRIEF DESCRIPTION OF THE DRAWINGS
[0041] Figure 1 This is a flowchart of the inversion of nearshore sea temperature using a multi-frequency microwave radiometer based on a neural network.
[0042] Figure 2 A model suitable for nearshore sea temperature inversion using a multi-frequency microwave radiometer is constructed for the embodiment.
[0043] Figure 3 The SMR global nearshore sea temperature distribution inverted by the radial basis function neural network model constructed for the embodiment.
[0044] Figure 4 Scatter plot comparing the SMR nearshore sea temperature inverted by the radial basis function neural network model constructed in the embodiment and the iQuam measured sea temperature.
[0045] Figure 5 This is a scatter plot comparing the SMR nearshore sea temperature inverted by the statistical regression method and the iQuam measured sea temperature. DETAILED DESCRIPTION
[0046] The present invention is further described in detail below with reference to specific embodiments.
[0047] Example
[0048] This embodiment introduces in detail the multi-frequency microwave radiometer nearshore sea temperature inversion method based on neural network. The multi-frequency microwave radiometer brightness temperature data of this embodiment is selected from the scanning microwave radiometer (SMR) multi-frequency brightness temperature data of the HY-2B satellite (https: / / osdds.nsoas.org.cn / # / ); the sea temperature measured auxiliary data is selected from the iQuam measured sea temperature data set provided by NOAA (http: / / www.star.nesdis.noaa.gov / sod / sst / iquam / index.html).
[0049] like Figure 1 As shown in FIG, the method is first based on the multi-frequency brightness temperature data of SMR, and uses the quality mark corresponding to the brightness temperature data to select the brightness temperature data with qualified quality, uses the sea ice mark and the rainfall mark to select the clear sky and ice-free ocean area, and uses the land percentage mark to select the coastal sea area with a land percentage of 0% to 50%, that is, there is land in the observation field of the scanning point, but the land ratio is less than 50%, so that each selected brightness temperature observation point is guaranteed to be a mixture of land and ocean and the ocean area is larger than the land area;
[0050] Secondly, the iQuam measured sea temperature auxiliary data was used to match the screened SMR multi-frequency brightness temperature data in time and space, with a time window of 30 minutes and a space window of 25 km to form a brightness temperature-sea temperature matching data set for the nearshore waters.
[0051] Then, the radial basis function neural network model is initialized, and the brightness temperature-sea temperature matching dataset of the nearshore waters is taken as the model input. The optimal parameters of the model are determined using grid search, and a model suitable for the nearshore sea temperature inversion of multi-frequency microwave radiometer is established based on the optimal model parameter settings.
[0052] Then, the new nearshore SMR brightness temperature data and related SST measured auxiliary data are substituted into the established nearshore SST inversion model to invert the nearshore SST.
[0053] Finally, the accuracy of SST inversion is evaluated, and the traditional statistical regression method is used as a comparative experiment to synchronously invert the SMR nearshore SST, and the inversion results of the two methods are compared and evaluated.
[0054] The time span of the data set used in this embodiment is from October 2020 to November 2020, of which the data in October 2020 is used as the training data and test data of the model, and the data in November 2020 is used as the verification data of the model. The results show that the method proposed in the present invention can effectively obtain nearshore sea surface temperature information.
[0055] The specific working process of the neural network-based multi-frequency microwave radiometer nearshore sea temperature inversion method of this embodiment is described in detail below.
[0056] Step (1) reads the SMR brightness temperature observation data, and selects the brightness temperature observation data with a land percentage greater than 0% and less than 50% and no sea ice and no rainfall according to the data quality mark, land percentage mark, sea ice mark, rainfall mark, etc.
[0057] Step (2) uses the SMR brightness temperature data selected in step (1) to perform spatiotemporal matching with the measured sea temperature auxiliary data iQuam to obtain a multi-frequency brightness temperature and sea temperature data matching dataset for the nearshore area. The time window is 30 minutes and the space window is 25 km.
[0058] The specific method for spatially matching the two data sources to be matched is to calculate the distance between the longitude and latitude of the SMR brightness temperature observation point and the longitude and latitude of the iQuam measured sea temperature according to the following calculation formula:
[0059]
[0060] a=Lat1-Lat2
[0061] b=Long1-Long2
[0062] Among them, Lat1 and Long1 are the longitude and latitude of the observation point corresponding to the brightness temperature data of SMR; Lat2 and Long2 are the longitude and latitude of the observation point corresponding to the iQuam measured sea temperature auxiliary data; R is the radius of the earth, which is 6378.137 kilometers.
[0063] After filtering out all matching points with S less than 25 km, these matching points are filtered by time window. The calculation formula is as follows:
[0064] T = Scantime - obser_time
[0065] Among them, scantime is the observation time corresponding to the SMR brightness temperature data, and obser_time is the observation time corresponding to the iQuam measured sea temperature auxiliary data. The observation locations with T less than 30min are matched data sets.
[0066] Step (3) initializes the radial basis neural network model and uses the matching data set obtained in step (2) as input data for model training.
[0067] The neural network model is a three-layer network structure, including input layer, hidden layer and output layer. The transformation from input layer to hidden layer is nonlinear, while the transformation from hidden layer to output layer is linear.
[0068]
[0069] point, assuming that the input vector is a d-dimensional vector The output is a real value, so the model can be constructed as:
[0070] Where q is the number of neurons in the hidden layer; c i is the center corresponding to the hidden layer neuron i, which can be obtained by; ω i is the weight corresponding to the hidden neuron i, ρ(x, c i ) is the radial basis function, defined as the square of the sample x and the center c i The Euclidean distance between them, in this embodiment, the function is defined as a Gaussian radial basis function, and the formula is as follows:
[0071]
[0072] β i is the expansion factor related to the variance.
[0073] The least squares loss function is expressed as:
[0074]
[0075] yi It represents the label value corresponding to the training set, that is, the true value, and P represents the number of training sets.
[0076] There are three parameters to solve the model: c i , β i And ω i .
[0077] Step (4) uses grid search to determine the key parameters of the model, such as the diffusion factor, the number of hidden layer neurons, etc., to obtain a radial basis function neural network model suitable for nearshore microwave radiometer sea temperature inversion.
[0078] Using the grid search algorithm, all possible parameter values are arranged and combined, and the results of all combinations are calculated to generate a grid, and the parameter combination with the highest accuracy is found: 6.925GHz-H, 6.925GHz-V, 10.7GHz-H, 10.7GHz-V, 18.7GHz-H, 18.7GHz-V, 23.8GHz-V, 37.0GHz-H, 37.0GHz-V, Lat_of_Observation_Point, Long_of_Observation_Point, Scan_Time. This parameter combination is the optimal model parameter for the nearshore sea temperature inversion of the multi-frequency microwave radiometer. Figure 2 The model constructed for the present invention and suitable for the inversion of nearshore sea temperature using a multi-frequency microwave radiometer has 12 input layers, 420 hidden layers, 1 output layer, and a diffusion factor set to 95.
[0079] Step 5) Use the new SMR brightness temperature data to verify the accuracy of the nearshore SST inversion of the radial basis function neural network model in step (4) and evaluate the accuracy of the nearshore SST inverted by the model.
[0080] Verification step (4) The best radial basis neural network model outputs the statistical information of the inverted nearshore sea temperature bias Bias, the inverted nearshore sea temperature root mean square error RMSE and the correlation coefficient R, where the nearshore sea temperature accuracy RMSE is:
[0081]
[0082] The average deviation Bias is:
[0083]
[0084] The correlation coefficient is R:
[0085]
[0086] N is the number of samples, A i represents the inversion value, B iRepresented as field measurement value or true value, represents the mean of N A, Represents the mean of N Bs.
[0087] Figure 3 The SMR global nearshore sea temperature distribution inverted by the radial basis function neural network model constructed by the present invention. Figure 4 This is a scatter plot comparing the SMR nearshore sea temperature inverted by the radial basis function neural network model constructed in the present invention and the iQuam measured sea temperature.
[0088] This embodiment uses the traditional statistical regression method as a comparative experiment to synchronously invert the SMR nearshore sea temperature, and compares and evaluates the inversion results of the two methods. Figure 5 This is a scatter plot comparing the SMR nearshore sea temperature inverted by the statistical regression method with the iQuam measured sea temperature. Figure 4 and Figure 5 It can be seen that the model suitable for nearshore sea temperature inversion constructed by the present invention is better than the statistical regression method. Table 1 gives the statistical information of the comparison of the two methods for nearshore sea temperature inversion.
[0089] Table 1 Statistical information of the model constructed by the embodiment of the present invention and the inversion of the nearshore sea temperature by the statistical regression method
[0090] index Model constructed by the embodiment Statistical regression Average deviation (Bias) -0.071℃ -0.454℃ Root mean square error (RMSE) 1.188℃ 2.529℃ Correlation coefficient (R) 0.99 0.955
[0091] Depend on Figure 4 , Figure 5 It can be concluded from Table 1 that the radial basis neural network model constructed in this paper is more suitable for nearshore SST inversion than the traditional statistical regression method, and its average deviation, root mean square error and correlation coefficient are better than those of the statistical regression method.
[0092] The embodiments described above are only descriptions of the preferred implementation modes of the present invention, and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should fall within the protection scope determined by the claims of the present invention.
Claims
1. A method for inverting nearshore sea temperature using a microwave radiometer based on a neural network model, characterized in that: The method includes: Step (1) reading the brightness temperature observation data of the multi-frequency microwave radiometer, and selecting the brightness temperature data with a land percentage greater than 0% and less than 50% and no sea ice and no rainfall; Step (2) uses the brightness temperature data selected in step (1) to perform spatiotemporal matching with the measured sea temperature auxiliary data to obtain a matching data set of multi-frequency brightness temperature and sea temperature data in the nearshore area; Step (3) initializing the radial basis neural network model, using the matching data set obtained in step (2) as input data for model training; Step (4) using a grid search to determine the key parameters of the model, obtaining a radial basis function neural network model suitable for inversion of nearshore microwave radiometer sea temperature, and based on the model, inverting the nearshore sea temperature using multi-frequency microwave radiometer brightness temperature observation data; The method for performing spatiotemporal matching in step (2) specifically includes: determining the size of the space window and the time window, screening out all matching points in the space window, and calculating the following formula: a=Lat1-Lat2 b=Long1-Long2 Among them, S represents the distance between the longitude and latitude of the observation points of the two data sources, Lat1 and Long1 are the longitude and latitude corresponding to the observation point of the brightness temperature data; Lat2 and Long2 are the longitude and latitude corresponding to the observation point of the measured sea temperature auxiliary data; R is the radius of the earth; Then, these matching points are screened by time window, and the calculation formula is as follows: T = Scantime - obser_time Among them, Scantime is the observation time corresponding to the brightness temperature data, and obser_time is the observation time corresponding to the measured sea temperature auxiliary data.
2. The method for inverting nearshore sea temperature using a microwave radiometer based on a neural network model according to claim 1, characterized in that: Also includes: The new microwave radiometer brightness temperature data is used to verify the nearshore sea temperature inversion accuracy of the radial basis function neural network model obtained in step (4).
3. The method for inverting nearshore sea temperature using a microwave radiometer based on a neural network model according to claim 1, characterized in that: In step (2), the spatial window is selected as 25 km, the time window is selected as 30 min, and all matching points with S less than 25 km are screened out. Then, from these matching points, the observation locations with T less than 30 min are screened out as the matching data set.
4. The method for inverting nearshore sea temperature using a microwave radiometer based on a neural network model according to claim 1, characterized in that: The step (3) specifically includes, according to the characteristics of the nearshore sea temperature inversion of the multi-frequency microwave radiometer, assuming that the input vector is a d-dimensional vector The output is a real value, so the initialization radial basis neural network model is constructed as: Where q is the number of neurons in the hidden layer; c i is the center corresponding to the hidden layer neuron i, which can be obtained by; ω i is the weight corresponding to the hidden neuron i, ρ(x, c i ) is the radial basis function, defined as the square of the sample x and the center c i The Euclidean distance between .
5. The method for inverting nearshore sea temperature using a microwave radiometer based on a neural network model according to claim 4 is characterized in that: The ρ(x, c i ) is the Gaussian radial basis function, formula: β i is the expansion factor related to the variance; The loss function using least squares is: y i represents the label value corresponding to the training set, that is, the true value, and P represents the number of training sets; There are three parameters to solve the model: c i , β i And ω i .
6. The method for inverting nearshore sea temperature using a microwave radiometer based on a neural network model according to claim 1, characterized in that: The step (4) specifically includes: using a grid search algorithm to arrange and combine all possible parameter values, calculating the results of all combinations to generate a grid, and finding the parameter combination with the highest accuracy. This parameter combination is the optimal model parameter suitable for multi-frequency microwave radiometer nearshore sea temperature inversion.
7. The method for inverting nearshore sea temperature using a microwave radiometer based on a neural network model according to claim 6 is characterized in that: The optimal model parameters are: 6.925GHz-H, 6.925GHz-V, 10.7GHz-H, 10.7GHz-V, 18.7GHz-H, 18.7GHz-V, 23.8GHz-V, 37.0GHz-H, 37.0GHz-V, Lat_of_Observation_Point, Long_of_Observation_Point and Scan_Time.
8. The method for inverting nearshore sea temperature using a microwave radiometer based on a neural network model according to claim 2 is characterized in that: Verification step (4) The best radial basis neural network model outputs the statistical information of the inverted nearshore sea temperature bias Bias, the inverted nearshore sea temperature root mean square error RMSE and the correlation coefficient R, where the nearshore sea temperature accuracy RMSE is: The average deviation Bias is: The correlation coefficient R is: N is the number of samples, A i represents the inversion value, B i Represented as field measurement value or true value, represents the mean of N A, Represents the mean of N Bs.
