Sea temperature inversion method based on K-Ka band microwave radiometer weak signal enhancement deep learning

Abstract

The invention belongs to the technical field of ocean remote sensing detection, and relates to a sea temperature inversion method based on K-Ka band microwave radiometer weak signal enhancement deep learning. Parameters of measured data of the microwave radiometer and reanalysis sea gas parameter data are selected; determining the size of a space-time window based on the brightness temperature characteristics of K and Ka wavebands, and carrying out space-time matching to establish a modeling data set; performing quality control on the modeling data set based on indexes such as quality identification, offshore distance, brightness temperature abnormity and air-sea parameter abnormity; normalizing the modeling data set and splitting the modeling data set into a training set, a verification set and a test set; and constructing a weak signal enhancement DNN model based on the K-Ka band microwave radiometer by using the training set and the verification set, and performing inversion by using the test set to obtain a sea surface temperature product. According to the method, the precision of performing sea surface temperature inversion by using the K and Ka wave band brightness temperatures of the microwave radiometer is improved, the problem of insufficient precision is solved, and a new possibility is provided for inversion of SST on a finer scale.

Description

Technical Field

[0001] The present invention belongs to the technical field of ocean remote sensing detection, and specifically relates to a sea temperature inversion method based on weak signal enhancement deep learning of K-Ka band microwave radiometer. Background Art

[0002] Sea surface temperature (SST) is a critical ocean and climate parameter that not only reflects the energy and material exchange between the ocean and the atmosphere, but also has an important impact on climate change, ocean circulation, and ecosystems. Accurately understanding the global SST distribution and monitoring its changes are of immeasurable value for forecasting weather, understanding climate change, and even guiding ship navigation and fishing activities.

[0003] Microwave radiometers are an important type of satellite remote sensing equipment that can penetrate clouds and the atmosphere (such as water vapor and aerosols) to provide all-weather and all-day SST observations. However, microwave radiometers face some design and application challenges, especially how to balance antenna size and weight while pursuing higher resolution. For example, the C-band (4~8 GHz) and X-band (8~12 GHz) microwave radiometers will cause the antenna size to be too large in pursuit of higher resolution, which puts higher requirements on the satellite's carrying capacity, not only increasing the cost of satellite launches, but also posing technical pressure on satellite design. In contrast, K-band (18~26.5 GHz) and Ka-band (26.5~40 GHz) microwave radiometers provide a better balance between spatial resolution and antenna weight.

[0004] However, traditionally, K-band and Ka-band microwave radiometers are not given priority for SST inversion because K- and Ka-band brightness temperatures (TBs) are less sensitive to SST and more susceptible to atmospheric interference. At present, the physical and statistical SST inversion methods of microwave radiometers are mainly designed for the characteristics of C- and X-band TBs, and these methods are difficult to meet the needs of SST inversion using K- and Ka-band TBs. The signal between C- and X-band TBs and SST is relatively strong, while the signal between K- and Ka-band TBs and SST is relatively weak, and is subject to complex interference from multiple sea and air parameters such as water vapor, sea breeze, and cloud liquid water. These factors are believed to reduce the effectiveness of K- and Ka-band TBs in SST inversion. How to effectively extract and enhance the weak relationship between K- and Ka-band TBs and SST from these interference signals has become a major problem and challenge in current technological development.

[0005] Deep learning methods provide a new approach to solving this technical problem. Deep learning methods can automatically learn meaningful patterns and structures from large amounts of data without making too many assumptions about data distribution or relationships. This enables deep neural networks (DNNs) with reasonable network structure design to more accurately extract SST information from the complex response of K and Ka band TB to SST, water vapor and sea breeze, thereby improving the accuracy of SST inversion using these frequencies.

[0006] The information disclosed in this background technology section is only intended to enhance the understanding of the overall background of the invention and should not be regarded as an acknowledgment or any form of suggestion that the information constitutes the prior art already known to a person skilled in the art. Summary of the invention

[0007] The purpose of the present invention is to solve the problem in the prior art that the signal between the K and Ka band brightness temperature and the sea temperature is relatively weak, and is subject to complex interference from a variety of sea and air parameters, resulting in reduced effectiveness and accuracy of sea temperature inversion. A sea temperature inversion method based on K-Ka band microwave radiometer weak signal enhancement deep learning is proposed, which improves the accuracy of sea surface temperature inversion using microwave radiometer K and Ka band brightness temperature, and provides new possibilities for inverting SST on a finer scale.

[0008] The technical solution of the present invention is: A sea temperature inversion method based on weak signal enhancement deep learning of K-Ka band microwave radiometer, comprising the following steps: (1) Select the parameters of microwave radiometer measured data and reanalysis of sea-air parameter data. The parameters of microwave radiometer measured data include brightness temperature, azimuth and incident angle in K and Ka bands, as well as the corresponding spatiotemporal information; the parameters of reanalysis of sea-air parameter data include sea breeze, water vapor and cloud liquid water; (2) Based on the brightness temperature characteristics of the K and Ka bands, a time and space window is selected to match the microwave radiometer measured data and the reanalysis of sea and air parameter data in time and space dimensions to form a modeling data set; (3) Quality control of the modeling datasets using a variety of indicators including quality identification, offshore distance, brightness temperature anomalies, and sea-air parameter anomalies; (4) Normalize the modeling data set and split the normalized data set into training set, validation set and test set according to a certain ratio; (5) Using the training set and validation set, a weak signal enhancement DNN model based on K-Ka band microwave radiometer was constructed; (6) The test set is input into the trained weak signal enhancement DNN model to invert and output the sea surface temperature product.

[0009] Furthermore, in step (1), latitude, longitude, annual accumulated day, K and Ka band brightness temperature, azimuth and incident angle are selected from the microwave radiometer measured data as model inputs, wherein the K band includes the frequency band range of 18~26.5 GHz, and the Ka band includes the frequency band range of 26.5~40 GHz; and sea breeze, water vapor and cloud liquid water are selected from the reanalysis of sea-air parameter data as model inputs.

[0010] Furthermore, in step (2), the appropriate spatiotemporal window size is determined based on the spatiotemporal resolution of the selected microwave radiometer K and Ka band measured data and the reanalysis of sea and air parameter data, and the microwave radiometer measured data and the reanalysis of sea and air parameter data are spatiotemporally matched according to the selected spatiotemporal window to form a modeling data set.

[0011] Furthermore, in step (3), the modeling data set is quality controlled by a variety of indicators, including quality identification filtering; offshore distance filtering to ensure that the study area is located in the offshore area; brightness temperature anomaly detection to eliminate abnormally high or low brightness temperature values; and sea-air parameter anomaly detection to eliminate abnormal reanalysis data.

[0012] Furthermore, in step (4), the split training set accounts for 70% of the total data set, the validation set accounts for 15% of the total data set, and the test set accounts for 15% of the total data set.

[0013] Furthermore, in the step (5), a weak signal enhancement DNN model is constructed by using an encoder-decoder structure and a skip link network architecture, the model is trained using a training set, and the model parameters are tested and adjusted using a validation set to prevent overfitting; The encoder-decoder structure reduces the number of nodes layer by layer in the encoding part to compress the information into an abstract feature representation, while the decoding part expands the number of nodes layer by layer to reconstruct detailed information from the feature representation; The skip link network architecture fuses the features extracted by the encoding part with the features of the decoding part, effectively reducing the information loss during the transmission process and improving the network's response to subtle signal changes between the K and Ka band brightness temperatures and the sea surface temperature.

[0014] Furthermore, in step (6), the test set that has not participated in training and verification is input into the trained weak signal enhancement DNN model, and the sea surface temperature product is inverted and output; and the performance and reliability of the model are evaluated by comparing with the known true value data.

[0015] Beneficial effects of the present invention: (1) The SST inversion method based on weak signal enhancement deep learning of K-Ka band microwave radiometer proposed in this invention can enhance the extraction capability of weak signals through the constructed weak signal enhancement DNN model, and improve the sensitivity of K and Ka band brightness temperature data to SST inversion, so as to more accurately extract the SST information in the complex response of K and Ka band brightness temperature to water vapor, sea breeze, cloud liquid water, etc., thereby improving the accuracy of SST inversion using K and Ka bands.

[0016] (2) The present invention can achieve SST inversion with reasonable accuracy without the need for C- and X-band brightness temperatures that have strong correlation with SST signals. It can effectively solve the technical problem of extracting and enhancing the weak relationship between K- and Ka-band brightness temperatures and SST from complex interference signals of various sea-atmosphere parameters such as water vapor, sea breeze, and cloud liquid water. It can further produce SST products with reasonable accuracy on microwave radiometers currently in orbit that lack C- and X-band channels. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a flow chart of the sea temperature inversion method based on K-Ka band microwave radiometer weak signal enhancement deep learning of the present invention; Figure 2 It is a sensitivity analysis diagram of the simulated TB and sea breeze, water vapor, cloud liquid water and sea temperature under different water vapor conditions in the present invention; Figure 3 This is a framework diagram of the weak signal enhancement DNN model based on the K-Ka band microwave radiometer in the present invention. DETAILED DESCRIPTION

[0018] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0019] In order to further understand the present invention, the present invention will be further described in conjunction with the accompanying drawings and embodiments.

[0020] The present invention provides a physical mechanism for determining K and Ka band brightness temperatures for sea temperature inversion based on radiation transfer simulation analysis, which is as follows: The microwave simulation TB of different bands is obtained through the radiation transmission model, and the sensitivity of the simulated TB under different water vapor conditions to four sea-air parameters: sea breeze, water vapor, cloud liquid water and sea temperature To analyze, TB p Represents the brightness temperature of different polarization modes, x Represents different sea and air parameters.

[0021] By adjusting the atmospheric transmittance to simulate different water vapor conditions, the research results show that changes in water vapor can make the K and Ka band brightness temperatures more sensitive to SST, thus providing new possibilities and a strong theoretical basis for using these bands for sea temperature inversion modeling.

[0022] like Figure 1 As shown, the present invention provides a sea temperature inversion method based on K-Ka band microwave radiometer weak signal enhancement deep learning, comprising the following steps: (1) Parameters for selecting microwave radiometer measured data and reanalysis of sea and air parameter data Based on the professional knowledge in the field of sea temperature inversion, key parameters are selected from the measured data of microwave radiometers, including the brightness temperature, azimuth and incident angle of the K and Ka bands, as well as the corresponding space-time information (including latitude, longitude and annual accumulated days); the K band includes the frequency range of 18~26.5 GHz, and the Ka band includes the frequency range of 26.5~40 GHz.

[0023] The reanalysis data selected include parameters of sea-air parameters such as sea breeze, water vapor and cloud liquid water.

[0024] These selected parameters can effectively characterize the sea surface environment and play an important guiding role in the modeling process. By rationally selecting and combining these input parameters, the accuracy of the model inverted sea temperature can be significantly improved.

[0025] (2) Based on the brightness temperature characteristics of the K and Ka bands, determine the size of the spatiotemporal window and perform data matching based on the spatiotemporal window to establish a modeling data set Specifically, the appropriate space-time window size is determined based on the space-time resolution of the selected microwave radiometer K and Ka band measured data and the reanalysis sea-air parameter data. The microwave radiometer measured data and the reanalysis sea-air parameter data are time-space matched according to the selected space-time window to form a modeling data set, laying the foundation for subsequent quality control and modeling.

[0026] (3) Quality control of modeling datasets based on indicators such as quality identification, offshore distance, brightness temperature anomaly, and sea and air parameter anomalies The quality of the modeling dataset is controlled through a variety of indicators, including but not limited to quality identification filtering (the quality identification can be derived from the inherent information of the satellite sensor), offshore distance filtering (ensuring that the study area is located in the far sea area to avoid interference from the complex offshore environment), brightness temperature anomaly detection (eliminating abnormally high or low brightness temperature values) and sea-air parameter anomaly detection (eliminating abnormal reanalysis data) to ensure the purity and reliability of the modeling dataset.

[0027] (4) Normalize the modeling dataset and split the dataset into training set, validation set, and test set The data is normalized using the maximum normalization method to eliminate the differences between different data scales. The normalized data set is split into training set, validation set, and test set according to a certain ratio to meet the needs of deep learning modeling. Usually, the training set accounts for 70% of the total data set, and the validation set and test set each account for 15%.

[0028] (5) Using the training set and validation set, a weak signal enhancement DNN model based on K-Ka band microwave radiometer was constructed. Based on deep learning technology, a weak signal enhancement DNN model is constructed through an encoder-decoder structure and a skip link network architecture. The goal of the model is to minimize the mean square error (MSE) loss function. In the process of model construction, the focus is on enhancing the ability to extract weak signals. Through the designed encoder-decoder structure and skip link network architecture, the sensitivity of K and Ka band brightness temperature data to sea temperature inversion is improved. The training set is used for model training, and the validation set is used to test and adjust the model parameters to prevent overfitting.

[0029] In the encoder-decoder structure, the encoding part gradually reduces the number of nodes and compresses the information into a more abstract feature representation; the subsequent decoding part gradually expands the number of nodes to reconstruct detailed information from the compressed feature representation. This structure is extremely effective in capturing and reconstructing the complex relationship between the model input and output, and also helps to reduce the risk of overfitting of the model.

[0030] Skip link network architecture,Considering the weak response between K- and Ka-band TB and SST, excessive losses during transmission,may hinder the model’s ability to fully utilize this response for inversion;,therefore, in order to improve the network’s response to subtle signal changes,,skip links are used to fuse the features extracted by the encoding part with the,features of the decoding part, effectively reducing the information loss during,transmission.

[0031] (6) The selected microwave radiometer test set is input into the weak signal enhancement DNN model based on the K-Ka band microwave radiometer to invert the sea surface temperature product The test set that was not involved in training and validation was input into the trained weak signal enhancement DNN model to invert and output the sea surface temperature product (SST value). By comparing with the known true value data, the performance and reliability of the model were evaluated to ensure that it had good generalization ability and prediction accuracy.

[0032] Example 1 The present invention is applicable to any microwave radiometer capable of observing brightness temperature in the K and Ka bands, wherein the AMSR-2 sensor is only used as an exemplary embodiment.

[0033] Taking the AMSR-2 sensor as an example, the specific implementation method is described, which specifically includes the following steps: S1. Based on the radiation transfer model The microwave simulation TB of different bands is obtained, where τ represents the atmospheric transmittance, e p represents the surface emissivity, T s represents the sea temperature, and Represents the upward and downward brightness temperature of the atmosphere. By adjusting the atmospheric transmittance τ to simulate the simulated TB under different water vapor conditions, and calculating its sensitivity to the four sea-air parameters of sea breeze, water vapor, cloud liquid water and sea temperature under different water vapor conditions, the generated Figure 2 The following points can be observed from the figure: First, the C-band (4~8 GHz) and X-band (8~12 GHz) are more sensitive to sea temperature.

[0034] The first point is also the reason why these two bands are widely used for sea temperature inversion. However, taking AMSR-2 as an example, to achieve a spatial resolution comparable to that of the K-band and Ka-band, the antenna size and weight of the C-band and X-band need to be three times that of the K-band and Ka-band, which greatly increases the cost of satellite launch and puts technical pressure on satellite design.

[0035] Second, the K-band (18~26.5 GHz) and Ka-band (26.5~40 GHz) are more sensitive to water vapor, sea breeze and cloud liquid water, but less sensitive to sea temperature.

[0036] Therefore, in the past, microwave radiometers using these two bands were usually designed to invert water vapor, sea breeze, and cloud liquid water. However, it can be seen from the figure that changes in water vapor conditions can enhance the sensitivity of K and Ka band brightness temperatures to SST. In particular, the sensitivity of the K band to SST is enhanced under the influence of water vapor (see Figure 2 The Ka band has a strong negative correlation with SST when there is no water vapor influence (see Figure 2 This provides important theoretical support for the use of K and Ka bands for SST inversion.

[0037] Example 2 Parameters used for SST inversion are selected based on domain knowledge of SST inversion.

[0038] The microwave radiometer used is AMSR-2, and its latitude, longitude and annual accumulated day and other temporal and spatial information as well as K and Ka band brightness temperature (6 channels of 18.7 GHz, 23.8 GHz and 37 GHz V and H), azimuth and incidence angle are selected as input; the reanalysis data uses ERA5 data, and its sea breeze, water vapor and cloud liquid water are selected as input.

[0039] Specifically, selecting longitude and latitude as input parameters not only provides geographical location information for each observation point, but also helps to reflect the distribution characteristics of water vapor changes, because the change of water vapor is not only closely related to longitude and latitude, but also directly affects the sensitivity of K and Ka band brightness temperature to SST. Annual accumulation days are used to characterize seasonal changes. K and Ka band brightness temperatures are the core observation data of microwave radiometers and have a complex nonlinear relationship with sea temperature. Azimuth and angle of incidence can correct the effect of observation angle on brightness temperature values. Sea breeze affects sea surface roughness and foam formation, thereby affecting brightness temperature. Water vapor and cloud liquid water affect observation results by attenuation and scattering of microwave signals, respectively. Therefore, introducing these parameters into the model can help improve the accuracy of sea temperature inversion and lay the foundation for subsequent modeling steps.

[0040] Determine the spatiotemporal window size and perform spatiotemporal matching to form a modeling dataset.

[0041] For AMSR-2 and ERA5, the scanning interval of AMSR-2's K and Ka band brightness temperatures is 25 km (which can be understood as the spatial resolution of the data), and the spatial resolution of ERA5 is also 25 km*25 km, so the spatial window is determined to be 50 km; and the time resolution of ERA5 is 1 hour, and the K and Ka band brightness temperatures are scanned instantaneously along the track, so the time window is determined to be 1 hour. The time-space matching method used is to match point by point based on the time-space information of the AMSR-2 radiometer observation data. With the observation point as the center, the time dimension is first matched to calculate the time difference Δt=|t between the radiometer and ERA5 data. AMSR-2 -t ERA5 |, if Δt < 1 hour, the matching point is retained. Then the spatial dimension is matched, and the distance between the data points is calculated based on the latitude and longitude information of the radiometer and ERA5 data d = R*arccos[sin(lat AMSR-2 )*sin(lat ERA5 )+cos(lat AMSR-2 )*cos(lat ERA5 )*cos(lat AMSR-2 -lat ERA5)], if d < 50 km, the matching point is retained, where R is the average radius of the earth, 6371 km. For the points that meet the above conditions, a modeling data set is formed, laying the foundation for subsequent quality control and modeling.

[0042] Perform quality control on the AMSR-2 modeling dataset.

[0043] First, quality identification filtering was performed. The TB data was evaluated according to the quality identification of the microwave radiometer data, and low-quality TB data was eliminated. Then, offshore distance filtering was performed. For the remote sensing data in the nearshore area, data points within 100 kilometers of the coast were eliminated to reduce the interference of land on the TB data. Then, brightness temperature anomaly detection was performed. For the 18-37 GHz TB data, the difference between the vertical (V) and horizontal (H) polarizations was negative, indicating that the oceanographic retrieval may be invalid, so such data was eliminated; for the 18 GHz V-polarized TB data, data points exceeding 240 K were excluded because this usually means the possibility of rainfall interference. Finally, sea and air parameter anomaly detection was performed. In order to eliminate the impact of sea ice on data quality, observation points with reanalysis sea ice values ​​not equal to 0 were eliminated; because high wind speeds will increase the roughness of the sea surface and may cause foam on the sea surface, thereby interfering with the TB data, data with reanalysis wind speed values ​​exceeding 20 m / s were eliminated. In the processing of SST data, only data with a temperature range of -2 to 34 degrees Celsius were retained, and outliers outside this range were eliminated.

[0044] The quality-controlled modeling dataset is normalized and split into training, validation, and test sets.

[0045] First, the data is normalized using the maximum value normalization method. The normalized value can be expressed as: X norm =(XX min ) / (X max -X min ). This ensures that the data features are on the same scale and reduces the impact of extreme values ​​in the data. After normalization, the data set is split into training set, validation set, and test set according to a certain ratio. In order to ensure the generalization of the results and avoid the risk of data leakage, the training set and validation set use data from 2020 to 2021, and the test set uses data from 2017, 2018, 2019, and 2022.

[0046] The split training set and validation set are used to build a weak signal enhancement DNN model based on the MSE loss function.

[0047] First, the basic network structure was determined, and the best loss values ​​of different models with different numbers of layers (1, 2, 3, 4, 5, 6, and 7) and different numbers of nodes in each hidden layer (8, 16, 32, 64, 128, 256, 512, 1024, and 2048) were trained and obtained, which are recorded in Table 1. As can be seen from the table, with the increase of the number of layers and nodes, the loss value shows a downward trend. This shows that the complexity and accuracy of the model increase with the increase of the number of layers and nodes. When the number of layers is 7 and the number of nodes in each layer is 1024, the loss value reaches the lowest point, which is about 1.18. Therefore, such a model structure was finally selected as the basic model.

[0048] Table 1 The impact of different neural network layers and nodes on SST inversion results Number of nodes and layers 8 16 32 64 128 256 512 1024 2048 1 7.29 5.58 5.44 5.42 5.40 5.39 5.39 5.39 5.34 2 3.50 2.92 3.11 2.79 2.67 2.65 2.61 2.60 2.62 3 2.52 2.39 2.02 1.92 1.88 1.85 1.85 1.85 1.85 4 2.22 2.02 1.80 1.61 1.53 1.51 1.51 1.52 1.51 5 2.17 1.73 1.51 1.43 1.37 1.35 1.35 1.33 1.34 6 1.92 1.62 1.38 1.33 1.29 1.27 1.25 1.24 1.24 7 2.04 1.58 1.36 1.24 1.22 1.21 1.19 1.18 1.19 In order to enhance the weak signal between K and Ka band TB and SST, the present invention also introduces two network structure optimizations, the encoder-decoder structure and the skip link network architecture, into the basic model (see Figure 3 ). Among them, the encoder-decoder compresses the information into a more abstract feature representation and then reconstructs the detailed information. This structure is extremely effective in capturing and reconstructing the complex relationship between the model input and output, and also helps to reduce the risk of overfitting of the model; and the skip link network architecture fuses the features extracted from different layers, effectively enhancing the weak signal between the K and Ka band TB and sea temperature. In order to evaluate the impact of this network structure optimization on the sea temperature inversion results, the training results of the model introducing the encoder-decoder structure and the skip link network architecture are compared with the basic optimal model, which is recorded in Table 2, where RMSE is the root mean square error, Bias is the bias, and STD is the standard deviation. It can be seen from the table that the introduction of these structures has a positive effect on the final result, and the accuracy is improved by 13.9% compared with the basic optimal model.

[0049] Table 2 The impact of introducing encoder-decoder structure and skip link network architecture on SST inversion results Model RMSE / K Bias / K STD / K Basic Optimal Model 0.79 -0.01 0.79 Introducing the encoding-decoding and skip-link model 0.68 -0.05 0.68 The split test data of 2017, 2018, 2019 and 2022 are input into the trained model, and the SST values ​​are inverted and compared with the known true value data. The statistical results are given and recorded in Table 3. It can be seen from the table that the accuracy of the SST products produced by the method of the present invention based on the K and Ka band brightness temperatures of AMSR-2 is highly consistent in each year, highlighting its reliability and accuracy in long time series.

[0050] Table 3 Analysis of SST inversion results in different years years RMSE / K Bias / K STD / K 2017 0.66 -0.07 0.65 2018 0.70 -0.02 0.70 2019 0.71 -0.08 0.71 2022 0.68 -0.05 0.68 In addition, the test data of 2022 are compared with the effects of the other three methods (extreme gradient boosting, random forest and linear fitting method), and the statistical results are given in a tabular form and recorded in Table 4. It can be seen from the table that the method of the present invention is significantly better than the other methods, and the accuracy improvement compared with the other three methods is 22.73%, 32.00% and 75.09% respectively.

[0051] Table 4 Comparative analysis of SST inversion results using different methods Method Type RMSE / K Bias / K STD / K Method of the present invention 0.68 -0.05 0.68 Extreme Gradient Boosting 0.88 -0.02 0.88 Random Forest 1.00 -0.03 1.00 Linear Fitting 2.73 0.02 2.73 The method of the present invention is not limited to the AMSR-2 sensor, and any microwave radiometer with K and Ka bands (18-26.5 GHz and 26.5-40 GHz) is also applicable to the method of the present invention. For example, other microwave radiometers such as SSMIS and GMI can also perform sea temperature inversion using the method of the present invention.

[0052] The above description is only a preferred embodiment of the present invention, and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art can still modify the technical solutions described in the above embodiments, or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, modification, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A sea temperature inversion method based on K-Ka band microwave radiometer weak signal enhancement deep learning, characterized in that: The following steps are involved: (1) Select the parameters of microwave radiometer measured data and reanalysis of sea-air parameter data. The parameters of microwave radiometer measured data include brightness temperature, azimuth and incident angle in K and Ka bands, as well as the corresponding spatiotemporal information; the parameters of reanalysis of sea-air parameter data include sea breeze, water vapor and cloud liquid water; (2) Based on the brightness temperature characteristics of the K and Ka bands, a time and space window is selected to match the microwave radiometer measured data and the reanalysis of sea and air parameter data in time and space dimensions to form a modeling data set; (3) Quality control of the modeling datasets using a variety of indicators including quality identification, offshore distance, brightness temperature anomalies, and sea-air parameter anomalies; (4) Normalize the modeling data set and split the normalized data set into training set, validation set and test set according to a certain ratio; (5) Using the training set and validation set, a weak signal enhancement DNN model based on K-Ka band microwave radiometer was constructed; (6) The test set is input into the trained weak signal enhancement DNN model to invert and output the sea surface temperature product.

2. The sea temperature inversion method based on weak signal enhancement deep learning of K-Ka band microwave radiometer according to claim 1 is characterized in that: In the step (1), latitude, longitude, annual accumulated day, K and Ka band brightness temperature, azimuth and incident angle are selected from the microwave radiometer measured data as model inputs, wherein the K band includes the frequency band range of 18~26.5 GHz, and the Ka band includes the frequency band range of 26.5~40 GHz; sea breeze, water vapor and cloud liquid water are selected from the reanalysis of sea and air parameter data as model inputs.

3. The sea temperature inversion method based on weak signal enhancement deep learning of K-Ka band microwave radiometer according to claim 1 is characterized in that: In the step (2), the appropriate space-time window size is determined according to the space-time resolution of the selected microwave radiometer K and Ka band measured data and the reanalysis sea-air parameter data, and the microwave radiometer measured data and the reanalysis sea-air parameter data are space-time matched according to the selected space-time window to form a modeling data set.

4. The sea temperature inversion method based on weak signal enhancement deep learning of K-Ka band microwave radiometer according to claim 1 is characterized in that: In step (3), the modeling data set is quality controlled by a variety of indicators, including quality identification filtering; offshore distance filtering to ensure that the study area is located in the offshore area; Brightness temperature anomaly detection, used to remove abnormally high or low brightness temperature values; and anomaly detection of air and sea parameters to remove abnormal reanalysis data.

5. The sea temperature inversion method based on weak signal enhancement deep learning of K-Ka band microwave radiometer according to claim 1 is characterized in that: In step (4), the split training set accounts for 70% of the total data set, the validation set accounts for 15% of the total data set, and the test set accounts for 15% of the total data set.

6. The sea temperature inversion method based on weak signal enhancement deep learning of K-Ka band microwave radiometer according to claim 1 is characterized in that: In the step (5), a weak signal enhancement DNN model is constructed by using an encoder-decoder structure and a skip link network architecture, the model is trained using a training set, and the model parameters are tested and adjusted using a validation set to prevent overfitting; The encoder-decoder structure reduces the number of nodes layer by layer in the encoding part to compress the information into an abstract feature representation, while the decoding part expands the number of nodes layer by layer to reconstruct detailed information from the feature representation; The skip link network architecture fuses the features extracted by the encoding part with the features of the decoding part, effectively reducing the information loss during the transmission process and improving the network's response to subtle signal changes between the K and Ka band brightness temperatures and the sea surface temperature.

7. The sea temperature inversion method based on weak signal enhancement deep learning of K-Ka band microwave radiometer according to claim 1 is characterized in that: In the step (6), the test set that has not participated in training and verification is input into the trained weak signal enhancement DNN model, and the sea surface temperature product is inverted and output; and the performance and reliability of the model are evaluated by comparing with the known true value data.

Images