Atmospheric temperature and humidity profile inversion method and device based on satellite hyperspectral data

Abstract

The invention provides an atmospheric temperature and humidity profile inversion method and device based on satellite hyperspectral data, and the method comprises the steps: constructing a background temperature and humidity profile through employing historical forecast data of a numerical mode, and calculating a background error between the background temperature and humidity profile and live observation; substituting the background profile into a radiation transmission model to simulate a satellite radiation value, and obtaining a radiation simulation error according to satellite hyperspectral observation; calculating a nonlinear distance correlation coefficient between the radiation simulation error of each channel and the background temperature and humidity error of each barometric layer, and taking the first M channels with the highest correlation coefficient as inversion preselected channels of the layer; the background error and related data are divided into first and second data sets. Establishing a relation from a radiation error to a background temperature and humidity error by using the first data set and a radial basis function neural network model with 1-M channels for each air pressure layer; and evaluating an inversion error under each channel combination by using the second data set, and selecting an optimal channel combination by taking the minimum error as a standard so as to obtain a final optimized neural network inversion model.

Description

Technical Field

[0001] This invention relates to the field of meteorological forecasting technology, and in particular to a method and apparatus for inverting atmospheric temperature and humidity profiles based on satellite hyperspectral data. Background Technology

[0002] Atmospheric temperature and humidity vertical profiles are key parameters for climate monitoring, weather analysis, and numerical weather prediction. Accurate acquisition of these profiles is crucial for improving forecast accuracy and understanding the mechanisms of weather system evolution. Traditionally, atmospheric temperature and humidity profiles have relied primarily on radiosonde observations. However, radiosonde stations are sparsely distributed and have uneven spatial coverage, with data being particularly scarce in ocean, plateau, and remote areas, making it difficult to meet the demands of modern high-precision, high spatiotemporal resolution meteorological operations. With the development of satellite remote sensing technology, especially the advent of hyperspectral infrared remote sensors, a revolutionary means has been provided for acquiring high-precision, continuous atmospheric temperature and humidity profiles globally.

[0003] The hyperspectral infrared vertical sounder can acquire infrared radiation information of the Earth-atmosphere system with extremely high spectral resolution. Its number of channels can reach thousands, far exceeding that of traditional sounders, enabling it to capture more refined atmospheric absorption and emission characteristics and significantly improving the inversion capability of temperature and humidity profiles. The spaceborne hyperspectral infrared vertical sounder receives infrared radiation signals from the Earth's atmospheric system, which need to be converted into temperature and humidity profiles through inversion algorithms. Early inversion methods mainly included statistical regression and physical inversion methods. Statistical regression methods establish statistical relationships between radiation and atmospheric parameters based on historical observation data, such as the eigenvector method and ridge regression solutions. These methods are fast and stable, but they ignore the physical processes of radiative transfer, resulting in limited inversion accuracy. Physical inversion methods seek optimal solutions by solving the radiative transfer equation and constructing a value function, such as one-dimensional variational methods. These methods offer high inversion accuracy, but are computationally complex and time-consuming, placing a significant burden on real-time operations.

[0004] In recent years, with the rapid development of artificial intelligence technology, machine learning algorithms have been gradually introduced into the field of atmospheric science. Backpropagation (BP) neural networks employ an error backpropagation mechanism, using sigmoid or tanh activation functions to achieve nonlinear mapping. They have the advantages of simple structure and ease of implementation, but training speed is slow, they are prone to getting trapped in local optima, and they are sensitive to the initial weights of the network. Convolutional neural networks use multi-layer convolution and pooling structures to extract spectral features and use convolutional kernels to process channel sequences. They have the advantages of strong automatic feature extraction capabilities and accurate capture of local spectral features, but the network structure is complex, parameter tuning is difficult, and training computational resources are large. Multi-objective genetic algorithms can construct dual objective functions to optimize long-wave and mid-wave channels separately and introduce physical constraints to invert temperature and humidity profiles. They have the advantages of clear physical meaning and not relying on a large amount of historical training data, but computational complexity is high and convergence speed is slow. Summary of the Invention

[0005] To address the technical problems existing in the prior art, this invention provides a method and apparatus for inverting atmospheric temperature and humidity profiles based on satellite hyperspectral data. Based on real-time information from satellite observations, the inversion model constructed using radial basis functions (RBF) can quickly invert multi-level atmospheric temperature and humidity information.

[0006] The difficulty of this invention is:

[0007] 1) Optimization of Hyperspectral Channel Selection. Hyperspectral infrared vertical sounders can have thousands of channels; for example, my country's FY4B-GIIRS has 1690 channels, and each channel exhibits varying sensitivities to temperature and water vapor at different pressure levels. Atmospheric radiative transfer is a nonlinear process. Therefore, establishing effective channel selection criteria and determining the optimal channel combination based on the inversion requirements of temperature and humidity profiles is the primary challenge of this method. Too many channels introduce redundant information and noise, significantly increasing computational complexity and hindering operational applications; while too few channels may result in the loss of crucial information, leading to decreased inversion accuracy. Therefore, it is essential to seek the optimal balance between inversion accuracy and computational efficiency, selecting the channel combination with the maximum information content and minimum redundancy.

[0008] 2) Determining the Basis Functions and Model Structure of the RBF Network. The choice of basis functions directly determines the network's approximation ability, generalization performance, and computational efficiency. For example, the Gaussian function has good local response characteristics and is suitable for capturing local patterns of atmospheric states in the feature space; while multiple quadratic functions, etc., have global characteristics, and changes in any data point may affect the entire model output. Therefore, choosing which basis functions to use and, based on this, determining key hyperparameters such as the number of centers through extensive experiments, is the core design challenge in building a high-performance inversion model.

[0009] 3) The physical design of input variables is a significant challenge for this invention. Atmospheric radiative transfer processes exhibit strong nonlinearity under various weather conditions, which greatly increases the difficulty of purely data-driven learning algorithms, easily leading to insufficient generalization ability and large errors in practical applications. Therefore, how to overcome the limitations of traditional "end-to-end" mapping and integrate prior physical knowledge of atmospheric radiative transfer (such as introducing physical quantities like model background field and brightness temperature deviation) into the design of input variables to construct input features with stronger constraints and clearer physical meaning, thereby reducing the nonlinear complexity of the learning problem, is a key challenge in improving the accuracy, robustness, and physical consistency of neural network inversion.

[0010] This invention provides a method for inverting atmospheric temperature and humidity profiles based on satellite hyperspectral data, comprising:

[0011] The temperature and humidity profiles from historical numerical model forecasts are used as background profiles, and the error between the background profiles and the observed profiles is calculated as the background error.

[0012] Substituting the background profile into the radiative transfer mode, we obtain the simulated radiative values ​​for all hyperspectral channels of the satellite. The error between the satellite-observed radiative values ​​and the simulated radiative values ​​is then calculated as the radiative simulation error.

[0013] Calculate the nonlinear distance correlation coefficient between the radiation simulation error of each hyperspectral channel of the satellite and the background error of each pressure layer. For each pressure layer, select M channels as pre-selected channels for temperature and humidity profile inversion according to the correlation coefficient from largest to smallest.

[0014] A portion of the background error and the corresponding radiation simulation error are selected to construct the first dataset, while another portion of the background error, the corresponding background profile, the actual observation profile, and the radiation simulation error are used to construct the second dataset.

[0015] A radial basis function neural network model is constructed using the first dataset and M pre-selected hyperspectral channels, and the radial basis function neural network model is optimized using the second dataset.

[0016] The optimized radial basis function neural network model is used to invert the ambient temperature and humidity profiles to obtain the final temperature and humidity profiles.

[0017] A method for inverting atmospheric temperature and humidity profiles based on satellite hyperspectral data according to the present invention includes:

[0018] Calculate the nonlinear distance correlation coefficient between the background error at each pressure layer in the dataset and the radiation simulation error of each hyperspectral channel of the satellite;

[0019] For the background error at each pressure layer, the satellite channels are sorted from largest to smallest according to the nonlinear distance correlation coefficient, and the top M channels are selected as pre-selected inversion channels.

[0020] Based on the background error and radiation simulation error of the pre-selected channel in the first dataset, a radial basis function neural network model is constructed.

[0021] According to the atmospheric temperature and humidity profile inversion method based on satellite hyperspectral data provided by the present invention, the nonlinear distance correlation coefficient between the background error at each pressure layer in the first dataset and the radiation simulation error of all hyperspectral channels of the satellite is calculated by the following formula:

[0022]

[0023] in, Construct a sequence for background error With radiation simulation error sequence The nonlinear distance correlation coefficient between them; , and represents the row average, column average, and overall average, respectively, and the same representation applies to b; A and B represent the bicentric distance matrix.

[0024] According to the present invention, an atmospheric temperature and humidity profile inversion method based on satellite hyperspectral data is provided, which constructs a radial basis function neural network model based on the background error in the first dataset and the selected radiation simulation error using the following formula:

[0025]

[0026]

[0027] in, Let k1 represent the vector formed by the radiation simulation errors of m pre-selected channels, 1≤m≤M; k1 and k2 are vector identifiers, and ||.|| represents the Euclidean distance between the two vectors; These are the Euler radial basis function values, and D is the radial basis function matrix. This is the first dataset. It is a vector composed of the background error of the j-th pressure layer in the first dataset. These are the coefficients of a radial basis function neural network model using m hyperspectral channels at the j-th pressure layer.

[0028] According to the present invention, a method for inverting atmospheric temperature and humidity profiles based on satellite hyperspectral data is provided, which optimizes the radial basis function neural network model using a second dataset, including:

[0029] Based on the second dataset, the temperature and humidity profiles of each pressure layer are inverted using radial basis function neural network models with different numbers of pre-selected channels, and the root mean square error of the inversion error is calculated.

[0030] For each pressure layer, the satellite channel combination corresponding to the minimum value of the root mean square error is selected as the optimal satellite channel combination for temperature (humidity) retrieval of that pressure layer.

[0031] According to the present invention, an atmospheric temperature and humidity profile inversion method based on satellite hyperspectral data is provided. The method utilizes a second dataset and different numbers of pre-selected inversion channels to construct a radial basis function neural network model to invert the temperature and humidity profiles of each pressure layer using the following formula:

[0032]

[0033] in, This is the first dataset. This is the second dataset, and the H matrix is... Radiation simulation error of the k-th profile and Medium radiation simulation error radial basis functions between The row vector formed, K s yes The number of midline contours The coefficients of the radial basis function neural network model are obtained by using the radiation simulation errors of the first m pre-selected channels as the basis correspondences and the background error of the j-th pressure layer. yes The k-th background profile is located at the temperature or humidity of the j-th pressure layer. It is the value after inversion.

[0034] The present invention also provides a device for inverting atmospheric temperature and humidity profiles based on satellite hyperspectral data, comprising:

[0035] The background error calculation module is used to use the historical temperature and humidity profiles of numerical models as background profiles, and calculate the error between the background temperature and humidity profiles and the actual observed profiles as background error.

[0036] The radiation simulation error calculation module is used to substitute the background profile into the radiation transfer mode to obtain the simulated radiation values ​​of all hyperspectral channels of the satellite, and calculate the error between the corresponding satellite-observed radiation value and the simulated radiation value to obtain the radiation simulation error.

[0037] The hyperspectral channel pre-selection module is used to calculate the nonlinear distance correlation coefficient between the temperature and humidity background error on each pressure layer and the radiation simulation error of each hyperspectral channel, and sort the hyperspectral channels from largest to smallest according to the correlation coefficient, and select the first M channels as pre-selected channels.

[0038] The sample dataset partitioning module is used to select a portion of the background error and the corresponding radiation simulation error to construct the first dataset, and another portion of the background error and the corresponding background profile, actual observation profile and radiation simulation error to construct the second dataset;

[0039] The construction and optimization module is used to construct a radial basis function neural network model using the first dataset and optimize the radial basis function neural network model using the second dataset.

[0040] The inversion module is used to invert the ambient temperature and humidity profiles using an optimized radial basis function neural network model to obtain the final temperature and humidity profiles.

[0041] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the atmospheric temperature and humidity profile inversion method based on satellite hyperspectral data as described above.

[0042] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the atmospheric temperature and humidity profile inversion method based on satellite hyperspectral data as described above.

[0043] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method for inverting atmospheric temperature and humidity profiles based on satellite hyperspectral data.

[0044] The present invention provides a method and apparatus for inverting atmospheric temperature and humidity profiles based on satellite hyperspectral data. The method uses radial basis function (RBF) to invert atmospheric temperature and humidity profiles based on satellite hyperspectral data. According to real-time information from satellite observations, the method rapidly inverts multi-level atmospheric temperature and humidity information based on the constructed inversion model.

[0045] The significance of this invention lies in:

[0046] 1) This invention has a significant advantage in inversion efficiency. Traditional variational assimilation methods require solving complex radiative transfer equations multiple times through iterative algorithms, resulting in a huge computational burden that is difficult to meet the demands of high-time-sensitivity applications. In contrast, this invention, based on a pre-trained neural network model, requires only one forward propagation in the inversion process, achieving a much faster inversion speed than variational methods, thus providing key technical support for real-time weather analysis and forecasting.

[0047] 2) This invention offers greater flexibility and specificity in its inversion strategy. Variational methods seek the globally optimal estimate of the overall atmospheric profile, making optimization for specific needs difficult. In contrast, this invention allows for optimization of inversion accuracy across all meteorological height layers during the training phase. Furthermore, through systematic hierarchical statistical verification of the inversion results, it is possible to clearly identify the pressure layers where this invention significantly improves upon the numerical model background field. This allows users to more confidently adopt higher-quality inversion information at these levels, enhancing the product's practical value.

[0048] 3) As the accuracy of numerical weather prediction continues to improve, higher requirements are placed on satellite remote sensing inversion products: the inversion error must be systematically better than the background field of the numerical model in order to produce a positive effect during assimilation. Existing purely data-driven neural network methods are constrained by the climatological statistics of the training set and fail to fully utilize real-time, high-precision model forecast information.

[0049] The innovation of this invention lies in proposing a hybrid inversion framework based on an RBF neural network. The core of this framework consists of two unique designs: first, it uses the temperature and humidity background profile errors provided by the numerical model, along with their simulated radiation errors, as network inputs; second, it explicitly establishes a statistical relationship model between background field errors and simulated radiation errors within the network structure. This design allows the network to not only learn the mapping of observed profiles but also, more importantly, learn how to use brightness-temperature bias to correct background field biases. This is equivalent to embedding the idea of ​​variational assimilation into the neural network, potentially enabling the final inversion profile to be statistically and consistently superior to the initial background field, significantly improving the value and reliability of the inversion product for model assimilation. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0051] Figure 1 This is a flowchart illustrating the atmospheric temperature and humidity profile inversion method based on satellite hyperspectral data provided by the present invention.

[0052] Figure 2 This is a schematic diagram of the distribution of the number of GIIRS channels used to retrieve temperature and water vapor mixing ratio at different pressure layers in the atmospheric temperature and humidity profile inversion method based on satellite hyperspectral data provided by this invention.

[0053] Figure 3 The present invention provides a method for inverting atmospheric temperature and humidity profiles based on satellite hyperspectral data, which includes (a) the hyperspectral channel distribution used for inverting temperature at various pressures and (b) the hyperspectral channel distribution used for inverting humidity at various pressures.

[0054] Figure 4 The present invention provides a method for inverting atmospheric temperature and humidity profiles based on satellite hyperspectral data, which compares the root mean square error (RMSE) of (a) the background profile and the temperature profile inverted by the RBF neural network, and (b) the ratio of the root mean square error (RMSE) of the temperature profile inverted by the RBF neural network to the root mean square error (RMSE) of the background profile.

[0055] Figure 5The comparison of the root mean square error (RMSE) of the water vapor back mixing ratio background profile and the water vapor profile inverted by the RBF neural network in the atmospheric temperature and humidity profile inversion method based on satellite hyperspectral data provided by this invention, and the ratio of the root mean square error (RMSE) of the water vapor profile inverted by the RBF neural network to the root mean square error (RMSE) of the background profile in (b) part.

[0056] Figure 6 This is a schematic diagram of the atmospheric temperature and humidity profile inversion device based on satellite hyperspectral data provided by the present invention. Detailed Implementation

[0057] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0058] The following is combined with Figure 1 This invention describes a method for inverting atmospheric temperature and humidity profiles based on satellite hyperspectral data, comprising:

[0059] Step 101: Use the historical temperature and humidity profiles of the numerical model as background profiles, and calculate the error between the background profile and the observed profile as background error.

[0060] Step 102: Substitute the background profile into the radiative transfer mode to obtain the simulated radiative values ​​of all hyperspectral channels of the satellite, and calculate the error between the hyperspectral radiative values ​​observed by the satellite and the simulated radiative values ​​as the radiative simulation error;

[0061] Step 103: Calculate the nonlinear distance correlation coefficient between the temperature (humidity) background error of each pressure layer and the radiation simulation error of each hyperspectral channel of the satellite, and select the top M channels from largest to smallest correlation coefficient as the inversion pre-selected channels for the temperature (humidity) of that pressure layer.

[0062] Step 104: Construct a first dataset based on a portion of the background error and the corresponding radiation simulation error, and construct a second dataset using another portion of the background error and the corresponding background profile, actual observation profile, and radiation simulation error.

[0063] Step 105: Using the first dataset, for each pressure layer, construct a radial basis function neural network model between the radiation simulation error and the background temperature and humidity profile error of each inversion preselected channel combination using the corresponding M inversion preselected channel combinations.

[0064] Step 106: Use the second dataset to statistically analyze the temperature and humidity inversion errors of M different inversion pre-selected channel combinations on each pressure layer, and select the inversion pre-selected channel combination with the smallest inversion error as the standard to select the best channel combination.

[0065] Step 107: Use the radial basis function neural network model corresponding to the selected optimal channel combination to perform atmospheric temperature and humidity profile inversion to obtain the final temperature and humidity profiles.

[0066] The relationship between satellite-observed radiation values ​​and the vertical profiles of physical variables in the Earth's atmosphere is defined as follows:

[0067]

[0068] in, This represents the outward radiation value of the atmospheric column at a specific latitude and longitude location on Earth, observed by a total of M0 hyperspectral channels of the satellite, and is the atmospheric temperature. and humidity Functions of physical parameters, which will be used uniformly here. This indicates that satellite data inversion is an inverse problem, namely, how to solve for atmospheric profiles using radiance values ​​observed by satellites. The values ​​at different pressure layers refer to the inverted temperature and humidity in this invention.

[0069] Among numerous machine learning methods, radial basis function (RBF) neural networks, with their unique advantages, achieve an excellent balance between accuracy and efficiency, becoming an important technical approach in hyperspectral temperature and humidity profile inversion. Employing radial basis functions as hidden layer activation functions, they possess the local approximation characteristic of single hidden layers, demonstrating unique value in temperature and humidity profile inversion and showcasing a balanced advantage in accuracy and efficiency. This method has a solid mathematical foundation, strong generalization ability, and excellent adaptability to unseen data. Its network structure effectively avoids the problem of BP networks easily getting trapped in local minima. Some studies have shown that compared to BP neural networks, training time is reduced to 1 / 4, significantly improving computational efficiency.

[0070] With the continuous development of numerical models, the accuracy of weather forecasts has reached a high level. This requires that the errors of satellite remote sensing inversion products be systematically superior to the background field of the numerical model in order to produce positive effects in weather analysis and data assimilation. Therefore, this embodiment introduces the temperature and humidity profiles predicted by the numerical model as the background profile for inversion. A mathematical network model of the radiation error simulated by the atmospheric radiative transfer model R(X) and the background profile error is established using radial basis functions (RBF) to achieve the inversion of temperature and humidity profiles.

[0071] First, prepare based on historical data from numerical model forecasts. The dataset consists of forecast profiles (such as 6-hour temperature forecast profiles) and corresponding observed profiles for the same time period, as well as radiation information from satellite observations (such as brightness temperature). The dataset is defined as follows:

[0072]

[0073] in, This represents the k-th real-world profile observed by the satellite. The corresponding radiation value vector is given by Composed of observed radiation values ​​from each channel; This is the background profile predicted by the corresponding numerical model. Both the actual and background profiles have observed and predicted values ​​at the same L pressure layers. This represents the background error of the k-th background profile at the j-th pressure layer.

[0074] The radiation simulation error of the background profile is calculated using the radiative transfer model. The background profile predicted by the numerical model is then compared. Substituting into the radiative transfer mode (such as the RTTOV mode) yields The simulated radiation values ​​of each satellite channel are used to obtain the simulated radiation error. = -R( Two datasets were constructed using background error and radiation simulation error. Among them, E s Used to construct a basis vector model, E t The following definitions are used to optimize the model:

[0075]

[0076] in: and These are the sample sizes of Es and Et, respectively. .

[0077] Radial basis function (RBF) is constructed using radiation simulation errors, and the coefficients of the inversion model are solved.

[0078] This embodiment introduces the temperature and humidity profiles predicted by numerical models as the background profiles for inversion, and uses radial basis functions to establish a mathematical network model between radiation simulation errors and background profile errors, thereby achieving rapid and accurate inversion of temperature and humidity profiles.

[0079] Based on the above embodiments, this embodiment constructs a radial basis function neural network model, including:

[0080] Calculate the nonlinear distance correlation coefficient between the background error at each pressure layer and the radiation simulation error of each satellite hyperspectral channel in the union of the first and second datasets;

[0081] For each pressure layer, the satellite channels are sorted from largest to smallest according to the nonlinear distance correlation coefficient, and the top M channels are selected as inversion pre-selected channels, with their simulation error serving as the basis vector of the inversion model;

[0082] A radial basis function neural network model is constructed based on the background error in the first dataset and the radiation simulation error of the first m channels of the pre-selected channels, where 1≤m≤M.

[0083] Based on the above embodiments, this embodiment calculates the nonlinear distance correlation coefficient between the background error at each pressure layer in the first dataset and the radiation simulation error of the satellite channel using the following formula:

[0084] (4)

[0085] in, Construct a sequence for background error With radiation simulation error sequence The nonlinear distance correlation coefficient between them; , and represents the row average, column average, and overall average, respectively, and the same representation applies to b; A and B represent the bicentric distance matrix.

[0086] Based on the above embodiments, this embodiment constructs a radial basis function neural network model according to the background error in the first dataset and the selected radiation simulation error using the following formula:

[0087]

[0088] (6)

[0089] Where 1≤m≤M, This represents the vector formed by the radiation simulation errors of the first m channels of the pre-selected channels; k1 and k2 are vector identifiers, and ||.|| represents the Euclidean distance between the two vectors; It is the value of the Euler radial basis function, D is Symmetric radial basis function matrix, This is the first dataset. It is the first dataset middle The length of the background error of the strip profile at the j-th pressure layer is The vector, These are the coefficients of the radial basis function neural network model, where the subscript j indicates the location of the pressure layer, and m indicates the number of hyperspectral channels used.

[0090] Based on the above embodiments, this embodiment optimizes the radial basis function neural network model using a second dataset, including:

[0091] Based on the second dataset, a radial basis function neural network model was constructed using different numbers of satellite channels m to invert the temperature and humidity profiles of each pressure layer (1≤m≤M).

[0092] Calculate the root mean square error of the inversion error of the second dataset when using different numbers m of preselected satellite channels;

[0093] For each pressure layer, the number of satellite channels corresponding to the minimum value of the root mean square error is selected as the optimal number of satellite channels for temperature (humidity) profile inversion of that pressure layer.

[0094] Based on the above embodiments, this embodiment uses the following formula to invert the temperature and humidity profiles of each pressure layer using a radial basis function neural network model with different numbers of satellite channels based on the second dataset, as shown in formula (7):

[0095]

[0096] in, This is the first dataset. This is the second dataset, and the H matrix is... Radiation simulation error of the k-th profile and Medium radiation simulation error radial basis functions between The row vectors formed yes The number of samples in the middle These are the coefficients of the radial basis function neural network model using m channels corresponding to the j-th pressure layer. yes The temperature or humidity value of the k-th background profile in the j-th pressure layer. This is the corresponding inversion result.

[0097] For the inversion of variables at layer j, using different numbers of satellite channels m, we obtain the result using formula (7). For all samples, calculate the RMSE (root mean square error) of the inversion error, and select the number of satellite channels corresponding to the minimum RMSE as the optimal number of channels for the inversion quantifier at the j-th layer.

[0098] This experiment used radiation data observed under clear-sky conditions by the FY4B-GIIRS satellite to retrieve atmospheric temperature and humidity profiles. GIIRS has a total of M0=1690 hyperspectral channels, of which channels 1 to 723 are long-wave channels and channels 724-1690 are mid-wave channels. 960 temperature and humidity profiles under clear-sky conditions from August to December 2024 within a certain region were selected as the actual data for modeling and validation. The background profile used a 6-hour forecast field from a numerical model, covering 41 vertical pressure layers from 1000 hPa to 0.01 hPa, consistent with the actual profiles; the radiative transfer model used was RTTOV-4.0. In the hyperspectral channel selection optimization section of the experiment, the design limited the use of a maximum of M=300 channels with the largest distance correlation coefficients for optimization.

[0099] Figure 2 The figure shows the number of hyperspectral channels used at different pressure levels to retrieve atmospheric temperature and humidity (mixing ratio) profiles from FY4B-GIIRS radiometric data for the RBF neural network model. The x-axis represents the number of channels, and the y-axis represents the pressure level. Red bars represent temperature, and green bars represent water vapor mixing ratio. For temperature retrieval, 300 channels were used only when retrieving 0.03 hPa temperature, and 299 channels were used for 600 hPa temperature retrieval. The number of satellite channels used at other pressure levels was far less than 300, with the fewest using only 12 channels of radiometric values. For humidity retrieval, the figure shows the channel number distribution from 1000 hPa to 100 hPa, because the water vapor content is almost zero at higher levels, making retrieval less meaningful. Seven pressure levels used 300 channels, and the number of channels used decreased significantly above 300 hPa. The number of channels used below 850 hPa was also close to 300.

[0100] Figure 3 The distribution of hyperspectral channels used for temperature and water vapor retrieval at each pressure level is presented. For temperature retrieval, channels 1-250 and 1450-1690 are used most frequently. For channels in the intermediate range, channels above 1 hPa and between 850 hPa and 700 hPa are used more often, while channels at other pressure levels are used relatively less. For humidity retrieval, channels below 400 hPa are mainly distributed between long-wave (channels 100-400) and medium-wave (channels 1100-1500).

[0101] Figure 4 The given value is the root mean square error (RMSE) of temperature inversion for 480 test samples. Figure 4 (a) and its relative change with background error Figure 4(See section (b)). The figure shows that the background error provided by the numerical model increases from low to high levels, especially above 10 hPa. The RMSE of the mid-to-low-level temperature is around 0.2 K, indicating that the temperature background profile predicted by the numerical model has reached a high level. The inversion error shows that, except for 30 and 200 hPa, the temperature RMSE is better than the background field, with an average of 0.3946 K. The improvement is most significant at 300 and 1 hPa, which are 0.65 and 0.56 times the background error, respectively. The overall RMSE of the temperature profile is 87.5% of the background error, a reduction of 12.5%.

[0102] Figure 5 The given value is the RMSE (Reference Mean Squared Score) derived from the water vapor mixing ratio inversion. Figure 5 (a) and its relative change with background error Figure 5 (See section (b)). The figure shows that the error in water vapor mixing ratio was improved from 100 hPa to 100 hPa. Water vapor is mainly distributed in the lower to middle layers of the atmosphere, and there is very little water vapor above 400 hPa. The figure shows that the RMSE improvement is particularly significant below 850 hPa, approximately 85% of the model background error, which is very meaningful for weather analysis and data assimilation. At higher pressure levels, the RMSE improvement is not significant, which is related to the very sparse water vapor at higher altitudes.

[0103] The atmospheric temperature and humidity profile inversion device based on satellite hyperspectral data provided by the present invention is described below. The atmospheric temperature and humidity profile inversion device based on satellite hyperspectral data described below can be referred to in correspondence with the atmospheric temperature and humidity profile inversion method based on satellite hyperspectral data described above.

[0104] like Figure 6 As shown, the device includes a background error calculation module 601, a radiation simulation error calculation module 602, an inversion channel selection module 603, a dataset partitioning module 604, a construction module 605, an optimization module 606, and an inversion module 607, wherein:

[0105] Background error calculation module 601 is used to use the temperature and humidity profiles of the numerical model’s historical forecasts as background profiles, and to calculate the error between the background profiles and the actual observed profiles as background error.

[0106] The radiation simulation error calculation module 602 is used to substitute the background profile into the radiation transfer mode to obtain the simulated radiation values ​​of all hyperspectral channels of the satellite, and calculate the error between the corresponding satellite observed radiation value and the simulated radiation value as the radiation simulation error;

[0107] The inversion channel screening module 603 is used to calculate the non-distance correlation coefficient between the background error on each pressure layer and the radiation error of each channel, and to screen out the pre-selected satellite channels on each pressure layer for inversion according to the correlation coefficient from large to small.

[0108] The dataset partitioning module 604 is used to construct a first dataset based on a portion of background errors and corresponding radiation simulation errors, and to construct a second dataset based on another portion of background errors and corresponding background profiles, actual observation profiles, and radiation simulation errors.

[0109] Module 605 is used to construct a radial basis function neural network model using the first dataset;

[0110] Optimization module 606 optimizes the radial basis function neural network model using the second dataset;

[0111] The inversion module 607 is used to invert the ambient temperature and humidity profiles using the optimized radial basis function neural network model to obtain the final temperature and humidity profiles.

[0112] This embodiment introduces the temperature and humidity profiles predicted by numerical models as the background profiles for inversion, and uses radial basis functions to establish a mathematical network model between the radiation error simulated by the atmospheric radiation transfer model R(X) and the background profile error, thereby achieving rapid and accurate inversion of temperature and humidity profiles.

[0113] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for retrieving atmospheric temperature and humidity profiles based on satellite hyperspectral data, characterized in that, The application comprises the following steps: The temperature and humidity profiles predicted by the numerical model are taken as background profiles, and the errors between the background profiles and the real observation profiles are calculated as background errors; The background profiles are substituted into a radiation transfer model to obtain simulated radiation values, and the errors between the high-spectral radiation values observed by the satellite and the simulated radiation values are calculated as radiation simulation errors; Nonlinear distance correlation coefficients between the radiation simulation errors of each high-spectral channel of the satellite and the temperature and humidity background errors on each pressure layer are calculated, and the first m channels are selected as the inversion preselected channel combination of the temperature and humidity on the pressure layer according to the correlation coefficients from large to small, wherein m is an integer from 1 to M, and M is a positive integer greater than 1; A part of the background error data and the corresponding radiation simulation error data are selected to construct a first data set, and another part of the background error and the corresponding background profile, the real observation profile and the radiation simulation error are constructed into a second data set; The radial basis function neural network model between the radiation simulation error and the background temperature and humidity profile error of each inversion preselected channel combination is constructed by using the first data set and using the corresponding M kinds of inversion preselected channel combinations for each pressure layer; The temperature and humidity inversion errors of the M kinds of different inversion preselected channel combinations on each pressure layer are counted by using the second data set, and the inversion preselected channel combination with the minimum inversion error is taken as the standard to select the best channel combination; The radial basis function neural network model corresponding to the selected best channel combination is used for atmospheric temperature and humidity profile inversion to obtain the final temperature and humidity profiles.

2. The atmospheric temperature and humidity profile retrieval method based on satellite hyperspectral data according to claim 1, characterized in that, The nonlinear distance correlation coefficients between the radiation simulation errors of each high-spectral channel of the satellite and the temperature and humidity background errors on each pressure layer are calculated by the following formula: ； wherein constitutes a sequence of background errors and a sequence of radiation simulation errors a non-linear distance correlation coefficient between , and represent the average of the rows, columns and overall average, respectively, in , , and represent the average of the rows, columns and overall average, respectively, in ; A and B denote bicoherence distance matrices.

3. The method according to claim 1, wherein the satellite-based hyperspectral data is selected from the group consisting of: Atmospheric Infrared Sounder (AIRS) data, Cross-track Infrared Sounder (CrIS) data, Infrared Atmospheric Sounding Interferometer (IASI) data, and Thermal Infrared Sensor (TIS) data. The radial basis function neural network model between the radiation simulation error and the background temperature and humidity profile error of each inversion preselected channel combination is constructed by the following formula: ； ； wherein, denotes a vector composed of the radiation simulation errors of m channels in each inversion pre-selected channel combination, 1≤m≤M; k1 and k2 are vector identifiers, ||.|| denotes the Euclidean distance between two vectors; is the Euler radial basis function value, D is the radial basis function matrix, is the first data set, is a vector composed of the background error of the jth pressure layer in the first data set, denotes the coefficient of the radial basis function neural network model corresponding to the inversion pre-selected channel combination containing m channels on the jth pressure layer.

4. The method according to claim 1, wherein, The inversion error is the root mean square error.

5. The method according to claim 1, wherein, The temperature and humidity profiles of each pressure layer are inverted according to the second data set and the radial basis function neural network model under different inversion preselected channel combinations by the following formula: ； wherein, is a first data set, is a second data set, H matrix is radiation simulation error of the kth profile in and radiation simulation error in radial basis function between row vector composed of K s is number of profile samples in is the coefficient of the radial basis function neural network model corresponding to the inversion pre-selected channel combination on the jth pressure layer, is value of the kth background profile in the jth pressure layer, is the inversion result.

6. An atmospheric temperature and humidity profile retrieval device based on satellite hyperspectral data, characterized in that, The application comprises the following steps: The background error calculation module is used to take the temperature and humidity profiles predicted by the numerical model as background profiles, and calculate the errors between the background profiles and the real observation profiles as background errors; The spectral radiation simulation error calculation module is used to substitute the background profiles into a radiation transfer model to obtain simulated radiation values, and calculate the errors between the high-spectral radiation values observed by the satellite and the simulated radiation values as radiation simulation errors; The high-spectral channel preselection module is used to calculate the nonlinear distance correlation coefficients between the radiation simulation errors of each high-spectral channel of the satellite and the temperature and humidity background errors on each pressure layer, and select the first m channels as the inversion preselected channel combination of the temperature and humidity on the pressure layer according to the correlation coefficients from large to small, wherein m is an integer from 1 to M, and M is a positive integer greater than 1 The sample data set division module is used to select a part of the background error and the corresponding radiation simulation error to construct a first data set, and another part of the background error and the corresponding background profile, the real observation profile and the radiation simulation error are constructed into a second data set; A construction module is configured to use the first data set to construct, for each air pressure layer, a radial basis function neural network model between radiative simulation errors and background temperature and humidity profile errors for each inversion pre-selected channel combination; An optimization module is configured to use the second data set to statistically determine temperature and humidity inversion errors of M different inversion pre-selected channel combinations at each air pressure layer, and select the best channel combination as a standard using the inversion pre-selected channel combination with the minimum inversion error; An inversion module is configured to use the radial basis function neural network model corresponding to the selected best channel combination to perform atmospheric temperature and humidity profile inversion, and obtain final temperature and humidity profiles.

7. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor implements the atmospheric temperature and humidity profile inversion method based on satellite hyperspectral data according to any one of claims 1 to 5 when executing the program.

8. A non-transitory computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program implements the atmospheric temperature and humidity profile inversion method based on satellite hyperspectral data according to any one of claims 1 to 5 when executed by the processor.

Images