Training method and prediction method of atmospheric parameter inversion model based on satellite radiation brightness temperature
By using a two-stage trained inversion model, combined with the multi-channel brightness temperature and observation geometric parameters of geostationary satellites, the accuracy problem of atmospheric parameter inversion from geostationary satellites was solved, and more accurate predictions of atmospheric temperature and humidity profiles were achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CENT SOUTH UNIV
- Filing Date
- 2026-06-11
- Publication Date
- 2026-07-14
AI Technical Summary
Existing technologies make it difficult to accurately invert atmospheric temperature and humidity profiles using microwave radiometers on geostationary satellites, mainly because the observation geometry differs from that of polar-orbiting satellites, rendering existing algorithms inapplicable and resulting in inaccurate inversion results.
An atmospheric parameter inversion model training method based on satellite radiative brightness temperature is adopted. Through a two-stage training process, the correlation and complementarity of brightness temperatures in different frequency bands are learned by combining multi-channel brightness temperature, observation geometric parameters and pressure layer parameters. The model is then corrected by a radiative transfer model to construct an atmospheric parameter inversion model.
It improves the accuracy of atmospheric parameter inversion models, enabling more accurate prediction of atmospheric temperature and humidity profiles, satisfying the microwave radiation transmission mechanism, and reducing the error of inversion results.
Smart Images
Figure CN122389971A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of meteorological observation technology, specifically to a training and prediction method for an atmospheric parameter inversion model based on satellite radiative brightness temperature. Background Technology
[0002] Atmospheric temperature and humidity profiles are key fundamental variables in numerical weather prediction, severe weather monitoring, typhoon track analysis, severe convective weather warning, and atmospheric data assimilation. Microwave radiometers can receive atmospheric radiation brightness temperature at different frequency bands. Different frequency bands have different sensitivities to different atmospheric altitudes, different absorption components, and different physical processes, and therefore can be used to retrieve atmospheric temperature and humidity profiles.
[0003] Currently, mature atmospheric temperature and humidity inversion algorithms are mainly geared towards polar-orbiting satellite observation data. Polar-orbiting satellites typically move along their orbits and scan across them, with changes in their observation angles primarily unfolding along the scan lines. Inversion algorithms often empirically correct the scan angle, observation zenith angle, or channel brightness temperature before performing profile inversion. In contrast, geostationary satellites are fixed relative to the Earth, allowing for high-frequency continuous observation of the same ground area. This gives them an advantage in capturing rapidly changing weather systems, such as mesoscale severe convection, typhoons, and rainstorm clouds. The main advantage of geostationary satellites is their high observation frequency, enabling them to capture weather phenomena that change rapidly over time, and making them suitable for weather analysis and early warning of mesoscale severe convective weather.
[0004] Geostationary microwave observations exhibit significantly different geometric characteristics from polar-orbiting observations. While the geostationary satellite platform remains approximately stationary relative to the Earth, and the radiometer's observation angle at the same surface point within its field of view remains stable over a long period, the zenith angle varies between different surface points; the farther the observation point is from the nadir, the larger the zenith angle typically is. For geostationary satellites, the zenith angle at the observation point on the surface varies considerably, making existing inversion algorithms for polar-orbiting satellite observation data unsuitable for the inversion process of geostationary microwave observations, thus hindering accurate inversion results. Therefore, there is an urgent need to propose an inversion algorithm suitable for microwave radiation acquired by geostationary satellites, enabling the retrieval of atmospheric temperature and humidity profiles based on this data. Summary of the Invention
[0005] To address the aforementioned technical problems, this application provides a training and prediction method for an atmospheric parameter inversion model based on satellite radiative brightness temperature.
[0006] The first aspect of this application provides a method for training an atmospheric parameter inversion model based on satellite radiative brightness temperature, including: The training data and labels of the model to be trained are obtained in the first and second training phases. The training data includes: multi-channel brightness temperature data and multi-channel attributes of the observation point, observation geometric parameters of the geostationary satellite, and pressure layer parameters representing different pressure layers. The labels include: actual atmospheric parameters of the observation point in different pressure layers. Based on the model to be trained, the multi-channel brightness temperature data and the corresponding channel attributes in the multi-channel attributes of the first training stage are jointly embedded, and the observation geometric parameters of the first training stage are mapped to conditional modulation parameters, with the pressure layer parameters as query terms, to obtain the first predicted atmospheric parameters of the observation point in different pressure layers. The first loss is constructed based on the first predicted atmospheric parameters and the labels of the first training phase, and then backpropagated to update the model to be trained to obtain the pre-trained model. Based on the pre-trained model, the multi-channel brightness temperature data and corresponding channel attributes in the multi-channel attributes of the second training stage are jointly embedded, and the observation geometric parameters of the second training stage are mapped to conditional modulation parameters. The pressure layer parameters are used as query terms to obtain the second predicted atmospheric parameters of the observation point in different pressure layers. Based on the second predicted atmospheric parameters, the multi-channel attributes of the training data of the second training stage and the observation geometric parameters, the multi-channel simulated brightness temperature is obtained through the forward radiative transfer model. A brightness temperature consistency loss is constructed based on the multi-channel simulated brightness temperature and the multi-channel brightness temperature data from the second training phase. A second loss is constructed based on the second predicted atmospheric parameters and the labels from the second training phase. The pre-trained model is then backpropagated to update the model to obtain the atmospheric parameter inversion model. The atmospheric parameters include atmospheric temperature and / or atmospheric humidity.
[0007] In some embodiments of this application, constructing a second loss based on the second predicted atmospheric parameters and the labels of the second training phase includes: determining a supervisory loss based on the second predicted atmospheric parameters and the actual atmospheric parameters of each pressure layer; determining a vertical smoothing loss based on the first-order difference and / or second-order difference of the second predicted atmospheric parameters between adjacent pressure layers; determining a physical boundary constraint loss based on the second predicted atmospheric parameters; and obtaining the second loss by weighted summation of the supervisory loss, the vertical smoothing loss, and the physical boundary constraint loss.
[0008] In some embodiments of this application, the second predicted atmospheric parameters include: a second predicted atmospheric temperature and a second predicted atmospheric humidity, and the vertical smoothing loss is: ; in, Indicates vertical smoothing loss. This indicates the total number of pressure layers. Indicates a pressure layer. Indicates the first Each pressure layer The second predicted atmospheric temperature, Indicates the first Each pressure layer The second predicted atmospheric humidity, This indicates the preset balancing weights.
[0009] In some embodiments of this application, the second predicted atmospheric parameter includes the second predicted atmospheric humidity, and the physical boundary constraint loss includes humidity non-negativity and saturation constraint loss, wherein the humidity non-negativity and saturation constraint loss is: ; in, This indicates that the humidity is non-negative and the saturation constraint loss is _____. This indicates the total number of pressure layers. Indicates the first Each pressure layer The second predicted atmospheric humidity, Indicates the first Each pressure layer Second predicted atmospheric temperature The saturated humidity below.
[0010] In some embodiments of this application, the second predicted atmospheric parameter includes a second predicted atmospheric temperature, and the physical boundary constraint loss includes a temperature decay rate constraint loss, which includes: ; in, This represents the loss due to the temperature lapse rate constraint. Indicates the first The height of each pressure layer The second predicted atmospheric temperature, This indicates the preset stability threshold.
[0011] In some embodiments of this application, the second predicted atmospheric parameters include: second predicted atmospheric humidity and second predicted atmospheric temperature; the actual atmospheric parameters include: actual atmospheric humidity and actual atmospheric temperature; and the supervised loss is: ; in, Indicates monitoring losses, This indicates the total number of pressure layers. Indicates the first Each pressure layer The second predicted atmospheric temperature, Indicates the first Each pressure layer The actual atmospheric temperature Indicates the first Each pressure layer The second predicted atmospheric humidity, Indicates the first Each pressure layer The actual atmospheric humidity, Indicates the first The weight of the temperature term in each pressure layer Indicates the first The humidity term weights for each pressure layer.
[0012] In some embodiments of this application, the multi-channel brightness temperature data includes: measured brightness temperature data from geostationary satellites, and / or simulated observed brightness temperature data generated by a forward radiative transfer model; the brightness temperature consistency loss is: ; in, This indicates a loss in brightness temperature uniformity. This indicates the total number of channels for multi-channel brightness temperature data. Indicates the observation point The Simulated brightness temperature of each channel, Indicates the observation point The first in multi-channel brightness temperature data Actual brightness temperature data for each channel This indicates the channel weight.
[0013] In some embodiments of this application, the second predicted atmospheric parameters of the observation point at different pressure layers are obtained by using multi-channel brightness temperature data, multi-channel attributes, and observation geometric parameters as model inputs and pressure layer parameters as query terms. This includes: for each channel, inputting the channel's brightness temperature data and channel attributes as parameter units into the encoder to obtain multi-band brightness temperature coding features; mapping the observation geometric parameters to conditional modulation parameters based on a multilayer perceptron, whereby the conditional modulation parameters characterize the modulation effect of the observation geometric parameters on the multi-band brightness temperature coding features; modulating the multi-band brightness temperature coding features according to the conditional modulation parameters to obtain modulated multi-band brightness temperature coding features; using the pressure layer parameters of the observation point at different pressure layers as query terms, extracting decoding features of different pressure layer parameters based on the modulated multi-band brightness temperature coding features; and outputting the second predicted atmospheric parameters of the observation point at different pressure layers based on the decoding features of different pressure layer parameters.
[0014] In some embodiments of this application, the second predicted atmospheric parameters include: a second predicted atmospheric temperature and a second predicted atmospheric humidity; outputting the predicted atmospheric parameters of the observation point at different pressure layers based on the decoding features of different pressure layer parameters includes: inputting the decoding features of different pressure layer parameters to a temperature output head to obtain the second predicted atmospheric temperature of different pressure layers; inputting the decoding features of different pressure layer parameters to a humidity output head to obtain the humidity output of different pressure layers; and processing the humidity output of different pressure layers based on the Softplus activation function to obtain the second predicted atmospheric humidity of different pressure layers.
[0015] In some embodiments of this application, the observation geometric parameters include at least one of the following: the longitude difference, latitude difference, observation zenith angle, and angle factor related to the tilt path length between the observation point and the nadir point of the geostationary satellite. The angle factor related to the tilt path length includes the secant value of the observation zenith angle.
[0016] The second aspect of this application provides an atmospheric parameter prediction method, comprising: acquiring multi-channel attributes of multiple acquisition channels of a geostationary satellite, multi-channel brightness temperature data acquired by the geostationary satellite for a target observation point, observation geometric parameters of the geostationary satellite, and pressure layer parameters of a target pressure layer set; using the multi-channel brightness temperature data acquired by the geostationary satellite, the multi-channel attributes of multiple acquisition channels, and the observation geometric parameters of the geostationary satellite as model inputs to an atmospheric parameter inversion model, and using the pressure layer parameters of the target pressure layer set as query terms to obtain atmospheric parameters of each pressure layer in the target pressure layer set, wherein the atmospheric parameter inversion model is trained based on the atmospheric parameter inversion model training method based on satellite radiation brightness temperature provided in the first aspect of this application.
[0017] The beneficial effects of this invention are as follows: The atmospheric parameter inversion model training method based on satellite radiative brightness temperature provided in this application can perform two-stage training on the model to be trained to obtain an atmospheric parameter inversion model. During the two-stage training process, the multi-channel brightness temperature, multi-channel attributes, and observation geometric parameters of the geostationary satellite collected from the observation points are used as model inputs, and the pressure layer parameters of different pressure layers are used as query terms. Thus, the predicted atmospheric parameters for different pressure layers are obtained through the model. The above training process enables the model to learn brightness temperature data from different channels, that is, to learn the correlation and complementarity between brightness temperatures in different frequency bands, and to correct the model during the processing of multi-channel brightness temperature data based on the observation geometric parameters of the geostationary satellite. Then, the predicted atmospheric parameters for different pressure layers can be obtained based on the pressure layer parameters of different pressure layers. In the above two-stage training process… The first training phase's first loss is constructed based on the first predicted atmospheric parameters and the labels from the first training phase. This first loss allows the first predicted atmospheric parameters obtained by the pre-trained model to approximate the actual atmospheric parameters, giving the pre-trained model basic inversion capabilities. The second training phase uses a brightness-temperature consistency loss constructed based on multi-channel simulated brightness temperature and multi-channel brightness-temperature data from the second training phase. The second predicted atmospheric parameters and the labels from the second training phase are used to construct a second loss, which is then backpropagated to update the pre-trained model, thereby obtaining an atmospheric parameter inversion model. This improves the consistency of radiative transfer between the atmospheric parameters predicted by the atmospheric parameter inversion model and the original multi-channel brightness-temperature data, ensuring that the inversion results satisfy the microwave radiative transfer mechanism. Consequently, the atmospheric parameter inversion model can obtain more accurate atmospheric parameter prediction results, leading to accurate atmospheric temperature and humidity profile inversion results. Attached Figure Description
[0018] Figure 1 A flowchart illustrating a training method for an atmospheric parameter inversion model based on satellite radiative brightness temperature, provided in an embodiment of this application; Figure 2 A flowchart illustrating a method for determining a second predicted atmospheric parameter provided in an embodiment of this application; Figure 3 This is a flowchart illustrating an atmospheric parameter prediction method provided in an embodiment of this application. Detailed Implementation
[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0020] Because the brightness temperature data obtained by microwave radiometers on geostationary satellites is essentially a spatially position-dependent multi-angle observation, the inversion problem of atmospheric parameters such as temperature and humidity differs from the angle variation along the scan line problem commonly encountered by polar-orbiting satellites. This places higher demands on the input design, angle representation, and physical constraints of the inversion algorithm. Furthermore, geostationary observations not only suffer from zenith angle differences between surface observation points but also from observation path offsets caused by atmospheric stratification. At the edge of the field of view or in areas with large observation angles, when the satellite's line of sight passes through the atmosphere, its actual radiation contribution path does not strictly correspond to the vertical profile of the surface point but may be influenced by the combined effects of atmospheric conditions and underlying surface conditions in the adjacent region. For example, when the observation path is relatively inclined and the atmospheric height reaches the upper troposphere or lower stratosphere, the spatial projection positions corresponding to different altitude layers will shift, resulting in a certain horizontal path mixing effect in the brightness temperature observation. This indicates that the inversion of atmospheric parameters such as temperature and humidity from geostationary orbit cannot simply follow the approximate vertical scanning or one-dimensional scanning angle processing methods used by polar-orbiting satellites. From the perspective of microwave radiation transmission, the apparent brightness temperature of upward atmospheric radiation is the result of the combined effects of the observation zenith angle, channel frequency, atmospheric absorption characteristics, and underlying surface boundary conditions. As the observation angle gradually tilts, the effective propagation path of electromagnetic waves in the atmosphere increases, and the contributions of atmospheric absorption, spontaneous emission, and surface reflection all change. Even under ideal conditions with identical atmospheric and underlying surface conditions, the observed brightness temperature at different observation zenith angles will exhibit systematic differences; in the region of large observation angles, this difference often manifests as a significant edge effect and may have different directional and amplitude effects on window channels, oxygen absorption channels, and water vapor absorption channels. Therefore, the core issue in geostationary orbit atmospheric temperature and humidity inversion is not simply establishing an empirical fitting relationship between brightness temperature and temperature / humidity.
[0021] Based on the above analysis, this application provides a training method for an atmospheric parameter inversion model based on satellite radiative brightness temperature. This training method can establish a physically consistent nonlinear mapping relationship between atmospheric parameters across multiple frequency bands, observation angles, and pressure layers. This mapping relationship should fully utilize the complementary sensitivity of different microwave channels to atmospheric states at different altitudes, explicitly describe the modulation effect of geostationary orbit observation geometry on brightness temperature, and ensure, through radiative transfer constraints, that the inverted atmospheric parameters can physically explain the original observed brightness temperature.
[0022] See Figure 1 This application provides a method for training an atmospheric parameter inversion model based on satellite radiative brightness temperature. The training method includes: S102. Obtain the training data and labels of the model to be trained in the first and second training phases. The training data includes: multi-channel brightness temperature data of the observation point, multi-channel attributes, observation geometric parameters of the geostationary satellite, and pressure layer parameters representing different pressure layers. The labels include: actual atmospheric parameters of the observation point in different pressure layers.
[0023] Specifically, the training data and labels used in the first and second training phases can be different, or the same. Multi-channel brightness temperature data can be acquired by a geostationary satellite from an observation point, which can be, for example, a point on the Earth's surface. Multi-channel attributes can include the channel attributes of different channels corresponding to the geostationary satellite when acquiring the aforementioned multi-channel brightness temperature data. Channel attributes can include center frequency, polarization, bandwidth, channel noise level, channel type identifier, and prior channel weighting functions. Observation geometric parameters can include at least one of the following: the longitude difference, latitude difference, observation zenith angle, and angle factors related to the tilt path length between the observation point and the nadir point of the geostationary satellite. The angle factors related to the tilt path length include the secant of the observation zenith angle.
[0024] S104. Based on the model to be trained, the multi-channel brightness temperature data and the corresponding channel attributes in the multi-channel attributes of the first training stage are jointly embedded, and the observation geometric parameters of the first training stage are mapped to conditional modulation parameters, and the pressure layer parameters are used as query terms to obtain the first predicted atmospheric parameters of the observation point in different pressure layers.
[0025] Specifically, the model to be trained can be constructed based on the Transformer Physical Information Neural Network (PINN) as the main model, and the pressure layer parameters are processed as query terms based on the constructed pressure layer query decoder to output the first predicted atmospheric parameters for different pressure layers.
[0026] S106. Construct the first loss based on the first predicted atmospheric parameters and the labels of the first training phase, and backpropagate to update the model to be trained to obtain the pre-trained model. S108. Based on the pre-trained model, the multi-channel brightness temperature data and corresponding channel attributes in the multi-channel attributes of the second training stage are jointly embedded, and the observation geometric parameters of the second training stage are mapped to conditional modulation parameters. The pressure layer parameters are used as query terms to obtain the second predicted atmospheric parameters of the observation point in different pressure layers. Based on the second predicted atmospheric parameters, the multi-channel attributes of the training data of the second training stage and the observation geometric parameters, the multi-channel simulated brightness temperature is obtained through the forward radiative transfer model.
[0027] Specifically, in the second training phase, the pre-trained model is fine-tuned, and the forward radiative transfer model can be obtained based on analytical radiative transfer formulas, DOTLRT models, RTTOV-type fast radiative transfer models, etc.
[0028] S110. Construct a brightness temperature consistency loss based on the multi-channel simulated brightness temperature and the multi-channel brightness temperature data of the second training stage, construct a second loss based on the second predicted atmospheric parameters and the labels of the second training stage, and backpropagate to update the pre-trained model to obtain the atmospheric parameter inversion model; wherein, the atmospheric parameters in this embodiment may include atmospheric temperature and / or atmospheric humidity.
[0029] Specifically, based on the prediction requirements, the actual atmospheric parameters in this application embodiment can be atmospheric humidity, atmospheric temperature, or atmospheric humidity and atmospheric temperature, etc.; correspondingly, the first predicted atmospheric parameter can correspond to the first predicted atmospheric humidity and the first predicted atmospheric temperature, or the first predicted atmospheric humidity and the first predicted atmospheric temperature, and the second predicted atmospheric parameter can correspond to the second predicted atmospheric humidity and the second predicted atmospheric temperature, or the second predicted atmospheric humidity and the second predicted atmospheric temperature. Atmospheric humidity profiles and atmospheric temperature profiles can be drawn based on the predicted atmospheric humidity and predicted atmospheric temperature of different pressure layers.
[0030] The atmospheric parameter inversion model training method based on satellite radiative brightness temperature provided in this application involves a two-stage training process to obtain the atmospheric parameter inversion model. During the two-stage training, multi-channel brightness temperature, multi-channel attributes, and observational geometric parameters of the geostationary satellite are used as model inputs, while pressure layer parameters for different pressure layers are used as query terms. This allows the model to obtain predicted atmospheric parameters for different pressure layers. This training process enables the model to learn brightness temperature data from different channels, i.e., to learn the correlation and complementarity between brightness temperatures in different frequency bands. Based on the observational geometric parameters of the geostationary satellite, the model is corrected during the processing of multi-channel brightness temperature data. Then, predicted atmospheric parameters for different pressure layers can be obtained based on the pressure layer parameters. In the two-stage training process described above, the first loss in the first training stage is constructed based on the first predicted atmospheric parameters and the labels from the first training stage. This first loss allows the first predicted atmospheric parameters obtained by the pre-trained model to approximate the actual atmospheric parameters, giving the pre-trained model basic inversion capabilities. The second training phase involves constructing a brightness temperature consistency loss based on multi-channel simulated brightness temperature and the multi-channel brightness temperature data from the second training phase. A second predicted atmospheric parameter and a second loss are then constructed using the labels from the second training phase. Backpropagation updates are performed on the pre-trained model to obtain an atmospheric parameter inversion model. This improves the consistency of radiative transfer between the atmospheric parameters predicted by the atmospheric parameter inversion model and the original multi-channel brightness temperature data, ensuring that the inversion results satisfy the microwave radiative transfer mechanism. This allows the atmospheric parameter inversion model to obtain more accurate atmospheric parameter predictions, leading to accurate atmospheric temperature and humidity profile inversion results.
[0031] See Figure 2 In some embodiments of this application, step S108 may include: S202. For each channel, the brightness temperature data and channel attributes of the channel are input to the encoder as parameter units to obtain multi-band brightness temperature coding features.
[0032] S204. Based on the multilayer perceptron, the observation geometric parameters are mapped to conditional modulation parameters, which are used to characterize the modulation effect of the observation geometric parameters on the multi-band brightness temperature coding features.
[0033] As an example, for the observation point Model input For example: ; in, , Indicates the first Each observation point is at Brightness temperature obtained under microwave channels The number of channels can be set according to the actual number of channels. Taking the acquisition of brightness temperature data by a geostationary satellite based on ATMS-type detectors as an example, the microwave channels... For example, it could be 22; , This represents a set of channel attributes, which may include center frequency, polarization, bandwidth, channel noise level, channel type identifier, and prior channel weighting function, etc. to Used to distinguish different channel attributes in a set of channel attributes; ; Indicates the first The observation geometric parameters corresponding to each observation point may include: the longitude difference between the observation point and the nadir point. and latitude difference Observation point Observation zenith angle , used to characterize the angle factor related to the length of the inclined path If a more precise representation of the observation geometry in the stratified atmosphere is required, equivalent observation angles corresponding to different pressure or altitude layers can be further introduced.
[0034] This application embodiment can treat the brightness temperature data of multiple channels at the same observation point as a single sample for joint characterization. Unlike splitting the 22 channels into 22 independent samples, this method assumes that the brightness temperatures of all channels at the same observation point jointly constrain an atmospheric temperature and humidity profile; therefore, the multi-channel brightness temperatures should be used as a joint input. For the first... The observation point and the first Each channel, with each parameter unit corresponding to a token, can be represented as: ; in, This represents the channel embedding function, which can be implemented by a linear layer, a multilayer perceptron, or by looking up an embedding table. This embedding not only includes the first... The brightness temperature data for each channel also includes the channel's brightness temperature data. Channel properties Channel attributes can include channel frequency, polarization, bandwidth, and noise level, enabling the model to distinguish the physical meaning of brightness temperature in different frequency bands.
[0035] Furthermore, all channel tokens at the same observation point are grouped into a sequence: ; Then input it into the encoder to get: ; This refers to an encoder, which can be a Transformer-based channel encoder. This indicates the multi-band brightness temperature coding characteristics.
[0036] Based on the above embodiments, nonlinear correlations can be established between window zone channels, oxygen absorption channels, and water vapor absorption channels; complementary sensitivities of different channels to temperature and humidity at different altitudes can be learned; the influence of single channel noise or abnormal brightness temperature on the inversion results can be suppressed; and multi-band joint constraint information can be provided for subsequent pressure layer decoding.
[0037] Because the observation angle of the microwave radiometer on a geostationary satellite varies with the location of the observation point, different observation angles will cause changes in atmospheric propagation path length, atmospheric absorption path, edge effect, and surface reflection contribution for the same atmospheric conditions. Therefore, based on step S204, the observation geometry can be explicitly introduced into the network as a conditional variable.
[0038] S206. Modulate the multi-band brightness temperature coding features according to the conditional modulation parameters to obtain the modulated multi-band brightness temperature coding features.
[0039] Specifically, the conditional modulation parameters include scaling parameters and offset parameters. Modulating the multi-band brightness temperature coding features according to the conditional modulation parameters includes: element-wise scaling of the multi-band brightness temperature coding features based on the scaling parameters, and element-wise offset correction based on the offset parameters. Regarding the observation geometry parameters, the longitude difference, latitude difference, and observation zenith angle between the observation point and the nadir point can be determined based on the longitude and latitude of the observation point, the longitude of the nadir point, the Earth's radius, and the orbital radius of the geostationary satellite. The aforementioned longitude difference, latitude difference, and / or observation zenith angle are then used as the first... Observation geometric parameters corresponding to each observation point ,Will Input angle-conditional embedding network to obtain embedding results : ; in, This represents a multilayer perceptron. To ensure that angle information is not simply concatenated to the input but dynamically influences the brightness temperature feature representation, embodiments of this application employ conditional modulation to map the angle embedding to scaling parameters. and offset parameters : ; And adaptive modulation is applied to the channel coding features:
[0040] in, This represents the angle embedding mapping function. Representation layer normalization, This represents element-wise multiplication. This indicates the brightness temperature coding characteristics of the modulated multi-band frequency band.
[0041] Based on the above examples, the systematic changes in brightness temperature caused by the observation angle of the geostationary orbit can be directly introduced into the inversion model; the same multi-band brightness temperature characteristics can have different inversion responses under different observation angle conditions; the inversion stability of the field of view edge and large zenith angle region can be improved; and the problem of simply treating observation samples from different angles as samples with the same distribution can be avoided.
[0042] S208. Using the pressure layer parameters of the observation point at different pressure layers as query items, the decoding features of different pressure layer parameters are extracted based on the modulated multi-band brightness temperature coding features. S210. Output the second predicted atmospheric parameters of the observation point in different pressure layers based on the decoding characteristics of different pressure layer parameters.
[0043] As an example, the output of the pre-trained model is the first... The second predicted atmospheric parameters for each observation point include the second predicted atmospheric temperature profile and the second predicted atmospheric humidity profile. Different pressure layers can be represented as follows: ; in, Indicates the first Pressure layer parameters of each pressure layer.
[0044] For each pressure layer Construct a pressure layer query token: ; in, This is the embedding function for the pressure layer. (Using...) This is because atmospheric pressure has an approximately exponential relationship with altitude, and logarithmic pressure is more suitable for expressing the hierarchical structure in the vertical direction.
[0045] Combining all pressure layer query tokens into a query sequence: ; Then, using the cross-attention mechanism in the Transformer decoder, the pressure layer query token is used as the query term, along with the modulated multi-band brightness temperature encoded features. Using these as keys and values, decoded features of parameters for different pressure layers are extracted. : ; in, , Indicates the first The observation point at the ... Decoding features on each pressure layer. The above example can transform the mapping from "multi-channel brightness temperature to the entire profile" into the process of "querying multi-channel brightness temperature features for each pressure layer"; enabling the model to output temperature and humidity on a specified pressure layer; facilitating adaptation to different vertical layers or different pressure layer settings; and enhancing the contextual relationship between the outputs of different pressure layers.
[0046] Based on the above steps, multi-channel brightness temperature data can be organized by channel, and the brightness temperature data of any channel and its corresponding channel attributes can be constructed as a parameter unit. This unit is then encoded using an encoder to obtain multi-band brightness temperature coding features. The observation geometry parameters of the geostationary satellite are converted into conditional modulation parameters, and the multi-band brightness temperature coding features are modulated based on these parameters, resulting in modulated multi-band brightness temperature coding features that incorporate the observation geometry parameters. By using the pressure layer parameters of the observation point at different pressure layers as query terms, decoding features for different pressure layer parameters can be extracted from the modulated multi-band brightness temperature coding features. Based on these decoding features, the second predicted atmospheric parameters for the observation point at different pressure layers can be output.
[0047] In some embodiments of this application, the second predicted atmospheric parameters include: a second predicted atmospheric temperature and a second predicted atmospheric humidity; outputting the predicted atmospheric parameters of the observation point at different pressure layers based on the decoding features of different pressure layer parameters includes: inputting the decoding features of different pressure layer parameters to a temperature output head to obtain the second predicted atmospheric temperature of different pressure layers; inputting the decoding features of different pressure layer parameters to a humidity output head to obtain the humidity output of different pressure layers; and processing the humidity output of different pressure layers based on the Softplus activation function to obtain the second predicted atmospheric humidity of different pressure layers.
[0048] As an example, in obtaining the first Decoding characteristics of different pressure layer parameters at each observation point Then, for any decoded feature The second predicted atmospheric temperature and the second predicted atmospheric humidity can be determined based on the following formula: ; ; in, Indicates pressure layer parameters The second predicted atmospheric temperature corresponding to the pressure layer. Indicates the temperature output head. Indicates pressure layer parameters The second predicted atmospheric humidity corresponding to the pressure layer. Indicates humidity output head, This represents the Softplus activation function. The temperature and humidity output heads can be composed of linear layers, one-dimensional convolutional layers, or multilayer sensing mechanisms. The second predicted atmospheric temperature is the second predicted atmospheric temperature profile, and the second predicted atmospheric humidity is the second predicted atmospheric humidity profile. Second predicted atmospheric parameters at each observation point for: ; ; ; in, Indicates pressure layer parameters The second predicted atmospheric temperature profile corresponding to the pressure layer. Indicates pressure layer parameters The second predicted atmospheric humidity profile corresponding to the pressure layer.
[0049] In some embodiments of this application, to reflect the physical coupling between temperature and humidity, the temperature branch and the humidity branch may share decoding features, or cross-attention or gating fusion may be introduced between the two branches. For example, the humidity branch may utilize intermediate features from the temperature branch to enhance the response of humidity prediction to temperature conditions, saturated vapor pressure, and vertical stability.
[0050] In some embodiments of this application, the process of obtaining the first predicted atmospheric parameters by processing the model input and query terms in the first training phase can be the same as the process of obtaining the second predicted atmospheric parameters by processing the model input and query terms in the second training phase.
[0051] In some embodiments of this application, the second predicted atmospheric parameters may include a second predicted atmospheric temperature profile and a second predicted atmospheric humidity profile, and the multi-channel simulated brightness temperature in step S108 can be obtained based on the following formula: ; in, , Indicates the first Each observation point is at Simulated brightness temperature under one microwave channel This represents the forward radiative transfer model. To enable the brightness-temperature consistency loss to be used for backpropagation to update the atmospheric parameter inversion model, the forward radiative transfer model can employ a differentiable analytical radiative transfer operator, a differentiable approximation operator built based on a fast radiative transfer model, or a differentiable surrogate forward model trained offline using non-automatic differential radiative transfer models such as DOTLRT and RTTOV. When using a non-automatic differential radiative transfer model, the gradient of the multi-channel simulated brightness temperature relative to the predicted atmospheric parameters can be approximated using a surrogate forward model, numerical difference gradient, or lookup table interpolation, thereby updating the atmospheric parameter inversion model parameters using the brightness-temperature consistency loss. Based on the above embodiments, step S108 can transform "whether the temperature and humidity profile is reasonable" into "whether the profile can reinterpret the original observed brightness temperature." If the predicted profile is close to the supervision label but the observed brightness temperature cannot be reconstructed through the forward model, it indicates that the prediction result is inconsistent in the sense of radiative transfer and needs to be constrained by the physical consistency loss.
[0052] In some embodiments of this application, the multi-channel brightness temperature data may include measured brightness temperature data from geostationary satellites and / or simulated observed brightness temperature data generated by a forward radiative transfer model; the brightness temperature consistency loss is: ; in, This indicates a loss in brightness temperature uniformity. This indicates the total number of channels for multi-channel brightness temperature data. Indicates the observation point The Simulated brightness temperature of each channel, Indicates the observation point In the above multi-channel brightness temperature data, the first Actual brightness temperature data for each channel This represents the channel weights. The brightness temperature consistency loss can constrain the predicted atmospheric parameters to be able to regenerate the original multi-band brightness temperatures through radiative transfer forward modeling, thereby improving the physical consistency of the inversion results.
[0053] Specifically, channel weight It can be determined based on channel noise level, channel reliability, channel sensitivity height, or observation angle, for example: ; in For the first The noise level or noise equivalent temperature difference (NEDT) of each channel. Channels with lower noise and higher signal-to-noise ratio can be assigned greater weight.
[0054] For regions with large observation angles, angle weights can be further introduced: ; in, This is a weighting function related to the observed zenith angle, used to adjust the contribution of samples from different observation angles during training.
[0055] In some embodiments of this application, step S110, which constructs the second loss based on the second predicted atmospheric parameters and the labels of the second training phase, may include: determining a supervisory loss based on the second predicted atmospheric parameters and the actual atmospheric parameters of each pressure layer; determining a vertical smoothing loss based on the difference in the second predicted atmospheric parameters of adjacent pressure layers; determining a physical boundary constraint loss based on the second predicted atmospheric parameters; and obtaining the second loss by weighted summation of the supervisory loss, the vertical smoothing loss, and the physical boundary constraint loss.
[0056] In some embodiments of this application, the second predicted atmospheric parameters may include: a second predicted atmospheric temperature and a second predicted atmospheric humidity, with the vertical smoothing loss being: ; in, Indicates vertical smoothing loss. This indicates the total number of pressure layers. Indicates a pressure layer. Indicates the first Each pressure layer The second predicted atmospheric temperature, Indicates the first Each pressure layer The second predicted atmospheric humidity, This indicates the preset balancing weights.
[0057] In some embodiments of this application, the second predicted atmospheric parameter may include a second predicted atmospheric humidity, and the physical boundary constraint loss may include a humidity non-negativity and saturation constraint loss, which is: ; in, This indicates that the humidity is non-negative and the saturation constraint loss is _____. This indicates the total number of pressure layers. Indicates the first Each pressure layer The second predicted atmospheric humidity, Indicates the first Each pressure layer Second predicted atmospheric temperature The saturated humidity below.
[0058] In some embodiments of this application, the second predicted atmospheric parameter may include a second predicted atmospheric temperature, and the physical boundary constraint loss may include a temperature decay rate constraint loss, which includes: ; in, This represents the loss due to the temperature lapse rate constraint. Indicates the first The height of each pressure layer The second predicted atmospheric temperature, This indicates the preset stability threshold.
[0059] In some embodiments of this application, the second predicted atmospheric parameters may include: a second predicted atmospheric humidity and a second predicted atmospheric temperature; the actual atmospheric parameters include: actual atmospheric humidity and actual atmospheric temperature; and the supervised loss is: ; in, Indicates monitoring losses, This indicates the total number of pressure layers. Indicates the first Each pressure layer The second predicted atmospheric temperature, Indicates the first Each pressure layer The actual atmospheric temperature Indicates the first Each pressure layer The second predicted atmospheric humidity, Indicates the first Each pressure layer The actual atmospheric humidity, Indicates the first The weight of the temperature term in each pressure layer Indicates the first The humidity term weights for each pressure layer.
[0060] In some embodiments of this application, the first loss and the second loss can be the same loss, and the first loss and the second loss can be, for example: ; ; in, Indicates the first loss and the second loss. Indicates the non-negativity of humidity and the saturation constraint loss. and temperature lapse rate constraint loss The physical boundary constraint loss obtained by weighted summation This indicates the weight of the temperature lapse rate constraint. and This indicates the weight of the corresponding loss term.
[0061] In some embodiments of this application, the total loss in the second training phase can be a weighted sum of the brightness-temperature consistency loss and the second loss: ; in, This represents the total loss for the second training phase, obtained by weighted summation of the brightness-temperature consistency loss and the second loss. The weight representing the monitoring loss, This indicates a loss in brightness temperature uniformity. The weights represent the brightness-temperature consistency loss. In some embodiments of this application, the second training phase may employ a smaller learning rate and trim the gradients to improve training stability.
[0062] See Figure 3 This application also provides an atmospheric parameter prediction method, including: S302. Obtain the multi-channel attributes of multiple acquisition channels of the geostationary satellite, the multi-channel brightness temperature data acquired by the geostationary satellite for the target observation point, the observation geometric parameters of the geostationary satellite, and the pressure layer parameters of the target pressure layer set. S304. The multi-channel brightness temperature data collected by the geostationary satellite, the multi-channel attributes of multiple acquisition channels, and the observation geometric parameters of the geostationary satellite are used as the model input of the atmospheric parameter inversion model. The pressure layer parameters of the target pressure layer set are used as the query items to obtain the atmospheric parameters of each pressure layer in the target pressure layer set. The atmospheric parameter inversion model is trained based on the atmospheric parameter inversion model training method based on satellite radiation brightness temperature provided in the embodiments of this application.
[0063] In summary, the atmospheric parameter inversion model training method based on satellite radiative brightness temperature provided in this application achieves the following technical effects: First, a direct inversion framework under multi-angle observation conditions of geostationary orbit is proposed. The model no longer corrects the brightness temperature of tilted observations to the vertical direction first, but instead uses the observation angle as a conditional variable to directly participate in the inversion, reducing the risk of angle correction errors being transmitted to the profile inversion process.
[0064] Second, a multi-band brightness temperature joint coding method is proposed. The model encodes the brightness temperature of all channels at the same observation point as a whole sample, rather than splitting different channels into independent samples, thereby making fuller use of the complementary sensitivity of different frequency bands to atmospheric conditions at different altitudes.
[0065] Third, the observation geometric parameters such as longitude difference, latitude difference, observation zenith angle, and path length correlation factor are mapped into modulation parameters to adaptively adjust the brightness temperature characteristics, so that the model can adapt to the changes in brightness temperature distribution under different observation angles.
[0066] Fourth, a pressure layer query-based decoding structure is proposed. The model uses the pressure layer as the query term, extracts atmospheric state information related to the specified pressure layer from multi-band brightness temperature features, and directly outputs the temperature and humidity at different pressure layers, avoiding the misuse of pressure layers as observation input.
[0067] Fifth, a physical consistency loss in radiative transfer is proposed. Predicted atmospheric parameters are input into the forward radiative transfer model to reconstruct the multi-channel brightness temperature, which is then compared with the input observed brightness temperature to ensure that the inversion results satisfy the microwave radiative transfer mechanism.
[0068] Sixth, introducing physical constraints such as vertical smoothing, non-negative humidity, saturated humidity, and temperature lapse rate can improve the physical rationality and stability of the inversion results.
[0069] This application also provides an atmospheric parameter prediction device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the atmospheric parameter prediction method provided in this application.
[0070] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the atmospheric parameter prediction method provided in this application.
[0071] The processor mentioned above can be a central processing unit (CPU), but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), or field-programmable gate arrays (FPGAs). The memory can be a programmable gate array (GGEA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor. In some embodiments, the memory can be an internal storage unit of the device, such as the device's hard drive or RAM. In other embodiments, the memory can be an external storage device of the device, such as a plug-in hard drive, smart media card (SMC), secure digital card (SD), flash card, etc. Furthermore, the memory can include both internal and external storage units of the device. The memory is used to store the operating system, applications, bootloader, data, and other programs, such as the program code of a computer program. The memory can also be used to temporarily store data that has been output or will be output.
[0072] This application provides a computer program product that, when run on a terminal device, enables the terminal device to implement the steps described in the various method embodiments above.
[0073] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of protection of this application is limited to these examples; under the concept of this application, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of one or more embodiments of this application as described above, which are not provided in detail for the sake of brevity.
[0074] One or more embodiments in this application are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of this application. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of one or more embodiments in this application should be included within the protection scope of this application.
Claims
1. A training method for an atmospheric parameter inversion model based on satellite radiative brightness temperature, characterized in that, The training method includes: The training data and labels of the model to be trained are obtained in the first and second training phases. The training data includes: multi-channel brightness temperature data, multi-channel attributes, observation geometric parameters of geostationary satellites, and pressure layer parameters representing different pressure layers. The labels include: actual atmospheric parameters of the observation points in different pressure layers. Based on the model to be trained, the multi-channel brightness temperature data and the corresponding channel attributes in the multi-channel attributes of the first training stage are jointly embedded, and the observation geometric parameters of the first training stage are mapped to conditional modulation parameters. The pressure layer parameters are used as query terms to obtain the first predicted atmospheric parameters of the observation point in different pressure layers. A first loss is constructed based on the first predicted atmospheric parameters and the labels from the first training phase, and the model to be trained is updated by backpropagation to obtain a pre-trained model. Based on the pre-trained model, the multi-channel brightness temperature data and corresponding channel attributes in the multi-channel attributes of the second training stage are jointly embedded, and the observation geometric parameters of the second training stage are mapped to conditional modulation parameters. The pressure layer parameters are used as query terms to obtain the second predicted atmospheric parameters of the observation point in different pressure layers. Based on the second predicted atmospheric parameters, the multi-channel attributes of the training data of the second training stage, and the observation geometric parameters, the multi-channel simulated brightness temperature is obtained through the forward radiative transfer model. A brightness temperature consistency loss is constructed based on the multi-channel simulated brightness temperature and the multi-channel brightness temperature data from the second training phase. A second loss is constructed based on the second predicted atmospheric parameters and the labels from the second training phase. The pre-trained model is then backpropagated to update the model to obtain an atmospheric parameter inversion model. The atmospheric parameters include atmospheric temperature and / or atmospheric humidity.
2. The method for training an atmospheric parameter inversion model based on satellite radiative brightness temperature according to claim 1, characterized in that, The construction of the second loss based on the second predicted atmospheric parameters and the labels of the second training phase includes: The monitoring loss is determined based on the second predicted atmospheric parameters and the actual atmospheric parameters of each of the aforementioned pressure layers; The vertical smoothing loss is determined based on the first-order and / or second-order difference of the second predicted atmospheric parameters between adjacent pressure layers; The physical boundary constraint loss is determined based on the second predicted atmospheric parameters; The second loss is obtained by weighted summation of the supervision loss, the vertical smoothing loss, and the physical boundary constraint loss.
3. The method for training an atmospheric parameter inversion model based on satellite radiative brightness temperature according to claim 2, characterized in that, The second predicted atmospheric parameters include: a second predicted atmospheric temperature and a second predicted atmospheric humidity, and the vertical smoothing loss is: ； in, This represents the vertical smoothing loss. This indicates the total number of pressure layers. Indicates a pressure layer. Indicates the first Each pressure layer The second predicted atmospheric temperature, Indicates the first Each pressure layer The second predicted atmospheric humidity, This indicates the preset balancing weights.
4. The method for training an atmospheric parameter inversion model based on satellite radiative brightness temperature according to claim 2, characterized in that, The second predicted atmospheric parameter includes the second predicted atmospheric humidity, and the physical boundary constraint loss includes humidity non-negativity and saturation constraint loss, wherein the humidity non-negativity and saturation constraint loss is: ； in, This indicates that the humidity is non-negative and the saturation constraint loss is... This indicates the total number of pressure layers. Indicates the first Each pressure layer The second predicted atmospheric humidity, Indicates the first Each pressure layer Second predicted atmospheric temperature The saturated humidity below.
5. The method for training an atmospheric parameter inversion model based on satellite radiative brightness temperature according to claim 2, characterized in that, The second predicted atmospheric parameter includes a second predicted atmospheric temperature, and the physical boundary constraint loss includes a temperature decay rate constraint loss, which includes: ； in, This represents the temperature lapse rate constraint loss. Indicates the first The height of each pressure layer The second predicted atmospheric temperature, This indicates the preset stability threshold.
6. The method for training an atmospheric parameter inversion model based on satellite radiative brightness temperature according to claim 2, characterized in that, The second predicted atmospheric parameters include: second predicted atmospheric humidity and second predicted atmospheric temperature; the actual atmospheric parameters include: actual atmospheric humidity and actual atmospheric temperature; and the supervised loss is: ； in, This indicates the monitoring loss. This indicates the total number of pressure layers. Indicates the first Each pressure layer The second predicted atmospheric temperature, Indicates the first Each pressure layer The actual atmospheric temperature Indicates the first Each pressure layer The second predicted atmospheric humidity, Indicates the first Each pressure layer The actual atmospheric humidity, Indicates the first The weight of the temperature term in each pressure layer Indicates the first The humidity term weights for each pressure layer.
7. The method for training an atmospheric parameter inversion model based on satellite radiative brightness temperature according to claim 1, characterized in that, The multi-channel brightness temperature data includes: measured brightness temperature data from geostationary satellites, and / or simulated observed brightness temperature data generated by the forward radiative transfer model; the brightness temperature consistency loss is: ； in, This indicates the loss of brightness temperature uniformity. This indicates the total number of channels for the multi-channel brightness temperature data. Indicates the observation point The Simulated brightness temperature of each channel, Indicates the observation point The first in the multi-channel brightness temperature data Brightness temperature data for each channel This indicates the channel weight.
8. The method for training an atmospheric parameter inversion model based on satellite radiative brightness temperature according to claim 1, characterized in that, The step of using the multi-channel brightness temperature data, multi-channel attributes, and the observed geometric parameters from the second training phase as model inputs, and the pressure layer parameters as query terms, to obtain the second predicted atmospheric parameters of the observation point at different pressure layers includes: For each channel, the brightness temperature data and channel attributes of the channel are input to the encoder as parameter units to obtain multi-band brightness temperature coding features; The observation geometric parameters are mapped to conditional modulation parameters based on a multilayer perceptron. The conditional modulation parameters are used to characterize the modulation effect of the observation geometric parameters on the multi-band brightness temperature coding features. The multi-band brightness temperature coding feature is modulated according to the conditional modulation parameters to obtain the modulated multi-band brightness temperature coding feature. Using the pressure layer parameters of the observation point at different pressure layers as query items, the decoding features of different pressure layer parameters are extracted based on the modulated multi-band brightness temperature coding features; Based on the decoding characteristics of the different pressure layer parameters, the second predicted atmospheric parameters of the observation point in different pressure layers are output.
9. The method for training an atmospheric parameter inversion model based on satellite radiative brightness temperature according to claim 8, characterized in that, The second predicted atmospheric parameters include: a second predicted atmospheric temperature and a second predicted atmospheric humidity; the step of outputting the predicted atmospheric parameters of the observation point at different pressure layers based on the decoding characteristics of the different pressure layer parameters includes: The decoding features of the different pressure layer parameters are input into the temperature output head to obtain the second predicted atmospheric temperature of the different pressure layers; The decoding features of the parameters of the different pressure layers are input to the humidity output head to obtain the humidity output of the different pressure layers; The humidity output of the different pressure layers is processed based on the Softplus activation function to obtain the second predicted atmospheric humidity of the different pressure layers.
10. The method for training an atmospheric parameter inversion model based on satellite radiative brightness temperature according to claim 1, characterized in that, The observation geometric parameters include at least one of the following: the longitude difference, latitude difference, observation zenith angle, and angle factor related to the tilt path length between the observation point and the nadir point of the geostationary satellite, wherein the angle factor related to the tilt path length includes the secant value of the observation zenith angle.
11. A method for predicting atmospheric parameters, characterized in that, include: The multi-channel attributes of multiple acquisition channels of a geostationary satellite are obtained, including the multi-channel brightness temperature data acquired by the geostationary satellite for the target observation point, the observation geometric parameters of the geostationary satellite, and the pressure layer parameters of the target pressure layer set. The multi-channel brightness temperature data acquired by the geostationary satellite, the multi-channel attributes of the multiple acquisition channels, and the observation geometric parameters of the geostationary satellite are used as the model inputs of the atmospheric parameter inversion model. The pressure layer parameters of the target pressure layer set are used as query items to obtain the atmospheric parameters of each pressure layer in the target pressure layer set. The atmospheric parameter inversion model is trained based on the atmospheric parameter inversion model training method based on satellite radiation brightness temperature as described in any one of claims 1-10.