Atmospheric temperature and humidity profile inversion method and system based on foundation infrared hyperspectral radiation

Abstract

The invention discloses an atmospheric temperature and humidity profile inversion method and system based on foundation infrared hyperspectral radiation. The method comprises the following steps: obtaining a predicted value of the nth time step of a temperature profile and a predicted value of the nth time step of a humidity profile from any input hyperspectral radiation vector of the nth time step by using a pre-trained bidirectional long-short-term memory network model; the bidirectional long-short-term memory network model comprises a feature extraction module, a temperature bidirectional long-short-term memory module, a humidity bidirectional long-short-term memory module, a temperature regression module and a humidity regression module, and the temperature regression module carries out regression on the bidirectional hidden state of the temperature sequence to obtain a predicted value of the nth time step of the temperature profile; and carrying out regression on the bidirectional hidden state of the humidity sequence through a humidity regression module to obtain a predicted value of the n-th time step of the humidity profile. The invention aims to improve the efficiency and precision of atmospheric temperature and humidity profile inversion.

Description

Technical Field

[0001] This invention relates to the field of weather forecasting technology, specifically to a method and system for inverting atmospheric temperature and humidity profiles based on ground-based infrared hyperspectral radiation. Background Technology

[0002] Continuous monitoring of atmospheric temperature and humidity profiles is crucial for understanding weather system evolution and advancing forecasts across spatial and temporal scales. These profiles form the fundamental inputs for initializing numerical weather prediction models, supporting applications ranging from short-term hazardous weather forecasts to long-term climate simulations. Research shows that assimilating atmospheric temperature and humidity profiles can significantly improve forecast accuracy. For example, integrating high-resolution temperature inversion improves the characterization of large-scale environmental steering flows, thus enhancing the prediction of phenomena such as hurricane tracks. Simultaneously, assimilating humidity profiles effectively constrains atmospheric humidity analysis, thereby mitigating errors in subsequent integrated water vapor and precipitation forecasts. Furthermore, as a key input to radiative transfer models, accurate atmospheric temperature and humidity profiles are indispensable for improving the inversion accuracy of key satellite inversion parameters (such as surface temperature and aerosol optical thickness). Therefore, developing atmospheric temperature and humidity profile inversion techniques that combine high accuracy and computational efficiency is crucial for continuously improving numerical weather prediction and enhancing satellite quantitative remote sensing capabilities.

[0003] Radiosondes have traditionally served as a benchmark for atmospheric thermodynamic profiles, providing high-precision in-situ observations. For example, the calibration uncertainties for the RS-90 and RS-92 series are well-defined: 0.5°C for temperature and 5%RH for relative humidity. However, their conventional deployment is inherently limited by coarse temporal resolution, typically restricting deployment to four radiosondes per day, spaced six hours apart. This sparse sampling introduces significant temporal discontinuities, hindering the characterization of rapidly evolving boundary layer processes. Consequently, many key weather phenomena cannot be effectively captured within the hourly observation intervals of radiosondes. This lack of temporal continuity imposes significant limitations on using radiosonde observations for minute-level atmospheric monitoring, nowcasting, and high-resolution model validation. To overcome these limitations, ground-based infrared hyperspectral observations, such as those from the Atmospheric Emitted Radiation Interferometer (AERI), offer a promising complementary approach. With its high temporal resolution and sensitivity to the thermal structure of the lower troposphere, AERI provides continuous radiometric measurements, enabling near real-time inversion of temperature and humidity profiles. However, deriving thermodynamic profiles from AERI radiation spectra is essentially an ill-posed inverse problem, in which maintaining a balance between computational efficiency and inversion accuracy poses a continuous challenge. Three main paradigms have been developed for inverting atmospheric thermodynamic profiles from hyperspectral radiation spectra: (1) Physical inversion methods solve nonlinear optimization problems within the Newton-Newton iteration framework, achieving atmospheric state estimation by minimizing the cost function. For example, the "peeling the onion" method used by the AERIprof algorithm only requires calculating the diagonal elements of the Jacobian matrix, which significantly improves computational efficiency compared to the full matrix method. The AERIoe algorithm uses the Gauss-Newton iteration framework combined with optimal estimation theory, stabilizing the iteration process by systematically adjusting the relative weights of the prior and observation covariance matrices. However, both of these methods rely on LBLRTM as the forward model, requiring repeated runs to update the Jacobian matrix in each iteration. This computationally intensive characteristic limits their real-time processing capabilities. (2) Statistical inversion methods establish an empirical mapping between radiation and atmospheric state parameters. Early linear models, such as eigenvector regression, used principal component analysis combined with least squares fitting to approximate the radiation-state relationship. However, simple linear models cannot represent the physical nature of atmospheric temperature and humidity profile inversion. (3) Deep learning methods typically learn the direct nonlinear mapping between observed radiation and the corresponding temperature and humidity profiles, bypassing the need for explicit forward or Jacobian modeling. They pioneered the application of artificial neural networks (implemented as fully connected architectures) to atmospheric temperature and humidity detection, proving the feasibility of directly inverting thermodynamic profiles from radiometric measurements without explicit radiative transfer modeling. However, despite improvements in computational efficiency, these fully connected networks are still limited by their limited nonlinear representation capabilities and cannot effectively capture the inherent spatial-spectral correlations of hyperspectral radiation data. These limitations subsequently promoted the application of convolutional neural networks in atmospheric thermodynamic profile inversion.While deep learning architectures like convolutional neural networks have significantly improved the efficiency and accuracy of atmospheric profile inversion, they remain inherently data-driven and largely ignore underlying physical constraints. This limitation of data-driven parameterization manifests primarily in insufficient structural generalization and a lack of physical constraints. The predicted convective-radiative heating rate exhibits a significantly reduced variance at the 900 hPa boundary layer. Snapshots of instantaneous heating rates indicate overly smooth predictions lacking vertical variability, directly undermining the model's ability to accurately distinguish the heights of inversion and mixing layers. Secondly, extrapolated predictions can become physically inconsistent or unreliable, violating fundamental laws (such as hydrostatic equilibrium), particularly outside the training distribution. This failure stems from the training objective only minimizing single-step errors without penalizing patterns that violate physical constraints, leading to amplified cumulative errors during extrapolation. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a method and system for inverting atmospheric temperature and humidity profiles based on ground-based infrared hyperspectral radiation, which aims to improve the efficiency and accuracy of atmospheric temperature and humidity profile inversion.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A method for inverting atmospheric temperature and humidity profiles based on ground-based infrared hyperspectral radiation includes inputting the hyperspectral radiation vector at any nth time step. The pre-trained bidirectional long short-term memory network model Phy-BiLSTM is used to obtain the predicted value of the temperature profile at the nth time step. Predicted values ​​of humidity profile at the nth time step The bidirectional long short-term memory network model Phy-BiLSTM includes: a feature extraction module. , used to input hyperspectral radiation vector Feature extraction Temperature bidirectional long and short term memory module Used for features Capturing the bidirectional dependence of temperature within the vertical atmospheric structure to obtain the bidirectional hidden state of the temperature sequence. Humidity bidirectional long and short term memory module Used for features Capturing the bidirectional dependence of humidity within the vertical atmospheric structure to obtain the bidirectional hidden state of the humidity sequence. Temperature regression module Used for bidirectional hidden states of temperature sequences. Regression is performed to obtain the predicted value of the temperature profile at the nth time step. Humidity regression module Used for bidirectional hiding of humidity sequences. Regression is performed to obtain the predicted value of the humidity profile at the nth time step. .

[0006] Optionally, the hyperspectral radiance vector at the nth time step The function expression is: ; in, ~ These represent the hyperspectral radiance values ​​of the 1st to Cth spectral channels, where C is the number of spectral channels in the hyperspectral radiance value, and the superscript T indicates the transpose operation; the predicted value of the temperature profile at the nth time step. Predicted values ​​of humidity profile at the nth time step The function expression is: ; ; in, ~ These are the predicted temperatures for the first to Lth vertical atmospheric layers. ~ These are the predicted humidity values ​​for the 1st to Lth vertical atmospheric layers, where L is the number of vertical atmospheric layers.

[0007] Optionally, the feature extraction module It includes multiple stacked sets of one-dimensional convolutional layers and max pooling layers. Each one-dimensional convolutional layer is followed by batch normalization and ReLU activation to capture local spectral dependencies and extract noise-resistant features from the hyperspectral radiative input. The max pooling layer is used to reduce spectral redundancy and preserve radiative features.

[0008] Optionally, the temperature bidirectional long short-term memory module Humidity bidirectional long and short term memory module All are bidirectional long short-term memory networks, and the function expressions for forward and backward propagation of the bidirectional long short-term memory network are as follows: ; ; in, For branches The hidden state of the k-th vertical atmospheric layer obtained from the forward propagation. For branches Forward-passing long short-term memory network, For branches The hidden state of the (k-1)th vertical atmospheric layer obtained from the forward propagation. This represents the input feature for the k-th vertical atmospheric layer. For branches The hidden state of the k-th vertical atmospheric layer obtained by backpropagation For branches Backpropagation Long Short-Term Memory Network For branches The hidden state of the (k-1)th vertical atmospheric layer obtained from the backpropagation, branch Temperature bidirectional long short-term memory module Corresponding temperature branch or humidity bidirectional long short-term memory module For the corresponding humidity branch, the bidirectional hidden state output by the bidirectional long short-term memory network for any k-th vertical atmospheric layer is: ; in, For branches The bidirectional hidden state output by the kth vertical atmospheric layer For splicing operations, The number of Long Short-Term Memory networks in each direction for forward and backward propagation; and the bidirectional hidden states of all vertical atmospheric outputs. Bidirectional hidden state of spliced ​​output temperature sequence Or the bidirectional hidden state of the humidity sequence. .

[0009] Optionally, the temperature regression module Humidity regression module Each layer includes two fully connected layers and one output layer, and the output layer outputs the predicted value of the temperature profile at the nth time step. Or the predicted value of the humidity profile at the nth time step. .

[0010] Optionally, the loss function used during training of the pre-trained bidirectional long short-term memory network model Phy-BiLSTM is expressed as follows: ; in, For loss function, For temperature loss, For humidity loss, The loss is physical, and we have: ; ; ; ; ; ; ; ; in, and For balance coefficient, For water vapor mixing ratio loss, and Let the predicted water vapor mixing ratio and the reference water vapor mixing ratio be the values ​​at the nth time step. For element-wise multiplication, Let be the vertical atmospheric pressure vector at the nth time step. This is the actual water vapor pressure. The predicted value at the nth time step of the temperature profile. The corresponding saturated vapor pressure, This is the vapor pressure of water at the triple point. and These are the fitting coefficients. Let be the temperature ratio vector, which represents the ratio of temperature to the triple point temperature. This is the triple point temperature of water; For potential temperature loss, and These are the predicted potential temperature vector and the reference potential temperature vector at the nth time step, respectively. For reference pressure constant, is Poisson's constant.

[0011] Optionally, the function expression for calculating the temperature loss is: ; The function expression for calculating the humidity loss is: ; in, For temperature loss, For humidity loss, Where L is the total number of time steps, and L is the number of vertical atmospheric divisions. and The first The first time step Predicted temperature values ​​and temperature label values ​​for each vertical atmospheric layer. and The first The first time step Predicted humidity values ​​and humidity label values ​​for each vertical atmospheric layer.

[0012] The present invention also provides a geothermal temperature and humidity profile inversion system based on ground-based infrared hyperspectral radiation, comprising a microprocessor and a memory interconnected thereto, wherein the microprocessor is programmed or configured to execute the geothermal temperature and humidity profile inversion method based on ground-based infrared hyperspectral radiation.

[0013] Compared with existing technologies, the present invention mainly achieves the following beneficial effects: The method of the present invention includes using a pre-trained bidirectional long short-term memory network model to obtain the predicted values ​​of the temperature profile and humidity profile at the nth time step by taking the hyperspectral radiation vector at any input time step n. The bidirectional long short-term memory network model includes a feature extraction module, a temperature bidirectional long short-term memory module, a humidity bidirectional long short-term memory module, a temperature regression module, and a humidity regression module. The temperature regression module regresses the bidirectional hidden state of the temperature sequence to obtain the predicted value of the temperature profile at the nth time step; the humidity regression module regresses the bidirectional hidden state of the humidity sequence to obtain the predicted value of the humidity profile at the nth time step. The present invention combines sequence modeling with physical constraints, which can capture the vertical dependence and coupling relationship between atmospheric layers, and improve the efficiency and accuracy of atmospheric temperature and humidity profile inversion. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the network structure of the Phy-BiLSTM bidirectional long short-term memory network model in an embodiment of the present invention.

[0015] Figure 2 This is a scatter plot of the CNN temperature inversion results in an embodiment of the present invention.

[0016] Figure 3 This is a scatter plot of the DReA temperature inversion results in an embodiment of the present invention.

[0017] Figure 4 This is a scatter plot of the Phy-BiLSTM temperature inversion results in an embodiment of the present invention.

[0018] Figure 5 This is a scatter plot of the CNN water vapor mixing ratio inversion results in an embodiment of the present invention.

[0019] Figure 6 This is a scatter plot of the DReA water vapor mixing ratio inversion results in an embodiment of the present invention.

[0020] Figure 7 This is a scatter plot of the Phy-BiLSTM water vapor mixing ratio inversion results in an embodiment of the present invention.

[0021] Figure 8 This shows the deviation distribution of the three models in temperature inversion in the embodiments of the present invention.

[0022] Figure 9 This shows the deviation distribution of the three models in relative humidity inversion in the embodiments of the present invention.

[0023] Figure 10 Box plots showing the temperature distributions obtained from the inversion of three models in this embodiment of the invention.

[0024] Figure 11 This is a box plot of the relative humidity deviation distribution obtained from the inversion of three models in this embodiment of the invention. Detailed Implementation

[0025] To enable those skilled in the art to better understand the technical solutions of the present invention, the technical solutions of the present invention will be further described in detail below with reference to the accompanying drawings in the embodiments of the present invention.

[0026] This embodiment of the atmospheric temperature and humidity profile inversion method based on ground-based infrared hyperspectral radiation includes inputting the hyperspectral radiation vector at any nth time step. The pre-trained bidirectional long short-term memory network model Phy-BiLSTM is used to obtain the predicted value of the temperature profile at the nth time step. Predicted values ​​of humidity profile at the nth time step The Phy-BiLSTM bidirectional long short-term memory network model aims to estimate vertical atmospheric temperature and humidity profiles directly from hyperspectral radiometric observations. It integrates deep sequence modeling with the principle of physical consistency to achieve physically meaningful inversion.

[0027] like Figure 1 As shown, the Phy-BiLSTM bidirectional long short-term memory network model includes: a feature extraction module. , used to input hyperspectral radiation vector Feature extraction Temperature bidirectional long and short term memory module Used for features Capturing the bidirectional dependence of temperature within the vertical atmospheric structure to obtain the bidirectional hidden state of the temperature sequence. Humidity bidirectional long and short term memory module Used for features Capturing the bidirectional dependence of humidity within the vertical atmospheric structure to obtain the bidirectional hidden state of the humidity sequence. Temperature regression module Used for bidirectional hidden states of temperature sequences. Regression is performed to obtain the predicted value of the temperature profile at the nth time step. Humidity regression module Used for bidirectional hiding of humidity sequences. Regression is performed to obtain the predicted value of the humidity profile at the nth time step. .

[0028] In this embodiment, the hyperspectral radiation vector was obtained using a high-spectral-resolution Fourier transform infrared spectrometer (AERI), but other instruments can also be used. The AERI passively records data from 520 to 3020 cm⁻¹. -1The downward spectral radiation corresponds to wavelengths of approximately 3.3–19.2 μm, with a spectral resolution better than 1 cm⁻¹. -1 The nominal sampling interval is 5 minutes. The instrument's spectral configuration allows for detailed sensing of atmospheric temperature and humidity through strong water vapor and carbon dioxide absorption characteristics. In this embodiment, radiation from 311 H2O absorption channels and 244 CO2 channels is used as the input feature set for the inversion network, retaining only the radiation spectrum filtered as cloudless based on synchronous cloud height observations to minimize contamination caused by cloud emission or scattering. AERI detection data can be filtered using a ground-based cloud height instrument. As an all-weather automatic observation device, the ground-based cloud height instrument continuously provides cloud base height information and can be used to identify cloud boundaries and filter data samples affected by cloud contamination. For example, the Vaisala CL31 cloud height sensor, as a ground-based cloud height instrument, operates based on the principle of pulsed lidar. It vertically emits infrared laser pulses with a wavelength of 905 nm (pulse width 110 ns, effective length 16.5 m) that interact with atmospheric components and receives backscattered light signals from cloud droplets and aerosol particles. Its detection range is up to 7.7 km, with a vertical resolution of 10 m. The height of the scattering layer is calculated using the time-of-flight of the echo signal. By combining backscattering intensity threshold determination and instrument background signal correction, it becomes possible to detect cloud base height (accuracy ±5m) and identify clear-sky conditions. An internal 30-second average interval ensures data consistency. To mitigate contamination from cloud emission or scattering in AERI observations, synchronous data from the CL31 ceilometer was used as auxiliary information for cloud screening. Specifically, only AERI observation samples where no clouds were detected by the CL31 within 3 hours before and after the target time period (cloud base height missing and signal-to-noise ratio SNR > 0.18) were retained. The resulting clear-sky radiation spectrum set served as input features for subsequent inversion models. This screening criterion effectively eliminated spurious echoes in the backscattered signal caused by instrument background noise and solar radiation interference.

[0029] The temperature and humidity profiles used during the pre-training of the Phy-BiLSTM bidirectional long short-term memory network model were detected using radiosondes. For example, the radiosonde at the SGP site used in this embodiment transmits four times daily at 05:30, 11:30, 17:30, and 23:30 UTC. Each radiosonde provides a temperature and humidity reference profile from the ground to approximately 3 km, for a total of 37 vertical layers. The vertical resolution is refined at the low boundary layer (8 layers below 0.1 km, 19 layers between 0.1 and 1 km, 6 layers between 1 and 2 km, and 4 layers between 2 and 3 km) to resolve near-surface thermodynamic gradients. The radiosondes used during the analysis included the Vaisala RS92-KL (February 2005 – February 2010) and the RS41-SGP (February 2010 – November 2017), ensuring instrument continuity and traceable calibration. The manufacturer specifies uncertainties of ±0.5K for temperature and ±5% for relative humidity. These radiosonde profiles are taken as ground truth values ​​from the temperature and humidity profiles.

[0030] In this embodiment, the hyperspectral radiation vector at the nth time step The function expression is: ; in, ~ These represent the hyperspectral radiance values ​​of the 1st to Cth spectral channels, where C is the number of spectral channels in the hyperspectral radiance value, and the superscript T indicates the transpose operation; the predicted value of the temperature profile at the nth time step. Predicted values ​​of humidity profile at the nth time step The function expression is: ; ; in, ~ These are the predicted temperatures for the first to Lth vertical atmospheric layers. ~ These represent the predicted humidity values ​​for the 1st to Lth vertical atmospheric layers, where L is the number of vertical atmospheric layers. The goal of the inversion task is to learn the hyperspectral radiance vector at the nth time step. The predicted value of the temperature profile at the nth time step Predicted values ​​of humidity profile at the nth time step Nonlinear mapping between: ; From the hyperspectral radiation vector at the corresponding nth time step The predicted value of the optimal temperature profile at the nth time step. Predicted values ​​of humidity profile at the nth time step . Let N be the hyperspectral radiation vectors at N time steps. The temperature and humidity profiles are for N time steps.

[0031] Feature extraction module in this embodiment Shared by both temperature and humidity branches, it is used to convert high-dimensional spectral inputs into a compact representation, which can be expressed as: ; in, Indicates feature extraction module Extracted features Indicates feature extraction module Learnable parameters. Feature extraction module. It includes multiple stacked sets of one-dimensional convolutional layers and max-pooling layers. Each one-dimensional convolutional layer is followed by batch normalization and ReLU activation to capture local spectral dependencies and extract noise-resistant features from the hyperspectral radiative input. The max-pooling layers are used to reduce spectral redundancy and preserve radiative features. Feature extraction module. A series of one-dimensional convolutional layers are employed to capture local spectral dependencies and extract robust features from the hyperspectral radiative input. Each convolutional layer is followed by batch normalization and ReLU activation to enhance feature stability and nonlinearity. Subsequently, max pooling is applied to progressively reduce spectral redundancy while preserving necessary radiative features. The number of groups of one-dimensional convolutional layers and max pooling layers can be configured as needed. The output feature map of the max pooling layer in the i-th group serves as the input feature map of the one-dimensional convolutional layer in the (i+1)-th group, and the output feature map of the max pooling layer in the last group serves as the feature extraction module. The final output feature map.

[0032] Feature extraction module Extracted It also sends data to two parallel branches: a temperature-based bidirectional long short-term memory module. Humidity bidirectional long and short term memory module Each contains a dedicated bidirectional long short-term memory network, which respectively captures the bidirectional dependence of temperature and humidity within the vertical atmospheric structure, and can be represented as: ; ; in, and These represent the bidirectional hidden states of the temperature and humidity sequences, respectively. and These are temperature bidirectional long short-term memory modules. Humidity bidirectional long and short term memory module Trainable parameters.

[0033] In this embodiment, the temperature bidirectional long short-term memory module Humidity bidirectional long and short term memory module Both are bidirectional long short-term memory (LSSM) networks, composed of forward and backward LSM layers to jointly model local and global spectral correlations. Originally designed for time series modeling, the bidirectional LSM network is used here to capture spatial dependencies along the vertical atmospheric column. Inspired by its ability to learn bidirectional contextual relationships in sequence data, the bidirectional LSM is applied to model the coupling interactions between adjacent atmospheric layers. For this purpose, the extracted spectral features... Reshaped into a sequence form: ; in, Let represent the spectral feature vector associated with the k-th atmospheric layer, where K represents the total number of vertical layers. Each bidirectional long short-term memory branch then processes this feature sequence in both upward and downward directions to effectively model the bidirectional coupling of thermodynamic processes in the vertical dimension. The functional expressions for the forward and backward propagation of the bidirectional long short-term memory network are: ; ; in, For branches The hidden state of the k-th vertical atmospheric layer obtained from the forward propagation. For branches Forward-passing long short-term memory network, For branches The hidden state of the (k-1)th vertical atmospheric layer obtained from the forward propagation. This represents the input feature for the k-th vertical atmospheric layer. For branches The hidden state of the k-th vertical atmospheric layer obtained by backpropagation For branches Backpropagation Long Short-Term Memory Network For branches The hidden state of the (k-1)th vertical atmospheric layer obtained from the backpropagation, branch Temperature bidirectional long short-term memory module Corresponding temperature branch or humidity bidirectional long short-term memory module For the corresponding humidity branch, the bidirectional hidden state output by the bidirectional long short-term memory network for any k-th vertical atmospheric layer is: ; in, For branches The bidirectional hidden state output by the kth vertical atmospheric layer For splicing operations, The number of Long Short-Term Memory networks in each direction for forward and backward propagation; and the bidirectional hidden states of all vertical atmospheric outputs. Bidirectional hidden state of spliced ​​output temperature sequence Or the bidirectional hidden state of the humidity sequence. The concatenated outputs of all K layers form a contextualized vertical representation. It encodes local and long-range dependencies along the vertical profile. This bidirectional mechanism enables the network to jointly represent upward and downward information flows, with the forward path (from the surface to the top) capturing upward turbulent transport and convective energy exchange, while the backward path (from the top to the surface) represents downward modulation related to radiative cooling and sinking effects, thus effectively modeling the coupled thermodynamic interactions that control the temperature and humidity structure in the atmosphere.

[0034] Bidirectional hidden state of temperature and humidity sequences and Using temperature regression module Humidity regression module Obtain the predicted value of the temperature profile at the nth time step. Predicted values ​​of humidity profile at the nth time step , can be represented as: ; ; in, and These are the temperature regression modules. Humidity regression module The trainable parameters. In this embodiment, the temperature regression module... Humidity regression module Each layer includes two fully connected layers and one output layer, and the output layer outputs the predicted value of the temperature profile at the nth time step. Or the predicted value of the humidity profile at the nth time step. .

[0035] In this embodiment, the Phy-BiLSTM bidirectional long short-term memory network model, through this hierarchical design, achieves a data-driven approximation of complex radiative transfer processes while maintaining physical interpretability through profile-level output. This structure enables the model to explicitly capture multi-scale dependencies and cross-layer interactions, which is crucial for accurately retrieving temperature and humidity in the lower atmosphere.

[0036] To explicitly and implicitly enforce the integration process of thermodynamic consistency constraints through a customized loss function, thereby guiding the network to produce physically reasonable and dynamically consistent inversion results, the pre-trained bidirectional long short-term memory network model Phy-BiLSTM in this embodiment uses the following loss function expression during training: ; in, For loss function, For temperature loss, For humidity loss, The loss is physical, and we have: ; ; ; ; ; ; ; ; in, and For balance coefficient, For water vapor mixing ratio loss, and Let the predicted water vapor mixing ratio and the reference water vapor mixing ratio be the values ​​at the nth time step. For element-wise multiplication, Let be the vertical atmospheric pressure vector at the nth time step. This is the actual water vapor pressure. The predicted value at the nth time step of the temperature profile. The corresponding saturated vapor pressure, This is the vapor pressure of water at the triple point. and These are the fitting coefficients. Let be the temperature ratio vector, which represents the ratio of temperature to the triple point temperature. This is the triple point temperature of water; For potential temperature loss, and These are the predicted potential temperature vector and the reference potential temperature vector at the nth time step, respectively. For reference pressure constant, is Poisson's constant.

[0037] In this embodiment, the function expression for calculating the temperature loss is: ; In this embodiment, the function expression for calculating the humidity loss is: ; in, For temperature loss, For humidity loss, Where L is the total number of time steps, and L is the number of vertical atmospheric divisions. and The first The first time step Predicted temperature values ​​and temperature label values ​​for each vertical atmospheric layer. and The first The first time step Predicted humidity values ​​and humidity label values ​​for each vertical atmospheric layer.

[0038] minimize The predicted profile is encouraged to match radiosonde observations, while also satisfying the conditions determined by physical loss. The applied thermodynamic consistency constraints achieve end-to-end optimization of prediction accuracy and physical interpretability, enabling the pre-trained bidirectional long short-term memory network model Phy-BiLSTM in this embodiment to approximate the nonlinear inversion of the radiative transfer process in a physically consistent manner. Although the Phy-BiLSTM network model learns the nonlinear mapping between hyperspectral radiation and atmospheric profiles in a data-driven manner, ensuring the physical consistency of the inverted temperature and humidity fields is crucial. Therefore, this embodiment introduces a set of soft thermodynamic constraints, denoted as physical loss. The inversion is regularized by penalizing violations of fundamental atmospheric relations, where the corresponding physical quantities are calculated and constrained as follows: (1) Water vapor pressure uniformity constraint. The water vapor pressure uniformity constraint forces the temperature to be inverted through the basic relationship between water vapor pressure. Thermodynamic consistency between the temperature and humidity profiles. This constraint ensures that the predicted temperature and relative humidity remain physically consistent according to atmospheric thermodynamics. Saturated vapor pressure Based on the predicted temperature profile The calculation is performed using a piecewise function that considers different temperature ranges. For temperatures above the triple point: ; ; in, for saturated water vapor pressure, The critical pressure of water. The critical temperature of water. , ; It is a dimensionless parameter representing the relative deviation of the current temperature from the critical temperature (i.e., ),coefficient ~ These are constants determined empirically. In this embodiment, we have: ; ; ; ; ; .

[0039] For temperatures at or below the triple point, the saturated vapor pressure is calculated as follows: ; ; in, This is a dimensionless temperature ratio, which is the ratio of the inversion temperature to the triple point temperature of water. This is the vapor pressure of water at the triple point. The triple point temperature of water. , and The coefficients are constants, and we have , .

[0040] Actual water vapor pressure Based on saturated water vapor pressure and predicted relative humidity Derivation: ; in, This indicates element-wise multiplication (consistent with tensor operations in the code).

[0041] Finally, the water vapor mixing ratio The calculation is as follows: ; The difference between the predicted and reference water vapor mixing ratios is minimized using the physical loss term, which is achieved through water vapor mixing ratio loss. : ; (2) Potential temperature constraint. Potential temperature reflects the thermodynamic state of the atmosphere and is closely related to temperature and water vapor content. To ensure that the retrieved temperature profile remains physically consistent with the thermodynamic equilibrium, the predicted temperature vector is constrained. Apply potential temperature constraints. Given the predicted temperature profile. and the corresponding pressure layer Temperature vector The expression for the computation function is as follows: ; in, This represents element-wise multiplication. This is the reference air pressure. It is Poisson's constant. Therefore, the potential temperature loss... It can be represented as: .

[0042] in This represents the reference potential temperature profile calculated based on radiosonde observations. This constraint encourages the predicted temperature field to be thermodynamically consistent with the atmospheric pressure structure. By combining these two constraints, the physical loss can be obtained. : ; in, and These are weighting coefficients used to balance the relative contributions of vapor pressure consistency and potential temperature constraint. The physical loss term in this combination ensures that the retrieved temperature and humidity profiles remain thermodynamically consistent and physically meaningful.

[0043] To verify the atmospheric temperature and humidity profile inversion method based on ground-based infrared hyperspectral radiation in this embodiment, the data preprocessing stage in this embodiment includes: (1) Quality control: Before model development, a strict quality control procedure was performed on the hyperspectral radiation vector to eliminate outliers and ensure the reliability of the dataset. Specifically, the Z-score statistical method was used for outlier detection, which evaluates the deviation of a single observation of the hyperspectral radiation vector from the overall distribution. Specifically, the Z-score is calculated as follows: ; in, Let i be the i-th observation of the hyperspectral radiance vector. Let represent the mean and standard deviation of the observed sequence of hyperspectral radiance vectors, respectively. Then... Observations exceeding a preset value (e.g., a value of 2) are considered anomalous and subsequently removed, effectively eliminating anomalous radiation values ​​caused by instrument noise or environmental interference, such as negative radiation and sudden spikes. (2) Cloud screening and data pairing: To ensure consistency between radiation observations and reference profiles, a Vaisala CL31 cloud height meter was used to measure and exclude spectra contaminated by clouds, with an accuracy of ±50 meters for detecting cloud base height. Only AERI and radiosonde clear-sky observations co-located in time (within ±5 minutes) were retained. The final dataset contains more than 12 years of high-quality co-located radiation-profile pairs, providing a solid foundation for training and evaluating the proposed inversion framework. This experiment used 555 AERI spectral channels, including 244 temperature-sensitive bands and 311 water vapor-sensitive bands. Following the data processing procedures, 8,153 clear-sky samples collected between 2006 and 2017 were retained, ensuring the reliability and representativeness of the dataset. Of these, 7351 samples from 2006 to 2016 were used for model training, and 802 samples from 2017 were retained for independent testing. The feature extraction module in this embodiment... The system consists of four stacked sets of one-dimensional convolutional layers and max-pooling layers. The one-dimensional convolutional layers use a kernel size of 5, a stride of 1 or 2, and 200, 400, 500, and 550 channels, respectively. The max-pooling layers have a kernel size of 2 and a stride of 2. The extracted features are flattened and fed into a temperature-controlled bidirectional long short-term memory module. Humidity bidirectional long and short term memory module In the middle, the temperature bidirectional long short-term memory module Humidity bidirectional long and short term memory module Each direction has 1000 hidden units, and the output is mapped through a fully connected layer to 37 vertical layers to obtain the temperature and humidity profile. All experiments were performed using PyTorch. The model was trained using a stochastic gradient descent optimizer with an initial learning rate of... The momentum is 0.9. To enhance generalization and promote sparsity of low-level feature extraction, an L1 regularization term was selectively applied to the first convolutional layer. Specifically, two parameter sets were defined in the optimizer: the convolutional layer weights had a weight decay of 0.01, while all other parameters had no weight decay. In this experiment, the bidirectional long short-term memory network model Phy-BiLSTM in this embodiment was comprehensively evaluated on the test set to assess its overall inversion accuracy. To put these results in context, the performance of the bidirectional long short-term memory network model Phy-BiLSTM in this embodiment was rigorously compared with two established pure data-driven baseline models: (1) Basic convolutional neural network model (CNN): This is a multi-task joint inversion framework based on a one-dimensional convolutional neural network. It shares a feature extraction backbone network to simultaneously invert temperature and humidity profiles from input radiation. Although this joint structure allows for the potential learning of some inter-variable relationships from the data, it remains a pure data-driven black box, lacking any explicit thermodynamic constraints or laws to guide and regularize its solutions. (2) Deep Retrieval Architecture (DReA): This model represents a single-task, purely data-driven approach. It relies entirely on a one-dimensional convolutional neural network for spectral feature extraction and profile inversion. Crucially, the temperature and humidity profiles are inverted in two completely independent and unrelated models. This architecture fails to capture the inherent thermodynamic coupling between temperature and humidity variables and lacks any embedded physical knowledge. To quantitatively evaluate the performance of the model in retrieving temperature and humidity profiles, this experiment uses radiosonde observations as a reference standard and employs the root mean square error (RMSE) and mean absolute error (MAE) as core metrics. The definitions of RMSE and MAE are as follows: ; ; in, and These represent the predicted temperature and the actual temperature of the nth sample and the lth layer, respectively. The values ​​represent the predicted and actual humidity, respectively. Calculations were performed for all N samples and all L vertical layers for each profile. The final comprehensive inversion performance evaluation results are shown in Table 1.

[0044]

[0045] As shown in Table 1, the bidirectional long short-term memory network model Phy-BiLSTM in this embodiment consistently provides the most accurate temperature and humidity profile inversion results across all metrics, demonstrating its significant advantages over purely data-driven methods and other deep learning-based baselines. For temperature inversion, the Phy-BiLSTM model in this embodiment achieves a root mean square error of 0.85 K and a mean absolute error of 0.58 K, outperforming joint inversion models based on convolutional neural networks and single-task DReA benchmarks. For humidity inversion (which is typically more challenging due to the nonlinear behavior of water vapor), the Phy-BiLSTM model in this embodiment also exhibits superior performance, achieving a root mean square error of 1.00 g / kg and a mean absolute error of 0.69 g / kg. These improvements not only highlight the effectiveness of explicitly modeling vertical dependencies but also emphasize the significant contribution of embedded thermodynamic constraints, which greatly enhances inversion accuracy and physical realism.

[0046] To further evaluate the inversion performance of the bidirectional long short-term memory network model Phy-BiLSTM and the CNN and DReA models in this embodiment, the predicted values ​​were compared with the co-located radiosonde observations in the entire test dataset. The scatter plots between the temperature inversion results and the radiosonde measurements, as well as the linear relationship between the water vapor mixing ratio inversion results and the radiosonde measurements, were calculated. Figure 2 , Figure 3 and Figure 4 This is a scatter plot showing the relationship between the temperature inversion results and radiosonde measurements in this embodiment. It is the coefficient of determination. It is the total number of samples. Figure 2 This is a scatter plot of the CNN temperature inversion results. Figure 3 A scatter plot of DReA. Figure 4 This is a scatter plot of the Phy-BiLSTM bidirectional long short-term memory network model in this embodiment. Figure 5 , Figure 6 and Figure 7 This is a scatter plot showing the relationship between the water vapor mixing ratio inversion results and radiosonde measurements in this embodiment. Figure 5 This is a scatter plot of the CNN water vapor mixing ratio inversion results. Figure 6 A scatter plot of DReA. Figure 7This is a scatter plot of the Phy-BiLSTM bidirectional long short-term memory network model in this embodiment. The results show that all models exhibit a strong linear correlation with the reference measurements, with CNN having a coefficient of determination of 0.986, DReA of 0.986, and Phy-BiLSTM of 0.990. For the linear relationship between the water vapor mixing ratio inversion results and radiosonde measurements, the coefficient of determination for CNN is 0.940, DReA is 0.944, and the Phy-BiLSTM model in this embodiment is 0.949. Similarly, the Phy-BiLSTM model in this embodiment achieves the strongest correlation. Notably, compared to purely data-driven models, it more effectively suppresses extreme biases under high temperature and humidity conditions, highlighting the benefits of incorporating thermodynamic constraints to avoid non-physical interpretations.

[0047] To quantify the systematic biases present in the Phy-BiLSTM bidirectional long short-term memory network model, CNN, and DReA models in this embodiment. Figure 8 and Figure 9 The diagram shows the deviation distribution of the three models in temperature and relative humidity inversion in this embodiment of the invention. Figure 8 The deviation distribution in temperature inversion, Figure 9 This represents the bias distribution in the relative humidity inversion. Figure 8 and Figure 9 The results clearly demonstrate that the Phy-BiLSTM bidirectional long short-term memory network model in this embodiment exhibits the smallest systematic bias among all methods. For temperature, the median bias of the Phy-BiLSTM model in this embodiment is -0.613K, slightly underestimating the true value, comparable to convolutional neural networks. In contrast, the DReA model shows a significant negative bias of -5.974K, indicating a severe underestimation. For humidity, the Phy-BiLSTM model in this embodiment again performs best, with a median bias of -0.101%, closest to zero. The median biases of convolutional neural networks and the DReA model are 0.115% and 1.455%, respectively, reflecting slight and significant overestimation. Overall, the bias analysis shows that the Phy-BiLSTM model in this embodiment significantly reduces systematic errors (especially for humidity) while maintaining temperature accuracy comparable to CNNs, highlighting the benefits of incorporating physical constraints into the Phy-BiLSTM model in this embodiment.

[0048] To evaluate the contribution of each model component, ablation experiments were conducted by selectively removing physical constraints and bidirectional long short-term memory layers. The results of the ablation experiments are shown in Table 2.

[0049]

[0050] Among them, "w / o physical constraints" refers to removing physical constraints (physical loss). The model, "w / o BiLSTM", removes the bidirectional long short-term memory module (temperature bidirectional long short-term memory module). Humidity bidirectional long and short term memory module The model was analyzed. Table 2 shows that removing physical constraints significantly degraded the temperature and humidity inversion performance, increasing the root mean square error by 14% and 7%, respectively. This confirms that without thermodynamic constraints, the model struggles to maintain the physical coherence between temperature, humidity, and pressure. Removing the bidirectional long short-term memory component also led to performance degradation. These results highlight the importance of the bidirectional long short-term memory mechanism in capturing vertical coupling and interlayer interactions.

[0051] Figure 10 and Figure 11 Box plots showing the temperature and relative humidity deviation distributions obtained from different models in this embodiment of the invention. Figure 10 This is a box plot showing the temperature deviation distribution. Figure 11 Box plot of relative humidity deviation distribution. Figure 10 and Figure 11 The results further demonstrate that only the complete Phy-BiLSTM network model in this embodiment can simultaneously achieve low error, reduced variance, and physically consistent vertical structure, highlighting the complementary role of physical constraints and the bidirectional long short-term memory architecture. The combination of physical constraints and the bidirectional long short-term memory architecture provides a mutually reinforcing mechanism: physical constraints enforce adherence to fundamental thermodynamic relationships, while bidirectional long short-term memory captures the inherent complex vertical dependencies of atmospheric profiles. This dual mechanism enables the Phy-BiLSTM network model in this embodiment to achieve performance levels that no single component can achieve independently, highlighting its value in model optimization and improving the representation of atmospheric processes.

[0052] Those skilled in the art will understand that the technical solutions provided by this invention can take the form of a method, a system, or a computer program product. For example, this invention can provide a system for inverting atmospheric temperature and humidity profiles based on ground-based infrared hyperspectral radiation, including a microprocessor and a memory interconnected, wherein the microprocessor is programmed or configured to execute the method for inverting atmospheric temperature and humidity profiles based on ground-based infrared hyperspectral radiation. This invention can provide a computer-readable storage medium storing a computer program or instructions programmed or configured to execute the method for inverting atmospheric temperature and humidity profiles based on ground-based infrared hyperspectral radiation via a processor. This invention can provide a computer program product including a computer program or instructions programmed or configured to execute the method for inverting atmospheric temperature and humidity profiles based on ground-based infrared hyperspectral radiation via a processor. Furthermore, this invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, this invention can take the form of a computer program product implemented on one or more computer-readable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 The computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The functions specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable apparatus for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0053] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for inverting atmospheric temperature and humidity profiles based on ground-based infrared hyperspectral radiation, characterized in that, This includes inputting the hyperspectral radiance vector at any nth time step. The pre-trained bidirectional long short-term memory network model Phy-BiLSTM is used to obtain the predicted value of the temperature profile at the nth time step. Predicted values ​​of humidity profile at the nth time step The bidirectional long short-term memory network model Phy-BiLSTM includes: a feature extraction module. , used to input hyperspectral radiation vector Feature extraction Temperature bidirectional long and short term memory module Used for features Capturing the bidirectional dependence of temperature within the vertical atmospheric structure to obtain the bidirectional hidden state of the temperature sequence. Humidity bidirectional long and short term memory module Used for features Capturing the bidirectional dependence of humidity within the vertical atmospheric structure to obtain the bidirectional hidden state of the humidity sequence. Temperature regression module Used for bidirectional hidden states of temperature sequences. Regression is performed to obtain the predicted value of the temperature profile at the nth time step. Humidity regression module Used for bidirectional hiding of humidity sequences. Regression is performed to obtain the predicted value of the humidity profile at the nth time step. .

2. The atmospheric temperature and humidity profile inversion method based on ground-based infrared hyperspectral radiation according to claim 1, characterized in that, The hyperspectral radiance vector at the nth time step The function expression is: ； in, ~ These represent the hyperspectral radiance values ​​of the 1st to Cth spectral channels, where C is the number of spectral channels in the hyperspectral radiance value, and the superscript T indicates the transpose operation; the predicted value of the temperature profile at the nth time step. Predicted values ​​of humidity profile at the nth time step The function expression is: ； ； in, ~ These are the predicted temperatures for the first to Lth vertical atmospheric layers. ~ These are the predicted humidity values ​​for the 1st to Lth vertical atmospheric layers, where L is the number of vertical atmospheric layers.

3. The atmospheric temperature and humidity profile inversion method based on ground-based infrared hyperspectral radiation according to claim 1, characterized in that, The feature extraction module It includes multiple stacked sets of one-dimensional convolutional layers and max pooling layers. Each one-dimensional convolutional layer is followed by batch normalization and ReLU activation to capture local spectral dependencies and extract noise-resistant features from the hyperspectral radiative input. The max pooling layer is used to reduce spectral redundancy and preserve radiative features.

4. The atmospheric temperature and humidity profile inversion method based on ground-based infrared hyperspectral radiation according to claim 1, characterized in that, The temperature bidirectional long short-term memory module Humidity bidirectional long and short term memory module All are bidirectional long short-term memory networks, and the function expressions for forward and backward propagation of the bidirectional long short-term memory network are as follows: ； ； in, For branches The hidden state of the k-th vertical atmospheric layer obtained from the forward propagation. For branches Forward-passing long short-term memory network, For branches The hidden state of the (k-1)th vertical atmospheric layer obtained from the forward propagation. This represents the input feature for the k-th vertical atmospheric layer. For branches The hidden state of the k-th vertical atmospheric layer obtained by backpropagation For branches Backpropagation Long Short-Term Memory Network For branches The hidden state of the (k-1)th vertical atmospheric layer obtained from the backpropagation, branch Temperature bidirectional long short-term memory module Corresponding temperature branch or humidity bidirectional long short-term memory module For the corresponding humidity branch, the bidirectional hidden state output by the bidirectional long short-term memory network for any k-th vertical atmospheric layer is: ； in, For branches The bidirectional hidden state output by the kth vertical atmospheric layer For splicing operations, The number of Long Short-Term Memory networks in each direction for forward and backward propagation; and the bidirectional hidden states of all vertical atmospheric outputs. Bidirectional hidden state of spliced ​​output temperature sequence Or the bidirectional hidden state of the humidity sequence. .

5. The atmospheric temperature and humidity profile inversion method based on ground-based infrared hyperspectral radiation according to claim 1, characterized in that, The temperature regression module Humidity regression module Each layer includes two fully connected layers and one output layer, and the output layer outputs the predicted value of the temperature profile at the nth time step. Or the predicted value of the humidity profile at the nth time step. .

6. The atmospheric temperature and humidity profile inversion method based on ground-based infrared hyperspectral radiation according to claim 1, characterized in that, The loss function used during training of the pre-trained bidirectional long short-term memory network model Phy-BiLSTM is expressed as follows: ； in, For loss function, For temperature loss, For humidity loss, The loss is physical, and we have: ； ； ； ； ； ； ； ； in, and For balance coefficient, For water vapor mixing ratio loss, and Let the predicted water vapor mixing ratio and the reference water vapor mixing ratio be the values ​​at the nth time step. For element-wise multiplication, Let be the vertical atmospheric pressure vector at the nth time step. This is the actual water vapor pressure. The predicted value at the nth time step of the temperature profile. The corresponding saturated vapor pressure, This is the vapor pressure of water at the triple point. and These are the fitting coefficients. Let be the temperature ratio vector, which represents the ratio of temperature to the triple point temperature. This is the triple point temperature of water; For potential temperature loss, and These are the predicted potential temperature vector and the reference potential temperature vector at the nth time step, respectively. For reference pressure constant, is Poisson's constant.

7. The atmospheric temperature and humidity profile inversion method based on ground-based infrared hyperspectral radiation according to claim 6, characterized in that, The expression for the calculation function of the temperature loss is: ； The function expression for calculating the humidity loss is: ； in, For temperature loss, For humidity loss, Where L is the total number of time steps, and L is the number of vertical atmospheric divisions. and The first The first time step Predicted temperature values ​​and temperature label values ​​for each vertical atmospheric layer. and The first The first time step Predicted humidity values ​​and humidity label values ​​for each vertical atmospheric layer.

8. A system for inverting atmospheric temperature and humidity profiles based on ground-based infrared hyperspectral radiation, comprising a microprocessor and a memory interconnected, characterized in that, The microprocessor is programmed or configured to execute the atmospheric temperature and humidity profile inversion method based on ground-based infrared hyperspectral radiation as described in any one of claims 1 to 7.

Images