Method for dynamically correcting bottom refractive index of occultation based on deep learning

Abstract

The invention provides a deep learning-based occultation bottom refractive index dynamic correction method. The method comprises the following steps of preprocessing input data through a data preprocessing module; constructing a neural network model; constructing a mixed loss function training neural network model; and adjusting a neural network model training strategy through an optimization module. The invention has the beneficial effects that the refractive index root mean square error (RMSE) is reduced from 0.185 to 0.041 in a 0-2km height layer; through the differentiable physical constraint layer, the second derivative smoothness of the refractive index profile is improved by 63.2%, and false oscillation caused by a traditional method is eliminated.

Description

Technical Field

[0001] The present invention belongs to the field of occultation bottom refractive index correction, and particularly relates to a method for dynamically correcting the occultation bottom refractive index based on deep learning. Background Art

[0002] Occultation detection retrieves atmospheric parameters by receiving the refraction effect when satellite navigation signals pass through the atmosphere, and is an important means to obtain global high vertical resolution atmospheric profile data. However, in the bottom region of 0 - 8 km (especially 0 - 2 km) near the ground, affected by multipath effects, atmospheric turbulence and surface reflection interference, there are significant systematic biases in the original refractive index data, seriously affecting the humidity retrieval accuracy, that is, the traditional empirical model cannot adaptively correct the non-linear refractive index bias under complex meteorological conditions, resulting in the diffusion of inversion errors in the bottom region; Currently, the mainstream correction methods include: Statistical empirical correction: relying on historical data to establish empirical formulas, but unable to adapt to non-linear biases under complex meteorological conditions; Variational assimilation iterative optimization: performing data assimilation through a numerical weather prediction model, and there is physical inconsistency between the assimilation result and the original data; The existing methods fail to effectively utilize the residual characteristics between the assimilated refractive index and the original data, and it is difficult to balance between computational efficiency and physical consistency. Summary of the Invention

[0003] In view of this, the present invention aims to propose a method for dynamically correcting the occultation bottom refractive index based on deep learning to solve the problem that the traditional empirical model cannot adaptively correct the non-linear refractive index bias under complex meteorological conditions, resulting in the diffusion of inversion errors in the bottom region.

[0004] To achieve the above object, the technical solution of the present invention is realized as follows: A method for dynamically correcting the occultation bottom refractive index based on deep learning, comprising the following steps: S1. Preprocess the input data through a data preprocessing module; S2. Construct a neural network model; S3. Construct a hybrid loss function to train the neural network model; S4. Adjust the training strategy of the neural network model through an optimization module.

[0005] Further, in step S1, the input data includes: Input the original occultation profile containing the refractive index; Input the data of the assimilated refractive index; Further, in step S1, the input data preprocessing includes: Set the data altitude layer: The bottom area is defined as 0 - 8 km; Interpolate the same altitude resolution through cubic splines: The specified resolution is 100 m; Construct the residual features: Within 100 m, subtract the original occultation profile from the assimilated refractive index data.

[0006] Furthermore, in step S2, the neural network model adopts the Hybrid CNN - LSTM model.

[0007] Furthermore, in step S3, the hybrid loss function is: (1); Among them, the MAE term is the mean absolute error, which is a loss function used to measure the difference between the predicted value and the true value; Among them, the MAE term is: (2); Among them, N is the total number of samples, that is, the number of data points considered when calculating the mean absolute error, n pred is the refractive index predicted by the model, which is the refractive index value calculated by the model according to the input, h i is the input variable of the model, n era is the refractive index obtained through the background field; L smooth is the smoothing constraint. The smoothing constraint ensures the smoothness of the refractive index profile by minimizing the sum of the squares of the second derivatives; The second derivative used is related to the curvature, and the smoothing constraint is: (3); Among them, H represents the total number of altitude layers, used to sum over all altitude layers, h is the altitude, representing the variable of different altitude layers in the atmosphere, n is the refractive index, which is the predicted output of the deep learning model and also the objective function of the smoothing constraint. The symbol d represents the differential symbol in the derivative, and the first derivative dn / dh represents the rate of change of the refractive index with altitude; The second derivative d 2 n / dh 2 represents the rate of change of the refractive index gradient; L phy is the physical constraint. The physical constraint is an indicator function, aiming to ensure that the refractive index n is not less than 1. Among them, the physical constraint is: (4); Among them, H represents the total number of altitude levels and is used to sum over all altitude levels. h is the altitude, representing the variable at different altitude levels in the atmosphere. n pred is the refractive index predicted by the model, which is the refractive index value calculated by the model based on the input. Ⅱ is the indicator function. Design the weight coefficient: According to the design α , β , γ of the numerical value, determine the effect verification.

[0008] Further, in step S4, adjust the training strategy through the optimization module, including: Select NAdam as the optimizer, learning rate scheduling: Cosine annealing + restart (CyclicLR); If an early stopping mechanism is involved, then when the validation set loss does not decrease for 5 consecutive times, terminate.

[0009] Compared with the prior art, the method for dynamically correcting the refractive index at the bottom of occultation based on deep learning according to the present invention has the following advantages: For the method for dynamically correcting the refractive index at the bottom of occultation based on deep learning according to the present invention, in the altitude range of 0 - 2 km, the root mean square error (RMSE) of the refractive index is reduced from 0.185 to 0.041; through the differentiable physical constraint layer, the smoothness of the second derivative of the refractive index profile is increased by 63.2%, eliminating the false oscillations caused by the traditional method. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] The drawings forming a part of the present invention are used to provide a further understanding of the present invention. The schematic embodiments and descriptions thereof of the present invention are used to explain the present invention and do not constitute an improper limitation to the present invention. In the drawings: Figure 1 is the schematic flow chart of the method described in the embodiment of the present invention; Figure 2 is the schematic data comparison diagram described in the embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0011] It should be noted that, without conflict, the embodiments in the present invention and the features in the embodiments can be combined with each other.

[0012] In the description of the present invention, it should be understood that the orientation or positional relationships indicated by the terms "center", "longitudinal", "transverse", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc. are based on the orientation or positional relationships shown in the drawings. These are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of the present invention. In addition, the terms "first", "second", etc. are only used for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, features defined with "first", "second", etc. may explicitly or implicitly include one or more of such features. In the description of the present invention, unless otherwise stated, the meaning of "a plurality" is two or more.

[0013] In the description of the present invention, it should be noted that unless otherwise clearly specified and limited, the terms "mounted", "connected", "coupled" should be understood in a broad sense. For example, it may be a fixed connection, a detachable connection, or an integral connection; it may be a mechanical connection or an electrical connection; it may be directly connected or indirectly connected through an intermediate medium, and it may be the communication inside two elements. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood through specific circumstances.

[0014] The present invention will be described in detail below with reference to the drawings and in conjunction with embodiments.

[0015] As Figure 1 shown, a method for dynamically correcting the refractive index at the bottom of occultation based on deep learning includes the following steps: S1. Data preprocessing module Input data: Input the original occultation profile containing the refractive index; Input the assimilated refractive index data; At the same time, set the height layer: the bottom area is defined as 0 - 8 km.

[0016] Through cubic spline interpolation, obtain the same height resolution: the specified resolution is 100 meters.

[0017] Construct a residual feature: that is, within 100 meters, subtract the original occultation profile from the assimilated refractive index data.

[0018] Construct some auxiliary features: which can give more conditions for determination.

[0019] In this embodiment, the auxiliary features may include: Height stratification coding: Divide 0 - 8 km into three layers (0 - 2 km, 2 - 5 km, 5 - 8 km), and each layer is independently normalized.

[0020] Temporal characteristics: Expand the data with a 10 - second time series as the temporal sequence.

[0021] Input variables: Such as temperature, air pressure, etc.

[0022] By defining the above - mentioned auxiliary features, more conditions can be given for judgment.

[0023] S2. Neural network architecture The Hybrid CNN - LSTM model is used here.

[0024] S3. Hybrid loss function (1); The MAE term (data precision constraint) is the mean absolute error, which is a common loss function used to measure the difference between the predicted value and the true value. Compared with the mean square error (MSE), MAE is more robust to outliers and can avoid the excessive influence of noisy data on model training.

[0025] Among them, the MAE term is: (2); L smooth The (smoothing constraint) is to ensure the smoothness of the refractive index profile by minimizing the sum of the squares of the second derivatives. The second derivative used is related to the curvature to avoid unnecessary fluctuations in the profile.

[0026] Among them, the smoothing constraint is: (3); L phy The (physical constraint) is an indicator function, aiming to ensure that the refractive index n is not less than 1. Because according to physical laws, the refractive index is 1 in a vacuum and usually slightly greater than 1 in the atmosphere. If the refractive index predicted by the model is less than 1, this constraint term will impose a penalty to ensure the physical rationality of the result. The symbol d represents the differential symbol in the derivative, and the first - order derivative dn / dh represents the rate of change of the refractive index with height; the second - order derivative d 2 n / dh 2 represents the rate of change of the refractive index gradient.

[0027] Among them, the physical constraint is: (4); Among them, H represents the total number of height layers, used to sum over all height layers.h is the altitude, representing the variable at different altitude levels in the atmosphere, n pred is the refractive index predicted by the model, which is the refractive index value calculated by the model based on the input. Ⅱ is the indicator function.

[0028] Design the weight coefficient: According to the design α , β , γ determine the final effect verification according to the numerical value.

[0029] S4. Optimization module The optimization mainly adjusts the training strategy, including that the optimizer may select NAdam (lr = 3e - 4, weight_decay = 1e - 5), and the learning rate scheduling: Cosine annealing + restart (CyclicLR).

[0030] It also involves an early stopping mechanism, which terminates when the validation set loss does not decrease for 5 consecutive times.

[0031] In view of the problem of large inversion error of the refractive index at the bottom of the occultation, the present invention proposes a deep - learning - based method to specifically solve the following technical pain points: The traditional empirical model cannot adaptively correct the non - linear refractive index deviation under complex meteorological conditions, resulting in the diffusion of the inversion error in the bottom area.

[0032] The present invention realizes the correction of the refractive index at the bottom of the occultation by integrating deep learning and data assimilation technology, and has the following significant advantages: In the altitude layer of 0 - 2 km, the root - mean - square error (RMSE) of the refractive index is reduced from 0.185 to 0.041; Through the differentiable physical constraint layer, the smoothness of the second - order derivative of the refractive index profile is increased by 63.2%, eliminating the false oscillation caused by the traditional method.

[0033] Example 1 1. Data construction: The data used are 10,000 data of the cloud - remote aerospace assimilated data and the occultation profile data before assimilation. Gaussian noise σ = 0.05 is added to generate a systematic deviation, and random pulse interference is added. After pre - processing, 10,000 assimilated independent profile data and 10,000 original occultation profile data are obtained.

[0034] 2. Model training, including: Residual feature construction: Calculate the residual between the assimilated data and the original data (Δn = n_assim - n_raw) as the key input; Altitude - layer encoding: Divide 0 - 8 km into three layers (0 - 2 km, 2 - 5 km, 5 - 8 km), and each layer is independently normalized; Temporal alignment: With a 10 - second time window, expand a single profile into a time series (length = 5 steps, step size = 2 seconds).

[0035] Dataset division: Among them, training set: 8,000 (80%) for both assimilation and original data; validation set: 1,000 (10%) for each; test set: 1,000 (10%) for each, strictly isolating the training process.

[0036] Hybrid CNN - LSTM model architecture: The network structure of the Hybrid CNN - LSTM model includes: Input layer: Single - sample dimension [number of time steps = 5, number of height layers = 3, number of features = 4] (features: original n, Δn, temperature, pressure); CNN module (spatial feature extraction): 1D convolutional layer × 2 (kernel = 3, channels = 64 → 128), sliding along the height dimension to extract local vertical correlations; Global average pooling to compress the spatial dimension, outputting [number of time steps, 128]; LSTM module (temporal modeling): Bidirectional LSTM (hidden units = 256), capturing the forward / backward atmospheric state evolution; Outputting [number of time steps, 512] temporal features; Physical constraint branch: Residual correction layer: A fully - connected network predicts Δn_correction, which is superimposed on the original data (n_corrected = n_raw + Δn_correction); Equation constraint layer: Enforcing the refractive index physical formula through a hybrid loss function.

[0037] The Hybrid CNN - LSTM model can achieve lightweight: Depth - wise separable convolution: Replacing the standard convolution, reducing the number of parameters by 72%; GRU replacing LSTM: With a precision loss < 1% on the validation set, the inference speed is increased by 35%.

[0038] 3. Validation results:

[0039] 4. Data comparison, as Figure 2 shown in the data comparison graph.

[0040] This method can achieve the upgrade of the global meteorological assimilation system: input the corrected data into numerical models such as ECMWF / CMA to improve the accuracy of the near-surface humidity field; This method can achieve extreme weather monitoring: optimize the impulse noise suppression module for typhoon and severe convection scenarios to enhance the robustness of the model; This method can achieve inter-satellite collaborative learning: use multi-satellite occultation data to construct a federated learning framework to break through the limitations of single-satellite data.

[0041] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for dynamically correcting the refractive index at the bottom of occultation based on deep learning, characterized in that: It includes the following steps: S1. Preprocess the input data through a data preprocessing module; S2. Build a neural network model; S3. Build a hybrid loss function to train the neural network model; S4. Adjust the training strategy of the neural network model through an optimization module; In step S3, the hybrid loss function is: （1）； Among them, L is the hybrid loss function, α , β , γ are the weight coefficients respectively. The MAE term is the mean absolute error, which is a loss function used to measure the difference between the predicted value and the true value. L smooth is a smoothing constraint. The smoothing constraint ensures the smoothness of the refractive index profile by minimizing the sum of the squares of the second derivatives; the second derivative used is related to the curvature; L phy is a physical constraint, and the physical constraint is an indicator function, aiming to ensure that the refractive index n is not lower than 1. Design weight coefficient: According to the design α , β , γ values, determine the effect verification; In step S4, adjusting the training strategy through the optimization module includes: Select NAdam as the optimizer and use Cosine annealing with restart for the learning rate scheduling; If an early stopping mechanism is involved, terminate when the validation set loss does not decrease for 5 consecutive times.

2. The method for dynamically correcting the bottom refractive index of occultation based on deep learning according to claim 1, wherein: In step S1, the input data includes: Input the original occultation profile containing the refractive index; Input the assimilated refractive index data.

3. A method for dynamically correcting the bottom refractive index of occultation based on deep learning according to claim 1, characterized in that: In step S1, the input data preprocessing includes: Set the data height layer: the bottom area is defined as 0 - 8 km; Interpolate to the same height resolution through cubic spline interpolation: the specified resolution is 100 meters; Construct the residual feature: within 100 meters, subtract the original occultation profile from the assimilated refractive index data.

4. A method for dynamically correcting the bottom refractive index of occultation based on deep learning according to claim 1, characterized in that: In step S2, the neural network model adopts a Hybrid CNN - LSTM model.

5. A method for dynamically correcting the bottom refractive index of occultation based on deep learning according to claim 1, characterized in that: In step S3, the MAE term is: （2）； Among them, N is the total number of samples, that is, the number of data points considered when calculating the mean absolute error, n pred is the refractive index predicted by the model, which is the refractive index value calculated by the model according to the input, h i is the input variable of the model, n era is the refractive index obtained through the background field.

6. A method for dynamically correcting the bottom refractive index of occultation based on deep learning according to claim 1, characterized in that: In step S3, the smooth constraint is: （3）； Among them, H represents the total number of altitude levels and is used to sum over all altitude levels, h is the altitude, representing the variable of different altitude levels in the atmosphere, n is the refractive index, which is the predicted output of the deep learning model and also the objective function of the smoothing constraint. The symbol d represents the differential symbol in the derivative, and the first derivative dn / dh represents the rate of change of the refractive index with altitude; the second derivative d 2 n / dh 2 represents the rate of change of the refractive index gradient.

7. A method for dynamically correcting the bottom refractive index of occultation based on deep learning according to claim 1, characterized in that: In step S3, the physical constraint is: （4）； Among them, H represents the total number of altitude levels and is used to sum over all altitude levels, h is the altitude, representing the variable of different altitude levels in the atmosphere, n pred is the refractive index predicted by the model, which is the refractive index value calculated by the model based on the input, and Ⅱ is the indicator function.

Images