Optical turbulence transferable feature construction method under heterogeneous observation condition

Abstract

The invention discloses an optical turbulence transferable feature construction method under heterogeneous observation conditions, and belongs to the technical field of atmospheric optics, boundary layer meteorology and data driving prediction. The method comprises the following steps: firstly, preprocessing heterogeneous observation data to obtain an effective observation sequence meeting a construction condition; secondly, introducing a unified reference height mapping item and a stability bounded correction item into a Monin-Obukhov similarity theory framework, constraining a measurement height item and the stability correction item, and constructing an optical turbulence characterization quantity; then isomorphic re-parameterization is carried out on the effective observation sequence according to constituent variables and a function relation of the optical turbulence characterization quantity, and transferable features with unified physical significance are generated; and finally, carrying out time sequence expansion on the transferable features to form a transferable feature set to be called by a future optical turbulence characterization quantity prediction model of the target scene. According to the method, feature space offset among different observation conditions can be reduced, and feature physical interpretability and migration consistency are improved.

Description

Technical Field

[0001] This invention belongs to the fields of atmospheric optics, boundary layer meteorology, and data-driven prediction technology, and particularly relates to a method for constructing transferable features of optical turbulence under heterogeneous observation conditions. More specifically, it relates to a method for constructing transferable features suitable for optical turbulence migration prediction by using a scenario-based physical reparameterization of near-surface observations under heterogeneous observation conditions based on the Monin-Obukhov similarity theory and introducing a unified reference height mapping term and a stability bounded correction term. Background Technology

[0002] Near-surface optical turbulence can affect the design of free-space optical communication links, the operation and scheduling of optical sites, the deployment of ground-based optical systems, and phase screen simulation. In newly established observation scenarios or scenarios with a short observation history, directly obtaining continuous, long-term, and in-situ optical turbulence measurement data is often costly and time-consuming. Therefore, there is a real need to use existing observation information from other observation scenarios to make short-term predictions of target scenarios.

[0003] In existing technologies, prediction methods for optical turbulence or seeing are mostly based on training in a single observation scenario or directly using raw meteorological observations as model input. Because observation scenarios differ significantly under different conditions of observation, measurement altitude, and stability distribution, the raw input features often exhibit strong distribution shifts when transferred to other application scenarios. This makes the model more likely to learn specific observation conditions rather than transferable patterns, leading to unstable prediction performance for the target scenario.

[0004] On the other hand, ordinary scale normalization methods typically only unify the dimensions or numerical ranges of variables. While this can improve numerical scale differences to some extent, it is difficult to reflect the physical constraints in the formation process of near-surface optical turbulence, especially in organizing air pressure, temperature, friction velocity, turbulent heat flux terms, and stability ratios into the same feature space according to a unified physical structure. Therefore, ordinary normalization cannot fundamentally solve the problem of inconsistent input representations under heterogeneous observation conditions.

[0005] Therefore, there is an urgent need for a feature construction method that can be applied to heterogeneous observation conditions, transform near-surface observations into a unified physical meaning space, and directly serve subsequent optical turbulence prediction models, in order to solve the problems of insufficient transferability, inconsistent input representation, and weak physical interpretability in existing technologies. Summary of the Invention

[0006] To address the aforementioned issues, this invention proposes a method for constructing transferable features of optical turbulence under heterogeneous observation conditions. The aim is to map near-surface original observations under different observation conditions to a unified physical meaning space by first constructing optical turbulence characterization quantities and then organizing input features based on the constituent variables and functional relationships of the characterization quantities, thereby forming a set of transferable features suitable for migration prediction.

[0007] The above objectives are achieved through the following technical solutions:

[0008] This invention provides a method for constructing transferable features of optical turbulence under heterogeneous observation conditions, the method comprising the following steps:

[0009] S1. Input near-surface observation data from multiple observation sources within a preset time range. The observation data includes air pressure, air temperature, friction velocity, turbulent heat flux, stability ratio, and measurement height. Then, perform time unification and field alignment on the observation data from each observation source to form a unified original dataset.

[0010] S2. Perform quality control on the unified original dataset formed in step 1, remove missing values, non-physical values ​​and samples with incomplete key fields, so that the remaining samples meet the conditions for constructing subsequent optical turbulence characterization quantities, thereby obtaining an effective observation sequence.

[0011] S3. Based on the effective observation sequence obtained in step S2, a unified reference height mapping term and a stability bounded correction term are introduced, and the height mapping coefficient α, the strong instability segmentation threshold τ, and the bounded attenuation coefficient β are determined. The measurement height term and stability correction term in the Monin-Obukhov similarity theory are constrained across observation conditions and scenarios to construct the optical turbulence characterization quantity corresponding to each time moment.

[0012] S4. Based on the constituent variables and functional relationships of the optical turbulence characterization quantities in step 3, perform isomorphic reparameterization on the effective observation sequence in step 2, and convert the original observations into transferable features consistent with the structure of the optical turbulence characterization quantities, wherein the transferable features include a unified reference height logarithmic term and a stability bounded correction term.

[0013] S5. Expand the transferable features obtained in step 4 according to the current time and historical time to construct a set of transferable features containing time lag information.

[0014] S6. Output the set of transferable features constructed in step S5 as input to the prediction model of optical turbulence characterization of the target scene.

[0015] Furthermore, the near-surface observation data mentioned in step 1 comes from raw observation records collected by multiple observation sources within the same time range; the raw observation records include air pressure, air temperature, friction velocity, turbulent heat flux, stability ratio, and measurement altitude. In step 1, the observation data from each observation source are unified in terms of time, field alignment, and units to eliminate differences in timestamp format, field naming methods, and unit representation methods among different observation sources, thus obtaining a unified raw dataset.

[0016] Further, the quality control in step 2 includes: screening for missing values, non-physical values, key field integrity, and anomalous mutation values ​​in the unified raw dataset obtained in step 1; the key fields include air pressure, air temperature, friction velocity, turbulent heat flux, stability ratio, and measurement height; the anomalous mutation value screening is used to delete samples where the change in any key field between two adjacent sampling times exceeds three times the median of the adjacent change in that field; the samples retained after quality control constitute a valid observation sequence to ensure that the variables required for the subsequent construction of optical turbulence characterization are complete and have clear physical meaning.

[0017] Furthermore, the Monin-Obukhov similarity theory is a boundary layer similarity theory used to describe the relationship between momentum transport, heat transport, and atmospheric stability in the near-surface layer, where the original measured altitude is denoted as... The length of Monin-Obukhov is denoted as The stability ratio is denoted as This invention characterizes the atmospheric stability at observation altitudes. The object of this invention is near-surface observation data, and the air pressure, air temperature, friction velocity, turbulent heat flux, stability ratio, and measurement altitude used satisfy the variable conditions required to construct a near-surface optical turbulence characterization. In step 3, a unified reference altitude mapping term and a bounded stability correction term are introduced into the basic characterization form obtained based on the Monin-Obukhov similarity theory. Cross-observation source constraints are applied to the original measurement altitude term and stability correction term to suppress the divergence of the characterization caused by differences in measurement altitudes from different observation sources and strongly unstable stratification conditions.

[0018] Furthermore, without introducing a unified reference height mapping term and a stability bounded correction term, the fundamental characterization of optical turbulence based on the Monin-Obukhov similarity theory can be expressed as: Among them, the basic stability correction function Represented as: In the basic representation form, the measured height item is directly derived from the original measured height. The determination is that the bounded stability correction term is directly derived from the fundamental stability correction function. The decision is as follows. In heterogeneous observation scenarios, this approach directly incorporates the differences in measured heights under different observation conditions into the characterization calculation process, and causes the correction term under strongly unstable stratification conditions to continuously increase, which is detrimental to unified characterization among observation sources. Therefore, in step 3, this invention introduces a unified reference height mapping term and a stability bounded correction term. A unified mapping constraint is applied to the height term, and a piecewise bounded constraint is applied to the stability correction term, thereby forming an optical turbulence characterization quantity suitable for migration prediction.

[0019] Furthermore, the optical turbulence characterization quantity described in step 3 is constructed according to the following formula: The temperature scale term is calculated using the following formula: The height term following the unified reference height mapping is calculated using the following formula: And the stability bounded correction term is: in, For the observation time, As a characterization quantity for optical turbulence, The pressure is expressed in hectopascals (hPa). Absolute temperature in Kelvin. The original measured height. To unify the height item after reference height mapping, For friction speed, For turbulent heat flux, The length of Monin-Obukhov. For stability ratio, To standardize reference height, For height mapping coefficients, The threshold for segmentation in cases of strong instability. The coefficient is a bounded attenuation factor. This is the refractive index conversion factor. and Here, is the stability correction constant, and is the empirical similarity coefficient related to the Monin-Obukhov similarity theory. This invention uses these empirical similarity coefficients in the bounded stability correction term. Step 3, through the above construction, introduces a unified reference height mapping term on the basic characterization form obtained based on the Monin-Obukhov similarity theory, implementing a unified mapping constraint on the observation source differences caused by the original measured height; simultaneously, a bounded stability correction term is introduced to implement piecewise amplitude limiting constraints on the stability correction growth process in the highly unstable section, thereby forming an optical turbulence characterization quantity oriented towards migration prediction. The unified reference height... Take the lowest valid measurement height corresponding to all valid samples.

[0020] Furthermore, the application of the Monin-Obukhov similarity theory in step 3 is as follows: first, based on the friction velocity... and turbulent heat flux term Constructing temperature scale terms Then introduce a unified reference height mapping term. A unified constraint is applied to the original measured height term under different observation conditions; then a stability bounded correction term is introduced. The stability correction function under strongly unstable stratification conditions is piecewise limited; finally, the pressure term is incorporated. Temperature item Calculate optical turbulence characterization .

[0021] The height mapping coefficient α, the strong instability segmentation threshold τ, and the bounded attenuation coefficient β are determined as follows: First, a set of candidate parameter combinations is constructed, consisting of candidate values ​​for the height mapping coefficient, candidate values ​​for the strong instability segmentation threshold, and candidate values ​​for the bounded attenuation coefficient. Each candidate parameter combination is substituted into a unified reference height mapping term and a stability bounded correction term to construct candidate optical turbulence characterization quantities corresponding to different observation conditions. The maximum mean difference statistics of candidate optical turbulence characterization quantities between each pair of different observation conditions are then calculated. The average value of the ten maximum mean difference statistics is then used as the inter-source distribution difference index corresponding to the candidate parameter combination. Finally, the parameter combination that minimizes the inter-source distribution difference index is selected as the final parameter combination (α, τ, β).

[0022] Step 4 is based on the above. The constituent variables and functional relationships are used to perform isomorphic reparameterization of the original observations in order to generate transferable features.

[0023] Furthermore, the constituent variables and functional relationships mentioned in step 4 refer to the pressure term, temperature term, temperature scale term, unified reference height mapping term, and stability bounded correction term used in step 3 to construct the optical turbulence characterization quantity, as well as the power, ratio, logarithm, and correction function relationships between these terms. The isomorphic reparameterization mentioned in step 4 represents the process of converting the original observations into feature variables with the same constituent logic as the optical turbulence characterization quantity, according to the aforementioned constituent variables and functional relationships. The transferable features generated in step 4 consist of the logarithmic term of friction velocity. Temperature scale item Stability bounded correction term Refractive index pressure related items Unified reference height logarithmic items Temperature item It consists of time period terms.

[0024] The time period term consists of an annual sine term, an annual cosine term, a daily sine term, and a daily cosine term, which are respectively represented as: in, For a moment The day number in a year, For a moment The hour sequence of the day, The number of days in a year. The number of hours in a day. Step 4, through the isomorphic reparameterization, ensures that the input features and the optical turbulence characterization quantities constructed in Step 3 maintain consistency in terms of constituent variables and functional relationships, while preserving the unified constraint effect introduced by the unified reference height mapping and the bounded stability correction.

[0025] Furthermore, in step 5, the transferable features obtained in step 4 are expanded temporally in chronological order to construct a set of transferable features containing time lag information. The time lag information in step 5 consists of features from the current time, 15 minutes before the current time, 30 minutes before the current time, and 60 minutes before the current time, so that the constructed feature set contains both a unified physical meaning and the short-term historical information required for the evolution of near-surface optical turbulence.

[0026] Furthermore, in step 6, the transferable feature set formed in step 5 is used as input to the prediction model of the future optical turbulence characterization of the target scene, so as to output the prediction results of the optical turbulence characterization of the target scene at future times. The prediction target in step 6 is set as follows: .

[0027] The prediction model for the future optical turbulence characterization of the target scene mentioned in step 6 is a tree model, a linear model, a neural network model, or a sequence model.

[0028] The present invention has the following advantages over the prior art:

[0029] This invention introduces a unified reference height mapping term and a stability bounded correction term on the basic characterization form obtained based on the Monin-Obukhov similarity theory. This can simultaneously constrain the amplification problem of the characterization quantity caused by the difference in measurement height and the strong unstable layer, thereby forming an optical turbulence characterization quantity suitable for migration prediction.

[0030] This invention weakens the direct impact of measurement height differences under different observation conditions on optical turbulence characterization quantities by using a unified reference height mapping term, and suppresses the excessive amplification of characterization quantities under strongly unstable stratification conditions by using a stability bounded correction term.

[0031] This invention uses the improved optical turbulence characterization parameters to perform isomorphic reparameterization of input features based on the variables and functional relationships of the components, thereby mapping data under different observation conditions to a unified physical meaning space and reducing feature distribution offset.

[0032] The present invention outputs a unified feature set that can be directly used by subsequent prediction models. Therefore, it does not depend on a specific prediction model structure and has good versatility and scalability. Attached Figure Description

[0033] Figure 1 This is a flowchart illustrating the overall process of the method of the present invention.

[0034] Figure 2 This is a flowchart of the construction process for optical turbulence characterization parameters in this invention.

[0035] Figure 3 This is a flowchart of the transferable feature generation process in this invention.

[0036] Figure 4 This is a flowchart of the construction process of the cross-time transferable feature set in this invention.

[0037] Figure 5 This is a comparison chart of the differences in the distribution of input features.

[0038] Figure 6 This is a comparison chart of model prediction errors. Detailed Implementation

[0039] The present invention will be further described in detail below with reference to the accompanying drawings, the invention description, and experimental data. It should be understood that the following embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0040] This embodiment uses heterogeneous observation data from five near-surface observation sources as input data sources, with an observation time resolution of 15 minutes. Each observation source records air pressure, air temperature, friction velocity, turbulent heat flux, stability ratio, and measurement altitude. Figure 1 The overall flow of this embodiment is shown. Figures 2 to 4 The processes of constructing representation parameters, constructing transferable features, and constructing feature sets across time periods are illustrated respectively. Figure 5 and Figure 6 The comparison verification results of this embodiment are shown.

[0041] Step 1: Input near-surface observation data from multiple observation sources within a preset time range, and perform time unification and field alignment on the near-surface observation data to form a unified original dataset.

[0042] In this step, observation records from five observation sources within the same time range are read, and each record is standardized to the same timestamp format, field name, and unit system. Barometric pressure is standardized to hectopascals, air temperature to Kelvin, measurement altitude to meters, and timestamps to 15-minute sampling intervals. For each observation source, the sample record corresponding to the lowest available measurement altitude is retained to eliminate direct structural differences caused by different measurement altitudes. After time standardization and field alignment, a unified original dataset containing samples from all observation sources is formed.

[0043] Step 2: Perform quality control on the unified original dataset to obtain valid observation sequences.

[0044] In this step, the unified raw dataset formed in step 1 is processed sequentially to remove missing values, non-physical values, key field integrity screening, and anomalous value removal. Missing value removal deletes records with null values ​​in the fields of air pressure, air temperature, friction velocity, turbulent heat flux, stability ratio, and measurement altitude. Non-physical value removal deletes records with air pressure less than zero, absolute temperature less than or equal to zero, measurement altitude less than or equal to zero, and friction velocity less than or equal to zero. Key field integrity screening ensures that each record contains air pressure, air temperature, friction velocity, turbulent heat flux, stability ratio, and measurement altitude. Anomalous value removal deletes samples where the change in any key field between two adjacent sampling times exceeds three times the median change in that field. After completing these processes, a valid observation sequence that meets the conditions for constructing the characterization parameters is obtained.

[0045] Step 3: Based on the effective observation sequence obtained in step S2, a unified reference height mapping term and a stability bounded correction term are introduced, and the height mapping coefficient α, the strong instability segmentation threshold τ, and the bounded attenuation coefficient β are determined. The measurement height term and stability correction term in the Monin-Obukhov similarity theory are constrained across observation conditions and scenarios to construct the optical turbulence characterization quantity corresponding to each time moment.

[0046] In this step, based on the Monin-Obukhov similarity theory, at each observation time... Calculate the corresponding optical turbulence characterization parameters. The Monin-Obukhov similarity theory is used to describe the relationship between momentum transport, heat transport, and atmospheric stability in the near-surface layer, where the original measurement altitude is denoted as . The length of Monin-Obukhov is denoted as The stability ratio is denoted as This invention characterizes the atmospheric stability at the observation altitude. The object of this invention is near-surface observation data, and the air pressure, air temperature, friction velocity, turbulent heat flux term, stability ratio, and measurement altitude used satisfy the variable conditions required to construct a near-surface optical turbulence characterization. Without introducing a unified reference altitude mapping term and a bounded stability correction term, a fundamental characterization form of optical turbulence can be obtained based on the aforementioned similarity theory. This fundamental characterization form directly uses the original measurement altitude. Form the height term and directly use the basic stability correction function. A stability correction term is formed. However, this method, under heterogeneous observation conditions, directly incorporates differences in measurement height and highly unstable stratification conditions into the characterization quantity, leading to increased distribution offset and uncontrolled amplification of the characterization quantity. To avoid these problems, this embodiment introduces a unified reference height mapping term and a bounded stability correction term, imposing unified constraints on the height and stability correction terms to suppress the divergence of the characterization quantity caused by differences in measurement height from different observation sources and highly unstable stratification conditions. In this embodiment, the effective observation sequence obtained in step 2 already includes air pressure, air temperature, friction velocity, turbulent heat flux, stability ratio, and measurement height, thus satisfying the variable conditions for constructing an optical turbulence characterization quantity based on the aforementioned theory. The basic characterization form can be expressed as: The basic stability correction function is expressed as follows: Based on the above construction method, this embodiment introduces a unified reference height mapping constraint for the measurement height term and a segmented amplitude limiting constraint for the highly unstable section for the stability correction term. The constructed optical turbulence characterization quantity is expressed by the following formula: Among them, the temperature scale term Calculate using the following formula: Unified Reference Height Mapping Item Calculate using the following formula: Stability bounded correction term Calculate using the following formula: in, For the observation time, As a characterization quantity for optical turbulence, The pressure is expressed in hectopascals (hPa). Absolute temperature in Kelvin. The original measured height. To unify the height item after reference height mapping, For friction speed, For turbulent heat flux, The length of Monin-Obukhov. For stability ratio, To standardize reference height, For height mapping coefficients, The threshold for segmentation in cases of strong instability. The coefficient is a bounded attenuation factor. This is the refractive index conversion factor. and The stability correction constant is a relevant empirical similarity coefficient based on the Monin-Obukhov similarity theory. This invention uses this empirical similarity coefficient in the bounded stability correction term. In this embodiment, a unified reference height is used. The lowest valid measurement height corresponding to the sample is retained from all five observation sources. In this step, firstly... and Calculate the temperature scale term Then introduce a unified reference height mapping term. Constraints are imposed on the height measurement term, and a stability bounded correction term is introduced. The stability correction function under strongly unstable stratification is piecewise limited, and finally combined with , Calculated optical turbulence characterization .

[0047] The height mapping coefficient α, the strong instability segmentation threshold τ, and the bounded attenuation coefficient β are determined as follows: First, a set of candidate parameter combinations is constructed, consisting of candidate values ​​for the height mapping coefficient, candidate values ​​for the strong instability segmentation threshold, and candidate values ​​for the bounded attenuation coefficient. Each candidate parameter combination is substituted into the unified reference height mapping term and the stability bounded correction term to construct candidate optical turbulence characterization quantities corresponding to each of the five observation sources. The maximum mean difference statistics of the candidate optical turbulence characterization quantities between each pair of the five observation sources are then calculated. The average value of the ten sets of maximum mean difference statistics is then used as the inter-source distribution difference index corresponding to the candidate parameter combination. Finally, the parameter combination that minimizes the inter-source distribution difference index is selected as the final parameter combination (α, τ, β).

[0048] The results obtained through this step It is an optical turbulence characterization quantity constrained by migration prediction, used as a target reference for subsequent feature construction.

[0049] Step 4: Based on the constituent variables and functional relationships of the optical turbulence characterization quantities obtained in step S3, perform isomorphic reparameterization on the effective observation sequence obtained in step S2 to obtain transferable features.

[0050] In this step, the original observations are not directly used as input to the prediction model. Instead, the effective observation sequence is converted into a model based on the constituent variables and functional relationships of the optical turbulence characterization quantities described in step 3. Transferable characteristics with the same structural logic. The structural variables and functional relationships refer to the construction... The pressure, temperature, temperature scale, uniform reference height mapping, and stability bounded correction terms used, as well as the square, ratio, logarithmic, and correction function relationships among these terms.

[0051] The transferable features generated in this step consist of the logarithmic term of friction velocity. Temperature scale item Stability bounded correction term Refractive index pressure related items Unified reference height logarithmic items Temperature item and time period items , , and constitute.

[0052] in, Indicates time The day number in a year, Indicates time The hourly sequence within a day. Through this step, the same physical mechanism is mapped to a unified input feature structure, thereby enabling samples from different observation sources to be expressed in a unified physical meaning space.

[0053] Step 5: Expand the transferable features obtained in step S4 according to the current time and historical time to construct a set of transferable features.

[0054] In this step, the transferable features obtained in step 4 are expanded temporally according to the current time and historical time. Specifically, for each time... ,extract , minute, minutes and The transferable features of each minute are concatenated in a fixed order to form a transferable feature sample. The resulting feature set contains both the unified physical structure provided in step 4 and short-term optical turbulence evolution information. The corresponding prediction target is set as the base-10 logarithmic value of the optical turbulence characterization for the next 60 minutes, i.e.: .

[0055] Step 6: Output the set of transferable features constructed in step S5 as input to the prediction model of optical turbulence characterization of the target scene.

[0056] In this step, the transferable feature set formed in step 5 is input into the prediction model for the future optical turbulence characteristics of the target scene, and the prediction results for the optical turbulence characteristics of the target scene for the next 60 minutes are output. The prediction model is a ridge regression model, a gradient boosting regression tree model, a multilayer perceptron model, or a gated recurrent unit model. Since the output of this invention is the upstream input feature constrained by a unified physical mechanism, it can be reused by multiple existing models without relying on a single model structure.

[0057] After completing steps 1 to 6 above, the method of the present invention is compared and verified.

[0058] In the comparative verification, the same five observation source data, the same time division method, and the same prediction target as the present invention were used to construct three types of input features. The first type is the original input features, which include air pressure, air temperature, wind speed, friction velocity, turbulent heat flux, stability ratio, measurement altitude, and time. The second type is the ordinary normalized features, which perform numerical scaling and logarithmic transformation on the wind speed, air pressure, friction velocity, turbulent heat flux, and measurement altitude in the original input features. The third type is the transferable features constructed in the present invention.

[0059] Subsequently, the three types of input features were fed into the ridge regression model, gradient boosting regression tree model, multilayer perceptron model, and gated recurrent unit model, respectively, for training and testing under the same training, validation, and test set partitioning. The training, validation, and test sets were constructed using a leave-one-out-of-observation-source approach, where one observation source was retained as the target scenario and the remaining four as source scenarios, with five cross-observation-source experiments conducted alternately. Evaluation metrics included mean absolute error, root mean square error, and the feature distribution difference metric between different observation conditions. The feature distribution difference metric between different observation conditions used the maximum mean difference statistic.

[0060] Figure 5The results show a comparison of feature distribution differences among three types of input features under different observation conditions in a cross-observation-source scenario. In the average results of five leave-one-out-of-observation-source experiments, the maximum mean difference statistic for the original input features is 0.1696, the maximum mean difference statistic for the ordinary normalized features is 0.1623, and the maximum mean difference statistic for the transferable features of this invention is 0.1092. These results demonstrate that the method of this invention can significantly reduce feature distribution shifts between different observation sources.

[0061] Figure 6 The comparison results of prediction errors of different models are shown. The figure uniformly compares the mean absolute errors of the ridge regression model, gradient boosting regression tree model, multilayer perceptron model, and gated recurrent unit model under the original input features, ordinary normalized features, and the transferable features of this invention. Under the same model conditions, the mean absolute error corresponding to the transferable features of this invention is generally lower than that of the original input features; among the four models, the mean absolute error corresponding to the transferable features of this invention is also lower than that of ordinary normalized features or better than the best-performing control term of ordinary normalized features. Specifically, under the multilayer perceptron model, the mean absolute error (MAO) of the transferable features of this invention is 0.7522, lower than 0.7755 for ordinary normalized features and 0.8693 for the original input features; under the gated recurrent unit model, the MAO of the transferable features of this invention is 0.6626, lower than 0.7820 for ordinary normalized features and 0.7869 for the original input features; and under the gradient boosting regression tree model, the MAO of the transferable features of this invention is 0.6742, lower than 0.6955 for the original input features and 0.6924 for the ordinary normalized features. These results demonstrate that the transferable features constructed in this invention not only reduce the distribution differences between observation sources but also stably improve cross-conditional transfer prediction performance across multiple existing prediction models.

Claims

1. A method for constructing transferable features of optical turbulence under heterogeneous observation conditions, characterized in that, Includes the following steps: S1. Input near-surface observation data from multiple observation sources within a preset time range, and perform time unification and field alignment on the near-surface observation data to form a unified original dataset; S2. Perform quality control on the unified raw dataset obtained in step S1 to obtain effective observation sequences; S3. Based on the effective observation sequence obtained in step S2, a unified reference height mapping term and a stability bounded correction term are introduced, and the height mapping coefficient α, the strong instability segmentation threshold τ, and the bounded attenuation coefficient β are determined. The measurement height term and stability correction term in the Monin-Obukhov similarity theory are constrained across observation conditions and scenarios to construct the optical turbulence characterization quantity corresponding to each time moment. S4. Based on the constituent variables and functional relationships of the optical turbulence characterization quantities obtained in step S3, perform isomorphic reparameterization on the effective observation sequence obtained in step S2 to obtain transferable features; S5. Expand the transferable features obtained in step S4 according to the current time and historical time to construct a set of transferable features; S6. Output the set of transferable features constructed in step S5 as input to the prediction model of optical turbulence characterization of the target scene.

2. The method for constructing transferable optical turbulence features under heterogeneous observation conditions according to claim 1, characterized in that, The near-surface observation data in step S1 includes air pressure, air temperature, friction velocity, turbulent heat flux, stability ratio, and measurement altitude; the time unification and field alignment include unifying all observation data to the same timestamp format, the same field name, and the same unit system. In the unit system, air pressure is unified to hPa, air temperature is unified to Kelvin, measurement altitude is unified to meters, and timestamps are unified to a 15-minute sampling interval.

3. The method for constructing transferable optical turbulence features under heterogeneous observation conditions according to claim 1, characterized in that, The quality control in step S2 includes screening for missing values, screening for non-physical values, screening for the integrity of key fields, and screening for abnormal mutation values; the key fields include air pressure, air temperature, friction velocity, turbulent heat flux, stability ratio, and measurement height.

4. The method for constructing transferable optical turbulence features under heterogeneous observation conditions according to claim 1, characterized in that, The Monin-Obukhov similarity theory described in step S3 is used to describe the relationship between momentum transport, heat transport, and atmospheric stability in the near-surface layer, where the original measured altitude is denoted as . The length of Monin-Obukhov is denoted as The stability ratio is denoted as In the process of constructing the fundamental characterization of optical turbulence based on the Monin-Obukhov similarity theory, a unified reference height mapping term is introduced. and stability bounded correction term Cross-observation source constraints are applied to the measured height term and the stability correction term, where: The unified reference height mapping term is represented as: The stability bounded correction term is: The optical turbulence characterization quantity Construct according to the following formula: in, This is the refractive index conversion factor. and This is the stability correction constant. For the observation time, The pressure is expressed in hectopascals (hPa). Absolute temperature in Kelvin. To standardize reference height, For height mapping coefficients, The threshold for segmentation in cases of strong instability. The bounded attenuation coefficient, Represents the temperature scale term, and , For friction speed, This is the turbulent heat flux term.

5. The method for constructing transferable optical turbulence features under heterogeneous observation conditions according to claim 4, characterized in that, The height mapping coefficient α, the strong instability segmentation threshold τ, and the bounded attenuation coefficient β are determined as follows: First, a set of candidate parameter combinations is constructed, consisting of candidate values ​​for the height mapping coefficient, candidate values ​​for the strong instability segmentation threshold, and candidate values ​​for the bounded attenuation coefficient. Each candidate parameter combination is then substituted into a unified reference height mapping term and a stability bounded correction term to construct candidate optical turbulence characterization quantities corresponding to different observation conditions. The maximum mean difference statistic between candidate optical turbulence characterization quantities for each pair of different observation conditions is then calculated. The average value of the maximum mean difference statistic is then used as the inter-source distribution difference index corresponding to the candidate parameter combination. Finally, the parameter combination that minimizes the inter-source distribution difference index is selected as the final parameter combination (α, τ, β).

6. The method for constructing transferable optical turbulence features under heterogeneous observation conditions according to claim 1, characterized in that, The transferable feature described in step S4 consists of the logarithmic term of the frictional velocity. Temperature scale item Stability bounded correction term Refractive index pressure related items Unified reference height logarithmic items Temperature item It consists of time period terms.

7. The method for constructing transferable optical turbulence features under heterogeneous observation conditions according to claim 6, characterized in that, The time period term consists of an annual sine term, an annual cosine term, a daily sine term, and a daily cosine term, which are respectively represented as: in, For a moment The day number in a year, For a moment The hour number in a day.

8. The method for constructing transferable optical turbulence features under heterogeneous observation conditions according to claim 1, characterized in that, The temporal unfolding includes extracting transferable features from the current time, 15 minutes before the current time, 30 minutes before the current time, and 60 minutes before the current time, and concatenating them in a fixed order to form feature samples in the set of transferable features.

9. The method for constructing transferable optical turbulence features under heterogeneous observation conditions according to claim 4, characterized in that, The target for predicting the future optical turbulence characterization of the target scene in step S6 is set as the logarithmic value of the optical turbulence characterization for the next 60 minutes, base 10. 。 10. The method for constructing transferable optical turbulence features under heterogeneous observation conditions according to claim 1, characterized in that, The set of transferable features serves as input to a ridge regression model, a gradient boosting regression tree model, a multilayer perceptron model, or a gated recurrent unit model.

Images