Method for predicting optical properties of sand dust aerosol and application

Abstract

This invention discloses a method and application for predicting the optical properties of dust aerosols, comprising: Step 1, establishing a triaxial ellipsoidal model for dust aerosol particles, obtaining the morphological and optical parameters of the dust aerosol particles as feature vectors, and establishing a dataset X; Step 2, obtaining the true values ​​of K types of optical property parameters corresponding to dataset X, and establishing a label set Y; Step 3, constructing a prediction model, which includes K parallel XGBoost-based sub-prediction modules, and training the prediction model; Step 4, testing the prediction model; Step 5, obtaining the feature vectors of the dust aerosol particles to be predicted, inputting them into the prediction model, and outputting the predicted values ​​of the K types of optical property parameters of the dust aerosol particles. This invention achieves rapid prediction of the K types of optical property parameters of dust aerosol particles through a prediction model, significantly improving the computational efficiency of dust aerosol optical properties and greatly reducing memory usage.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, and more specifically, to a method and application for predicting the optical properties of dust aerosols. Background Technology

[0002] Dust aerosols, as an important component of the troposphere, play a crucial regulatory role in the Earth system's radiation balance, cloud microphysical processes, and biogeochemical cycles. The optical properties of dust aerosols, such as extinction efficiency factor, scattering efficiency factor, single-scattering albedo, and asymmetry factor, are core foundational data for atmospheric radiative transfer simulation and global climate model construction. They are also key basis for developing algorithms for inverting passive and active aerosol remote sensing products.

[0003] Currently, the mainstream method for obtaining optical property parameters of dust aerosols relies on numerical calculations of electromagnetic scattering to accurately calculate the optical properties of particles with complex shapes such as triaxial ellipsoids. Based on these precise calculation methods, the optical characteristics of sampled samples are pre-calculated, and a lookup table database of optical properties covering different sizes, shapes, and refractive indices is constructed. In practical applications, the required optical properties can be quickly obtained by querying this database and combining it with interpolation methods.

[0004] However, the method of constructing and using optical property lookup table databases has significant limitations. First, to ensure applicability to real dust aerosols with spatiotemporal heterogeneity, the database needs to cover the widest possible range of morphological parameters. This results in a huge storage space, reaching GB levels, severely limiting portability and sharing efficiency. Second, the sampled samples have a fixed parameter sampling interval, meaning the database cannot cover all points in the continuous parameter space, and the interpolation process itself introduces a certain degree of accuracy loss.

[0005] Therefore, there is an urgent need for a new method that can both ensure computational accuracy and significantly reduce storage costs and improve call efficiency. Summary of the Invention

[0006] The technical problem to be solved by this application is to provide a method and application for predicting the optical properties of dust aerosols, addressing the shortcomings of the prior art. The method has a simple structure and reasonable design. Based on the morphological and optical parameters of the triaxial ellipsoidal model, a prediction model is established based on XGBoost to achieve rapid prediction of K-type optical property parameters of dust aerosol particles. This significantly improves the computational efficiency of dust aerosol optical properties and greatly reduces memory usage, providing an efficient and accurate optical property prediction solution for scenarios such as atmospheric radiative transfer models and aerosol remote sensing inversion.

[0007] To solve the above-mentioned technical problems, the technical solution adopted in this application is: In a first aspect, the present invention provides a method for predicting the optical properties of dust aerosols, characterized by comprising the following steps: Step 1: Establish a triaxial ellipsoidal model for dust aerosol particles, obtain the morphological and optical parameters of dust aerosol particles as feature vectors, and establish dataset X; Step 2: Obtain the true values ​​of K types of optical characteristic parameters corresponding to dataset X, and establish a label set Y, where K is a positive integer; Step 3: Construct a prediction model. The prediction model includes K parallel XGBoost-based sub-prediction modules. K types of optical characteristic parameters are used as the output labels of the K sub-prediction modules. Each sub-prediction module is iteratively trained using the dataset X. The optimal parameters under the current task are saved to obtain K trained sub-prediction modules. Step 4: Test the predictive performance of the prediction model using a test set; Step 5: Obtain the feature vector of the dust aerosol particles to be predicted, input it into the prediction model, and output the predicted values ​​of the K-type optical characteristic parameters of the dust aerosol particles.

[0008] The aforementioned method for predicting the optical properties of dust aerosols is characterized by the following morphological parameters: particle size x and aspect ratio. and the aspect ratio Optical parameters include refractive index m; dataset X = (x, , (m) , a represents the short semi-axis, b represents the middle semi-axis, and c represents the long semi-axis. .

[0009] The above-mentioned method for predicting the optical properties of dust aerosols is characterized by: refractive index ,in Indicates the real part, The imaginary part is represented by 'i', which represents the imaginary unit. ∈[1.10, 2.10], ∈[0.0005, 0.50].

[0010] The aforementioned method for predicting the optical properties of dust aerosols is characterized by: K=4, Y=(Q ext Q sca SSA, g), Q ext Q represents the extinction efficiency factor. sca denoted by scattering efficiency factor, SSA by single scattering albedo, and g by asymmetry factor.

[0011] The aforementioned method for predicting the optical properties of dust aerosols is characterized by: the loss function of the sub-prediction module, ,in Represents the learning parameters of the model. This represents the true value of the k-th type of optical characteristic parameter of the j-th dust aerosol particle. Compared with the predicted value The error between them, where n represents the total number of samples; Let represent the regularization term of the t-th decision tree, where T represents the total number of decision trees, and 1 ≤ k ≤ K.

[0012] The aforementioned method for predicting the optical properties of dust aerosols is characterized in that: the loss function of the sub-prediction module is an improved loss function. , ,in Represents the learning parameters of the model. This represents the true value of the k-th type of optical characteristic parameter of the j-th dust aerosol particle. Compared with the predicted value The error between them, where n represents the total number of samples; Let T represent the regularization term of the t-th decision tree, and T represent the total number of decision trees; Indicates the penalty coefficient. This represents the physical mechanism penalty term set for the k-th type of optical characteristic parameter, where 1≤k≤K.

[0013] The aforementioned method for predicting the optical properties of dust aerosols is characterized by the following specific method for testing the prediction performance of the prediction module using a test set: Input the test set T into the prediction model; The prediction model outputs an extinction efficiency factor. The predicted value, scattering efficiency factor The predicted value, single-scattering albedo The predicted value, and the asymmetric factor The predicted value; Calculate the root mean square error (RMSE) and the coefficient of determination respectively. ; Root Mean Square Error (RMSE) and Coefficient of Determination If all values ​​fall within the threshold range, the prediction model passes the test.

[0014] The above-mentioned method for predicting the optical properties of dust aerosols is characterized in that: the test set T includes a first test set T1 and a second test set T2, the first test set T1 = ( , The second test set T2 = (x, m), m).

[0015] The above-mentioned method for predicting the optical properties of dust aerosols is characterized by: , This represents the true value of the k-th type of optical characteristic parameter of the h-th dust aerosol particle in the test set. This represents the predicted value of the k-th type of optical characteristic parameter of the h-th dust aerosol particle in the test set, and M represents the number of samples in the test set. , Let K represent the true average value of the k-th type of optical characteristic parameter of all samples in the test set, where 1 ≤ k ≤ K.

[0016] Secondly, the present invention provides an application of a method for predicting the optical properties of dust aerosols, characterized in that: the method for predicting the optical properties of dust aerosols is applied to an atmospheric radiation transfer simulation model, and the predicted optical properties are transmitted to the atmospheric radiation transfer simulation model as input items of the atmospheric radiation transfer simulation model.

[0017] This application has the following advantages compared with the prior art: 1. This application has a simple structure, reasonable design, and is easy to implement and use.

[0018] 2. This application uses a triaxial ellipsoidal model to characterize the morphology of dust aerosol particles. By accurately restoring the non-spherical nature of dust aerosol particles through three independent semi-axis a, b and c, the influence of morphological characterization deviation on optical property calculation is eliminated, laying a physical foundation for subsequent high-precision prediction.

[0019] 3. This application uses the morphological and optical parameters of dust aerosol particles as a benchmark and establishes a prediction model based on XGBoost to achieve rapid prediction of K-type optical characteristic parameters of dust aerosol particles, replacing the original GB-sized optical characteristic lookup table database, and has good performance.

[0020] 4. This application constructs an incomplete test set to verify the model's ability to extrapolate uncovered parameters, and verifies that the prediction model has strong robustness and generalization ability, which can effectively cope with the complex and variable dust parameter scenarios in the real atmosphere, and provide stable and reliable optical property data support for downstream applications such as global climate models and aerosol remote sensing inversion.

[0021] 5. This application can replace the original GB-sized optical characteristic lookup table database with the lightweight code of the prediction model, which greatly reduces the storage space required; at the same time, it can achieve millisecond-level prediction of optical characteristic parameters, which significantly improves the computational efficiency compared with the time taken by traditional methods in the seconds or even minutes; it breaks through the efficiency and storage bottlenecks of traditional methods and has good performance.

[0022] In summary, this application uses the morphological and optical parameters of the triaxial ellipsoidal model as a benchmark and establishes a prediction model based on XGBoost to achieve rapid prediction of K-type optical characteristic parameters of dust aerosol particles. This significantly improves the computational efficiency of dust aerosol optical characteristics and greatly reduces memory usage, providing an efficient and accurate optical characteristic prediction solution for scenarios such as atmospheric radiative transfer models and aerosol remote sensing inversion.

[0023] The technical solution of this application will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0024] Figure 1 This is a flowchart of the method in this application.

[0025] Figure 2 This is a flowchart of the method for training a prediction model.

[0026] Figure 3 This is a triaxial ellipsoidal model of the dust aerosol particles in this application.

[0027] Figure 4 This is a scatter plot of the predicted and actual values ​​of the extinction efficiency factor.

[0028] Figure 5 This is a scatter plot showing the predicted and actual values ​​of the scattering efficiency factor.

[0029] Figure 6 A scatter plot showing the predicted and actual values ​​of single-scatter albedo.

[0030] Figure 7 A scatter plot showing the predicted and actual values ​​of the asymmetric factor. Detailed Implementation

[0031] The present application will now be described in further detail with reference to the accompanying drawings and embodiments thereof.

[0032] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0033] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0034] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented, for example, in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0035] For ease of description, spatial relative terms such as "above," "on top of," "on the upper surface of," "above," etc., are used herein to describe the spatial positional relationship of a device or feature as shown in the figures to other devices or features. It should be understood that spatial relative terms are intended to encompass different orientations in use or operation beyond the orientation of the device as described in the figures. For example, if the device in the figures were inverted, a device described as "above" or "on top of" other devices or structures would subsequently be positioned as "below" or "under" other devices or structures. Thus, the exemplary term "above" can include both "above" and "below." The device may also be positioned in other different ways (rotated 90 degrees or in other orientations), and the spatial relative descriptions used herein will be interpreted accordingly.

[0036] Example 1 like Figure 1 and Figure 2 As shown, the optical property prediction method for sand and dust aerosols of the present invention includes the following steps: Step 1: Establish a triaxial ellipsoidal model for dust aerosol particles, obtain the morphological and optical parameters of the dust aerosol particles as feature vectors, and establish a dataset X.

[0037] In practical applications, real dust aerosol particles typically appear as irregular ellipsoids or plates with varying curvatures under a microscope. Therefore, this application employs a triaxial ellipsoidal model to simulate dust aerosol particles. By using three independent semi-axial parameters a, b, and c, the non-spherical nature of dust aerosol particles is accurately reproduced, reducing errors caused by morphological approximation and thus ensuring the accuracy of the calculation of class k optical property parameters.

[0038] Morphological parameters include particle size x and aspect ratio. and the aspect ratio To simplify the model and reduce parameter redundancy, particle size x and aspect ratio were defined. and the aspect ratio Particle size x reflects the equivalent spherical volume of dust aerosol particles; aspect ratio and the aspect ratio Used to characterize the flatness and asymmetry of particles.

[0039] Optical parameters include refractive index m; dataset X = (x, , (m) , a represents the short semi-axis, b represents the middle semi-axis, and c represents the long semi-axis. .

[0040] Each sample in dataset X corresponds to a set of feature vectors (x, , The optical properties of dust aerosol particles are strongly correlated with their morphological and optical parameters. Morphological parameters determine the spatial geometry of the particles, while optical parameters reflect the nature of the interaction between the particles and light. Therefore, morphological and optical parameters are used to characterize the physical features of dust aerosol particles.

[0041] Figure 3 For a triaxial ellipsoidal model of dust aerosol particles, such as Figure 3 As shown, a, b, and c represent three mutually perpendicular semi-axes. To reduce parameter redundancy and simplify the analysis process, without loss of generality, it is assumed that the three semi-axes satisfy... Size relationships, morphological parameters include particle size x, aspect ratio and the aspect ratio Using the description of the current morphological parameters, various geometries can be determined by specifying the values ​​of the major and minor diameter ratios.

[0042] Optical parameters include the refractive index (m). Refractive index ,in Indicates the real part, The imaginary part is represented by 'i', which represents the imaginary unit. ∈[1.10, 2.10], ∈[0.0005, 0.50]. For real particulate matter with absorption capabilities, such as sand and dust aerosol particles, the refractive index m adopts the complex refractive index to match the optical properties of sand and dust aerosol particles with different compositions, such as quartz and clay. The real part n characterizes the ability of sand and dust aerosol particles to refract or scatter light, reflecting the change in the speed of light propagation inside the sand and dust aerosol particles; the imaginary part... It characterizes the ability of dust aerosol particles to absorb light, reflecting the energy loss of light as it propagates inside the dust aerosol particles.

[0043] Optical parameters may also include particle size distribution, optical thickness, polarization characteristics and / or wavelength λ.

[0044] In another possible embodiment, the optical parameters also include wavelength λ, and the dataset X = (x, , ,m,λ).

[0045] Step 2: Obtain the true values ​​of K types of optical characteristic parameters corresponding to dataset X, and establish label set Y.

[0046] Each sample in dataset X corresponds to a set of k types of optical property parameters.

[0047] In this embodiment, K=4, Y=(Q ext Q sca SSA, g), Q ext Q represents the extinction efficiency factor. sca denoted by scattering efficiency factor, SSA by single scattering albedo, and g by asymmetry factor.

[0048] Extinction efficiency factor Q ext Scattering efficiency factor Q sca The single scattering albedo (SSA) and the asymmetry factor g constitute a minimal complete set that can fully describe the fundamental physical processes of the interaction between a single dust aerosol particle and radiation. Therefore, in this embodiment, these four parameters are used as the optical characteristic parameters of the dust aerosol particles.

[0049] In another possible embodiment, the optical characteristic parameters of the tag set also include single scattering characteristics, which include the absorption efficiency factor Q. abs Quality extinction efficiency, linear depolarization ratio, and lidar ratio.

[0050] Step 3: Construct a prediction model. The prediction model includes K parallel XGBoost-based sub-prediction modules. K types of optical characteristic parameters are used as the output labels of the K sub-prediction modules. Each sub-prediction module is trained iteratively using the dataset X. The optimal parameters for the current task are saved to obtain K trained sub-prediction modules.

[0051] In this embodiment, the prediction model includes four parallel sub-prediction modules: an extinction efficiency prediction model, a scattering efficiency prediction module, a single albedo prediction module, and an asymmetry factor prediction module.

[0052] If the optical characteristic parameters of the tag set include the extinction efficiency factor Q ext Scattering efficiency factor Q scaSingle-scattering albedo (SSA), asymmetry factor g, and single-scattering characteristics, including absorption efficiency factor Q. abs When considering factors such as quality extinction efficiency, linear depolarization ratio, and lidar ratio, the prediction model comprises eight parallel sub-prediction modules.

[0053] The dataset and label set of dust aerosol particles exhibit a highly complex nonlinear relationship, which XGBoost effectively captures by integrating multiple decision trees. The XGBoost-based sub-prediction module is built upon decision trees, using them as base classifiers. Features are split by continuously adding decision trees to the model, and the new function formed by the newly added trees is used to fit the residuals of previous predictions. Finally, the predictions from all trees are summed to obtain the final prediction. Parameters that need to be set include the number of trees and the maximum depth.

[0054] In this embodiment, the number of trees is 8000-12000, and the maximum depth is 12-18. Other parameters use default parameters.

[0055] The optimization objective of the XGBoost-based sub-prediction module is to train each new decision tree by minimizing the loss function. The loss function design balances fitting error and model complexity constraints, ensuring training accuracy while avoiding overfitting. The loss function for the t-th iteration is defined as: , Represents the learning parameters of the model. This represents the true value of the k-th type of optical characteristic parameter of the j-th dust aerosol particle. Compared with the predicted value The error between the two samples, where n represents the total number of samples. Let represent the regularization term of the t-th decision tree, and T represent the total number of decision trees.

[0056] It is the XGBoost general loss function. The general loss function combines minimizing training loss and regularization constraints, and can accurately capture the nonlinear mapping between the true value and the predicted value. The general loss function can learn the inherent physical laws of the optical property parameters of sand and dust aerosol particles, achieve a high fit between the predicted value and the true value, and optimize the residual from a statistical perspective, which can solve the current engineering pain points.

[0057] Let represent the regularization term of the t-th decision tree. , This represents the total number of leaf nodes in the t-th decision tree. and These represent the regularization parameters, This represents the output weight of the p-th leaf node. The complexity used for constraint tree models.

[0058] The dataset X and label set Y are normalized. The extinction efficiency prediction model takes dataset X as input and the extinction efficiency factor Q corresponding to each sample in label set Y as input. ext The output labels are used for iterative training, and the optimal parameters are saved to obtain a trained extinction efficiency prediction model.

[0059] Similarly, the scattering efficiency prediction module takes dataset X as input and the scattering efficiency factor Q corresponding to each sample in the label set Y as input. sca The output labels are used for iterative training, and the optimal parameters are saved to obtain the trained scattering efficiency prediction module.

[0060] The single albedo prediction module takes dataset X as input and the single scattering albedo (SSA) corresponding to each sample in label set Y as output label, performs iterative training, saves the optimal parameters, and obtains the trained single albedo prediction module.

[0061] The asymmetric factor prediction module takes dataset X as input and the asymmetric factor g corresponding to each sample in label set Y as output label, performs iterative training, saves the optimal parameters, and obtains the trained asymmetric factor prediction module.

[0062] The XGBoost-based sub-prediction module optimizes the prediction results by progressively adding decision trees; its iterative prediction formula is as follows: ,in This indicates that the j-th dust aerosol particle sample is obtained after the (t-1)-th iteration. The prediction results of the k-th type of optical characteristic parameter, This indicates that the model trained in the t-th iteration is effective on the samples. The predicted value.

[0063] Step 4: Use the test set to test the predictive performance of the prediction model.

[0064] The specific method for testing the prediction performance of the prediction module using a test set is as follows: Input the test set T into the prediction model.

[0065] Test set T includes the first test set T1 and the second test set T2, where the first test set T1 = ( , The second test set T2 = (x, m), m).

[0066] To verify the predictive model's ability to extrapolate parameters not covered by the database and to verify its robustness, a first test set T1 and a second test set T2 were designed. The particle size x parameter from the database X was removed from the training set to form the first test set T1, where T1 = ( , The first test set T1 is missing particle size x, designed to simulate encountering new particle sizes not covered in the training set in real-world applications, and to obtain predicted values ​​for K types of optical property parameters. The minor axis ratio from the database X is used... The parameters are removed from the training set to form the second test set T2, where T2 = (x, ..., ...) The second test set T2, which lacks the minor aspect ratio, is designed to simulate encountering particles of new forms and to obtain predicted values ​​of K-type optical property parameters.

[0067] Both the first test set T1 and the second test set T2 are incomplete test sets. If the prediction model only interpolates on the training data, its performance on these incomplete test sets will drop sharply. However, the prediction model of this patent can still maintain a high coefficient of determination on these two incomplete test sets. The low root mean square error (RMSE) demonstrates that the prediction model has indeed learned the physical laws between the dataset X and the label set Y of dust aerosol particles.

[0068] The first test set T1 is used to test the prediction performance of the model under unseen particle size combinations. The second test set T2 is used to test the prediction performance of the model under unseen axis ratio combinations. By using the first test set T1 and the second test set T2, it can be verified whether the prediction model can still maintain high accuracy under parameter combinations not covered by the training data. This ensures that the prediction model has effectively learned the universal physical laws of the optical properties of dust aerosols, can cope with the spatiotemporal variability of dust particle morphology and size in the atmosphere, avoids the problem of performing well only within the training data coverage, but failing when encountering new parameters in practical applications, and improves the robustness of the prediction model.

[0069] The prediction model outputs an extinction efficiency factor. The predicted value, scattering efficiency factor The predicted value, single-scattering albedo The predicted value, and the asymmetric factor The predicted value; Calculate the root mean square error (RMSE) and the coefficient of determination respectively. The root mean square error (RMSE) reflects the absolute error level between the predicted value and the benchmark value, and is a coefficient of determination. It reflects the degree of linear fit between the predicted value and the benchmark value.

[0070] , This represents the true value of the k-th type of optical characteristic parameter of the h-th dust aerosol particle in the test set. This represents the predicted value of the k-th type of optical characteristic parameter of the h-th dust aerosol particle in the test set, where k = 1, 2, 3, 4, and M represents the number of samples in the test set.

[0071] , This represents the average of the true values ​​of the k-th type of optical characteristic parameter across all samples in the test set.

[0072] Root Mean Square Error (RMSE) and Coefficient of Determination If all values ​​fall within the threshold range, the prediction model passes the test. It should be noted that the root mean square error... The threshold range is [0, 0.045], and the coefficient of determination is... The threshold range is [0.998, 1].

[0073] Step 5: Obtain the feature vector of the dust aerosol particles to be predicted, input it into the prediction model, and output the predicted values ​​of the K-type optical characteristic parameters of the dust aerosol particles.

[0074] The morphological and optical parameters of the dust aerosol particles to be predicted are obtained and used as feature vectors. These vectors are then input into the prediction model, which outputs the extinction efficiency factor of the dust aerosol particles. The predicted value, scattering efficiency factor The predicted value, single-scattering albedo The predicted value, and the asymmetric factor The predicted value.

[0075] This application introduces a triaxial ellipsoidal model and a prediction model to replace the optical property lookup table database, achieving dual optimization of data storage and computational efficiency. It can not only output high-precision prediction results, but also achieve millisecond-level prediction of optical property parameters, which significantly improves computational efficiency compared to traditional methods that take seconds or even minutes. Furthermore, by replacing the original GB-sized optical property lookup table database with lightweight code of the prediction model, the required storage space is greatly reduced.

[0076] Example 2 Unlike Embodiment 1, this embodiment introduces a physical mechanism penalty term for optical feature parameters on top of the general XGBoost loss function. That is, constructing a loss function guided by physical information can further improve the prediction accuracy and physical rationality of the prediction model.

[0077] The loss function of the sub-prediction module is an improved loss function. , ,in Represents the learning parameters of the model. This represents the true value of the k-th type of optical characteristic parameter of the j-th dust aerosol particle. Compared with the predicted value The error between them, where n represents the total number of samples; Let T represent the regularization term of the t-th decision tree, and T represent the total number of decision trees; Indicates the penalty coefficient. This represents the physical mechanism penalty term set for the k-th type of optical characteristic parameter. The physical mechanism penalty term is related to... A function that is related to at least one parameter.

[0078] It should be noted that the physical mechanism penalty term is set according to the specific sub-prediction model. The physical mechanism penalty term includes the physical boundary penalty term and the physical trend penalty term.

[0079] In one possible embodiment, for single-scattering albedo In terms of its physical meaning, the true value must strictly fall within the [0,1] interval. Predictive models trained using a general loss function may output out-of-bounds values ​​under extreme inputs; therefore, when targeting single-scattering albedo... A physical boundary penalty term is introduced into the loss function of the single-shot albedo prediction module. In one possible embodiment, The two constraints are non-negativity and an upper bound of 1, respectively, to adapt to the actual physical range of values ​​for dust aerosols. Represents the single-scattering albedo of the j-th dust aerosol particle. The true value.

[0080] In one possible embodiment, for the extinction efficiency factor Generally speaking, the extinction efficiency factor increases monotonically with particle size x and refractive index m, then tends to plateau. A physical trend penalty term is introduced into the loss function of the extinction efficiency prediction model. , This refers to the extinction efficiency factor. Physical mechanism penalty items are set.

[0081] This indicates an indicator function. The function takes the value 1 if the condition within the parentheses is true, and takes the value 0 otherwise. Represents the particle size of the j-th dust aerosol particle, and the... This indicates that the particle size is increasing. Let represent the complex refractive index of the j-th dust aerosol particle. This indicates that the negative refractive index is increasing, and both of these conditions are met simultaneously. The value is 1. For , This represents the extinction efficiency factor of the j-th dust aerosol particle. The true value is that when the extinction efficiency of the (j+1)th sample increases relative to the jth sample, it conforms to physical laws. =0, Physical trend penalty term When the extinction efficiency of the (j+1)th sample decreases relative to the jth sample, it violates the laws of physics. The value is taken as the difference itself, which generates a penalty term, thereby guiding the extinction efficiency prediction model to learn the correct physical evolution law.

[0082] In one possible embodiment, for the scattering efficiency factor In other words, the core physical constraints are The prediction model trained using a general loss function may output different values ​​under extreme input conditions. The value is inconsistent with physical meaning, therefore, when considering the scattering efficiency factor... A physical boundary penalty term is introduced into the loss function of the scattering efficiency prediction module. In one possible embodiment, . This represents the scattering efficiency factor of the j-th dust aerosol particle. The true value, This represents the extinction efficiency factor of the j-th dust aerosol particle. The true value.

[0083] In one possible embodiment, for the asymmetry factor In other words, the core physical constraints are For dust aerosols, the actual value of the asymmetry factor is... Predictive models trained using a general loss function may output out-of-bounds values ​​under extreme inputs. Therefore, when dealing with asymmetric factors... A physical boundary penalty term is introduced into the loss function of the asymmetric factor prediction module. In one possible embodiment, The two constraints are an upper bound of 0.8 and a lower bound of 0, respectively, to fit the actual physical range of dust aerosol values. The asymmetry factor of the j-th dust aerosol particle The true value.

[0084] It should be noted that, in another possible embodiment, the penalty term is squared in order to strengthen the physical mechanism penalty.

[0085] Example 3 The optical property prediction method based on dust aerosols is applied to an atmospheric radiative transfer simulation model. The predicted optical property values ​​are transmitted to the atmospheric radiative transfer simulation model as input. The predicted optical property values ​​replace table lookup inputs, improving computational speed and accuracy.

[0086] The prediction model can directly output the four core optical characteristic parameters required for atmospheric radiative transfer simulation models, including the extinction efficiency factor. Scattering efficiency factor Single scattering albedo and asymmetric factors No additional interpolation calculations are required. This facilitates the integrated application of atmospheric radiative transfer simulation in global climate models, aerosol remote sensing inversion, and other scenarios.

[0087] Performance testing The system is structured around four parallel sub-prediction modules built on the XGBoost framework: an extinction efficiency prediction module, a scattering efficiency prediction module, a single albedo prediction module, and an asymmetry factor prediction module. These modules enable accurate prediction of various optical characteristic parameters. A low-level C language library interface is used to construct data transfer pointers, ensuring efficient computational resources during model iteration.

[0088] Configuration: The basic loss function is configured as regression squared error according to Example 1. The regularization term constrains the complexity of the decision tree. The maximum tree depth ranges from 12 to 18. The number of iterations ranges from 8000 to 12000. All other hyperparameters use the default configuration of the framework.

[0089] Each sub-prediction module employs differentiated parameter configurations and is trained independently. The XGBoost framework's input includes an externally customizable parameter structure and a maximum number of iterations, providing the foundation for the differentiated parameter configurations of the four sub-prediction models. Each sub-prediction model can be configured with different parameters based on the physical characteristics of its corresponding optical feature parameters, without requiring global uniformity.

[0090] For example, the extinction efficiency factor Q ext The evolutionary patterns are complex. The maximum tree depth for the extinction efficiency prediction model is configured as 15, and the number of iterations is configured as 10,000. The maximum tree depth for the scattering efficiency prediction module is configured as 14, and the number of iterations is configured as 9,500. The single scattering albedo (SSA) boundary constraints are clear. The maximum tree depth for the single albedo prediction module is configured as 12, and the number of iterations is configured as 9,000. The maximum tree depth for the asymmetric factor prediction module is configured as 13, and the number of iterations is configured as 9,000 based on the convergence.

[0091] Training: The dataset used for model training and testing contains a total of 320,400 samples. Each sample covers the morphological and optical parameters of dust aerosols, and calculates the corresponding four types of optical property parameters.

[0092] For example, some values ​​of x are 0.26, 1.46, 6.53, 15.38, 29.24, 68.84, 180.4, and 651.71. The partial values ​​are 1.00, 1.10, 1.13, 1.16, 1.20, 1.50, 1.80, 2.10, 2.40, 2.70, and 3.30. The partial values ​​are 1.00, 1.10, 1.13, 1.16, 1.20, 1.50, 1.80, 2.10, 2.40, 2.70, and 3.30. The partial values ​​are 1.10, 1.20, 1.30, 1.40, 1.50, 1.60, 1.70, 1.90, and 2.10. The partial values ​​are 0.0005, 0.001, 0.005, 0.01, 0.02, 0.05, 0.08, 0.1, 0.2, and 0.5.

[0093] To ensure the objectivity and generalizability of the evaluation, the dataset adopts a random partitioning strategy: 70% of the samples are used for training the prediction model and 30% of the samples are used for testing the prediction model.

[0094] In each iteration, the prediction model optimizes the objective function, continuously minimizing the loss, until all iterations are completed and convergence is achieved.

[0095] Test: Divide the test set into a first test set and a second test set. Remove the particle size x parameter from the first test set to form the first test set T1, where T1 = ( , (m). The minor axis ratio of the second test set. The parameters are removed from the set to form the second test set T2, T2 = (x, ..., ...) m).

[0096] The prediction model was tested using the first and second test sets.

[0097] Figure 4 A scatter plot showing the predicted and actual values ​​of the extinction efficiency factor. Figure 5 A scatter plot showing the predicted and actual values ​​of the scattering efficiency factor. Figure 6 A scatter plot showing the predicted and actual values ​​of single-scatter albedo. Figure 7 A scatter plot showing the predicted and actual values ​​of the asymmetric factor.

[0098] The extinction efficiency factor was calculated. coefficient of determination Extinction efficiency factor The root mean square error (RMSE) is 0.03607; the scattering efficiency factor... coefficient of determination Scattering efficiency factor The root mean square error (RMSE) is 0.04319; single scattering albedo. coefficient of determination Single scattering albedo The root mean square error (RMSE) is 0.00734; asymmetry factor. coefficient of determination asymmetric factor The root mean square error (RMSE) is 0.00591.

[0099] Therefore, the calculated root mean square error (RMSE) of the prediction model is (0.03607 + 0.04319 + 0.00734 + 0.00591) / 4 = 0.02313, and the coefficient of determination of the prediction model is... = (0.99956 + 0.99894 + 0.99975 + 0.9999) / 4 = 0.99954. The prediction model passes the test.

[0100] like Figures 4 to 7 As shown, the true values ​​are represented by a straight line with a slope of 1, and the predicted values ​​closely surround the true values, with the overall deviation between the predicted and true values ​​being extremely small. This demonstrates that the prediction model of this application can accurately capture the physical laws between optical properties and the feature vectors of triaxial ellipsoidal dust aerosol particles, proving that the prediction model has indeed learned the physical laws between the dataset X and the label set Y of dust aerosol particles. The prediction model possesses excellent learning ability and demonstrates outstanding predictive ability for all four types of optical property parameters.

[0101] The above description is merely an embodiment of this application and does not constitute any limitation on this application. Any simple modifications, alterations, or equivalent structural changes made to the above embodiments based on the technical essence of this application shall still fall within the protection scope of the technical solution of this application.

Claims

1. A method for predicting the optical properties of dust aerosols, characterized in that, Includes the following steps: Step 1: Establish a triaxial ellipsoidal model for dust aerosol particles, obtain the morphological and optical parameters of dust aerosol particles as feature vectors, and establish dataset X; Step 2: Obtain the true values ​​of K types of optical characteristic parameters corresponding to dataset X, and establish a label set Y, where K is a positive integer; Step 3: Construct a prediction model. The prediction model includes K parallel XGBoost-based sub-prediction modules. K types of optical characteristic parameters are used as the output labels of the K sub-prediction modules. Each sub-prediction module is iteratively trained using the dataset X. The optimal parameters under the current task are saved to obtain K trained sub-prediction modules. Step 4: Test the predictive performance of the prediction model using a test set; Step 5: Obtain the feature vector of the dust aerosol particles to be predicted, input it into the prediction model, and output the predicted values ​​of the K-type optical characteristic parameters of the dust aerosol particles.

2. The method for predicting the optical properties of dust aerosols according to claim 1, characterized in that: Morphological parameters include particle size x and aspect ratio. and the aspect ratio Optical parameters include refractive index m. Dataset X = (x, , (m) , a represents the short semi-axis, b represents the middle semi-axis, and c represents the long semi-axis. .

3. The method for predicting the optical properties of dust aerosols according to claim 2, characterized in that: Refractive index ,in Indicates the real part, The imaginary part is represented by 'i', which represents the imaginary unit. ∈[1.10, 2.10], ∈[0.0005, 0.50].

4. The method for predicting the optical properties of dust aerosols according to claim 1, characterized in that: K=4, Y=(Q ext Q sca SSA, g), Q ext Q represents the extinction efficiency factor. sca denoted by scattering efficiency factor, SSA by single scattering albedo, and g by asymmetry factor.

5. The method for predicting the optical properties of dust aerosols according to claim 1, characterized in that: Loss function of sub-prediction module ,in Indicates the learning parameters of the model. This represents the true value of the k-th type of optical characteristic parameter of the j-th dust aerosol particle. Compared with the predicted value The error between them, where n represents the total number of samples; Let represent the regularization term of the t-th decision tree, where T represents the total number of decision trees, and 1 ≤ k ≤ K.

6. The method for predicting the optical properties of dust aerosols according to claim 1, characterized in that: The loss function of the sub-prediction module is an improved loss function. , ,in Indicates the learning parameters of the model. This represents the true value of the k-th type of optical characteristic parameter of the j-th dust aerosol particle. Compared with the predicted value The error between them, where n represents the total number of samples; Let T represent the regularization term of the t-th decision tree, and T represent the total number of decision trees; Indicates the penalty coefficient. This represents the physical mechanism penalty term set for the k-th type of optical characteristic parameter, where 1≤k≤K.

7. The method for predicting the optical properties of dust aerosols according to claim 1, characterized in that: The specific method for testing the prediction performance of the prediction module using a test set is as follows: Input the test set T into the prediction model; The prediction model outputs an extinction efficiency factor. The predicted value, scattering efficiency factor The predicted value, single-scattering albedo The predicted value, and the asymmetric factor The predicted value; Calculate the root mean square error (RMSE) and the coefficient of determination respectively. ; Root Mean Square Error (RMSE) and Coefficient of Determination If all values ​​fall within the threshold range, the prediction model passes the test.

8. The method for predicting the optical properties of dust aerosols according to claim 7, characterized in that: Test set T includes the first test set T1 and the second test set T2, where the first test set T1 = ( , The second test set T2 = (x, m), m).

9. A method for predicting the optical properties of dust aerosols according to claim 7, characterized in that: , This represents the true value of the k-th type of optical characteristic parameter of the h-th dust aerosol particle in the test set. This represents the predicted value of the k-th type of optical characteristic parameter of the h-th dust aerosol particle in the test set, and M represents the number of samples in the test set. , Let K represent the true average value of the k-th type of optical characteristic parameter of all samples in the test set, where 1 ≤ k ≤ K.

10. An application of a method for predicting the optical properties of dust aerosols, characterized in that: The method for predicting the optical properties of dust aerosols as described in any one of claims 1-9 is applied to an atmospheric radiation transfer simulation model, and the predicted optical properties are transmitted to the atmospheric radiation transfer simulation model as input to the model.

Images