Ground-based photometer aerosol parameter inversion algorithm based on integrated machine learning

Abstract

This invention discloses a ground-based photometer aerosol parameter inversion algorithm based on ensemble machine learning. The invention uses a forward radiative transfer simulation model to obtain a training dataset for training the ensemble machine learning model. This training dataset is independent of existing aerosol products and observations affected by noise. The algorithm exhibits good generalization ability and performs well after being applied to real photometer observations. The inverted aerosol parameters are capable of characterizing aerosol types. The aerosol inversion algorithm based on the ensemble machine learning model, developed for ground-based photometer observations, completes an aerosol inversion in milliseconds. It can simultaneously utilize multi-band, multi-angle radiance observations to invert aerosol parameters across multiple bands at once. The algorithm does not require any prior assumptions or smoothing constraints when inverting aerosol parameters and, to some extent, avoids the loss of inversion results caused by gradient descent non-convergence in numerical inversion algorithms, reducing the waste of effective observations.

Description

Technical Field

[0001] This invention relates to atmospheric composition observation technology, specifically to a ground photometer aerosol parameter inversion algorithm based on integrated machine learning. Background Technology

[0002] Ground-based solar photometers are among the most widely used instruments for aerosol observation. They can measure direct solar radiation and diffuse sky radiation across multiple visible light bands within atmospheric windows. Based on these visible light observations, aerosol optical depth (AOD), single scattering albedo (SSA), scattering asymmetry factor (g), and effective radius (r) are also measured. effAerosol optical and microphysical parameters, such as those for aerosols, can be inverted, which is crucial for determining the total amount, type, and radiative effects of aerosol columns. The CE-318 ground-based photometer performs direct solar radiation observations in the 440, 675, 870, 1020, 340, 380, 500, and 1640 nm bands, but only performs diffuse sky radiation observations in the first four bands. The photometer has three basic scanning modes for observing diffuse radiance: almucantar, principal plane, and hybrid scanning. Almucantar scanning maintains the viewing zenith angle (VZA) equal to the solar zenith angle (SZA) and measures at 23 fixed relative azimuth angles (RAA) in the horizontal plane; it is the most widely used scanning mode. AERONET (AErosol Robotic Network) is the most successful global ground-based photometer network. Each site is equipped with a CE-318 (or similar Cimel Electronique) photometer, providing nearly 30 years of relatively stable aerosol monitoring. Its aerosol products are widely used for satellite data validation, air quality monitoring, and aerosol climate forcing research. AERONET uses a unified official algorithm to invert aerosol parameters from direct observations of AOD and mean latitude scans of the sky diffuse radiance. This algorithm is a traditional aerosol numerical inversion algorithm, and its core idea is statistical optimization, i.e., iteratively adjusting the aerosol particle size distribution and complex refractive index to minimize the difference between the radiation simulated in forward modeling and the radiation observed by the photometer. After convergence, aerosol optical parameters such as AOD and SSA are calculated by a particle scattering model. AERONET aerosol products are mainly available in Level 1.5 and Level 2.0. The latter undergoes quality control measures such as AOD > 0.4 and SZA > 50°, with an SSA uncertainty of approximately 0.03 and a g uncertainty of approximately 0.02. Similar ground-based photometer networks include CARSNET (China Aerosol Remote Sensing Network) in China, SKYNET in Eurasia, AEROCAN (AERosol CANada) in Canada, and AGSNet (Aerosol Ground Station Network) in Australia, providing supplementary information on regional aerosol characteristics.

[0003] While numerical inversion algorithms like AERONET can achieve high-precision aerosol inversion, they also have some significant drawbacks. First, forward modeling mainly consists of particle scattering and radiative transfer models, resulting in high computational costs with multiple runs, making it the most time-consuming part of numerical inversion. Furthermore, inverting aerosol particle size distribution from sky diffuse radiation observations is an ill-posed problem, meaning it has a non-unique solution. The algorithm's matrix solving relies on numerous prior assumptions and smoothing constraints. Especially when the initial guess deviates significantly from the true value, the algorithm may fail to converge, losing potentially valid observations. Improvements to these algorithms focus on optimizing radiative transfer calculations, such as converting radiative transfer models from scalar to vector, updating gas absorption and spectral solar constant libraries, and considering the scattering of non-spherical particles. However, these improvements cannot overcome the high computational cost of iterative forward modeling and gradient descent in numerical inversion. Recently, machine learning (ML) technology has rapidly developed, outperforming numerical methods in capturing nonlinear relationships and computational speed, and it does not require initial guesses or prior constraints. Its applications in atmospheric composition remote sensing, including aerosol remote sensing, are increasing. For satellite aerosol remote sensing, machine learning-based (ML-based) inversion algorithms can be broadly categorized into two types based on the source of their training sets: one type matches satellite observations with ground-based photometer aerosol inversion products; the other type uses forward modeling to extensively simulate satellite observations under different aerosol scenarios. The former's advantage is that the training set more closely resembles real atmospheric conditions, but its disadvantage is insufficient representativeness. For example, areas with long-term, stable photometer observations will have more matching data and play a dominant role in model training. The latter's advantage is that the training set is independent and can contain various types of aerosols, meeting the requirement of large data volumes in machine learning training sets. Its disadvantage is that it places higher demands on the model's generalization ability; a well-trained model should not exhibit significant performance degradation when applied to noisy real-world observations. In contrast, only a few ML-based algorithms have been used for ground-based aerosol inversion, such as those for observations from the All-Sky Imager and CM21 radiometer, which mostly use photometer aerosol products as the ground truth for model training. However, currently, no ML-based inversion algorithm is widely used for ground-based photometer aerosol remote sensing. Summary of the Invention

[0004] To address the problems existing in the prior art, this invention proposes an algorithm for inverting aerosol parameters of ground photometers based on integrated machine learning.

[0005] The EML-based (Ensemble Machine Learning) algorithm for inverting aerosol parameters from a ground-based photometer, as described in this invention, includes the following steps:

[0006] 1) Obtain the training dataset:

[0007] a) Establish a forward radiative transfer simulation model: The forward radiative transfer simulation model includes a particle scattering mode and a radiative transfer mode;

[0008] b) Based on the complex refractive index and particle size distribution of the aerosol, the single scattering albedo (SSA), scattering phase matrix, and scattering asymmetry factor (g) of the aerosol are obtained using a particle scattering model.

[0009] c) Based on the single scattering albedo (SSA), scattering phase matrix, scattering asymmetry factor (g), aerosol optical thickness (AOD), surface albedo, and atmospheric temperature and pressure profiles output by the particle scattering mode, the radiative transfer mode is used to obtain the radiance at each observation angle when the ground photometer performs mean latitude scanning in multiple different set observation bands.

[0010] d) Based on the above forward radiative transfer simulation model, the photometer observations under multiple atmospheric scenarios were simulated to obtain multiple sets of aerosol optical thickness (AOD), single scattering albedo (SSA), scattering asymmetry factor (g), particle size distribution, and corresponding photometer radiance, forming a dataset.

[0011] e) Add Gaussian white noise to the dataset to obtain the training dataset for training the ensemble machine learning model;

[0012] 2) Obtain an ensemble machine learning model:

[0013] a) Constructing an ensemble machine learning model: An ensemble machine learning model consists of multiple independent base learners and meta-learners; each base learner has a different regression prediction method, and its output will be different when faced with the same set of inputs; the outputs of multiple base learners are linearly weighted and combined by the meta-learners to give the final prediction result of the ensemble machine learning model for the input.

[0014] b) The ensemble machine learning model is trained using the training dataset obtained in step 1). The model input includes radiance, observation angle and aerosol optical thickness (AOD) of multiple observation bands, and the output includes aerosol single albedo and scattering asymmetry factor and particle size distribution of the corresponding multiple bands. Multiple base learners in the ensemble machine learning model are trained in parallel.

[0015] 3) Aerosol parameters are obtained through inversion algorithm:

[0016] Using a trained ensemble machine learning model, the radiance, observation angle, and aerosol optical thickness (AOD) measured by a ground photometer are input, and the inversion algorithm obtains the aerosol single-path albedo, scattering asymmetry factor, and particle size distribution for multiple bands.

[0017] In step 1)a), the particle scattering mode is a mapping from aerosol microphysical parameters to aerosol optical parameters, and the present invention uses the T-Matrix model; the radiative transfer mode is a mapping from aerosol optical parameters, gas molecule absorption and Rayleigh scattering, and surface reflectance characteristics to the radiance observed by the photometer, and the present invention uses the VLIDORT radiative transfer mode.

[0018] In step 1)b), the complex refractive index and particle size distribution of the aerosol are sampled from existing aerosol products and randomly combined as input to the particle scattering mode. The particle scattering mode uses a continuous particle size distribution, which is set to satisfy a bimodal log-normal distribution function. The coarse and fine modes of the aerosol are determined by the size radius (r) and size, respectively. vc and r vf The two peaks at () represent

[0019] In step 1)c), the aerosol optical thickness (AOD) is a direct observation product from a ground photometer, the surface albedo is obtained from satellite surface observation products, and the atmospheric temperature and pressure profile is obtained from reanalysis data.

[0020] In step 1), d), different atmospheric scenarios refer to different aerosol optical depth (AOD), surface albedo, and atmospheric temperature and pressure profiles; particle size distribution specifically refers to the change in particle number concentration with respect to its radius, expressed as effective radius r. eff The two parameters, FMF and fine particles, represent this.

[0021] In step 1)e), appropriate noise enhancement is crucial to the training effectiveness and generalization ability of the model. This invention utilizes the difference between the radiance actually measured by the photometer and the radiance simulated by the forward radiative transfer model under the same atmospheric scene to fit a set of white noise that conforms to a Gaussian normal distribution, representing the noise characteristics of the photometer in actual observation.

[0022] In step 2)a), an ensemble machine learning model can employ multiple base learners, including random forests, gradient boosting and multilayer perceptrons, support vector machines, and convolutional neural networks; the meta-learner uses RidgeCV, LassoRegression, and ElasticNet to uniformly process the outputs of multiple base learners. Base learners allow for multivariate inputs and multivariate outputs.

[0023] In step 2)b), a 10-fold cross-validation is performed before training the ensemble machine learning model to evaluate its generalization ability. The results show that the prediction scores at each validation fold remain stable, with fluctuations controlled within 0.01. The training objects of the ensemble machine learning model include the built-in parameters of each base learner and the parameters of the meta-learner. The built-in parameters differ between base learners; the built-in parameters of the random forest include splitting features, splitting thresholds, and leaf node values, while the built-in parameters of the multilayer perceptron include the weight matrix and bias vector of each neuron in the layer. The parameters of the meta-learner include weights, intercepts, and regularization parameters. The training objective of the ensemble machine learning model is to make the model's final prediction as close as possible to the true value.

[0024] In step 3), the model simultaneously inverts aerosol parameters across multiple bands. The output aerosol parameters specifically include optical parameters, namely the single scattering albedo (SSA) and the scattering asymmetry factor (g), and microphysical parameters, namely the effective radius (r). eff Fine particle size distribution (FMF) is the ratio of fine particles to light particles. Optical parameters distinguish wavelengths, while particle size distribution does not.

[0025] Advantages of this invention:

[0026] 1. This invention is an aerosol inversion algorithm based on integrated machine learning developed for observations by the CE-318 ground photometer. It has a fast calculation speed, completing one aerosol inversion in the millisecond range, which is five orders of magnitude faster than the traditional numerical iterative optimization inversion algorithm.

[0027] 2. The aerosol algorithm in this invention can simultaneously utilize multi-band, multi-angle radiance observations to retrieve aerosol optical parameters across multiple bands in one operation: single scattering albedo (SSA), scattering asymmetry factor (g), and microphysical parameters: effective radius (r). eff Compared to fine-particle inversion (FMF), existing numerical inversion algorithms often only allow for single-band inversion.

[0028] 3. The algorithm in this invention does not require any prior assumptions or smoothing constraints when inverting aerosol parameters, and to a certain extent avoids the loss of inversion results caused by gradient descent non-convergence in numerical inversion algorithms, thus reducing the waste of effective observations;

[0029] 4. The training of the machine learning model in this invention is entirely based on the dataset simulated by the self-built forward radiative transfer model, which is independent of noisy instrument observations and existing aerosol products, and has good robustness and generalization ability.

[0030] 5. The aerosol inversion algorithm in this invention has shown satisfactory aerosol inversion capability after being used in real observations by a photometer. The output results are highly consistent with existing aerosol products, with equal uncertainty, and the inverted aerosol parameters have the ability to characterize aerosol types. Attached Figure Description

[0031] Figure 1 This is a flowchart of the ground photometer aerosol parameter inversion algorithm based on integrated machine learning according to the present invention.

[0032] Figure 2 This is a scatter plot comparing the aerosol parameters retrieved according to an embodiment of the ground photometer aerosol parameter inversion algorithm based on integrated machine learning according to the present invention with the official aerosol product of AERONET. In this plot, (a) to (d) are the results of the single scattering albedo (SSA) inversion algorithm for different observation bands, (e) to (h) are the results of the scattering asymmetry factor (g) inversion algorithm for different observation bands, (i) is the result of the effective radius inversion algorithm, and (j) is the result of the fine particle ratio inversion algorithm. Detailed Implementation

[0033] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0034] like Figure 1 As shown, the EML-based (Ensemble Machine Learning) ground photometer aerosol parameter inversion algorithm of this embodiment includes the following steps:

[0035] 1) Obtain the training dataset:

[0036] a) Constructing a forward radiative transfer simulation model: The forward radiative transfer simulation model includes a particle scattering mode and a radiative transfer mode; the particle scattering mode is a mapping from aerosol microphysical parameters to aerosol optical parameters, and this embodiment uses the T-Matrix model; the radiative transfer mode is a mapping from aerosol optical parameters, gas molecule absorption and Rayleigh scattering, and surface reflectance characteristics to the radiance observed by the photometer, and this embodiment uses the VLIDORT radiative transfer mode;

[0037] b) Download Level 2.0 aerosol products from long-term observations at various AERONET sites worldwide. Resample and freely combine the aerosol optical thickness (AOD), particle size distribution parameters, and complex refractive index to obtain the complex refractive index and particle size distribution. Using a particle scattering model, calculate the single scattering albedo (SSA), scattering phase matrix, and scattering asymmetry factor (g) based on the aerosol's complex refractive index and particle size distribution. The particle size distribution in the particle scattering model uses a continuous distribution and is set to satisfy a bimodal logarithmic distribution. The bimodal logarithmic distribution corresponds to the coarse mode and the fine mode, respectively, which are determined by the large and small radii (r). vc and r vf The two peaks at () represent that FMF can represent the proportion of fine particles and include aerosol type information;

[0038] c) Aerosol optical thickness (AOD) was obtained from direct observation products of the ground photometer; surface albedo was obtained from BRDF products observed by the MODIS satellite detector; and atmospheric temperature and pressure profiles were obtained from reanalysis data ERA5. Using the linearized vector radiative transfer mode VLIDORT 2.8.1, based on the single scattering albedo (SSA), scattering phase matrix, and scattering asymmetry factor (g) output by the particle scattering mode, as well as the aerosol optical thickness (AOD), surface albedo, and atmospheric temperature and pressure profiles, the radiance of the ground photometer at various observation angles during mean latitude scanning in multiple different set observation bands was calculated.

[0039] d) Based on the above forward radiative transfer simulation model, photometer observations under multiple atmospheric scenarios were simulated. Different atmospheric scenarios refer to different aerosol optical thickness (AOD), surface albedo, and atmospheric temperature and pressure profiles. Multiple sets of aerosol optical thickness (AOD), single-scattering albedo (SSA), scattering asymmetry factor (g), and particle size distribution (based on the effective aerosol radius r) were obtained. eff (e) Using the difference between the actual radiance measured by the photometer and the radiance simulated by the forward radiative transfer model under the same atmospheric scene, a set of white noise conforming to a Gaussian normal distribution is fitted to represent the noise characteristics of the photometer in real observation; Gaussian white noise is added to the dataset to obtain the training dataset for training the ensemble machine learning model.

[0040] 2) Obtain an ensemble machine learning model:

[0041] a) Constructing an ensemble machine learning model: The ensemble machine learning model consists of three independent base learners and a meta-learner. The three independent base learners employ random forest, gradient boosting, and multilayer perceptron, respectively. The meta-learner employs...

[0042] RidgeCV; Random Forest is a collection of decision trees, and its output is the average of the predictions from all trees, making it a typical example of stacked decision-making (bagging) methods; Gradient Boosting is also based on decision trees, but it uses a step-by-step iterative training method. Each new tree is used to fit the residual of the previous model and continuously correct the prediction results, making it a typical example of boosting methods; Multilayer Perceptron is a feedforward neural network, consisting of an input layer, multiple hidden layers, and an output layer. It is trained using nonlinear activation functions and backpropagation algorithms and can learn complex nonlinear mapping relationships between inputs and outputs; RidgeCV linearly weights and combines the outputs of multiple base learners to give the final prediction result of the ensemble machine learning model for the input;

[0043] b) Before training the ensemble machine learning model, a 10-fold cross-validation was performed to assess its generalization ability. Based on a training dataset of 100,000 samples, the dataset was shuffled and divided into ten equal subsets. Nine subsets were used for training each time, with the remaining subset used for validation. This process was repeated for all subsets. The results showed that the prediction scores at each validation fold remained stable, with fluctuations controlled within 0.01. The training included all built-in parameters of the three base learners and one meta-learner, with the training objective being to make the model's final predictions as close as possible to the true values. Table 1 shows the prediction scores at each validation fold: average coefficient of determination R0 2 The three parameters—root mean square error (RMSE), mean absolute deviation (MAD)—remained stable throughout ten rounds of training and validation, with fluctuations controlled within 0.01; the average coefficient of determination (R²)... 2 The root mean square error (RMSE) and mean absolute deviation (MAD) were 0.773, 0.43, and 0.282, respectively. Although the RMSE exceeded the typical inversion uncertainty of a single aerosol parameter (e.g., 0.03 for SSA and 0.02 for g), this was expected because these metrics are calculated based on the combined bias of all inversion variables, rather than for a single output parameter.

[0044] After cross-validation, the ensemble machine learning model was trained using all 100,000 samples. The model's input included radiance, observation angle, and aerosol optical thickness (AOD) at four observation bands: 440, 675, 870, and 1020 nm. The output included the aerosol single albedo (SSA) and scattering asymmetry factor (g) and effective radius (r) for the corresponding four bands. eff The FMF ratio compared to fine particles is shown in Table 2; the three base learners in the ensemble machine learning model are trained in parallel;

[0045] To verify whether the ensemble machine learning (ENL-based)-based inversion algorithm conforms to physical laws, this invention couples a SHapley Additive exPlanation module to the ensemble machine learning model. The contribution of each input feature to predicting the output target is evaluated using the SHAP value; a higher SHAP value indicates a greater influence of the feature on the model's prediction. The importance of each input feature in the aerosol parameter inversion algorithm of this invention when inverting specific aerosol parameters is summarized in Table 3. It can be seen that aerosol optical thickness (AOD) contributes the most to the inversion of single scattering albedo (SSA), while scattering asymmetry factor g and effective radius r... eff Compared to fine particles, FMF relies more on multi-band and multi-angle radiance observations, and these results are consistent with the physical laws of radiative transfer.

[0046] 3) Aerosol parameters are obtained through inversion algorithm:

[0047] Using a pre-trained ensemble machine learning model, and inputting aerosol optical thickness (AOD) measured by direct sunlight from a ground photometer, radiance measured by Almucantar scanning at the mean latitude circle, and the corresponding observation angles, the single aerosol albedo (SSA), scattering asymmetry factor (g), and effective radius (r) for the corresponding four spectral bands are obtained. eff Compared to fine particles, FMF.

[0048] Figure 2 The solid line is the equation obtained by least-squares fitting of the scattered points, where x is the abscissa and y is the ordinate; the dashed line is a 1:1 line. Figure 2 This paper presents a comparison between the aerosol parameters obtained by the inversion algorithm of this invention and the official AERONET aerosol product. Both algorithms used the same observations. It should be noted that the data points in the scatter plot were sparsified tenfold for easier visualization. The aerosol parameters obtained by the inversion algorithm of this invention show a high degree of consistency with the AERONET product. Except for the g value in the 440nm band, the correlation coefficient R of all variables exceeds 0.9. The root mean square error (RMSE) of the single aerosol albedo (SSA) and the scattering asymmetry factor g remains within 0.03, while the effective radius r... eff The error of FMF compared to fine particles is approximately 0.1. The main advantage of the ensemble machine learning-based inversion algorithm in this invention lies in its computational efficiency. In this test, a single inversion only takes an average of 0.18 milliseconds, while traditional numerical inversion algorithms typically require several minutes to process a single case. The algorithm improves efficiency by approximately 10% by avoiding iterative radiative transfer mode operation. 5 Times. Effective radius r eff Compared to fine particles, FMF can serve as a key indicator of aerosol particle size distribution: when r eff When the particle size is <0.3 μm and FMF>0.5, fine-mode aerosols such as sulfate, nitrate, and biomass combustion particles dominate; conversely, when r eff When the particle size is >1.0μm and FMF<0.3, coarse-mode aerosols generated by natural sources such as mineral dust and sea salt are dominant. Figure 2 The results show that the fine particle size distribution (FMF) exhibits a bimodal distribution around 0.3 and 0.7, corresponding to coarse and fine modal characteristics at 0.6 μm and 0.28 μm, respectively. Therefore, this algorithm can effectively classify aerosol types based on SSA, scattering asymmetry factor g, and particle size distribution.

[0049] Table 1. Results of 10-fold cross-validation for ensemble machine learning models

[0050]

[0051] Table 2: Inputs and Outputs of the Ensemble Machine Learning Model

[0052]

[0053] Table 3. Feature importance (%) of each input variable in ensemble machine learning models for retrieving specific aerosol parameters.

[0054]

[0055] Finally, it should be noted that the purpose of disclosing the embodiments is to help further understand the present invention. However, those skilled in the art will understand that various substitutions and modifications are possible without departing from the spirit and scope of the present invention and the appended claims. Therefore, the present invention should not be limited to the content disclosed in the embodiments, and the scope of protection of the present invention is defined by the claims.

Claims

1. A ground photometer aerosol parameter inversion algorithm based on ensemble machine learning, characterized in that, The inversion algorithm includes the following steps: 1) Obtain the training dataset: a) Establish a forward radiative transfer simulation model: The forward radiative transfer simulation model includes a particle scattering mode and a radiative transfer mode; b) Based on the complex refractive index and particle size distribution of the aerosol, the single scattering albedo, scattering phase matrix and scattering asymmetry factor of the aerosol are obtained using the particle scattering model. c) Based on the output of the particle scattering mode and the aerosol optical thickness, surface albedo, and atmospheric temperature and pressure profile, the radiative transfer mode is used to obtain the radiance at each observation angle when the ground photometer performs mean latitude scanning in multiple different observation bands. d) Based on the forward radiative transfer simulation model, the photometer observations under multiple atmospheric scenarios were simulated to obtain multiple sets of aerosol parameters and the corresponding photometer radiance, forming a dataset; e) Add Gaussian white noise to the dataset to obtain a training dataset for training an ensemble machine learning model; 2) Obtain an ensemble machine learning model: a) Constructing an ensemble machine learning model: An ensemble machine learning model consists of multiple independent base learners and meta-learners; b) The ensemble machine learning model is trained using the training dataset obtained in step 1). The inputs are the radiance, observation angle and aerosol optical thickness of multiple observation bands, and the outputs are the aerosol single scattering albedo and scattering asymmetry factor and particle size distribution of the corresponding multiple bands. Multiple base learners in the ensemble machine learning model are trained in parallel. 3) The inversion algorithm obtains aerosol parameters: Using a trained ensemble machine learning model, the radiance, observation angle, and aerosol optical thickness measured by a ground photometer are input, and the inversion algorithm obtains the aerosol single scattering albedo, scattering asymmetry factor, and particle size distribution for multiple bands.

2. The inversion algorithm as described in claim 1, characterized in that, In step 1)b), the complex refractive index and particle size distribution of the aerosol are sampled from existing aerosol products and randomly combined as input to the particle scattering mode.

3. The inversion algorithm as described in claim 1, characterized in that, In step 1)b), the aerosol particle size distribution in the particle scattering mode adopts a continuous particle size distribution, and the continuous particle size distribution is set to satisfy a bimodal log-normal distribution function. The coarse mode and fine mode of aerosol are represented by two peaks at the large and small radii, respectively.

4. The inversion algorithm as described in claim 1, characterized in that, In step 1)c), the aerosol optical thickness is a direct observation product from a ground photometer, the surface albedo is obtained from a satellite surface observation product, and the atmospheric temperature and pressure profile is obtained from reanalysis data.

5. The inversion algorithm as described in claim 1, characterized in that, In step 1), d), the aerosol parameters include aerosol optical thickness, single scattering albedo, scattering asymmetry factor, and particle size distribution.

6. The inversion algorithm as described in claim 1, characterized in that, In step 1) e), a set of white noise conforming to a Gaussian normal distribution is obtained by fitting the difference between the actual radiance measured by the photometer and the radiance simulated by the forward radiative transfer model, in order to represent the noise characteristics of the photometer during actual observation.

7. The inversion algorithm as described in claim 1, characterized in that, In step 2)a), the regression prediction methods of each base learner are different, and their outputs are different when faced with the same set of inputs; by linearly weighting the outputs of multiple base learners through meta-learners, the final prediction result of the ensemble machine learning model for the input is given.

8. The inversion algorithm as described in claim 1, characterized in that, In step 2)b), the ensemble machine learning model is subjected to a 10-fold cross-validation test before training to assess its generalization ability.

Images