All-sky imager aerosol optical parameter inversion algorithm constructed based on machine learning

Abstract

The invention discloses an all-sky imager aerosol optical parameter inversion algorithm constructed based on machine learning, and belongs to the field of meteorological observation. According to the method, an iterative threshold segmentation method and threshold self-optimization are used, a clear sky extraction quantitative standard is established for the first time, and the problems of manual dependence and poor consistency are solved; according to the method, an equal zenith angle scanning mode is simulated, multi-dimensional feature extraction is carried out from a sky area picture, aerosol scattering information reflecting multi-angle and multi-position information related to an aerosol scattering phase function and multi-angle information of total brightness is obtained, and key support is provided for SSA inversion; according to the method, an integrated regression model is constructed by taking the multi-dimensional features as input and AOD and SSA as output, a spatial difference and time difference matching sample is set, synchronous output is realized, a sample standard is defined, and efficiency and precision are greatly improved. The method is applied to the fields of atmosphere remote sensing, climate mode research, air quality evaluation and the like.

Description

Technical Field

[0001] This invention belongs to the field of meteorological observation, specifically involving a machine learning-based algorithm for retrieving aerosol optical parameters from an all-sky imager. Background Technology

[0002] Aerosols, as a common atmospheric component, play a crucial role in Earth's radiation budget, cloud and precipitation development, ecosystems, and air quality. However, estimating their contribution to radiative forcing is considered the primary source of uncertainty in quantifying current climate change. Therefore, accurate monitoring of aerosol spatial distribution and temporal dynamics is essential for understanding the Earth's climate system. Aerosol optical depth (AOD) is a dimensionless indicator that measures the scattering and absorption effects of solar radiation by aerosols within the air column from the Earth's surface to the top of the atmosphere. Simply put, it represents the degree to which light is attenuated by the presence of aerosols as it passes through the atmosphere. Therefore, AOD can be used to assess the climate impacts of aerosols and air pollution, and significant efforts have been devoted to measuring AOD at different scales in recent years. Since aerosol distribution, chemical composition, and radiation characteristics vary with location and time, detailed monitoring of aerosol spatiotemporal variations is necessary. Currently, many aerosol observations rely on satellites or space-based platforms, which, while enabling long-term, continuous monitoring in the temporal dimension, typically have low spatial resolution. However, due to factors such as geographical location, instrument costs, and maintenance difficulties, setting up more atmospheric measurement stations is not practically feasible.

[0003] However, relying solely on aerosol optical thickness cannot fully characterize the radiative effects of aerosols. The role of aerosols in solar radiation also depends on parameters such as Single Scattering Albedo (SSA). Single Scattering Albedo, defined as the ratio of the scattering coefficient to the extinction coefficient, directly reflects whether aerosols are primarily scattering (e.g., sea salt) or primarily absorbing (e.g., black carbon). It is a key parameter determining the sign and intensity of aerosol direct radiative forcing, contributing over 30% to the uncertainty of aerosol radiative forcing. Therefore, more precise observations of SSA are a necessary prerequisite for quantifying the radiative effects of aerosols.

[0004] All-sky imagers are instruments used for automated monitoring of ground cloud cover, designed to replace traditional manual observation. Represented by the Total Sky Imager first developed by Yankee in the United States, these instruments have evolved into various improved designs. All-sky imagers typically use CCD cameras equipped with fisheye lenses to capture high spatiotemporal resolution images of the entire sky in the visible light spectrum during the day, obtaining high-quality cloud cover monitoring data, which is then automatically uploaded to a computer for storage and processing. In recent years, all-sky imagers have seen significant improvements in quality, resolution (spatial and temporal), and calibration, making them suitable for environmental research and measurement. Based on the ability of all-sky imagers to capture high-resolution images of the entire sky in the visible light spectrum, researchers have attempted to use them to invert optical properties such as aerosol optical thickness.

[0005] The high spatiotemporal resolution of all-sky imagers (AS-ASAs) offers new possibilities for aerosol observation, but existing AOD retrieval methods still have significant limitations. Traditional AOD retrieval methods based on AS-ASAs typically require calculating the radiance of each pixel, iteratively adjusting the AOD value input to a radiative transfer model, and finally selecting the AOD result that best matches the observed data. This algorithm has high computational complexity, low computational efficiency, and requires significant computing power. Furthermore, research based on AS-ASAs has largely focused on AOD retrieval and has not yet achieved simultaneous AOD and SSA retrieval. This prevents the coordinated analysis of total aerosol-scattering characteristics at a fine scale, severely limiting the comprehensive analysis of their radiative effects and climate impacts.

[0006] Aerosol optical thickness (AOD) has traditionally been obtained primarily by measuring the attenuation (extinction ratio) of direct solar radiation. For all-sky imagers, it depends on the overall level of sky background radiance. Even with traditional radiative transfer models, AOD inversion involves adjusting the AOD value to best match the measured total radiance level. Since all-sky imagers directly reflect sky brightness, using them to invert AOD is physically natural. Single-scatter albedo (SSA) has traditionally been obtained by measuring the angular distribution of sky radiation (i.e., the distribution of scattered light), requiring radiance information from multiple angles and locations. SSA inversion is extremely sensitive to the geometric information of the image and the relative distribution (not just absolute values) of sky radiance. Therefore, SSA inversion using all-sky imagers presents a significant technical challenge in this field. Summary of the Invention

[0007] To overcome the problems of high equipment prices, low spatial density of instrument installation, and time delay of solar photometers in existing technologies, this invention proposes a machine learning-based algorithm for inverting aerosol optical parameters of an all-sky imager. By using the machine learning algorithm, aerosol optical thickness and single-scattering albedo are inverted based on all-sky imager data, thereby enabling the measurement of aerosol optical parameters of the all-sky imager at any location worldwide.

[0008] The present invention provides a machine learning-based algorithm for inverting aerosol optical parameters of all-sky imagers, comprising the following steps: 1) Collect raw sky images with time stamps; 2) Perform image processing on the original sky photo to remove all areas that affect the extraction of cloudless sky radiance. Specifically, the iterative threshold segmentation method is used to remove clouds and light spots generated by the lens module, resulting in a sky area image that retains only the clean sky. 3) Imitating the equal zenith angle scanning mode, feature extraction is performed from the sky region image. The average values ​​of R, G and B, the red-blue ratio rbr, the zenith angle of the sun and the scattering information of aerosols of all colored pixels are extracted. The above features are called multidimensional features and are represented by x. 4) Obtain the aerosol parameters for the set wavelength from the publicly available database, including aerosol optical thickness (AOD) and single scattering albedo (SSA). These parameters are called target variables and are represented by y. The geographical locations of the sites corresponding to the aerosol parameters in the publicly available database are consistent with the locations of the instruments that acquired the original sky images. 5) Match the multidimensional features with the target variable according to time, use them as a dataset, and divide them into training and test sets for model training and for evaluating the model's generalization ability. Store the data in a file and convert the data into the DMatrix format, which is specific to ensemble regression models. 6) Using an ensemble regression model, the multidimensional features are used as the input of the ensemble regression model, and the target variable is used as the output of the ensemble regression model. The data from step 5) are input into the ensemble regression model, and the training set is used for training and the test set is used for validation and consistency verification to obtain a well-trained ensemble regression model for this wavelength. 7) Repeat steps 4) to 6) multiple times by changing the wavelength to obtain multiple trained ensemble regression models for each wavelength; 8) Collect original sky images for the new time period, and input the multidimensional features obtained from the image processing in step 2) and feature extraction in step 3) into the integrated regression model to obtain the inverted aerosol optical thickness (AOD) and aerosol single scattering albedo (SSA).

[0009] In step 1), the instrument for acquiring the original sky images is fixed in position, the lens is perpendicular to the ground, the instrument is a full-sky imager, the original sky images include images of the entire sky, and the number of original sky photos is more than 100,000.

[0010] In step 2), the original sky photo is processed to remove all areas that affect the extraction of cloudless sky radiance, including the following steps: a) The relative position of the instrument to the surrounding buildings is fixed, and the surrounding buildings are removed by using a black mask; b) Traverse the image and find the brightest circular area, which is considered to be the sun and its surrounding bright areas. Use a black mask to remove the sun and its surrounding bright areas. c) Use an iterative thresholding method to repeatedly binarize the image until the threshold stabilizes. Use a black mask to cover the brighter areas and remove the light spots generated by the cloud and lens module.

[0011] Existing technologies use computer vision to extract edges for image segmentation. This invention uses an iterative threshold segmentation method, including the following steps: randomly setting an initial threshold; dividing pixels in the image into two categories based on their brightness compared to the threshold: above the threshold and below the threshold; calculating the average brightness of each category of pixels, and then averaging these two average brightness values ​​to obtain a new threshold; reclassifying pixels according to the new threshold, and repeating the above steps until a stable threshold is obtained; pixels above the stable threshold are considered to not meet the clear sky condition and are masked to remove clouds and light spots generated by the lens module, resulting in an image that retains only the clear sky area. Existing technologies rely on edge recognition and require manual parameter tuning. This invention uses a self-optimizing threshold and establishes a quantification standard for clear sky extraction, solving the problems of manual dependence and poor consistency.

[0012] In step 3), all colored pixels are extracted from the sky region image. Each colored pixel is divided into three channels: red (R), green (G), and blue (B). The average values ​​of R, G, and B for all colored pixels are calculated. The red-to-blue ratio (rbr) is calculated by dividing the average R value by the average B value. The zenith angle of the sun is calculated by combining the average R, G, and B values ​​and the rbr value corresponding to the sky region image with time and latitude and longitude. To mimic the optical parameter source of the AERONET website, namely the solar photometer, the scanning mode is the equal zenith angle (Almucantar) mode, which is to fix the zenith angle and change the azimuth angle. A radius of 0.6 to 0.8 times the radius of the sky region image is taken, and the center of the circle is located at the center point of the sky region image. A circle is drawn, and more than 10 pixels are taken on average on the circumference. The RGB values ​​of these multiple pixels are extracted and called aerosol scattering information. Aerosol scattering information reflects the multi-angle, multi-position information and total brightness related to the aerosol scattering phase function. Existing technologies only extract one-dimensional features, which cannot reflect the scattering angle distribution. This invention is the first to obtain multi-angle information, providing key support for SSA inversion.

[0013] In step 4), the spatial difference (horizontal distance) between the geographical location of the site corresponding to the aerosol parameters in the publicly available database and the location of the instrument that acquired the original sky image is within a spatial threshold, and is considered to be consistent in spatial location representativeness. The spatial threshold is 3~10 km. Aerosol optical thickness reflects the degree of atmospheric pollution. The measurement of single scattering albedo (SSA) in the prior art is inaccurate and has great uncertainty; the present invention obtains accurate SSA, thereby reducing the uncertainty of atmospheric models.

[0014] In step 5), the time of the aerosol parameters in the publicly available database does not perfectly match the time of the original sky image acquisition. If the time difference between the acquisition of the original sky image and the aerosol parameters in the publicly available database is within the time threshold, it is considered a time match; the time threshold is three to ten minutes.

[0015] In step 6), the ensemble regression model employs the extreme gradient boosting regressor XGBoost, which includes multiple base learners. The mean squared error (RMSE) is used as the loss function, and the optimal number of boosting iterations N is determined through K-fold cross-validation and an early stopping strategy, where K is a natural number ≥ 5. Subsequently, the optimal number of boosting iterations N is used as the number of base learners, and hyperparameter combinations are set for maximum tree depth, learning rate, subsampling ratio, column sampling ratio, and minimum loss reduction for leaf node splits. This combination is then used to train the ensemble regression model to obtain the joint output. The ensemble regression model is used to predict the test set, obtaining predicted vector pairs, and a scatter plot is drawn with the true value on the x-axis and the predicted value on the y-axis. An ensemble regression model is constructed using multidimensional features as input and AOD and SSA as outputs, with RMSE as the loss function and K-fold cross-validation used to determine the parameters. Spatial difference and 5-minute time difference are set to match samples. Existing techniques struggle with synchronous iterative inversion; this scheme achieves synchronous output, clarifies sample standards, and significantly improves efficiency and accuracy.

[0016] Advantages of this invention: Existing technologies use computer vision to extract edges for image segmentation; this invention uses an iterative threshold segmentation method with self-optimized thresholds, and establishes a quantification standard for clear sky extraction, solving the problems of reliance on manual methods and poor consistency. All colored pixels were extracted from the sky region image. Each colored pixel was divided into three channels: red (R), green (G), and blue (B). The average values ​​of R, G, and B for all colored pixels were calculated. The red-to-blue ratio (rbr) was calculated by dividing the average R value by the average B value. The solar zenith angle was calculated by combining the average R, G, and B values ​​and the rbr value of the sky region image with time and latitude / longitude. To mimic the optical parameter source of the AERONET website, an equal zenith angle scanning mode, i.e., a solar photometer, was used. A circle was drawn with a radius of 0.6 to 0.8 times the radius of the sky region image, centered at the center of the sky region image. At least 10 pixels were averaged on the circumference of the circle, and these pixels were extracted. The invention extracts RGB values ​​from multiple pixels. Existing technologies only extract one-dimensional features, which cannot reflect the scattering angle distribution. This invention mimics the equal zenith angle scanning mode to extract features from sky region images, thereby obtaining aerosol scattering information that reflects multi-angle, multi-location information and total brightness related to the aerosol scattering phase function, providing key support for SSA inversion. This invention uses multi-dimensional features as input and AOD and SSA as outputs to construct an integrated regression model, uses RMSE as the loss function, K-fold cross-validation to determine parameters, and sets spatial and temporal differences to match samples. Existing technologies have difficulty in synchronizing iterative inversion, while this invention achieves synchronous output, clarifies sample standards, and significantly improves efficiency and accuracy. Attached Figure Description

[0017] Figure 1This is a flowchart of the algorithm for inverting aerosol optical parameters of an all-sky imager based on machine learning, as described in this invention. Figure 2 A density scatter plot comparing the SSA and AOD at 440nm and 675nm with data from the Beijing site, obtained by an embodiment of the machine learning-based algorithm for inverting aerosol optical parameters of an all-sky imager according to the present invention. Figure 3 A density scatter plot comparing the SSA and AOD at 870nm and 1020nm obtained with data from the Beijing site, based on an embodiment of the machine learning-based algorithm for inverting aerosol optical parameters of an all-sky imager according to the present invention. Figure 4 This is a density scatter plot comparing the SSA and AOD at 440nm and 675nm with the SGP site, obtained from an embodiment of the machine learning-based algorithm for inverting aerosol optical parameters of an all-sky imager according to the present invention. Figure 5 This is a density scatter plot comparing the SSA and AOD at 870nm and 1020nm with the SGP site, obtained from an embodiment of the machine learning-based algorithm for inverting aerosol optical parameters of an all-sky imager. Detailed Implementation

[0018] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0019] This embodiment uses a machine learning-based algorithm to invert aerosol optical parameters of an all-sky imager, such as... Figure 1 As shown, it includes the following steps: 1) The instrument position for image acquisition (39°59'N, 116°18'E) is fixed, the lens acquisition angle is perpendicular to the ground, the instrument uses an all-sky imager, and raw sky images with time stamps are acquired. The raw sky images include images of the entire sky, the number of raw sky photos is 200,000, the image acquisition time is from April 2018 to January 2020, the resolution is 180dpi, and the image shooting time is extracted from the image name; 2) Image processing of the original sky photo: a) The relative position of the instrument to the surrounding buildings is fixed, and the surrounding buildings are removed by using a black mask; b) Traverse the image and find the brightest circular area, which is considered to be the sun and its surrounding bright areas. Use a black mask to remove the sun and its surrounding bright areas. The radius of the circular area of ​​the black mask should be 0.3 times the radius of the circular colored area obtained after removing the surrounding buildings in step a). c) The iterative threshold segmentation method includes the following steps: randomly set an initial threshold of 150; divide the pixels in the image into two categories based on their brightness compared to the threshold: above the threshold and below the threshold; calculate the average brightness of each of these two categories of pixels, and then average these two average brightness values ​​to obtain a total average value, which is used as the new threshold; divide the pixels into two categories again according to the new threshold, and repeat the above steps until the threshold stabilizes; for pixels that are greater than the stable threshold, they are considered to be pixels that do not meet the clear sky condition and are masked to remove clouds and light spots generated by the lens module, thereby removing all areas that affect the extraction of the brightness of the cloudless sky, and obtaining a sky area image that retains only the clean sky. 3) Mimicking the equal zenith angle scanning mode, all colored pixels are extracted from the sky region image. Each colored pixel is divided into three channels: red (R), green (G), and blue (B). The average values ​​of R, G, and B for all colored pixels are calculated. The red-to-blue ratio (rbr) is calculated by dividing the average R value by the average B value. The average R, G, and B values ​​and rbr value corresponding to the sky region image are combined with time and latitude / longitude to calculate the solar zenith angle. To mimic the optical parameter source of the AERONET website, namely the solar photometer, the scanning mode is equal zenith angle mode, i.e., fixed. With a fixed zenith angle and varying the azimuth angle, a circle is drawn with a radius of 0.6 to 0.8 times the radius of the sky region image, centered at the center of the sky region image. At least 10 pixels are averaged on the circumference, and the RGB values ​​of these pixels are extracted. This is called aerosol scattering information, which reflects multi-angle, multi-position information related to the aerosol scattering phase function and the total brightness. The average values ​​of R, G, and B of all extracted colored pixels, the red-to-blue ratio (rbr), the solar zenith angle, and the aerosol scattering information are called multidimensional features, denoted by x. 4) Obtain the aerosol parameters for the set wavelength collected by the solar photometer from the publicly available database AERONET website. These parameters include aerosol optical thickness (AOD) and single scattering albedo (SSA), referred to as the target variable y, with latitude and longitude of (40.0N, 116.3E). The data quality level is Level 2.0. The horizontal distance between the geographical location of the site corresponding to the aerosol parameters in the publicly available database and the location of the instrument that collected the original sky image is less than 5 km, which is considered to be consistent in spatial location representativeness. 5) Match the multidimensional features with the target variable according to time, create a dataset, and divide it into training and test sets for model training and evaluation of model generalization ability. Store the dataset in a single file and convert it to the DMatrix format specific to ensemble regression models. The time of aerosol parameters in publicly available databases cannot perfectly match the time of acquisition of original sky images. A time match is considered to be within 5 minutes of the time error between the acquisition of original sky images and aerosol parameters in publicly available databases. Time matching is performed using the same wavelength, i.e., using the aerosol optical thickness AOD_440nm and single scattering albedo SSA_440nm simultaneously measured by the AERONET solar photometer at a wavelength of 440 nm as ground truth labels to establish time-matched sample pairs with the input multidimensional feature x. The matching time window is no greater than ±5 minutes, forming a sample set. Divide the dataset into training and independent test sets in an 8:2 ratio. 6) The ensemble regression model employs the extreme gradient boosting regressor XGBoost, which includes multiple base learners. The mean squared error (RMSE) is used as the loss function. The optimal number of boosting iterations N is determined through K-fold cross-validation (K=10) and an early stopping strategy. Subsequently, the optimal number of boosting iterations N is used as the hyperparameter combination for the number of base learners, maximum tree depth (7), learning rate (0.08), subsampling ratio (0.75), column sampling ratio (1.0), and minimum loss reduction for leaf node splits (0). Multidimensional features are used as input to the ensemble regression model, and the target variable is used as the output. The data from step 5) is input into the ensemble regression model. Training is performed using the training set, and validation and consistency checks are performed using the test set. The joint output (SSA_440nm, ...) is obtained through training. An ensemble regression model for AOD_440nm was developed; the ensemble regression model was used to predict the test set, and the predicted vector pairs were obtained and a scatter plot of the density was drawn with the true values ​​on the horizontal axis and the predicted values ​​on the vertical axis; the point cloud density was calculated using Gaussian kernel density estimation, and the unit slope reference line y=x and the linear regression fitting line y=kx+b were simultaneously superimposed in a visual form; the proportional consistency was quantitatively evaluated by the regression coefficient k; and a well-trained ensemble regression model for this wavelength was obtained. 7) Change the wavelength and repeat steps 4) to 6) for 440nm, 675nm, 870nm and 1020nm respectively to obtain the trained ensemble regression model for each wavelength; 8) Collect original sky images for the new time period. After image processing in step 2) and feature extraction in step 3), input them into the ensemble regression model to obtain the inverted aerosol optical thickness (AOD) and aerosol single scattering albedo (SSA). This verifies that the ensemble regression model meets the requirements for high-precision inversion of multiple optical parameters of aerosols.

[0020] The original sky images from the ARM_SGP site (36.6N, 97.5W) from July 2024 to July 2025, matched with AERONET data, were processed in step 2) and feature extraction was performed in step 3). The resulting multidimensional features were then input into the ensemble regression model. The ensemble regression model was tested and verified again using this site, proving that the method of the present invention is universally applicable to all sites.

[0021] like Figure 2 and 3 As shown, the horizontal axis represents the actual values ​​of aerosol parameters, namely the single scattering albedo (SSA) and aerosol optical thickness (AOD) at the target wavelength, obtained from the AERONET website; the vertical axis represents the predicted values ​​of aerosol parameters, obtained from the XGBoost model inversion results after image processing in step 2) and feature extraction processing in step 3) of the all-sky imager data from the Beijing site. Sample basis: The dataset is divided into training and test sets in an 8:2 ratio. Figure 2 and 3 All data in this dataset consists of independent test set samples.

[0022] All subplots contain two types of key reference lines: dashed lines, i.e., the y=x reference line, representing the ideal matching state of "predicted value = actual value", used to intuitively determine the direction of prediction deviation; solid lines, i.e., regression lines: the subplots are labeled with the linear regression equation, such as the SSA870nm subplot labeled "regression line y=0.83x-0.15" and the AOD 1020nm subplot labeled "regression line: y=0.98x+0.01", used to quantitatively assess the consistency between the predicted and actual values.

[0023] Numerical range: The horizontal axis scale is adjusted according to the parameter type. The horizontal axis range of the SSA subplot is concentrated between 0.80 and 1.00, which conforms to the physical meaning of SSA: 0 ≤ SSA ≤ 1. The horizontal axis range of the AOD subplot varies with wavelength, reaching 3.5 at 440nm and approximately 1.75 at 870nm and 1020nm. The numerical ranges of the horizontal and vertical axes are consistent with the scale.

[0024] Subplot grouping: Divided into SSA and AOD groups according to parameter type, listed in two columns; each group is arranged sequentially by wavelength (440nm, 675nm, 870nm, 1020nm). Figure 2 and 3 In the middle, each row has two subgraphs, with SSA on the left and AOD on the right.

[0025] Model performance evaluation based on Beijing site: 1. Figure 2 and 3 The SSA group in the middle includes four sub-plots in the left column, covering 440nm, 675nm, 870nm, and 1020nm: All wavelength SSA data values ​​are concentrated in the range of 0.80-1.00, which is in line with the physical range of SSA. The data points are distributed along the y=x reference line as a whole: there is no obvious deviation at 440nm, 675nm, and 1020nm; although there is a slight low prediction deviation of "y=0.83x-0.15" at 870nm, the accuracy meets the standard.

[0026] 2. Figure 2 and 3 The middle AOD group includes four sub-plots in the right column, covering 440nm, 675nm, 870nm, and 1020nm: At low AOD (<1.0), all AOD data points at each wavelength closely follow the y=x reference line, while at high AOD (>1.0), they are slightly underestimated, which is consistent with the commonality of "slight systematic underestimation of high AOD" and the overall correlation is high.

[0027] like Figure 4 and 5 As shown, the vertical axis represents the predicted aerosol parameters, derived from the XGBoost model inversion results of all-sky imager data from the SGP site after image processing in step 2) and feature extraction in step 3); the descriptions of the remaining image elements are as follows. Figure 2 and 3 All are consistent.

[0028] Model Performance Evaluation Based on SGP Sites 1. Figure 4 and Figure 5 The SSA group in the middle includes four sub-plots in the left column, covering 440nm, 675nm, 870nm, and 1020nm: The actual and predicted values ​​of the single-scattering albedo (SSA) at each wavelength are concentrated in the range of 0.2-1.0, which is consistent with the physical range of SSA. The data points are distributed along the reference line y=x, with no obvious systematic bias. The regression results for 440nm (regression line y=0.84x+0.14), 675nm (y=0.84x+0.14), 870nm (y=0.85x+0.13), and 1020nm (y=0.87x+0.11) all show high correlation.

[0029] 2. Figure 4 and Figure 5 The middle AOD group includes four sub-plots in the right column, covering 440nm, 675nm, 870nm, and 1020nm: The data points for aerosol optical thickness (AOD) at various wavelengths are densely distributed, with the low AOD range closely following the y=x reference line. The coefficients for 440nm (y=0.94x+0.01), 675nm (y=0.96x+0.00), 870nm (y=0.95x+0.00), and 1020nm (y=0.94x+0.00) are close to 1.0, indicating high correlation. In particular, the data points for wavelengths such as 675nm almost perfectly match the reference line. Overall, the model's AOD inversion performance is satisfactory.

[0030] The inversion results of this invention are good. In the Beijing site validation set, the simultaneous inversion of AOD and SSA parameters outperforms the separate inversion of these two parameters. When using the XGBoost model to simultaneously invert AOD and SSA, the correlation coefficient is 0.93 for the Beijing site validation set and 0.94 for the SGP site validation set. This demonstrates that the inversion model has good spatiotemporal transferability. Although there is a slight systematic underestimation at high AOD values, the overall performance of the model is still satisfactory, and it can be considered that the accuracy of extracting aerosol optical parameters from all-sky imagers using this algorithm is comparable to AERONET. These findings collectively demonstrate the feasibility of using all-sky imagers for aerosol optical parameter inversion, providing a new technical path for future atmospheric remote sensing research. The experimental design is shown in the table below: 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 machine learning-based algorithm for inverting aerosol optical parameters of an all-sky imager, characterized in that, The inversion algorithm includes the following steps: 1) Collect raw sky images with time stamps; 2) Perform image processing on the original sky photo to remove all areas that affect the extraction of cloudless sky radiance. Specifically, the iterative threshold segmentation method is used to remove clouds and light spots generated by the lens module, resulting in a sky area image that retains only the clean sky. 3) Imitating the equal zenith angle scanning mode, feature extraction is performed on the sky region image, and the average values ​​of R, G and B, red-blue ratio rbr value, solar zenith angle and aerosol scattering information of all colored pixels are extracted, which are called multidimensional features; 4) Obtain the aerosol parameters for the set wavelength from a publicly available database, including aerosol optical thickness (AOD) and single scattering albedo (SSA), which are referred to as target variables; 5) Match the multidimensional features with the target variable according to time, use them as a dataset, and divide them into training and test sets for model training and for evaluating the model's generalization ability. 6) Using an ensemble regression model, the multidimensional features are used as the input of the ensemble regression model, and the target variable is used as the output of the ensemble regression model. The data from step 5) are input into the ensemble regression model, and the training set is used for training and the test set is used for validation and consistency verification to obtain a well-trained ensemble regression model for this wavelength. 7) Repeat steps 4) to 6) multiple times by changing the wavelength to obtain multiple trained ensemble regression models for each wavelength; 8) Collect original sky images for the new time period, and input the multidimensional features obtained from the image processing in step 2) and feature extraction in step 3) into the integrated regression model to obtain the inverted aerosol optical thickness (AOD) and aerosol single scattering albedo (SSA).

2. The inversion algorithm according to claim 1, characterized in that, In step 1), the instrument for acquiring the original sky image is fixed in position, the lens is perpendicular to the ground, and the original sky image includes the entire sky.

3. The inversion algorithm according to claim 1, characterized in that, In step 2), the original sky photograph undergoes image processing, including the following steps: a) Use a black mask to cut out surrounding buildings; b) Traverse the image and find the brightest circular area, which is considered to be the sun and its surrounding bright areas. Use a black mask to remove the sun and its surrounding bright areas. c) Use an iterative thresholding method to repeatedly binarize the image until the threshold stabilizes. Use a black mask to cover the brighter areas and remove the light spots generated by the cloud and lens module.

4. The inversion algorithm according to claim 1, characterized in that, In step 3), all colored pixels are extracted from the sky region image. Each colored pixel is divided into three channels: red (R), green (G), and blue (B). The average values ​​of R, G, and B for all colored pixels are calculated. The red-to-blue ratio (rbr) is calculated by dividing the average R value by the average B value. The zenith angle of the sun is calculated by combining the average R, G, and B values ​​and the rbr value corresponding to the sky region image with time and latitude and longitude. A circle is drawn with a radius of 0.6 to 0.8 times the radius of the sky region image and the center of the circle located at the center point of the sky region image. At least 10 pixels are taken on average on the circumference of the circle, and the RGB values ​​of these multiple pixels are extracted, which are called aerosol scattering information.

5. The inversion algorithm according to claim 1, characterized in that, In step 4), if the spatial difference between the geographical location of the site corresponding to the aerosol parameters in the public database and the location of the instrument that acquired the original sky image is within the spatial threshold, it is considered to be consistent in spatial location representativeness.

6. The inversion algorithm according to claim 1, characterized in that, In step 5), a time match is considered to be achieved if the time difference between the acquired original sky image and the aerosol parameters from the publicly available database is within a time threshold.

7. The inversion algorithm according to claim 1, characterized in that, In step 6), the ensemble regression model uses the extreme gradient boosting regressor XGBoost, which includes multiple base learners. The mean squared error (RMSE) is used as the loss function. The optimal number of boosting iterations N is determined by K-fold cross-validation and an early stopping strategy, where K is a natural number ≥ 5. Then, the optimal number of boosting iterations N is used as the number of base learners, and the hyperparameter combination of maximum tree depth, learning rate, subsampling ratio, column sampling ratio, and minimum loss reduction for leaf node splitting is set to train the ensemble regression model with joint output.

Images