A method and system for measuring spectral distribution of a PAR

Abstract

The application discloses a kind of measurement method and measurement system of the spectral distribution of PAR, including by the geometric calibration and radiation calibration of all-sky imager, the observation zenith angle and observation azimuth of each pixel of all-sky image and RGB waveband radiance value are acquired;Through cloud detection algorithm, all-sky image is divided into clear sky and cloudy condition;For clear sky all-sky image, combined with radiation transfer model, PAR spectral radiation distribution algorithm reconstruction is carried out to obtain the spectral distribution of PAR;For the image with cloud, a deep learning model is used to obtain the spectral distribution of PAR, and finally the spectral distribution of the whole PAR is obtained.The application has the advantages that: the spectral distribution of PAR is reconstructed using all-sky image, and by combining physical radiation model and deep learning model, the spectral irradiance of visible light band reaching the ground, i.e.PAR spectral distribution information, is obtained from the image captured by all-sky imager.

Description

Technical Field

[0001] This invention relates to the technical field of image algorithms, and in particular to a method and system for measuring the spectral distribution of PAR. Background Technology

[0002] Photosynthetically Active Radiation (PAR) refers to the visible light radiation from the sun incident on the Earth's surface that can be utilized by plants for photosynthesis, with wavelengths ranging from 400 to 700 nm. PAR is a fundamental energy source for plant life activities, organic matter synthesis, and yield formation, directly impacting plant growth, development, yield, and yield quality. Therefore, PAR is considered a major ecological factor influencing plant growth and is widely used in terrestrial ecosystem research, vegetation productivity calculation models, and CO2 exchange models. However, due to different stages of vegetation development, the intensity of photosynthesis is not uniformly distributed within the PAR spectral range. By observing and studying plant growth and development under different PAR spectral distributions, vegetation growth models with high predictive accuracy can be established, providing valuable information and support for agricultural production, forest management, and environmental protection. Therefore, it is essential to measure PAR, especially its spectral distribution.

[0003] A typical ground-based PAR measurement device is a photosynthetically active radiation meter (PAR meter), which uses a silicon photodetector and a visible light optical filter to convert the received light signal into an electrical signal, thereby measuring PAR. This device can measure the total radiation within its wavelength band, but it cannot measure the corresponding spectral distribution. Measuring its spectral distribution primarily relies on spectrometers. With high-quality observation and maintenance, reference radiation data can usually be provided; however, spectrometers are expensive and require precise maintenance, limiting their widespread use and failing to meet the growing demand. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention proposes a method and system for measuring the spectral distribution of PAR, the specific technical solution of which is as follows:

[0005] A method for measuring the spectral distribution of PAR, comprising:

[0006] By performing geometric and radiometric calibrations on the all-sky imager, the observation zenith angle, observation azimuth angle, and RGB band radiance values ​​of each pixel in the all-sky image are obtained.

[0007] The cloud detection algorithm is used to divide the entire sky image into clear sky and cloudy sky conditions. For clear sky images, the PAR spectral distribution is reconstructed by combining the radiative transfer mode algorithm. For images with clouds, a deep learning model is used to obtain the PAR spectral distribution, and finally the spectral distribution of PAR for the entire sky is obtained.

[0008] Optionally, for clear-sky all-sky images, the PAR spectral distribution is reconstructed using a PAR spectral radiative distribution algorithm combined with the radiative transfer mode, including:

[0009] Based on the radiometric calibration results, the grayscale values ​​of the input all-sky image pixels are converted into radiance values; based on the geometric calibration results, the observed zenith angle and azimuth angle of each pixel are obtained.

[0010] Based on the all-sky radiance value and the observation angle, the aerosol optical characteristic parameters under clear sky are obtained by inversion. The aerosol optical characteristic parameters include aerosol optical thickness, single scattering rate, and asymmetry factor parameter information.

[0011] Using the aerosol optical properties parameters under clear skies and the local surface visible light albedo as inputs to the radiative transfer mode, the radiative transfer calculation outputs the visible light band solar irradiance incident on the surface under clear skies, i.e., the PAR spectral distribution.

[0012] Optionally, based on the all-sky radiance values ​​and the observation angle, the optical properties parameters of aerosols under clear skies can be retrieved, including:

[0013] An inversion lookup table is established using radiative transfer modes. The input parameters of the lookup table are solar zenith angle, observed zenith angle, relative azimuth angle, aerosol optical thickness, single irradiance, and asymmetry factor. The output parameter is the RGB band radiance value.

[0014] Linear interpolation is performed on the established inversion lookup table using the actual measured solar zenith angle, observed zenith angle, and relative azimuth angle;

[0015] Using the root mean square error between the interpolated RGB band radiance value and the actual observed RGB band radiance value as the cost, the optimization algorithm iterates over this cost value. The aerosol optical thickness, single scattering rate, and asymmetry factor at the minimum cost value are the required aerosol optical characteristic parameters.

[0016] Optionally, for images with clouds, a deep learning model is used to obtain the PAR spectral radiation distribution, including:

[0017] Build a multi-source input neural network framework;

[0018] A dataset for training and testing a multi-source input neural network was established. First, all-sky images were collected and the corresponding shooting time information was saved. The solar zenith angle was calculated based on the shooting time information, and the spectrometer observation results at the corresponding time were obtained. The training set and the test set were created using all-sky images and solar zenith angle as input data and spectrometer observation results as output data.

[0019] The multi-source input neural network framework is trained using a training dataset, and the trained network model is saved. The training process iteratively optimizes the loss value and calculates the accuracy of the test set in each iteration until the loss value and accuracy approach a set range, at which point the model weights are saved.

[0020] By inputting real-time all-sky images and solar zenith angles into a trained neural network model, PAR spectral irradiance data under cloud conditions can be obtained.

[0021] Optionally, a multi-source input neural network framework can be built, including:

[0022] First, the entire sky image is processed by three convolutional and pooling layers. Then, the input is flattened by a flattening layer. The solar zenith angle at the corresponding time is then stitched and fused with the flattened data after passing through a fully connected layer. After passing through two fully connected layers, the data enters the output layer, and finally, the PAR spectral irradiance data is output.

[0023] Optionally, through geometric calibration of the all-sky imager, including:

[0024] Place the checkerboard calibration plate in front of the all-sky imager and take multiple all-sky images with the checkerboard calibration plate attached.

[0025] The corner points of the chessboard grid were obtained, and the distortion parameters k1, k2, k3, k4 of the all-sky imager and the camera intrinsic parameter f were calibrated using the Kannala-Brandt lens model from the OpenCV library. x f y c x c y ;

[0026] Based on the distortion parameters k1, k2, k3, k4 and the camera intrinsic parameter f x f y c x c y Obtain the observed zenith angle and relative azimuth angle for each pixel;

[0027] The relative azimuth angle is further corrected based on the solar azimuth angle to obtain the true observed azimuth angle distribution of each pixel in the all-sky image.

[0028] Optionally, based on the distortion parameters k1, k2, k3, k4 and the camera intrinsic parameter fx f y c x c y Obtain the observed zenith angle and relative azimuth angle for each pixel, including:

[0029] Suppose the pixel coordinates of a certain point are (u, v), and the camera intrinsic parameters f are known. x f y c x c y According to the following formula (1), its position (x, y) on the imaging plane can be obtained.

[0030]

[0031] By solving the following formulas (2) and (3), the observed zenith angle θ and relative azimuth angle can be obtained.

[0032]

[0033]

[0034] Optionally, radiometric calibration of the all-sky imager can be performed, including:

[0035] An all-sky imager was used to photograph an integrating sphere under different lighting intensities. The multiple images were then flattened to obtain the flattened image.

[0036] Establish a linear relationship between grayscale values ​​and radiance for the flat-field corrected image;

[0037] Based on the linear relationship and the grayscale values ​​obtained from actual observation, the radiance values ​​of each pixel in the entire sky image are calculated.

[0038] Optionally, the step of using a cloud detection algorithm to classify the entire sky image into clear sky and cloudy conditions includes:

[0039] Based on the input all-sky image, for the area around the sun in the image, if the maximum value of the corresponding channel is less than a certain threshold T1, it is considered a cloudy image; otherwise, it is used for subsequent spectral testing.

[0040] Remove the area around the sun and detect the ratio RBR of the remaining pixel channels R to B. If the average RBR is greater than a certain threshold T2, it is a cloudy image; otherwise, it continues to participate in subsequent spectral tests.

[0041] For the full-sky images taken at the corresponding time and the previous time, remove the area around the sun and calculate the grayscale difference between the two image channels R. If the proportion of pixels with a grayscale difference greater than a certain threshold T3 exceeds the threshold T4, it is a cloudy image; otherwise, it is a clear sky image.

[0042] A system for measuring the spectral distribution of PAR, comprising:

[0043] Calibration module; used for geometric and radiometric calibration of the all-sky imager;

[0044] The cloud detection module is used to divide the entire sky image into clear sky and cloud cover conditions;

[0045] The clear-sky radiation distribution detection module is used to reconstruct the PAR spectral distribution by combining the radiative transfer mode with the PAR spectral radiation distribution algorithm to obtain the PAR spectral distribution.

[0046] The cloud radiation distribution detection module uses a deep learning model to obtain the PAR spectral radiation distribution.

[0047] The synthesis module is used to synthesize the spectral distribution of PAR throughout the day based on the PAR spectral distribution output by the clear sky radiation distribution detection module and the PAR spectral radiation distribution output by the cloudy radiation distribution detection module.

[0048] The advantages of this invention are:

[0049] (1) This invention provides a method and system for measuring the spectral distribution of PAR. It uses all-sky images to reconstruct the spectral distribution of PAR. By combining physical radiation models and deep learning models, it obtains the visible light band spectral irradiance reaching the Earth's surface from the images captured by the all-sky imager, i.e., the PAR spectral distribution information.

[0050] (2) This invention uses only an all-sky imager to observe PAR and its spectral distribution. Compared with spectrometer observation, the all-sky imager has lower cost, higher stability, and better universality.

[0051] (3) The present invention uses a physical radiation model to reconstruct real-time PAR and its spectral distribution from clear sky images. During the reconstruction process, not only can aerosol optical property information be obtained, but also the PAR spectral distribution can be obtained quickly and conveniently based on the principle of radiative transfer.

[0052] (4) This invention uses a deep learning neural network model to reconstruct real-time PAR and its spectral distribution from cloud-covered all-sky images, solving the problem of difficulty in PAR reconstruction caused by changes in the three-dimensional structure of clouds. With the help of the powerful learning ability of deep learning, the spectral distribution of PAR under clouds can be obtained more accurately. Attached Figure Description

[0053] Figure 1 This is a flowchart of the measurement method of the present invention.

[0054] Figure 2 This is a flowchart of the cloud detection algorithm.

[0055] Figure 3 This is a framework diagram of the PAR spectral component reconstruction algorithm under clear sky conditions.

[0056] Figure 4 This is a diagram of a multi-source input neural network framework.

[0057] Figure 5 This is a schematic diagram of the measurement system of the present invention. Detailed Implementation

[0058] In existing technologies, typical photosynthetically active radiation (PAR) meters cannot measure the corresponding spectral distribution. Measurement of this distribution primarily relies on spectrometers. High-quality observation and maintenance can typically provide baseline radiation data; however, spectrometers are expensive and require precise maintenance, limiting their widespread development and failing to meet the growing demand. All-sky imagers, using RGB three-channel photosensitive pixels, receive radiation from all directions of the sky and rapidly capture the entire sky's condition through a specific exposure time. This equipment is low-cost, highly stable, and is currently widely deployed at various meteorological stations. The captured all-sky images can observe the optical properties of clouds and aerosols and, to some extent, represent the background radiance of the sky, providing feasibility for reconstructing PAR and its spectral distribution.

[0059] as follows Figure 1 As shown, a method for measuring the spectral distribution of PAR is characterized by comprising:

[0060] S1. By performing geometric and radiometric calibration on the all-sky imager, obtain the observation zenith angle and observation azimuth angle of each pixel in the all-sky image and its RGB band radiance value.

[0061] S2. Using a cloud detection algorithm, the entire sky image is divided into clear sky and cloudy conditions;

[0062] S3. For clear sky images, the PAR spectral distribution is reconstructed using a PAR spectral distribution algorithm based on the radiative transfer mode. For images with clouds, a deep learning model is used to obtain the PAR spectral distribution, and finally the spectral distribution of PAR throughout the day is obtained.

[0063] In detail, step S1, which involves geometric calibration of the all-sky imager, includes:

[0064] SA11. Place the checkerboard calibration plate in front of the all-sky imager and take multiple all-sky images with the checkerboard calibration plate attached.

[0065] SA12. Obtain the corner points of the chessboard grid. Using the Kannala-Brandt lens model from the OpenCV library, calibrate and obtain the distortion parameters k1, k2, k3, k4 of the all-sky imager and the camera intrinsic parameter f.x f y c x c y ;

[0066] SA13, based on the distortion parameters k1, k2, k3, k4 and the camera intrinsic parameter f x f y c x c y Obtain the observed zenith angle and relative azimuth angle for each pixel;

[0067] Specifically, suppose the pixel coordinates of a certain point are (u, v), and the camera intrinsic parameters f are known. x f y c x c y According to the following formula (1), its position (x, y) on the imaging plane can be obtained.

[0068]

[0069] By solving the following formulas (2) and (3), the observed zenith angle θ and relative azimuth angle can be obtained.

[0070]

[0071]

[0072] SA14. Further correct the relative azimuth angle based on the solar azimuth angle. This yields the true distribution of the observed azimuth angles for each pixel in the entire sky image. Specifically, this includes:

[0073] SA141. For a full-sky image where the sun is not obscured, obtain the pixel coordinates (u) of the center point of the solar region on the image. s v s Based on the image imaging time and location, the solar zenith angle θ at the corresponding moment is calculated. s With solar azimuth

[0074] SA142, the obtained pixel coordinates (u s v s The relative azimuth angle can be obtained by formulas (1), (2), and (3).

[0075] SA143, Calculate the relative azimuth angle With solar azimuth The difference between

[0076]

[0077] For any pixel, its observation azimuth angle is the relative azimuth angle. and The sum of the differences It is a constant. For the pixel at the location of the sun, the sun's azimuth angle θ is... s Sum and Difference The sum of the two is indeed the solar azimuth. However, for other non-solar pixels in the all-sky image, the relative azimuth angle is added to a fixed value. It is for observing the azimuth angle, not the solar azimuth angle.

[0078] In detail, the purpose of radiometric calibration of the all-sky imager in step S1 is to establish the relationship between the grayscale values ​​of the RGB channels of the all-sky image and the radiance values, so that the all-sky radiation information can be obtained later based on the image grayscale values. The principle of radiometric calibration is as follows:

[0079]

[0080] Among them, E Δλ ΔΩ represents the irradiance acquired within the pixel's field of view, ΔΩ is the pixel's solid angle, and Δλ is the photosensitive wavelength range of that pixel. Meanwhile, E... Δλ Related to image grayscale value DN, exposure time Δt, and pixel area A in Relevant. For a stable all-sky imager, the exposure time, pixel area, and solid angle distribution are generally constant, and the observed illumination I... Δλ It has a linear relationship with its image grayscale value DN, which is set to C here. 2λ .

[0081] The main content of radiometric calibration for the all-sky imager is to obtain C 2λ Because all-sky images generated by all-sky imagers contain numerous pixels, there are differences in radiometric perception among these pixels. Therefore, radiometric calibration of all-sky imagers mainly includes two parts: first, flat-field correction to eliminate response differences between pixels in the all-sky image; and second, establishing a linear relationship between grayscale values ​​and radiance. The main steps are as follows:

[0082] SB11. Use an all-sky imager to photograph the integrating sphere under different lighting intensities. Perform planar correction on the multiple images obtained to acquire the planar corrected image. The specific steps are as follows:

[0083] SB111. Subtract the dark field image from each image. The dark field image is obtained by imaging the darkroom with the all-sky imager.

[0084] SB112. Divide the result of SB111 by the grayscale ratio matrix to obtain the flat-field corrected image. The grayscale ratio matrix is ​​obtained as follows: The all-sky imager images a uniformly gray object (such as professional white balance paper) under uniform illumination to obtain a uniform field image; subtract the darkroom image from SB111 from the uniform field image to obtain a difference image; using the grayscale value at the zenith of the difference image as a reference, calculate the grayscale ratio of each pixel to the reference value to obtain the grayscale ratio matrix.

[0085] SB12. Establish a linear relationship between grayscale values ​​and radiance for the flat-field corrected image, including:

[0086] SB121. Convolve the spectral radiance provided by the integrating sphere using the camera's RGB response curve to obtain the RGB channel radiance at the zenith.

[0087] SB122. Establish a linear relationship between the RGB grayscale value at the zenith and the channel radiance value;

[0088] SB13. Based on the linear relationship and the grayscale values ​​obtained from actual observation, the radiance values ​​of each pixel in the entire sky image are calculated.

[0089] like Figure 2 As shown, the cloud detection algorithm divides the entire sky image into clear sky and cloudy conditions. The specific steps are as follows:

[0090] S21. Based on the input all-sky image, for the area surrounding the sun in the image, if the maximum value of the corresponding channel is less than a certain threshold T1, it is considered a cloudy image; otherwise, it participates in subsequent spectral testing. In the test, the area surrounding the sun can refer to the area where the pixel scattering angle is less than 15°. Generally, when there are no clouds, this area will be overexposed, and the gray value will reach its maximum value (255 for an 8-bit camera). Therefore, the threshold T1 can be set to a number close to the maximum value (250 for an 8-bit camera).

[0091] S22. Remove the area around the sun and detect the ratio of channel R to channel B (RBR) of the remaining pixels. If the average RBR is greater than a certain threshold T2, the image is considered cloudy; otherwise, it continues to participate in subsequent spectral tests. In this scheme, the threshold T2 can be adjusted based on the RBR values ​​of multiple manually selected clear-sky images.

[0092] S23. For the full-sky images taken at the corresponding time and the previous time, remove the area around the sun and calculate the grayscale difference between the two image channels R. If the proportion of pixels with a grayscale difference greater than a certain threshold T3 exceeds a certain threshold T4, then it is a cloudy image; otherwise, it is a clear sky image. This step is mainly based on the fact that clouds change rapidly, while the radiance of clear sky images remains basically unchanged. The thresholds T3 and T4 can be adjusted and determined based on historical data.

[0093] When there are no clouds, the main uncertainty affecting PAR is aerosols, and obtaining the optical properties of aerosols is a prerequisite for reconstruction. For example... Figure 3 As shown in Figure 3, for clear sky images, the PAR spectral distribution is reconstructed using a PAR spectral distribution algorithm based on the radiative transfer mode, including:

[0094] SA31. Based on the radiometric calibration results, convert the grayscale values ​​of the input all-sky image pixels into radiance values; based on the geometric calibration results, obtain the observed zenith angle and azimuth angle of each pixel.

[0095] SA32. Based on the all-sky radiance values ​​and observation angle, the aerosol optical properties parameters under clear skies are retrieved, including aerosol optical thickness, single-shot scattering rate, and asymmetry factor parameters. This step mainly includes the following:

[0096] SA321. An inversion lookup table is established through the radiative transfer mode. The input parameters of the lookup table are solar zenith angle, observed zenith angle, relative azimuth angle, aerosol optical thickness, single irradiance, and asymmetry factor. The output parameter is the RGB band radiance value.

[0097] SA322: Linear interpolation is performed on the established inversion lookup table using the actual measured solar zenith angle, observed zenith angle, and relative azimuth angle;

[0098] SA323. Using the root mean square error between the interpolated RGB band radiance value and the actual observed RGB band radiance value as the cost, the optimization algorithm is used to iterate the cost value continuously. The aerosol optical thickness, single scattering rate, and asymmetry factor at the minimum cost value are the required aerosol optical characteristic parameters.

[0099] SA33. Using the aerosol optical characteristic parameters under clear sky and the local surface visible light albedo as inputs to the radiative transfer mode, the radiative transfer calculation outputs the visible light band solar irradiance incident on the surface under clear sky, i.e., the PAR spectral distribution. By integrating the spectral results of the output, the PAR can be further obtained.

[0100] When clouds are present, the solar spectral radiation incident on the Earth's surface is strongly affected by the clouds, making it difficult to calculate. This patent employs a deep learning method to build a multi-source input neural network framework. When clouds are present, it takes a full-sky image and the corresponding solar zenith angle as input, and outputs PAR spectral irradiance data. Detailed steps are as follows:

[0101] SB31. Construct a multi-source input neural network framework model as follows: Figure 4As shown, the steps for building the multi-source input neural network framework model are as follows: First, the entire sky image is processed by three convolutional and pooling layers. Then, the input flattening layer is used to flatten the image. The solar zenith angle at the corresponding moment is then stitched and fused with the flattened data after passing through a fully connected layer. After passing through two fully connected layers, the image enters the output layer, and finally, the PAR spectral irradiance data is output.

[0102] SB32. Establish a dataset for training and testing the neural network model in step SB31. First, collect all-sky images and save the corresponding shooting time information. Calculate the solar zenith angle based on the shooting time information and obtain the spectrometer observation results at the corresponding times. Using the all-sky images and solar zenith angle as input data and the spectrometer observation results as output data, create training and testing sets.

[0103] SB33. The multi-source input neural network framework is trained using the training dataset, and the trained network model is saved. The training process iteratively optimizes the loss value and calculates the accuracy of the test set in each iteration until the loss value and accuracy approach the set range, at which point the model weights are saved.

[0104] SB34. By inputting real-time all-sky images and solar zenith angles into a trained neural network model, PAR spectral irradiance data under cloud conditions can be obtained.

[0105] like Figure 5 As shown, this application also discloses a system for measuring the spectral distribution of PAR, comprising:

[0106] Calibration module 1; used for geometric and radiometric calibration of the all-sky imager;

[0107] Cloud detection module 2 is used to divide the entire sky image into clear sky and cloud conditions;

[0108] Clear-sky radiation distribution detection module 3 is used to reconstruct the PAR spectral distribution by combining the radiative transfer mode with the PAR spectral radiation distribution algorithm to obtain the PAR spectral distribution.

[0109] Cloud radiation distribution detection module 4 uses a deep learning model to obtain the PAR spectral radiation distribution;

[0110] Synthesis module 5 is used to synthesize the spectral distribution of PAR throughout the day based on the PAR spectral distribution output by clear sky radiation distribution detection module 3 and the PAR spectral radiation distribution output by cloudy radiation distribution detection module 4.

[0111] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for measuring the spectral distribution of PAR, characterized in that, include: By performing geometric and radiometric calibrations on the all-sky imager, the observation zenith angle, observation azimuth angle, and RGB band radiance values ​​of each pixel in the all-sky image are obtained. The cloud detection algorithm is used to divide the entire sky image into clear sky and cloudy sky conditions. For clear sky images, the PAR spectral distribution is reconstructed by combining the radiative transfer mode algorithm. For images with clouds, a deep learning model is used to obtain the PAR spectral distribution, and finally the spectral distribution of PAR for the entire sky is obtained. For images with clouds, a deep learning model is used to obtain the PAR spectral radiance distribution, including: Build a multi-source input neural network framework; A dataset for training and testing a multi-source input neural network was established. First, all-sky images were collected and the corresponding shooting time information was saved. The solar zenith angle was calculated based on the shooting time information, and the spectrometer observation results at the corresponding time were obtained. The training set and the test set were created using all-sky images and solar zenith angle as input data and spectrometer observation results as output data. The multi-source input neural network framework is trained using a training dataset, and the trained network model is saved. The training process iteratively optimizes the loss value and calculates the accuracy of the test set in each iteration until the loss value and accuracy approach a set range, at which point the model weights are saved. By inputting real-time all-sky images and solar zenith angles into a trained neural network model, PAR spectral irradiance data under cloud conditions can be obtained.

2. The measurement method according to claim 1, characterized in that, For clear-sky full-sky images, the PAR spectral distribution is reconstructed using a PAR spectral radiative distribution algorithm combined with radiative transfer modes, including: Based on the radiometric calibration results, the grayscale values ​​of the input all-sky image pixels are converted into radiance values; based on the geometric calibration results, the observed zenith angle and azimuth angle of each pixel are obtained. Based on the all-sky radiance value and the observation angle, the aerosol optical characteristic parameters under clear sky are obtained by inversion. The aerosol optical characteristic parameters include aerosol optical thickness, single scattering rate, and asymmetry factor parameter information. Using the aerosol optical properties parameters under clear skies and the local surface visible light albedo as inputs to the radiative transfer mode, the radiative transfer calculation outputs the visible light band solar irradiance incident on the surface under clear skies, i.e., the PAR spectral distribution.

3. The measurement method according to claim 2, characterized in that, Based on the all-sky radiance values ​​and observation angle, the optical properties of aerosols under clear skies were retrieved, including: An inversion lookup table is established using radiative transfer modes. The input parameters of the lookup table are solar zenith angle, observed zenith angle, relative azimuth angle, aerosol optical thickness, single irradiance, and asymmetry factor. The output parameter is the RGB band radiance value. Linear interpolation is performed on the established inversion lookup table using the actual measured solar zenith angle, observed zenith angle, and relative azimuth angle; Using the root mean square error between the interpolated RGB band radiance value and the actual observed RGB band radiance value as the cost, the optimization algorithm iterates over this cost value. The aerosol optical thickness, single scattering rate, and asymmetry factor at the minimum cost value are the required aerosol optical characteristic parameters.

4. The measurement method according to claim 1, characterized in that, Building a multi-source input neural network framework includes: First, the entire sky image is processed by three convolutional and pooling layers. Then, the input is flattened by a flattening layer. The solar zenith angle at the corresponding time is then stitched and fused with the flattened data after passing through a fully connected layer. After passing through two fully connected layers, the data enters the output layer, and finally, the PAR spectral irradiance data is output.

5. The measurement method according to claim 1, characterized in that, Through geometric calibration of the all-sky imager, including: Place the checkerboard calibration plate in front of the all-sky imager and take multiple all-sky images with the checkerboard calibration plate attached. The corner points of the chessboard grid were obtained, and the distortion parameters of the all-sky imager were calibrated using the Kannala-Brandt lens model from the OpenCV library. With camera internal parameters ; Based on distortion parameters With camera internal parameters Obtain the observed zenith angle and relative azimuth angle for each pixel; The relative azimuth angle is further corrected based on the solar azimuth angle to obtain the true observed azimuth angle distribution of each pixel in the all-sky image.

6. The measurement method according to claim 5, characterized in that, Based on distortion parameters With camera internal parameters Obtain the observed zenith angle and relative azimuth angle for each pixel, including: Assume that the pixel coordinates of a certain point are Given camera intrinsic parameters According to the following formula (1), its position on the imaging plane can be obtained. ; （1） The observed zenith angle can be obtained by solving the following formulas (2) and (3). and relative azimuth ; （2） （3）。 7. The measurement method according to claim 1, characterized in that, Radiometric calibration of the all-sky imager, including: An all-sky imager was used to photograph an integrating sphere under different lighting intensities. The multiple images were then flattened to obtain the flattened image. Establish a linear relationship between grayscale values ​​and radiance for the flat-field corrected image; Based on the linear relationship and the grayscale values ​​obtained from actual observation, the radiance values ​​of each pixel in the entire sky image are calculated.

8. The measurement method according to claim 1, characterized in that, The cloud detection algorithm categorizes the entire sky image into clear sky and cloudy conditions, including: Based on the input all-sky image, for the area around the sun in the image, if the maximum value of the corresponding channel is less than a certain threshold T1, it is considered a cloudy image; otherwise, it is used for subsequent spectral testing. Remove the area around the sun and detect the ratio RBR of the remaining pixel channels R to B. If the average RBR is greater than a certain threshold T2, it is a cloudy image; otherwise, it continues to participate in subsequent spectral tests. For the full-sky images taken at the corresponding time and the previous time, remove the area around the sun and calculate the grayscale difference between the two image channels R. If the proportion of pixels with a grayscale difference greater than a certain threshold T3 exceeds the threshold T4, it is a cloudy image; otherwise, it is a clear sky image.

9. A measurement system for implementing the method for measuring the spectral distribution of PAR as described in claim 1, characterized in that, include: Calibration module; Used for geometric and radiometric calibration of all-sky imagers; The cloud detection module is used to divide the entire sky image into clear sky and cloud cover conditions; The clear-sky radiation distribution detection module is used to reconstruct the PAR spectral distribution by combining the radiative transfer mode with the PAR spectral radiation distribution algorithm to obtain the PAR spectral distribution. The cloud radiation distribution detection module uses a deep learning model to obtain the PAR spectral radiation distribution. The synthesis module is used to synthesize the spectral distribution of PAR throughout the day based on the PAR spectral distribution output by the clear sky radiation distribution detection module and the PAR spectral radiation distribution output by the cloudy radiation distribution detection module.

Images