A ground-based all-sky image cloud cover calculation method and system

Abstract

The present application relates to a kind of ground-based all-sky image cloud amount calculation method and system, comprising the following steps: acquisition ground-based all-sky image and pre-processing, the image data obtained after pre-processing is labeled and data enhancement;The image set is trained using cloud detection model based on channel weighting-feature fusion (Channel Weighting-Feature Fusion, CWFF) structure, and prediction matrix is obtained;The prediction matrix is segmented to determine the proportion of cloud pixels, to determine the corresponding cloud amount.The present application effectively enhances the representation learning ability of cloud detection model and cloud detection effect, greatly improves the calculation efficiency about cloud amount, and has certain significance to the important evaluation of ground-based photoelectric telescope site selection.

Description

Technical Field

[0001] This invention belongs to the field of astronomy, specifically relating to a ground-based method and system for calculating cloud cover in all-sky images. Background Technology

[0002] Currently, cloud cover calculation mainly employs satellite detection and ground-based all-sky cloud cover camera monitoring. Satellite detection offers advantages such as convenient data acquisition, multi-band detection, and long-term data accumulation. However, satellite-based cloud cover calculation is limited by the spatial and temporal resolution of the detected images; typically, cloud cover over a site is only covered by a few pixels in the image, and the detection time interval is on the order of days or hours. Ground-based all-sky cameras, on the other hand, offer stable performance and real-time monitoring, and are therefore widely used for cloud cover calculation and evaluation at various sites. With the development of computer technology and big data processing technology, cloud cover calculation methods are gradually shifting from traditional manual observation to automated and intelligent processing. These methods mainly include thresholding, clustering, machine learning, and deep learning.

[0003] Deep learning is an algorithm that attempts to perform high-level abstraction of data using multiple processing layers containing complex structures or consisting of multiple nonlinear transformations. To date, several deep learning frameworks have been applied in fields such as computer vision, image segmentation, and natural language processing, achieving excellent results. In recent years, some domestic and international experts and scholars have conducted research on cloud detection based on deep learning-related technologies.

[0004] Deep learning is an algorithm used in the cloud detection layer of this invention, aiming to extract high-level cloud features through multi-layered complex structures and nonlinear transformations. Deep learning has achieved significant results in computer vision, image segmentation, and natural language processing. However, for cloud images captured by all-sky cameras, the segmentation accuracy of deep learning is greatly affected by the complexity of cloud layers and the non-uniformity of light intensity. As cloud complexity increases, segmentation accuracy becomes increasingly difficult, potentially leading to over-segmentation or under-segmentation. Furthermore, existing semantic segmentation models only achieve high accuracy for specific scenes and lack generalization ability for cloud detection in all-sky images. In convolutional operations, uniformly processing features across the entire image reduces attention to important features. Therefore, the advantages of deep learning in this invention need to be described in detail in the specific implementation section, and the above problems need to be addressed to improve the accuracy and generalization ability of cloud detection. This invention relates to the field of cloud detection technology, specifically a cloud detection model based on deep learning. This model can effectively distinguish between clouds and the sky background, achieving binary classification of clouds, i.e., cloud category and non-cloud category. Compared to traditional methods, deep learning technology offers advantages in cloud detection due to its low cost and high efficiency, further improving the accuracy and efficiency of target object recognition. Therefore, this invention has broad application prospects in the field of cloud detection. Summary of the Invention

[0005] To overcome the aforementioned problems in the existing technology, the present invention provides a ground-based method and system for calculating cloud cover in all-sky images, which solves the problems existing in the prior art.

[0006] A ground-based method for calculating cloud cover in all-sky images includes the following steps:

[0007] S1. Acquire ground-based all-sky images and perform preprocessing, annotation, and enhancement;

[0008] S2. Train the enhanced image data using a cloud detection model that includes the improved U-Net model to obtain the prediction matrix;

[0009] S3. The prediction matrix is ​​segmented to determine the proportion of cloud pixels, thereby determining the corresponding cloud amount.

[0010] In addition to the aspects and any possible implementations described above, a further implementation is provided, wherein the preprocessing in S1 includes: cropping the cloud objects within the all-sky image with an inscribed circle, so that the top, bottom, left, and right boundaries of the cloud objects are tangent to the boundaries of the image, the alignment depth is uniformly adjusted to 24, and the size is uniformly adjusted to 256*256.

[0011] In addition to the aspects and any possible implementations described above, a further implementation is provided, wherein the annotation in S1 specifically includes: marking the cloud objects in the preprocessed image with closed curves, where the area inside the curve is the cloud object and the other part is the sky object, and setting the cloud object to white and the sky object to black.

[0012] In addition to the aspects and any possible implementations described above, a further implementation is provided in which the enhancement in S1 specifically includes: performing rotation, translation and shear transformation operations on the labeled image to obtain enhanced image data.

[0013] In addition to the aspects and any possible implementations described above, a further implementation is provided in which the cloud detection model of the improved U-Net model in S2 is a cloud detection model based on the CWFF feature extraction structure.

[0014] In addition to the aspects described above and any possible implementations, a further implementation is provided, wherein the cloud detection model based on the CWFF feature extraction structure comprises ten layers. The first four layers are encoding networks, wherein the first layer is the input layer, comprising a traditional convolutional layer and two CWFF feature extraction structures; the second to fourth layers each comprise two CWFF feature extraction structures and a max pooling layer; the fifth layer comprises two CWFF feature extraction structures and a random deactivation layer; the sixth to ninth layers are decoding networks, each comprising two CWFF feature extraction structures and an upsampling layer; and the tenth layer converts the previous feature layer into a prediction matrix.

[0015] In addition to the aspects and any possible implementations described above, a further implementation is provided, wherein the operation of the CWFF feature extraction structure specifically includes the following steps:

[0016] (1) Perform global average pooling on the image data input to the input layer;

[0017] (2) Compress the image data after global average pooling into a one-dimensional feature vector;

[0018] (3) Using a fully connected layer, the one-dimensional feature vector is mapped to the channel correlation weights. The fully connected layer encodes the correlation between each channel as a weight value.

[0019] (4) Multiply the input feature layer with the weight value to assign weights, thereby enhancing the feature representation of important channels;

[0020] (5) Three feature extractions are performed on the weighted feature layer, each including a convolution and regularization operation to extract image features;

[0021] (6) Add the initial unweighted input feature layer to the final extracted image features to obtain the extracted feature matrix.

[0022] In addition to the aspects and any possible implementations described above, an implementation is further provided in which S3 includes:

[0023] S31. Each cell in the prediction matrix includes two columns. The first column represents the probability of a sky pixel, and the second column represents the probability of a cloud pixel. Each cell in the prediction matrix is ​​thresholded, and elements in the prediction matrix with cloud pixel probabilities greater than the threshold are set to (0 1), and elements with probabilities less than or equal to the threshold are set to (1 0).

[0024] S32. Setting a two-dimensional identity column matrix Multiply by each cell in the prediction matrix;

[0025] S33. Each cell of the prediction matrix after processing in S32 has only one column, which is either 0 or 1, where 0 represents sky pixels and 1 represents cloud pixels;

[0026] S34. Calculate the percentage of cloud pixels in the whole sky image.

[0027] In addition to the aspects described above and any possible implementation, a further implementation is provided in which the percentage is calculated by: calculating the proportion of the cloud pixel to the inscribed circle, summing all cloud pixels in the feature matrix, and then dividing by the number of pixels in the inscribed circle of the square matrix.

[0028] The present invention also provides a ground-based all-sky image cloud cover calculation system, the system being used to implement the aforementioned calculation method, comprising:

[0029] The preprocessing module acquires ground-based all-sky images for preprocessing, annotation, and data augmentation;

[0030] The cloud detection module is used to train the enhanced image data to obtain the prediction matrix;

[0031] The cloud cover calculation module is used to segment the prediction matrix to determine the proportion of cloud pixels, thereby determining the corresponding cloud cover.

[0032] Beneficial effects of the present invention

[0033] Compared with the prior art, the present invention has the following beneficial effects:

[0034] This invention provides a ground-based method for calculating cloud cover in all-sky images, comprising the following steps: In the preprocessing stage, cloud objects in the image are first cropped with inscribed circles to ensure their top, bottom, left, and right boundaries are tangent to the image boundaries, and the alignment depth is uniformly adjusted to 24. The sky and clouds in the dataset are then labeled. This labeling includes marking cloud objects in the image with closed curves, identifying other parts as sky objects, and setting cloud objects to white and sky objects to black. Image data is obtained by rotating, translating, and shearing the labeled image. In the cloud detection stage, a cloud detection model based on a CWFF feature extraction structure is trained in the dataset to obtain a prediction matrix. The prediction matrix is ​​segmented to determine the proportion of cloud pixels, thereby calculating the corresponding cloud cover. This invention constructs a cloud cover calculation model for all-sky images based on CWFF feature extraction and a cloud cover calculation model based on a cloud detection model. The cloud detection model is based on a CWFF feature extraction structure and uses an improved U-Net neural network to enhance the model's extraction of deep semantics, making the segmentation results more reflective of the feature information of the all-sky image. Meanwhile, the cloud cover calculation model method in this invention improves the accuracy of cloud cover calculation. This has significant implications for the important evaluation of ground-based photoelectric telescope site selection. Attached Figure Description

[0035] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0036] Figure 2-1 This is a schematic diagram of the image before preprocessing.

[0037] Figure 2-2 This is a schematic diagram of the image after preprocessing;

[0038] Figure 3 This is a schematic diagram of image annotation processing;

[0039] Figure 4 This is a schematic diagram of image data enhancement according to the present invention, wherein (a) is the original image; (b) is the image after rotation and stretching; (c) is the image after rotation and displacement; and (d) is the image after shear displacement.

[0040] Figure 5 for Figure 1 A schematic diagram of the cloud detection model of the present invention is shown in the figure.

[0041] Figure 6 This is a schematic diagram of the CWFF feature extraction structure of the present invention. Detailed Implementation

[0042] To better understand the technical solution of this invention, the content of this invention includes, but is not limited to, the specific embodiments described below. Similar technologies and methods should be considered within the scope of protection of this invention. To make the technical problems to be solved, the technical solutions, and advantages of this invention clearer, a detailed description will be provided below in conjunction with the accompanying drawings and specific embodiments.

[0043] It should be understood that the embodiments described in this invention are merely some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without inventive effort are within the scope of protection of this invention.

[0044] The terminology used in the embodiments of this invention is for the purpose of describing particular embodiments only and is not intended to limit the invention. The singular forms “a,” “the,” and “the” as used in the embodiments of this invention and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise.

[0045] like Figure 1 As shown, the ground-based all-sky image cloud cover calculation method provided by this invention includes the following steps:

[0046] S1. Acquire ground-based all-sky images and perform preprocessing, annotation, and enhancement;

[0047] S2. Train the enhanced image data using a cloud detection model that includes the improved U-Net model to obtain the prediction matrix;

[0048] S3. The prediction matrix is ​​segmented to determine the proportion of cloud pixels, thereby determining the corresponding cloud amount.

[0049] For S1, the present invention specifically includes the following steps:

[0050] (1) This invention processes the entire sky image into a 256*256 PNG image with a bit depth of 24. Secondly, LabelMe software is used to label the cloud-containing pixels in the entire sky image. During the labeling process, cloud objects are first marked using multiple closed curves, with cloud objects marked inside the curves and sky objects marked outside the curves; and cloud objects are set to white and sky objects to black, such as... Figure 2-1 , Figure 2-2 and Figure 3 As shown.

[0051] (2) Appropriate data augmentation techniques can help alleviate overfitting, improve the robustness and generalization ability of the model, and solve the problem of imbalanced samples. To enhance data diversity, this invention uses random rotation, translation, and shearing transformations to transform and augment the dataset preprocessed in step (1) to obtain the image dataset img, such as... Figure 4 As shown in (a), (b), (c), and (d) in the figure. Among them, rotational displacement refers to rotating and translating an object, causing a change in its position and orientation in a plane; shear displacement is a shear transformation that performs oblique translation and shape changes on an object.

[0052] (3) Create a label set mask that corresponds one-to-one with the img image dataset. The specific steps are as follows: use normalization to scale each pixel in the img image dataset to the range of 0-1 to facilitate processing and analysis; secondly, set the non-zero values ​​in the normalized pixels to 1; finally, use One-Hot encoding to convert the values ​​of 1 in the img image dataset to (0,1) and 0 to (1,0).

[0053] (4) Generate an image array by mapping the img image dataset to its corresponding label mask. For S2, this specifically includes:

[0054] (1) After processing by S1, the size of the image array is 256*256 pixels, and the top, bottom, left and right boundaries of the cloud object are tangent to the boundaries of the processed image.

[0055] (2) The image array is trained using a cloud detection model based on a Channel Weighting-Feature Fusion (CWFF) structure to obtain the prediction matrix. The cloud detection model based on the CWFF feature extraction structure consists of ten layers, each executed stepwise. Layers six to nine have skip connections with layers four to one, fusing features from different levels of the encoder and decoder to enrich the feature representation of the decoder. The first four layers are the encoder: the first layer is the input layer, which includes a traditional convolutional layer and two CWFF feature extraction structures; the second to fourth layers each include two CWFF feature extraction structures and a max-pooling layer; the fifth layer includes two CWFF feature extraction structures and a random deactivation layer; and the sixth to ninth layers are the decoder, each including two CWFF feature extraction structures and an upsampling layer. The tenth layer converts the features extracted from the previous layer into a prediction matrix through an activation function.

[0056] The processing steps using the CWFF feature extraction structure include the following:

[0057] (1) Perform global average pooling on the input feature layer;

[0058] (2) Compress the result obtained in (1) into a one-dimensional feature vector;

[0059] (3) Using a fully connected layer, the obtained one-dimensional feature vector is mapped to the channel correlation weights. The fully connected layer can encode the correlation between each channel into a weight value.

[0060] (4) Multiply the input feature layer with the weight coefficients to assign weights, thereby enhancing the feature representation of important channels;

[0061] (5) Perform three feature extractions on the weighted feature layer, each of which includes a convolution and regularization operation;

[0062] (6) Add the initial unweighted input feature layer to the final extracted features as the output.

[0063] The cloud detection model in S2 consists of an improved U-Net model, which replaces the traditional feature extraction layer of U-Net with a CWFF feature extraction structure to form a new cloud detection model based on the CWFF feature extraction structure. The model training cycle is 150 rounds; each batch of training images consists of 8 images; the model optimizer is Adam; the initial learning rate is 0.0001; and the factor for reducing the learning rate each time is 0.5.

[0064] The cloud detection model based on CWFF feature extraction consists of ten layers. The entire network is executed layer by layer. Layers six through nine have skip connections with layers four through one, allowing the fusion of features from different levels of the encoder and decoder to enrich the feature representation capabilities of the decoder. The first four layers are the encoder network. The first layer is the input layer, consisting of a traditional convolutional layer and two CWFF feature extraction structures. The second through fourth layers each consist of two CWFF feature extraction structures and a max-pooling layer. The fifth layer consists of two CWFF feature extraction structures and a random deactivation layer. Layers six through nine are the decoder network, each consisting of two CWFF feature extraction structures and an upsampling layer. The tenth layer converts the features extracted from the previous layer into a prediction matrix using an activation function. See the table below:

[0065] Structure table of cloud detection model based on CWFF feature extraction structure

[0066]

[0067]

[0068] like Figure 5As shown, each CWFF feature extraction structure includes a feature fusion unit and a channel weighting unit. The channel weighting unit comprises a feature input layer, a global average pooling (GAP) layer, a first fully connected layer (FC), and a second fully connected layer (FC), connected in sequence. The first fully connected layer uses the ReLU activation function, and the second fully connected layer uses the Sigmoid activation function. The feature fusion unit consists of three sequentially connected feature extraction layers and one output layer. Each feature extraction layer uses batch normalization. The specific execution steps of the CWFF internal feature extraction structure are as follows:

[0069] The feature vector x input to the previous feature extraction structure i The dimension is [h, w, c], where h is x i The height of w is x i The width, c is x i The channels are weighted. In the channel weighting unit, global average pooling is used to compress it into a feature vector F of [1,1,c]. sq As shown in formula (1):

[0070]

[0071] Among them, u k Let i represent the k-th channel in the feature layer, i represent the height of the feature layer, j represent the width of the feature layer, and k∈[1,c]. The weights for each channel are generated through a fully connected layer, outputting c weight coefficients s, as shown in formula (2):

[0072] s=σ(w2δ(w1F sq (2)

[0073] Where w1 and w2 are fully connected operations, and w1 is multiplied by F sq For the first fully connected layer operation, w1∈c / r×c, where r is a dimensionality reduction coefficient, it is mapped to an output more suitable for the next layer through the δ(ReLU) activation function. The mathematical expression of ReLU is shown in Equation (3):

[0074] y = max(0,x) (3)

[0075] Where y represents the output of ReLU, and x represents w1F. sq The output is then used as input. A fully connected operation is then performed, multiplied by w2, where w2∈c / r×c passes through the activation function σ(sigmoid) to obtain the weights s. The mathematical expression for sigmoid is shown in formula (4):

[0076]

[0077] Among them, M (x)The output of the sigmoid function is given, and x represents the output after passing through the fully connected layers w1 and w2, which is then used as input. The obtained weights s are then compared with the initial feature layer x. i Multiplication yields the feature layer F with channel weights. w The feature layer F with channel weights w The input is fed into the feature fusion unit, where it undergoes three rounds of convolution and batch normalization to obtain the feature vector F(x). i Finally, the input feature vector x i With F(x) i The sums are used to obtain the final output vector x of this layer. i+1 As shown in formula (5).

[0078] x i+1 =f(H(x) i )+F(x i (5)

[0079] Where F represents the residual feature, indicating the learned feature vector; f is the ReLU activation function. H(x) i H(x) represents the difference between the actual value and the predicted value. i )=x i When, it indicates an identity mapping, such as Figure 6 As shown.

[0080] S3 specifically includes:

[0081] (1) Each cell in the prediction matrix includes two columns. The first column represents the probability of a sky pixel, and the second column represents the probability of a cloud pixel. Each cell in the prediction matrix is ​​thresholded. Elements in the prediction matrix with a cloud pixel probability greater than the threshold are set to (0 1), and elements with a probability less than or equal to the threshold are set to (1 0).

[0082] (2) Setting up a two-dimensional identity column matrix Multiply by each element in the prediction matrix to determine whether it is a sky pixel or a cloud pixel;

[0083] (3) Each cell of the processed prediction matrix has only one column, which is either 0 or 1, where 0 represents sky pixels and 1 represents cloud pixels. The percentage of cloud pixels in the entire sky image is calculated by summing all cloud pixels in the prediction matrix of the cloud detection layer and then dividing by the number of pixels in the inscribed circle to obtain the cloud amount in the entire sky image. The prediction matrix generated by the tenth layer activation function is shown in formula (6):

[0084]

[0085] Where m, n∈[0,255], each a i,j Both consist of two terms, x and y, i.e., a i,j = (xy), where x represents the probability that the current pixel is a sky pixel, y represents the probability that the current pixel is a cloud pixel, i∈[0,m-1], j∈[0,n-1], x,y∈[0,1].

[0086] The prediction matrix is ​​segmented by a threshold, as shown in formula (7). λ is the segmentation threshold. In this invention, the threshold is set to 0.5 at a time, with a step size of 0.05 and a stop value of 0.95. When the threshold λ is 0.9, the cloud cover statistics accuracy is closest to the standard value. Therefore, values ​​of y greater than λ are segmented by a. i,j Set it to (0 1), and set a for y less than or equal to λ. i,j Set to (1 0).

[0087] a i,j =(a i,j ) y >λ? (1 0):(0 1) (7)

[0088] To determine the pixel values, a two-row, one-column matrix is ​​introduced. By using a i,j Multiply to determine each a i,j The result determines whether the pixel belongs to the cloud or the sky. If the result is 0, it is determined to be a sky pixel, and if the result is 1, it is determined to be a cloud pixel, as shown in formula (8):

[0089]

[0090] The percentage of cloud pixels in the sky image is calculated as follows: the proportion of cloud pixels in the inscribed circle is calculated by summing all cloud pixels in the feature matrix and then dividing by the number of pixels in the inscribed circle of the square matrix. Since the image data has been processed into a rectangular image of size m×n in the preprocessing stage, it is preferred to be a square image when m=n in this invention. At this time, the entire sky area is inscribed in the square image, and the prediction matrix output by the cloud detection model is also a square prediction matrix of size m×n and m=n. Therefore, the calculation of the percentage of cloud pixels in the sky image is transformed into the calculation of the proportion of cloud pixels in the inscribed circle of the square matrix. Therefore, formula (9) is used to obtain the percentage of cloud pixels in the sky image by summing each cloud pixel in the prediction matrix and dividing by the number of pixels in the inscribed circle of the square matrix. The cloud amount calculation result is obtained based on this proportion. The expression of formula (9) is as follows:

[0091]

[0092] The present invention also provides a ground-based all-sky image cloud cover calculation system, the system being used to implement the aforementioned calculation method, comprising:

[0093] The preprocessing module acquires ground-based all-sky images for preprocessing, annotation, and data augmentation;

[0094] The cloud detection module is used to train the enhanced image data to obtain the prediction matrix;

[0095] The cloud cover calculation module is used to segment the prediction matrix to determine the proportion of cloud pixels, thereby determining the corresponding cloud cover.

[0096] The foregoing description illustrates and describes several preferred embodiments of the present invention. However, as previously stated, it should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the inventive concept described herein through the foregoing teachings or techniques or knowledge in related fields. Any modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A ground-based method for calculating cloud cover in all-sky images, characterized in that, Includes the following steps: S1. Acquire ground-based all-sky images and perform preprocessing, annotation, and enhancement; S2. The enhanced image data is trained using a cloud detection model including an improved U-Net model to obtain a prediction matrix. The improved U-Net model is a cloud detection model based on a CWFF feature extraction structure. The cloud detection model based on the CWFF feature extraction structure includes ten layers. The first four layers are encoding networks. The first layer is the input layer, which includes a traditional convolutional layer and two CWFF feature extraction structures. The second to fourth layers each include two CWFF feature extraction structures and a max pooling layer. The fifth layer includes two CWFF feature extraction structures and a random deactivation layer. The sixth to ninth layers are decoding networks, each including two CWFF feature extraction structures and an upsampling layer. The tenth layer converts the previous feature layer into a prediction matrix. The operation of the CWFF feature extraction structure specifically includes the following steps: (1) Perform global average pooling on the image data input to the input layer; (2) Compress the image data after global average pooling into a one-dimensional feature vector; (3) Using a fully connected layer, the one-dimensional feature vector is mapped to the channel correlation weights. The fully connected layer encodes the correlation between each channel as a weight value. (4) Multiply the input feature layer with the weight value to assign weights, thereby enhancing the feature representation of important channels; (5) Three feature extraction operations are performed on the weighted feature layer, each including a convolution and regularization operation to extract image features; (6) Add the initial unweighted input feature layer to the final extracted image features to obtain the extracted feature matrix; S3. The prediction matrix is ​​segmented to determine the proportion of cloud pixels, thereby determining the corresponding cloud amount.

2. The ground-based all-sky image cloud cover calculation method according to claim 1, characterized in that, The preprocessing in S1 includes: cropping the cloud objects in the all-sky image with an inscribed circle so that the top, bottom, left, and right boundaries of the cloud objects are tangent to the boundaries of the image, with a bit depth of 24 and a size of 256*256 pixels.

3. The ground-based all-sky image cloud cover calculation method according to claim 2, characterized in that, The annotation in S1 specifically includes: marking the cloud objects in the preprocessed image with closed curves, where the area inside the curve is the cloud object and the other part is the sky object, and setting the cloud object to white and the sky object to black.

4. The ground-based all-sky image cloud cover calculation method according to claim 3, characterized in that, The enhancement in S1 specifically includes: performing rotation, translation, and shearing transformations on the labeled image to obtain enhanced image data.

5. The ground-based all-sky image cloud cover calculation method according to claim 1, characterized in that, S3 includes: S31. Each cell in the prediction matrix includes two columns. The first column represents the probability of a sky pixel, and the second column represents the probability of a cloud pixel. Each cell in the prediction matrix is ​​thresholded, and elements in the prediction matrix with cloud pixel probabilities greater than the threshold are set to (0 1), and elements with probabilities less than or equal to the threshold are set to (1 0). S32. Setting a two-dimensional identity column matrix Multiply by each element in the prediction matrix; S33. Each cell of the prediction matrix after processing in S32 has only one column, which is either 0 or 1, where 0 represents sky pixels and 1 represents cloud pixels. S34. Calculate the percentage of cloud pixels in the whole sky image.

6. The ground-based all-sky image cloud cover calculation method according to claim 5, characterized in that, The percentage is calculated as follows: calculate the proportion of the cloud pixel to the inscribed circle, sum all cloud pixels in the feature matrix, and then divide by the number of pixels in the inscribed circle of the square matrix.

7. A ground-based all-sky image cloud cover calculation system, characterized in that, The system is used to implement the calculation method according to any one of claims 1-6, including: The preprocessing module acquires ground-based all-sky images for preprocessing, annotation, and data augmentation; The cloud detection module is used to train the enhanced image data to obtain the prediction matrix; The cloud cover calculation module is used to segment the prediction matrix to determine the proportion of cloud pixels, thereby determining the corresponding cloud cover.

Images