Full-solar-surface image cloud pollution detection method, system, equipment and medium

Abstract

The invention relates to a full-solar-surface image cloud pollution detection method, system and device and a medium. The method comprises the following steps: preprocessing a full-solar-surface image sequence in a solar-surface effective area to obtain a plurality of image blocks and corresponding DCT feature vectors; based on the spatial adjacency relation between the image blocks and the precedence relation of the image blocks at the same position in the time sequence, constructing a space-time diagram in combination with DCT feature vectors; inputting the space-time diagram into a pre-trained DCT-based graph neural network model to obtain a node advanced feature vector; decoding the advanced feature vector of each node to obtain a cloud pollution probability value, and generating a cloud pollution probability map in a full-solar-surface range based on the cloud pollution probability values of all the image blocks; and performing time sequence smoothing and adaptive threshold binarization processing on the cloud pollution probability map to generate a full-solar-surface cloud pollution binary mask, and performing calculation based on the full-solar-surface cloud pollution binary mask to obtain a cloud coverage rate quantitative index. By adopting the method, the cloud pollution area can be accurately identified and quantified.

Description

Technical Field

[0001] This invention belongs to the field of cloud contamination analysis technology, and in particular relates to a method, system, equipment and medium for detecting cloud contamination in full-area images. Background Technology

[0002] With the development of technology in the field of solar physics observation, continuous imaging technology of ground and space solar telescopes has emerged. By stably acquiring full-disk solar images, it provides a core basic data source for solar activity research, magnetic field evolution analysis and space weather early warning. However, the presence of clouds will attenuate, scatter or even completely block solar radiation to varying degrees, resulting in local or overall brightness reduction, contrast loss, blurring of details or even complete loss of data in the acquired full-disk images. This phenomenon is called cloud contamination.

[0003] In traditional technologies, cloud contamination detection for full-surface solar images mainly employs two methods: one is manual visual inspection, where staff screen images one by one to determine whether there is contamination such as decreased brightness, loss of contrast, or blurred details caused by clouds; the other is a simple global brightness threshold method, which sets a uniform brightness standard and identifies areas below the threshold as suspected cloud contamination areas.

[0004] However, the above methods are extremely inefficient for manual inspection and cannot meet the actual needs of automated processing pipelines for massive solar observation data. The simple global threshold method is not robust enough, making it difficult to distinguish between real activity features such as sunspots and dark streaks and local darkening caused by clouds. It is also easily affected by the inherent brightness difference between the center and edge of the solar surface, resulting in large detection errors. This directly affects the extraction accuracy of key parameters such as sunspot number and magnetic flux, and cannot guarantee the quality and usability of solar observation data. Summary of the Invention

[0005] Therefore, it is necessary to provide a method, system, device, and medium for detecting cloud pollution in full-day images that can accurately identify and quantify cloud pollution areas, addressing the aforementioned technical problems.

[0006] Firstly, this application provides a method for detecting cloud contamination in a full-area image, including:

[0007] A full-surface image sequence is acquired, and the full-surface image sequence is preprocessed within the effective area of ​​the solar surface to obtain multiple image patches and corresponding DCT feature vectors; the preprocessing corresponds to solar surface spatial grid division and selection of low-frequency coefficients of two-dimensional discrete cosine transform;

[0008] Based on the spatial adjacency between image patches and the temporal sequence of image patches at the same location, a spatiotemporal graph is constructed by combining DCT feature vectors. The spatiotemporal graph includes graph nodes, temporal edges, and spatial edges. Graph nodes are represented by the DCT feature vectors of each image patch at each observation time. Spatial edges connect graph nodes that are spatially adjacent at the same observation time. Temporal edges connect graph nodes at the same spatial location at different observation times.

[0009] The spatiotemporal graph is input into a pre-trained DCT-based graph neural network model to obtain the high-level feature vectors of each graph node; the DCT-based graph neural network model performs multiple rounds of information propagation and aggregation along the spatial and temporal edges to update the DCT feature vectors of each graph node.

[0010] The high-level feature vectors of each node are decoded to obtain the cloud contamination probability value of each corresponding image patch. Based on the cloud contamination probability values ​​of all image patches, a cloud contamination probability map of the entire solar surface is generated.

[0011] The cloud contamination probability map is subjected to temporal smoothing and adaptive threshold binarization to generate a full-surface cloud contamination binary mask, and the cloud coverage rate quantification index is calculated based on the full-surface cloud contamination binary mask.

[0012] In one embodiment, a spatiotemporal map is constructed based on the spatial adjacency between image patches and the temporal sequence of image patches at the same location, combined with DCT feature vectors, including:

[0013] Each image patch is defined as a graph node at each observation time, and each graph node is assigned a corresponding DCT feature vector as the initial node feature.

[0014] For each set of graph nodes at each observation time, all graph node pairs are traversed. If the image patch corresponding to the graph node pair satisfies the preset adjacency condition in the solar space grid, an undirected edge is established between the two graph nodes of the graph node pair as a spatial edge.

[0015] For each image patch corresponding to a spatial location, a directed edge is established between the corresponding graph nodes at adjacent times, pointing from an earlier time to a later time, according to the observation time order;

[0016] By merging spatial edges and temporal edges, and combining the graph nodes and their corresponding initial node features, a spatiotemporal graph is obtained.

[0017] In one embodiment, the spatiotemporal graph is input into a pre-trained DCT-based graph neural network model to obtain high-level feature vectors for each graph node, including:

[0018] The DCT feature vectors of the graph nodes are used as the embedding representation of the zeroth layer nodes in the DCT-based graph neural network model.

[0019] Based on the zero-level node embedding representation, multi-level forward propagation is performed, and the final feature representation obtained in the last layer is determined as the high-level feature vector representation of the graph nodes;

[0020] The forward propagation process includes the following steps:

[0021] Based on the Layered graph neural networks spatially aggregate graph nodes to obtain intermediate feature representations; the intermediate feature representations are as follows: ,in, For the observation time Spatial location The graph node passes through the first Intermediate features after layer space aggregation; For the graph node at the th Features of the layer; For graph nodes The set of neighboring nodes; This is a normalization constant based on node degree; For the first Layer-learnable weight matrix; It is a non-linear activation function;

[0022] Based on temporal edges, a gating mechanism is used to fuse intermediate feature representations with the final feature representations of graph nodes at adjacent time points at the same location, resulting in the final feature representation of the graph node. The gating mechanism is represented as follows: , , ,in, To update the gate vector, For graph nodes At any moment No. The final feature representation of the layer, For candidate feature vectors, For element-wise multiplication, and The weight matrix is ​​a learnable matrix. For graph nodes In the same location Time Node Characteristics of the layer.

[0023] In one embodiment, the high-level feature vectors of each node are decoded to obtain the probability value of cloud contamination for each corresponding image patch, including:

[0024] The probability of cloud contamination can be obtained using the following formula:

[0025]

[0026] in, For high-level feature vectors; This is the weight matrix; For bias; and Image blocks At any moment The probability of being classified as clean and the probability of being classified as cloud-polluted, and .

[0027] In one embodiment, the cloud contamination probability map is subjected to temporal smoothing and adaptive threshold binarization to generate a full-surface cloud contamination binary mask, including:

[0028] Obtain the cloud contamination probability map at the current moment, the cloud contamination probability map at the adjacent previous moment, and the cloud contamination probability map at the adjacent subsequent moment.

[0029] For the probability value of each spatial location in the probability map at the current time, a weighted median filter is applied to the probability value of the same location in the cloud pollution probability maps of adjacent preceding and subsequent times to obtain the time-smoothed probability value, and a smoothed probability map is generated.

[0030] A gray-level histogram is constructed based on all valid probability values ​​in the smooth probability map, and the gray-level histogram is analyzed using the maximum inter-class variance method to obtain the optimal threshold for segmenting contaminated and clean areas.

[0031] Traverse each position in the smooth probability map. If the probability value is greater than or equal to the optimal threshold for separating contaminated and clean areas, mark it as 1 at the corresponding position in the output mask. If the probability value is less than the optimal threshold for separating contaminated and clean areas, mark it as 0 at the corresponding position in the output mask to obtain a binary mask of cloud contamination on the entire solar surface.

[0032] Secondly, this application also provides a full-plane image cloud contamination detection system, comprising:

[0033] The full-surface image module is used to acquire a full-surface image sequence and preprocess the full-surface image sequence within the effective area of ​​the solar surface to obtain multiple image blocks and corresponding DCT feature vectors; the preprocessing corresponds to solar surface spatial grid division and selection of low-frequency coefficients of two-dimensional discrete cosine transform;

[0034] The spatiotemporal graph module is used to construct a spatiotemporal graph based on the spatial adjacency between image patches and the temporal sequence of image patches at the same location, combined with DCT feature vectors. The spatiotemporal graph includes graph nodes, temporal edges, and spatial edges. Graph nodes are represented by the DCT feature vectors of each image patch at each observation time. Spatial edges connect graph nodes that are spatially adjacent at the same observation time. Temporal edges connect graph nodes at the same spatial location at different observation times.

[0035] The spatiotemporal aggregation and propagation module is used to input the spatiotemporal graph into a pre-trained DCT-based graph neural network model to obtain the high-level feature vectors of each graph node; the DCT-based graph neural network model performs multiple rounds of information propagation and aggregation along the spatial and temporal edges to update the DCT feature vectors of each graph node.

[0036] The cloud contamination probability module is used to decode the high-level feature vectors of each node to obtain the cloud contamination probability value of each corresponding image patch, and generate a cloud contamination probability map of the entire solar surface based on the cloud contamination probability values ​​of all image patches.

[0037] The cloud coverage module is used to perform temporal smoothing and adaptive threshold binarization on the cloud pollution probability map, generate a full-area cloud pollution binary mask, and calculate the cloud coverage rate quantification index based on the full-area cloud pollution binary mask.

[0038] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of any of the above-described methods for detecting cloud contamination in full-area images.

[0039] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the above-described methods for detecting cloud contamination in full-plane images.

[0040] The aforementioned method, system, device, and medium for detecting cloud contamination in full-solar images acquire and preprocess full-solar image sequences to obtain image patches and their DCT feature vectors. This constructs a spatiotemporal graph that integrates spatial adjacency and temporal sequence. Then, a pre-trained DCT-based graph neural network is used to perform multiple rounds of information propagation and aggregation along the spatiotemporal edges. After updating node features, the cloud contamination probability values ​​of each image patch are decoded to form a full-solar probability map. Finally, by performing temporal smoothing and adaptive threshold binarization on the map, a full-solar cloud contamination binary mask and a quantitative index of cloud coverage are output. This achieves a full-solar image cloud contamination detection method that can automatically and accurately distinguish between the sun's own structure and cloud contamination, adapt to complex observation conditions, and provide a quantitative assessment of the degree of contamination. Attached Figure Description

[0041] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0042] Figure 1 This is a flowchart illustrating the full-surface image cloud contamination detection method of the present invention;

[0043] Figure 2 This is a flowchart illustrating the steps of step S105.

[0044] Figure 3 This is a structural diagram of the full-plane image cloud contamination detection system of the present invention. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0046] In one embodiment, such as Figure 1 As shown, a method for detecting cloud contamination in a full-area image is provided. This embodiment illustrates the application of this method to a terminal. It is understood that this method can also be applied to a server, or to a system including both a terminal and a server, and implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0047] S101. Obtain the full solar surface image sequence and preprocess the full solar surface image sequence within the effective area of ​​the solar surface to obtain multiple image blocks and corresponding DCT feature vectors; the preprocessing corresponds to the solar surface spatial grid division and the selection of low-frequency coefficients of the two-dimensional discrete cosine transform.

[0048] Indicatively, the acquired full-plane image sequence undergoes flat-field correction, dark-field correction, and precise registration. Flat-field correction eliminates the effects of optical system inhomogeneities, dark-field correction subtracts the instrument's own dark current noise, and precise registration eliminates displacement deviations between images in the sequence through image registration algorithms, ensuring that solar features at the same spatial location are aligned across different frames. A binary solar mask is generated based on the heliocentric coordinates and solar radius. This mask clearly separates the effective solar region from the background region, with the effective solar region marked as valid pixels and the background region marked as invalid pixels, thus limiting the scope of preprocessing. Furthermore, local photometric normalization is performed on each frame, calculating the global median of the image under the solar binary mask. The grayscale value of each valid pixel in the image is divided by this global median. This operation unifies the overall brightness benchmark between the images in the sequence while preserving the relative brightness variations in local areas, effectively eliminating the inherent brightness differences between the solar center and edges caused by edge dimming.

[0049] For example, the effective solar disk region is divided into multiple non-overlapping image blocks according to a preset partitioning rule. The partitioning process must balance detail preservation with computational efficiency, ensuring that each image block can fully capture the solar feature information of the local region, and that there is no pixel overlap or omission between adjacent image blocks. A two-dimensional discrete cosine transform (DCT) is performed on each image block, converting the spatial domain pixel value matrix of the image block into a frequency domain DCT coefficient matrix. The transformation formula is as follows: ,in, Represents the frequency domain coordinates, corresponding to the row and column positions in the DCT coefficient matrix; and These are the width and height of the image patch, respectively; Represents the spatial coordinates of an image patch The pixel grayscale value at that location; These are normalization coefficients; when the variable is 0, for the width direction... Or for the height direction When the variable is greater than or equal to 1, for the width direction Or for the height direction The normalization coefficients ensure that the DCT transform satisfies energy conservation, guaranteeing a consistent energy distribution in the image before and after the transform. After the DCT transform, the energy of image blocks is mainly concentrated in the low-frequency coefficient region. These low-frequency coefficients contain the main contour and brightness information of the image, making them particularly sensitive to local brightness changes caused by cloud contamination. In contrast, the high-frequency coefficients mainly correspond to detailed noise in the image and contribute less to cloud contamination detection. To form a fixed-length node feature vector, a zigzag scanning sequence is used to traverse the DCT coefficient matrix, starting from the DC component in the upper left corner. Begin by sequentially scanning along paths that are horizontally to the right, diagonally downwards to the left, vertically downwards, and diagonally upwards to the right, prioritizing the areas with concentrated energy. A low-frequency coefficient, Each coefficient includes both the DC component reflecting the overall brightness and contour of the image patch and the most critical low-frequency AC component, thus retaining the core feature information required for cloud contamination detection in a compact form, ultimately forming... The dimension vector is the DCT feature vector of the corresponding image patch.

[0050] S102. Based on the spatial adjacency relationship between image patches and the temporal sequence relationship of image patches at the same location, a spatiotemporal graph is constructed by combining DCT feature vectors. The spatiotemporal graph includes graph nodes, temporal edges, and spatial edges. Graph nodes are represented by the DCT feature vectors of each image patch at each observation time. Spatial edges connect graph nodes that are spatially adjacent at the same observation time. Temporal edges connect graph nodes at the same spatial location at different observation times.

[0051] In a schematic manner, during the construction of the spatiotemporal graph, the definition and representation method of graph nodes are clearly defined. Each image patch is assigned an independent graph node identity at its corresponding observation time. The initial feature vector of this node directly adopts the low-frequency coefficient vector of DCT, so that the node naturally carries the spatial frequency feature information of the corresponding image patch, laying the data foundation for subsequent feature propagation and aggregation. For the construction of spatial edges, for all graph nodes at the same observation time, the adjacency is determined based on the actual positional relationship of the image patches in the solar space grid. The adjacency condition can adopt the 4-neighborhood or 8-neighborhood rule. The 4-neighborhood rule means that two image patches are directly adjacent in the four directions of up, down, left, and right. The 8-neighborhood rule adds the adjacency determination in the diagonal direction on the basis of the 4-neighborhood rule. An undirected spatial edge is established between two graph nodes that meet the adjacency condition. This edge is used to characterize the spatial correlation between different regions within the same frame of the image, so that the feature information of adjacent regions can be transferred to each other. For the construction of temporal edges, focusing on image patches at the same spatial location, directed temporal edges are established between graph nodes corresponding to adjacent observation times, with the edge direction pointing from the earlier observation time to the later observation time. This edge is used to capture the feature evolution patterns of the same region at different time points, realizing the correlation of temporal information across frames. All graph nodes, spatial edges, and temporal edges are integrated to form a complete spatiotemporal graph. This graph contains both the spatial connections between different regions of the solar surface at the same time and the temporal evolution relationships of the same region at different times, providing structured data support for the subsequent spatiotemporal feature learning of the graph neural network.

[0052] S103. Input the spatiotemporal graph into the pre-trained DCT-based graph neural network model to obtain the high-level feature vectors of each graph node; the DCT-based graph neural network model performs multiple rounds of information propagation and aggregation along the spatial and temporal edges to update the DCT feature vectors of each graph node.

[0053] Indicatively, the constructed spatiotemporal graph is input into a pre-trained DCT-based graph neural network model. This model is an end-to-end deep learning architecture specifically designed for cloud contamination detection in full-surface solar images. It integrates DCT frequency features with the spatiotemporal modeling capabilities of graph neural networks, and adapts its customized network layer structure to the characteristics of solar images and the dynamic changes in cloud contamination. For example, the model consists of an input layer, multiple alternating stacked DCT-spatial graph convolutional modules and temporal gating propagation modules, and a feature output layer. The input layer receives DCT feature vectors from graph nodes, the intermediate modules perform deep aggregation of spatiotemporal features, and the output layer outputs high-level feature vectors of nodes fused with the global context. It captures local features of cloud contamination by leveraging the sensitivity of DCT features to brightness changes, models spatial neighborhood relationships through graph structure, and distinguishes between normal evolution and abnormal changes in time series using gating mechanisms. These three elements work together to achieve accurate identification of cloud contamination.

[0054] Optionally, the training data is prepared by constructing a labeled dataset containing diverse cloud contamination scenarios. Data sources can be categorized into two types: manual annotation and semi-automatic generation. Manual annotation involves domain experts classifying image patches as clean / contaminated based on visual characteristics of cloud contamination. Semi-automatic generation involves selecting images from extremely stable observation periods in long-term sequences as absolutely clean samples and images with obvious cloud cover characteristics as contaminated samples. The dataset is then expanded using data augmentation techniques such as random cropping and brightness perturbation, ultimately forming training sample pairs containing spatiotemporal subgraph inputs and node labels. Each spatiotemporal subgraph is centered on the target node and includes its spatial neighborhood nodes and temporal nodes from the past few frames. For example, during the training phase, a weighted cross-entropy loss function is used to address the sample imbalance problem. The loss function formula is as follows: ,in This represents the total number of nodes within the batch. Represents cleaning category, Represents the pollution category; As a class weight, when the number of clean samples is much greater than the number of contaminated samples, Value greater than The specific value is determined by inversely proportional calculation of the sample size; For nodes The actual label is encoded using one-hot encoding. If the node belongs to the corresponding category, the value is 1; otherwise, it is 0. For model prediction nodes Category The probability of the loss is calculated using a logarithmic function to ensure sensitive feedback to prediction errors, while the weights balance the contribution of different classes of samples to the loss. The training strategy employs mini-batch stochastic gradient descent, dividing the dataset into training, validation, and test sets, typically in a 7:2:1 ratio. During training, fixed-length spatiotemporal subgraphs are input into the model in batches. Forward propagation is first performed to calculate the prediction probability and loss value for each node. Then, backpropagation is used to calculate the partial derivatives of the loss function with respect to the weight parameters of each layer. The Adam optimizer is used to update the weights. The initial learning rate of the optimizer is set based on experimental tuning, and a learning rate decay strategy is used to reduce the learning rate when the validation set accuracy stagnates, thus promoting model convergence. During training, the training set loss and validation set detection accuracy, F1 score, and other metrics are monitored in real time. When the validation set metrics do not improve for a preset number of rounds, the model is considered converged, and the current weights are saved as a pre-trained model to ensure good generalization ability on unseen data.

[0055] The information propagation and aggregation process proceeds alternately or sequentially along spatial and temporal edges. Each round of propagation corresponds to one layer of the model's network structure. After a preset number of iterations, deep feature fusion is achieved. Specifically, the DCT feature vectors of graph nodes are used as the node embedding representation of the zeroth layer of the model. This embedding directly reflects the original frequency features of the image patch. Entering the forward propagation stage of each network layer, for example, a spatial aggregation operation is performed. Through an improved spatial graph convolution operation, the features of each node and all its neighboring nodes in its spatial neighborhood are fused. ,in Indicates the observation time Spatial location The graph node passes through the first Intermediate feature representation after layer space aggregation; This indicates that the graph node is at the [number]th [node]. Layer feature representation; This represents the set of spatial neighbors of a node in the graph. This represents a normalization constant based on node degree, used to balance the influence of nodes of different degrees on the aggregation result. Its value is the node's... Degree and node The square root of the product of degrees; and All are the first The learnable weight matrices of the layer network are used to perform linear transformations on the features of the node itself and the features of its neighboring nodes, respectively. This represents a non-linear activation function, used to introduce non-linear feature mapping capabilities and improve the model's ability to fit complex patterns.

[0056] After spatial aggregation is completed, a temporal fusion operation is performed. A gating mechanism dynamically fuses the intermediate feature representation of the current moment with the feature representation of the previous moment at the same spatial location to distinguish between the slow changes caused by normal solar evolution and the rapid changes introduced by cloud pollution. The gating mechanism's operation includes three core steps: first, calculating and updating the gate vector... ,in This represents the update gate vector, whose elements take values ​​in the range [0,1], and is used to control the retention ratio of feature information from the previous time step and the update ratio of intermediate feature information at the current time step. This represents the learnable weight matrix of the update gate; This represents the joint feature vector formed by concatenating the intermediate feature vector at the current time step with the feature vector at the previous time step; This represents the sigmoid activation function, used to map the elements of the update gate vector to the interval [0,1]. Subsequently, candidate feature vectors are calculated. ,in, Represents the candidate feature vector; This represents the learnable weight matrix calculated from the candidate features; This represents element-wise multiplication, used to weight the feature vector from the previous time step based on the updated gate vector; This represents the hyperbolic tangent activation function, used to map the element values ​​of the candidate feature vector to the interval [-1, 1], enhancing the non-linear expressive power of the features. The final update yields the current time step [i]. Layer Node Feature Representation ,in Represents graph nodes At the observation time No. The final feature representation of the layer is achieved through dynamic weighted fusion of features from the previous time step and current candidate features by updating the gate vector. When cloud pollution causes rapid changes in regional features, the updated gate vector element values ​​approach 1, and the model adopts more current candidate features; when the solar surface region features are stable, the updated gate vector element values ​​approach 0, and the model retains more features from the previous time step. Through the stacking of multiple layers of the above spatial aggregation and temporal fusion operations, the feature representation of each graph node gradually fuses the feature information of all nodes in its spatial neighborhood as well as its own feature information from multiple historical time steps, ultimately outputting the first... Layer feature representation This refers to the high-level feature vector of a node, which can comprehensively characterize the spatiotemporal context features of the corresponding image patch, providing accurate feature support for cloud contamination detection.

[0057] S104. Decode the high-level feature vectors of each node to obtain the cloud contamination probability value of each corresponding image patch, and generate a cloud contamination probability map of the entire solar surface based on the cloud contamination probability values ​​of all image patches.

[0058] Decoding is performed on the high-level feature vector of each graph node. This decoding process is achieved through a combination of fully connected layers and the Softmax activation function, mapping the high-dimensional node feature vector to a two-dimensional probability distribution. ,in Represents image blocks At the observation time The probability of being identified as a clean area. This indicates the probability that the image patch is identified as a cloud-contaminated area, and the sum of the two is 1; This represents the learnable weight matrix of the fully connected layer, used to map high-dimensional high-level feature vectors to a two-dimensional feature space; Represents graph nodes High-level feature vectors; This represents the bias term of the fully connected layer, used to adjust the offset of the feature map; This represents the normalization exponential function, used to convert the output values ​​in the two-dimensional feature space into a probability distribution that conforms to the probability axioms, such that each output value is in the interval [0,1] and the sum is 1. After calculating the contamination probability of all image patches, based on the actual position of each image patch in the solar disk space, the corresponding... Probability values ​​are filled into the corresponding spatial regions to form an initial cloud contamination probability map with the same size as the effective area of ​​the solar disk but with the same resolution as the image patch division. To obtain a refined probability map with the same resolution as the original full-solar image, a bilinear interpolation algorithm is used to upsample the initial probability map. Bilinear interpolation calculates the contamination probability value for each pixel in the original image by weighting the probability values ​​of four adjacent image patches, ensuring the spatial continuity and refinement of the probability map. Finally, a high-resolution cloud contamination probability map covering the entire solar disk is generated, in which the value of each pixel directly reflects the probability of cloud contamination at that location.

[0059] S105. Perform time-series smoothing and adaptive threshold binarization on the cloud contamination probability map to generate a full-surface cloud contamination binary mask, and calculate the cloud coverage rate quantification index based on the full-surface cloud contamination binary mask.

[0060] This illustrates how temporal smoothing is performed on high-resolution cloud contamination probability maps to suppress single-frame misjudgments and improve the temporal consistency of detection results by utilizing probability map information from adjacent observation times. Specifically, the current observation time is obtained... The cloud pollution probability map is obtained, and adjacent preceding time steps are extracted. and adjacent subsequent time steps The cloud contamination probability map is processed by combining the probability value of each spatial location at the current time with the probability values ​​of the same spatial location at previous and subsequent times using a weighted median filter. The weighting coefficients are set according to the principle that the probability value at the current time has a higher weight than that at adjacent times, to ensure the dominance of the information at the current time. At the same time, the motion trajectory of the cloud region is smoothed by introducing information from adjacent times, reducing false detections caused by single-frame noise or instantaneous interference. Furthermore, after the temporal smoothing process is completed, adaptive threshold binarization is performed on the obtained smoothed probability map. The probability values ​​of all effective pixels in the smoothed probability map, i.e., the pixels within the effective area of ​​the solar surface, are counted. A grayscale histogram is constructed based on these probability values. The horizontal axis of the histogram represents the probability value interval, and the vertical axis represents the number of pixels in the corresponding interval. The Otsu algorithm, a method involving maximum inter-class variance (MOV) analysis, is employed to analyze the grayscale histogram. By traversing all possible candidate threshold values, the inter-class variance between the foreground (contaminated area) and the background (clean area) corresponding to each candidate value is calculated. The candidate value with the largest inter-class variance is determined as the optimal segmentation threshold. This threshold can adapt to the distribution characteristics of the probability map under different weather conditions, effectively distinguishing between contaminated and clean areas. Based on the optimal segmentation threshold, the smoothed probability map is binarized. Each valid pixel in the smoothed probability map is traversed. If the probability value of a pixel is greater than or equal to the optimal segmentation threshold, its corresponding position in the output mask is marked as 1, representing that position as a cloud contaminated area. If the probability value of a pixel is less than the optimal segmentation threshold, its corresponding position is marked as 0, representing that position as a clean area. This generates a full-surface cloud contamination binary mask. Based on this binary mask, a quantitative index of cloud coverage is calculated. The cloud coverage is calculated as the ratio of the total number of pixels marked as 1 in the binary mask to the total number of pixels in the effective area of ​​the solar surface. This ratio intuitively reflects the proportion of the solar surface image that is polluted by clouds, providing a quantitative basis for the quality assessment of solar observation data.

[0061] In the aforementioned method for detecting cloud contamination in full-surface images, the process begins by acquiring a sequence of full-surface images. Preprocessing is then performed within the effective area of ​​the solar surface, involving spatial grid division and selection of low-frequency coefficients using two-dimensional discrete cosine transform (DCT). This yields multiple image patches and their corresponding DCT feature vectors. Based on the spatial adjacency between image patches and the temporal sequence of image patches at the same location, a spatiotemporal graph containing graph nodes, spatial edges, and temporal edges is constructed using the DCT feature vectors. This spatiotemporal graph is then input into a pre-trained DCT-based graph neural network model, which performs multiple rounds of information propagation along the spatial and temporal edges. The method aggregates and updates node features, then decodes the high-level feature vectors of each node to obtain the cloud contamination probability value of the corresponding image patch and generates a cloud contamination probability map of the entire solar surface. Finally, the probability map is subjected to temporal smoothing and adaptive threshold binarization to generate a binary mask of cloud contamination on the entire solar surface and calculate the cloud coverage quantitative index. This method effectively overcomes the problems of low efficiency, insufficient robustness and susceptibility to interference in traditional detection methods, achieves accurate detection of cloud contamination in the entire solar surface image, improves the stability and adaptability of the detection results, and provides a reliable quantitative basis for the quality assessment of solar observation data.

[0062] In one embodiment, a spatiotemporal map is constructed based on the spatial adjacency between image patches and the temporal sequence of image patches at the same location, combined with DCT feature vectors, including:

[0063] S11. Define each image patch as a graph node at each observation time, and assign each graph node a corresponding DCT feature vector as the initial node feature.

[0064] Each image patch is assigned an independent graph node identity at each corresponding observation time, ensuring that image patches at different times and spatial locations have a unique node mapping. The initial features of each graph node directly adopt the DCT feature vector of the corresponding image patch, so that the node carries core information that can characterize the spatial frequency features of the image patch from the initial stage, providing basic data support for the subsequent propagation and aggregation of spatiotemporal features.

[0065] S12. For the set of graph nodes at each observation time, traverse all graph node pairs. If the image block corresponding to the graph node pair satisfies the preset adjacency condition in the solar space grid, then establish an undirected edge between the two graph nodes of the graph node pair as a spatial edge.

[0066] For each observation time, a traversal algorithm is used to check the correspondence of all graph node pairs in turn. For any pair of graph nodes, it is determined whether the positions of the two corresponding image blocks in the solar space grid meet the preset adjacency conditions. These adjacency conditions are preset based on the solar space grid division rules and can accurately define the spatial adjacency relationship between image blocks. When the adjacency conditions are met, an undirected edge is immediately established between the two graph nodes. This undirected edge is the spatial edge connecting adjacent spatial regions at the same time.

[0067] S13. For each image patch corresponding to a spatial location, according to the observation time order, establish directed edges between corresponding graph nodes at adjacent times, pointing from the earlier time to the later time, as time edges.

[0068] For each image patch corresponding to a fixed spatial location, the images are sorted out according to the order of observation time. The graph nodes corresponding to the image patch at two adjacent observation times are located. A directed edge is established between the graph nodes at these two adjacent observation times. The direction of the directed edge strictly follows the direction from the graph node corresponding to the earlier observation time to the graph node corresponding to the later observation time. In this way, a temporal edge that can reflect the temporal evolution correlation of the same spatial region is constructed.

[0069] S14. Merge the spatial edges and temporal edges, and combine the graph nodes and their corresponding initial node features to obtain the spatiotemporal graph.

[0070] All spatial edges and all temporal edges are combined to ensure that all spatial and temporal related edges are integrated into a unified edge set without omission or repetition. This edge set is then associated and combined with all defined graph nodes and their corresponding initial node features to form a complete spatiotemporal graph containing both spatiotemporal and temporal related information.

[0071] In one embodiment, the spatiotemporal graph is input into a pre-trained DCT-based graph neural network model to obtain high-level feature vectors for each graph node, including:

[0072] S21. The DCT feature vectors of the graph nodes are used as the embedding representation of the zeroth layer nodes of the DCT-based graph neural network model.

[0073] Indicatively, the DCT feature vector corresponding to the graph node is directly determined as the zeroth layer node embedding representation of the DCT-based graph neural network model. This embedding representation fully carries the core spatial frequency information of the corresponding image block, providing initial data support for feature propagation, aggregation, and deep processing in subsequent layers of the model.

[0074] S22. Based on the zero-level node embedding representation, perform multi-level forward propagation and determine the final feature representation obtained from the last layer as the high-level feature vector representation of the graph nodes.

[0075] The forward propagation process includes the following steps:

[0076] Based on the Layered graph neural networks spatially aggregate graph nodes to obtain intermediate feature representations; the intermediate feature representations are as follows: ,in, For the observation time Spatial location The graph node passes through the first Intermediate features after layer space aggregation; For the graph node at the th Features of the layer; For graph nodes The set of neighboring nodes; This is a normalization constant based on node degree; For the first Layer-learnable weight matrix; It is a non-linear activation function;

[0077] Based on temporal edges, a gating mechanism is used to fuse intermediate feature representations with the final feature representations of graph nodes at adjacent time points at the same location, resulting in the final feature representation of the graph node. The gating mechanism is represented as follows: , , ,in, To update the gate vector, For graph nodes At any moment No. The final feature representation of the layer, For candidate feature vectors, For element-wise multiplication, and The weight matrix is ​​a learnable matrix. For graph nodes In the same location Time Node Characteristics of the layer.

[0078] Specifically, starting with the zero-th layer node embedding representation, the model performs a preset number of multi-layer forward propagation operations. The feature output of each layer serves as the input to the next layer. After iterative processing through all preset layers, the feature representation output of the last layer is directly determined as the high-level feature vector representation of the graph nodes. This vector has fully incorporated spatiotemporal context information. The forward propagation process includes spatial aggregation operations and temporal feature fusion based on a gating mechanism. The spatial aggregation operation aggregates features of the graph nodes based on the network parameters of the l-th layer to obtain intermediate feature representations. Subsequently, a gating mechanism is used to achieve feature fusion in the time dimension. By utilizing the features of adjacent time nodes at the same location associated with time edges, temporal information is dynamically integrated. Specifically, the gate vector is calculated and updated. Next, the candidate feature vectors are calculated. Finally, the final feature representation of the l-th layer is obtained through updating. Through this gating fusion process, dynamic adaptation of features at different times in the same spatial location can be achieved, accurately distinguishing the feature differences brought about by normal evolution and abnormal changes.

[0079] In one embodiment, the high-level feature vectors of each node are decoded to obtain the probability value of cloud contamination for each corresponding image patch, including:

[0080] The probability of cloud contamination can be obtained using the following formula:

[0081]

[0082] in, For high-level feature vectors; This is the weight matrix; For bias; and Image blocks At any moment The probability of being classified as clean and the probability of being classified as cloud-polluted, and .

[0083] In one embodiment, such as Figure 2 As shown, temporal smoothing and adaptive threshold binarization are performed on the cloud contamination probability map to generate a full-day cloud contamination binary mask, including:

[0084] S201. Obtain the cloud contamination probability map at the current time, the cloud contamination probability map of the adjacent preceding time, and the cloud contamination probability map of the adjacent following time.

[0085] Accurately obtain the cloud pollution probability map at the current observation time. Simultaneously, cloud contamination probability maps corresponding to adjacent previous observation times are extracted. Cloud contamination probability map corresponding to subsequent observation times This ensures that the three maps maintain consistency in spatial resolution and effective area of ​​the solar disk, providing complete temporal data support for subsequent temporal smoothing processing.

[0086] S202. For the probability value of each spatial location in the probability map at the current time, perform weighted median filtering on the probability values ​​of the same location in the cloud pollution probability maps of adjacent preceding and subsequent times to obtain the time-smoothed probability value, and generate a smoothed probability map.

[0087] For the probability graph at the current time The probability value of each spatial location is compared with the probability map of the same spatial location at previous time steps. Probability graph of subsequent time steps The corresponding probability values ​​are combined and processed using a weighted median filtering algorithm. By reasonably setting the weight coefficients, the dominant role of the probability value at the current moment is highlighted, while the constraints of the probability values ​​at adjacent moments are incorporated to filter out single-frame noise interference, thereby obtaining the probability value of each spatial location after temporal smoothing, and then integrating them to generate a smooth probability map of the entire domain.

[0088] S203. Construct a gray-level histogram based on all valid probability values ​​in the smooth probability map, and analyze the gray-level histogram using the maximum inter-class variance method to obtain the optimal threshold for segmenting contaminated and clean areas.

[0089] Indicatively, all valid probability values ​​within the effective area of ​​the solar disk are extracted from the smoothed probability map. A grayscale histogram is constructed based on the distribution characteristics of these probability values, clearly showing the pixel quantity distribution corresponding to different probability intervals. The maximum inter-class variance method is used to analyze this grayscale histogram. By calculating the inter-class variance between the foreground and background under different thresholds, the threshold with the largest inter-class variance is selected as the optimal segmentation threshold, achieving adaptive differentiation between contaminated and clean areas.

[0090] Specifically, statistical analysis is performed on all probability values ​​belonging to the effective region of the solar disk in the smoothed probability map. The natural range of these probability values ​​is determined to be [0,1]. To achieve accurate histogram distribution statistics, this interval is uniformly discretized into... There are 10 levels, each corresponding to the same probability interval width, where... Representing the The number of pixels corresponding to the probability values ​​included in each level was calculated, and invalid pixels outside the valid area of ​​the solar surface were strictly excluded during the statistical process to ensure that the histogram distribution could accurately reflect the distribution characteristics of the pollution probability. Then, all possible threshold levels were iterated. , The range of values ​​is from arrive Each threshold level A dividing point in the range of corresponding probability values Using this as a boundary, all valid pixels are divided into two categories, with category C0 containing pixels with a probability value less than 1. pixels, The class contains a probability value greater than or equal to The pixels. Based on this classification, the inter-class variance between the two classes is calculated. ,in This represents the total number of pixels within the effective area of ​​the solar disk. express The proportion of pixel-like structures is calculated as follows: The ratio of the sum of the number of pixels in each level to the total number of pixels N, i.e. ; express The proportion of each pixel type is such that, since the sum of the proportions of the two pixel types is 1, therefore... ; for The average probability value of the pixel class. for The average probability value of a pixel class and the physical meaning of inter-class variance are to quantify the degree of separation between two classes of pixels in terms of probability values. The larger the value, the more significant the difference between the two classes and the better the segmentation effect.

[0091] After calculating the inter-class variance for all possible threshold levels T, select those that minimize the inter-class variance. Threshold level for reaching the maximum value Since this threshold level is based on discretization The values ​​obtained from each level need to be mapped back to the original probability value range to obtain the optimal segmentation threshold that is actually usable. The mapping method is as follows: This mapping process is achieved by multiplying the level number by the probability interval width, ensuring that the threshold accurately corresponds to the actual probability distribution in the probability map.

[0092] Optionally, to avoid significant fluctuations in the optimal segmentation threshold due to anomalies in a single frame image, the optimal segmentation threshold at the current moment is obtained... Next, time smoothing constraints need to be applied. Specifically, the threshold at the current time step needs to be calculated. Compared with the optimal segmentation threshold at the previous time step The difference is pre-set with a maximum allowable threshold change. If the difference exceeds a preset reasonable range, then the threshold at the current time will be adjusted. Restricted to That is, when hour, Pick ;when hour, Pick This constraint mechanism ensures the stability of the threshold over time, thereby improving the continuity and reliability of the cloud contamination binary mask.

[0093] S204. Traverse each position in the smooth probability map. If the probability value is greater than or equal to the optimal threshold for separating contaminated and clean areas, mark it as 1 at the corresponding position in the output mask. If the probability value is less than the optimal threshold for separating contaminated and clean areas, mark it as 0 at the corresponding position in the output mask to obtain the full-surface cloud contamination binary mask.

[0094] Each spatial location in the smoothed probability map is iterated through one by one, and the relationship between the probability value of that location and the optimal segmentation threshold is determined. If the probability value of that location is greater than or equal to the optimal segmentation threshold, it is marked as 1 at the corresponding location in the output mask, representing that the location belongs to the cloud-contaminated area; if the probability value of that location is less than the optimal segmentation threshold, it is marked as 0 at the corresponding location in the output mask, representing that the location belongs to the clean area. Finally, a full-surface cloud-contaminated binary mask that completely covers the effective area of ​​the solar surface is formed.

[0095] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0096] Based on the same inventive concept, this application also provides a full-plane image cloud contamination detection system for implementing the above-described full-plane image cloud contamination detection method. The solution provided by this system is similar to the implementation described in the above method; therefore, the specific limitations in one or more embodiments of the full-plane image cloud contamination detection system provided below can be found in the limitations of the full-plane image cloud contamination detection method described above, and will not be repeated here.

[0097] In one exemplary embodiment, such as Figure 3 As shown, a full-area image cloud contamination detection system is provided, comprising:

[0098] The full-surface image module 301 is used to acquire a full-surface image sequence and preprocess the full-surface image sequence within the effective area of ​​the solar surface to obtain multiple image blocks and corresponding DCT feature vectors; the preprocessing corresponds to solar surface spatial grid division and selection of low-frequency coefficients of two-dimensional discrete cosine transform;

[0099] The spatiotemporal graph module 302 is used to construct a spatiotemporal graph based on the spatial adjacency relationship between image patches and the temporal sequence relationship of image patches at the same location, combined with DCT feature vectors. The spatiotemporal graph includes graph nodes, temporal edges, and spatial edges. Graph nodes are represented by the DCT feature vectors of each image patch at each observation time. Spatial edges connect graph nodes that are spatially adjacent at the same observation time. Temporal edges connect graph nodes at the same spatial location at different observation times.

[0100] The spatiotemporal aggregation and propagation module 303 is used to input the spatiotemporal graph into a pre-trained DCT-based graph neural network model to obtain the high-level feature vectors of each graph node; the DCT-based graph neural network model performs multiple rounds of information propagation and aggregation along the spatial and temporal edges to update the DCT feature vectors of each graph node.

[0101] The cloud contamination probability module 304 is used to decode the high-level feature vectors of each node to obtain the cloud contamination probability value of each corresponding image patch, and generate a cloud contamination probability map of the entire solar surface based on the cloud contamination probability values ​​of all image patches.

[0102] The cloud coverage module 305 is used to perform temporal smoothing and adaptive threshold binarization on the cloud pollution probability map, generate a full-surface cloud pollution binary mask, and calculate the cloud coverage rate quantification index based on the full-surface cloud pollution binary mask.

[0103] In one embodiment, the spatiotemporal graph module 302 is further configured to:

[0104] Each image patch is defined as a graph node at each observation time, and each graph node is assigned a corresponding DCT feature vector as the initial node feature.

[0105] For each set of graph nodes at each observation time, all graph node pairs are traversed. If the image patch corresponding to the graph node pair satisfies the preset adjacency condition in the solar space grid, an undirected edge is established between the two graph nodes of the graph node pair as a spatial edge.

[0106] For each image patch corresponding to a spatial location, a directed edge is established between the corresponding graph nodes at adjacent times, pointing from an earlier time to a later time, according to the observation time order;

[0107] By merging spatial edges and temporal edges, and combining the graph nodes and their corresponding initial node features, a spatiotemporal graph is obtained.

[0108] In one embodiment, the spatiotemporal aggregation propagation module 303 is further configured to:

[0109] The DCT feature vectors of the graph nodes are used as the embedding representation of the zeroth layer nodes in the DCT-based graph neural network model.

[0110] Based on the zero-level node embedding representation, multi-level forward propagation is performed, and the final feature representation obtained in the last layer is determined as the high-level feature vector representation of the graph nodes;

[0111] The forward propagation process includes the following steps:

[0112] Based on the Layered graph neural networks spatially aggregate graph nodes to obtain intermediate feature representations; the intermediate feature representations are as follows: ,in, For the observation time Spatial location The graph node passes through the first Intermediate features after layer space aggregation; For the graph node at the th Features of the layer; For graph nodes The set of neighboring nodes; This is a normalization constant based on node degree; For the first Layer-learnable weight matrix; It is a non-linear activation function;

[0113] Based on temporal edges, a gating mechanism is used to fuse intermediate feature representations with the final feature representations of graph nodes at adjacent time points at the same location, resulting in the final feature representation of the graph node; the gating mechanism is expressed as follows: , , ,in, To update the gate vector, For graph nodes At any moment No. The final feature representation of the layer, For candidate feature vectors, For element-wise multiplication, and The weight matrix is ​​a learnable matrix. For graph nodes In the same location Time Node Characteristics of the layer.

[0114] In one embodiment, the cloud coverage module 305 is further configured to:

[0115] Obtain the cloud contamination probability map at the current moment, the cloud contamination probability map at the adjacent previous moment, and the cloud contamination probability map at the adjacent subsequent moment.

[0116] For the probability value of each spatial location in the probability map at the current time, a weighted median filter is applied to the probability value of the same location in the cloud pollution probability maps of adjacent preceding and subsequent times to obtain the time-smoothed probability value, and a smoothed probability map is generated.

[0117] A gray-level histogram is constructed based on all valid probability values ​​in the smooth probability map, and the gray-level histogram is analyzed using the maximum inter-class variance method to obtain the optimal threshold for segmenting contaminated and clean areas.

[0118] Traverse each position in the smooth probability map. If the probability value is greater than or equal to the optimal threshold for separating contaminated and clean areas, mark it as 1 at the corresponding position in the output mask. If the probability value is less than the optimal threshold for separating contaminated and clean areas, mark it as 0 at the corresponding position in the output mask to obtain a binary mask of cloud contamination on the entire solar surface.

[0119] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps in the above method embodiments.

[0120] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0121] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0122] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A method for detecting cloud contamination in full-solar surface images, characterized in that, The method includes: A full-surface image sequence is acquired, and the full-surface image sequence is preprocessed within the effective area of ​​the solar surface to obtain multiple image blocks and corresponding DCT feature vectors; the preprocessing corresponds to solar surface spatial grid division and selection of low-frequency coefficients of two-dimensional discrete cosine transform; Based on the spatial adjacency between the image patches and the temporal sequence of image patches at the same location, a spatiotemporal graph is constructed using the DCT feature vectors. The spatiotemporal graph includes graph nodes, temporal edges, and spatial edges. The graph nodes are represented by the DCT feature vectors of each image patch at each observation time. The spatial edges connect graph nodes that are spatially adjacent at the same observation time. The temporal edges connect graph nodes at the same spatial location at different observation times. The spatiotemporal graph is input into a pre-trained DCT-based graph neural network model to obtain the node high-level feature vectors of each graph node; the DCT-based graph neural network model performs multiple rounds of information propagation and aggregation along the spatial edge and the temporal edge to update the DCT feature vectors of each graph node. The high-level feature vectors of each node are decoded to obtain the cloud contamination probability value of each image patch, and a cloud contamination probability map of the entire solar surface is generated based on the cloud contamination probability values ​​of all image patches. The cloud contamination probability map is subjected to temporal smoothing and adaptive threshold binarization to generate a full-surface cloud contamination binary mask, and the cloud coverage rate quantification index is calculated based on the full-surface cloud contamination binary mask.

2. The method according to claim 1, characterized in that, The construction of a spatiotemporal map based on the spatial adjacency relationships between the image patches and the temporal sequence relationships of image patches at the same location, combined with the DCT feature vectors, includes: Each image block is defined as a graph node at each observation time, and each graph node is assigned a corresponding DCT feature vector as an initial node feature. For each set of graph nodes at each observation time, all graph node pairs are traversed. If the image block corresponding to the graph node pair satisfies the preset adjacency condition in the solar space grid, then an undirected edge is established between the two graph nodes of the graph node pair as the space edge. For each image patch corresponding to a spatial location, a directed edge pointing from an earlier time to a later time is established between the corresponding graph nodes at adjacent times, according to the observation time order, as the time edge; The spatial edge and the temporal edge are merged, and the spatiotemporal graph is obtained by combining the graph nodes and their corresponding initial node features.

3. The method according to claim 2, characterized in that, The step of inputting the spatiotemporal graph into a pre-trained DCT-based graph neural network model to obtain the node high-level feature vectors of each graph node includes: The DCT feature vector of the graph node is used as the zeroth layer node embedding representation of the DCT-based graph neural network model; Based on the zero-level node embedding representation, multi-level forward propagation is performed, and the final feature representation obtained in the last layer is determined as the high-level feature vector representation of the graph node; The forward propagation corresponds to the following steps: Based on the A layered graph neural network performs spatial aggregation on the graph nodes to obtain intermediate feature representations; the intermediate feature representations are as follows: ,in, For the observation time Spatial location The graph node passes through the first Intermediate features after layer space aggregation; For the graph node in the th Features of the layer; For graph nodes The set of neighboring nodes; This is a normalization constant based on node degree; For the first Layer-learnable weight matrix; It is a non-linear activation function; Based on the aforementioned time edge, the intermediate feature representation is fused with the final feature representation of the graph node at the same adjacent time point using a gating mechanism to obtain the final feature representation of the graph node; the gating mechanism is expressed as follows: , , ,in, To update the gate vector, For graph nodes At any moment No. The final feature representation of the layer, For candidate feature vectors, For element-wise multiplication, and The weight matrix is ​​a learnable matrix. For graph nodes In the same location Time Node Characteristics of the layer.

4. The method according to claim 1, characterized in that, The process of decoding the high-level feature vectors of each node to obtain the cloud contamination probability value for each corresponding image patch includes: The probability value of cloud contamination can be obtained using the following formula: in, This refers to the high-level feature vector; This is the weight matrix; For bias; and Image blocks At any moment The probability of being classified as clean and the probability of being classified as cloud-polluted, and .

5. The method according to claim 1, characterized in that, The step of performing temporal smoothing and adaptive threshold binarization on the cloud contamination probability map to generate a full-day cloud contamination binary mask includes: Obtain the cloud contamination probability map at the current moment, the cloud contamination probability map at the adjacent previous moment, and the cloud contamination probability map at the adjacent subsequent moment. For the probability value of each spatial location in the probability map at the current time, a weighted median filter is applied to the probability value of the same location in the cloud pollution probability maps of the adjacent preceding time and the adjacent subsequent time to obtain the probability value after time-series smoothing, and a smooth probability map is generated. A gray-level histogram is constructed based on all valid probability values ​​in the smooth probability map, and the gray-level histogram is analyzed using the maximum inter-class variance method to obtain the optimal threshold for segmenting contaminated and clean areas. Traverse each position in the smooth probability map. If the probability value is greater than or equal to the optimal threshold for segmenting contaminated and clean areas, mark it as 1 at the corresponding position in the output mask. If the probability value is less than the optimal threshold for segmenting contaminated and clean areas, mark it as 0 at the corresponding position in the output mask to obtain the full-surface cloud contamination binary mask.

6. A full-plane image cloud contamination detection system, characterized in that, The system includes: The full-surface image module is used to acquire a full-surface image sequence and preprocess the full-surface image sequence within the effective area of ​​the solar surface to obtain multiple image blocks and corresponding DCT feature vectors; the preprocessing corresponds to solar surface spatial grid division and selection of low-frequency coefficients of two-dimensional discrete cosine transform; The spatiotemporal graph module is used to construct a spatiotemporal graph based on the spatial adjacency relationships between the image patches and the temporal sequence relationships of image patches at the same location, combined with the DCT feature vectors. The spatiotemporal graph includes graph nodes, temporal edges, and spatial edges. The graph nodes are represented by the DCT feature vectors of each image patch at each observation time. The spatial edges connect graph nodes that are spatially adjacent at the same observation time. The temporal edges connect graph nodes at the same spatial location at different observation times. The spatiotemporal aggregation and propagation module is used to input the spatiotemporal graph into a pre-trained DCT-based graph neural network model to obtain the node high-level feature vectors of each graph node; the DCT-based graph neural network model performs multiple rounds of information propagation and aggregation along the spatial edge and the temporal edge to update the DCT feature vectors of each graph node. The cloud contamination probability module is used to decode the high-level feature vectors of each node to obtain the cloud contamination probability value of each image patch, and generate a cloud contamination probability map of the entire solar surface based on the cloud contamination probability values ​​of all image patches. The cloud coverage module is used to perform temporal smoothing and adaptive threshold binarization on the cloud pollution probability map to generate a full-surface cloud pollution binary mask, and calculate the cloud coverage rate quantification index based on the full-surface cloud pollution binary mask.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 5.

Images