Annual cloud coverage prediction method and device based on graph neural network and monthly features

Abstract

The application discloses an annual cloud coverage prediction method and device based on a graph neural network and monthly features, and belongs to the technical field of remote sensing information intelligent processing, atmospheric science and deep learning. The method comprises the following steps: acquiring historical cloud coverage data of a geographical sampling point, and constructing a 12-dimensional monthly feature vector as an input feature for each sampling point; constructing an adjacency matrix representing the topological relationship of a space according to the geographical coordinates of the sampling point and based on a spherical distance and a K nearest neighbor algorithm; inputting the input feature and the adjacency matrix into a pre-trained spatio-temporal graph neural network model, and outputting a daily cloud coverage grade prediction result of a target year; and the model is composed of a cascaded residual graph convolution network module, a bidirectional long short-term memory network module, a multi-head attention module and a plurality of independent monthly classifiers. The application realizes annual scale cloud coverage prediction only by relying on historical data, and has significant precision advantages and generalization capabilities in complex terrain areas and scenes with insufficient historical data.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary fields of remote sensing information intelligent processing, atmospheric science and deep learning, and specifically relates to an annual cloud cover prediction method and device based on graph neural networks and monthly features. Background Technology

[0002] Cloud cover is a key factor affecting the quality of optical remote sensing data acquisition. In applications such as agricultural monitoring, ecological environment assessment, and disaster emergency response, high cloud cover often leads to the loss of effective observation windows, causing data gaps. Therefore, predicting future cloud cover conditions in advance is of great significance for optimizing satellite mission scheduling and improving the continuity of remote sensing products.

[0003] Currently, cloud coverage prediction mainly relies on the following two technical approaches:

[0004] The first category consists of numerical weather prediction models, such as the Integrated Forecasting System of the European Centre for Medium-Range Weather Prediction or the Weather Research and Forecasting Model (MRTM) in the United States. These methods simulate cloud formation processes by solving atmospheric physical equations, providing high-precision short-term forecasts. However, these models have significant drawbacks: first, they are extremely computationally expensive, requiring vast computational resources to solve large-scale dynamic equations; second, they are complex systems involving precise settings of numerous physical parameters; and third, they are difficult to extend to annual forecasts, as their physical modeling logic limits their application to short-term forecasts of a few hours to a few days in the future, failing to meet the annual planning needs of remote sensing Earth observation missions.

[0005] The second category comprises statistical methods based on historical remote sensing data, a typical example being the historical climate mode prediction method. This method utilizes the frequency of cloud cover on the same day and month over many years to statistically determine the most frequent category as the basis for prediction. This type of method has advantages such as simple computation and no need for real-time observation data. However, its core drawback lies in its assumption that each geographical sampling point is independent, completely ignoring the spatial continuity and propagation characteristics of the cloud system. In areas with complex terrain, such as the Sichuan Basin or the karst landforms of Guangxi, isolated point prediction is easily affected by local noise, leading to limited prediction accuracy and insufficient generalization ability.

[0006] In recent years, deep learning technology has been widely applied in the field of remote sensing information processing. Methods such as convolutional neural networks (CNNs) and long short-term memory (LSM) networks have made some progress in short-term forecasting. However, existing deep learning methods still have the following limitations: First, CNNs typically rely on regular grid structures, making them difficult to adapt to irregularly distributed geographic sampling points; second, while LSM networks can capture temporal dependencies, they cannot explicitly model the geospatial adjacency relationships between sites; third, these methods heavily rely on recent continuous observation sequences as input, making them unsuitable for long-term planning tasks where no observations are available for the year in question; fourth, deep networks are prone to gradient vanishing or loss of original physical statistical signals during training, leading to a lack of physical plausibility in the prediction results.

[0007] In summary, existing technologies still lack a cloud cover prediction method that can take into account spatial topological perception, multi-scale temporal modeling, and good generalization ability, which is suitable for practical applications such as regional long-term climate prediction and annual planning of remote sensing Earth observation missions. Summary of the Invention

[0008] To address the aforementioned technical problems, this invention provides an annual cloud cover prediction method and apparatus based on graph neural networks and monthly features. By constructing a spatiotemporal graph neural network architecture that integrates residual graph convolutional networks and bidirectional long short-term memory networks, and by aggregating spatial topological features between geographic sampling points using a K-nearest neighbor spatial adjacency graph, and by combining a multi-head attention mechanism to adaptively weight historical year-on-year time series features, high-precision prediction of annual daily cloud cover levels is achieved by relying solely on historical monthly statistical data.

[0009] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0010] An annual cloud cover prediction method based on graph neural networks and monthly features, the method comprising:

[0011] Step 1: Obtain historical cloud cover data for geographical sampling points, and construct a 12-dimensional monthly input feature and an output label representing the daily cloud cover level vector for each sampling point;

[0012] Step 2: Based on the geographical coordinates of each sampling point, construct an adjacency matrix representing spatial topological relationships using spherical distance and the K-nearest neighbor algorithm;

[0013] Step 3: Input the 12-dimensional monthly input features and adjacency matrix into the pre-trained spatiotemporal graph neural network model, and output the daily cloud cover level prediction results for the target prediction year; wherein, in the spatiotemporal graph neural network model, the residual graph convolutional network layer aggregates the spatial neighborhood features defined by the adjacency matrix based on the input features and outputs enhanced spatial features; the bidirectional long short-term memory network layer performs long-range dependency modeling on the enhanced spatial features in the historical year-on-year time dimension and outputs latent features that fuse spatiotemporal dependencies; the multi-head self-attention weighted layer adaptively weights the time-series features of each date throughout the year based on the latent features and outputs the four-class daily cloud cover level probability.

[0014] Furthermore, step 1 includes: the input feature is a 12-dimensional monthly feature vector obtained based on the daily cloud coverage level statistics of historical years; the output label is a four-category daily cloud coverage level vector consisting of sunny, cloudy, mixed, and unknown, defined by the cloud mask status bits of the MODIS product.

[0015] Furthermore, each dimension of the 12-dimensional monthly feature vector is obtained by calculating the arithmetic mean of the cloud cover levels of all valid observation days in the corresponding month and rounding it to the nearest integer. If there is no valid observation data for the month, it is marked as a preset missing value.

[0016] Furthermore, step 2 includes: calculating the physical distance between any two sampling points based on the spherical distance formula, using the K nearest neighbor algorithm, and constructing an adjacency matrix representing the spatial topological relationship based on the calculated physical distance. The rule for the value of the elements in the adjacency matrix is: if a sampling point belongs to the K nearest neighbor nodes of another sampling point, the corresponding element is set to one; otherwise, it is set to zero.

[0017] Furthermore, in step 3, the residual graph convolutional network layer includes a first graph convolutional layer to a fourth graph convolutional layer; the first graph convolutional layer is used to map the input 12-dimensional monthly feature vector to a 128-dimensional high-dimensional feature space; the second graph convolutional layer to the fourth graph convolutional layer are used to extract deep spatial topological features through three consecutive graph convolution operations; starting from the second graph convolutional layer, a residual connection mechanism is introduced to superimpose the features output from the previous layer with the output of the current layer after graph convolution, batch normalization and LeakyReLU nonlinear activation.

[0018] Furthermore, in step 3, the bidirectional long short-term memory network layer is composed of three cascaded bidirectional long short-term memory networks, which are used to perform bidirectional long-range time dependency modeling on the enhanced spatial features output by the residual graph convolutional network layer along the historical year dimension, and output implicit features that fuse forward and backward information.

[0019] Furthermore, the multi-head self-attention weighted layer adopts an eight-head cross-attention mechanism, using 365 learnable date query vectors as the query basis, and directly using the latent features output by the bidirectional long short-term memory network layer as key vectors and value vectors. After attention calculation, it outputs daily features. The daily features are divided into 12 groups according to the month, and are respectively input to 12 independent monthly classifiers. Each classifier is composed of fully connected layers, and outputs the predicted probability of four-class daily cloud coverage level through the Softmax function.

[0020] On the other hand, the present invention provides an annual cloud cover prediction device based on graph neural networks and monthly features, comprising:

[0021] The acquisition module is used to acquire historical cloud cover data of geographical sampling points, and construct a 12-dimensional monthly input feature and an output label representing the daily cloud cover level vector for each sampling point;

[0022] The matrix construction module is used to construct an adjacency matrix representing spatial topological relationships based on the geographic coordinates of each sampling point and the spherical distance and K-nearest neighbor algorithm.

[0023] The prediction output module is used to input the 12-dimensional monthly input features and the adjacency matrix into a pre-trained spatiotemporal graph neural network model, and output the daily cloud cover level prediction results for the target prediction year. In the spatiotemporal graph neural network model, the residual graph convolutional network layer aggregates spatial neighborhood features defined by the adjacency matrix based on the input features and outputs enhanced spatial features; the bidirectional long short-term memory network layer models the long-range dependency of the enhanced spatial features in the historical year-on-year time dimension and outputs latent features that fuse spatiotemporal dependencies; and the multi-head self-attention weighted layer adaptively weights the temporal features of each date throughout the year based on the latent features and outputs the four-class daily cloud cover level probability.

[0024] Thirdly, the present invention provides an electronic device, comprising: one or more processors; and a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned annual cloud coverage prediction method based on graph neural networks and monthly features.

[0025] Fourthly, the present invention provides a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, enable the processor to implement the aforementioned annual cloud coverage prediction method based on graph neural networks and monthly features.

[0026] The beneficial effects of this invention are as follows:

[0027] First, it achieves explicit modeling of geospatial topological attributes. This invention constructs a graph convolutional network based on K-nearest neighbors to quantitatively aggregate the geographic association features of the target point and its neighboring nodes. This enables the model to identify the spatial continuity and propagation characteristics of cloud systems. In areas with complex terrain, it utilizes the collaborative information of surrounding sampling points to suppress local observation noise, effectively improving the prediction accuracy in irregularly distributed scenarios.

[0028] Second, a residual information transfer mechanism that balances low-level statistical features with deep abstraction was established. This invention introduces a residual connection path into the spatial coding module, directly transmitting the historical year-on-year features of the original input to the deep network. While extracting high-order spatiotemporal features, it retains the underlying meteorological statistical patterns, thus solving the gradient vanishing problem in deep network training and ensuring the physical rationality of the prediction results.

[0029] Third, it possesses adaptive feature filtering capabilities for non-stationary meteorological data. The multi-head self-attention mechanism introduced in this invention can adaptively allocate weights for each moment in historical years according to forecast target requirements. By capturing spatiotemporal correlations of different dimensions in parallel through multiple independent attention heads, it can accurately locate key evolutionary nodes in complex historical sequences and automatically reduce the impact of anomalous climate fluctuations.

[0030] Fourth, it solves the problem of dependence on recent continuous observation sequences in long-term mission planning. This invention establishes a forecasting logic based on "historical pattern extrapolation" rather than "recent state extrapolation" by spatial topological aggregation and bidirectional temporal evolution modeling of historical year-on-year characteristics, and can still output high-precision annual forecast results even without real-time observation data access.

[0031] Fifth, it achieves a balance between high-precision forecasting and low computational cost. Compared to numerical weather prediction models that require solving large-scale dynamic equations, this invention can quickly generate regional-scale cloud cover forecasts through forward computation after training, significantly reducing computational overhead and facilitating practical deployment and application. Attached Figure Description

[0032] Figure 1 This is a flowchart of the annual cloud coverage prediction method based on graph neural networks and monthly features of the present invention.

[0033] Figure 2 This is a structural diagram of the spatiotemporal graph neural network model of the present invention. Detailed Implementation

[0034] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0035] This application relates to an annual cloud cover prediction method and device based on graph neural networks and monthly features. Its core is a spatiotemporal graph neural network architecture that integrates spatial topological features and temporal evolution laws. This architecture consists of a residual graph convolutional network layer, a bidirectional long short-term memory network layer, and a multi-head self-attention weighting mechanism, which are cascaded together to achieve hierarchical processing of meteorological feature vectors.

[0036] The system constructs a residual graph convolutional module based on KNN topology to aggregate the geographic correlation features of the target sampling point and its neighboring nodes. This module introduces residual connection paths to pass the historical year-on-year features of the original input across layers to the convolutional output for superposition. This mechanism effectively alleviates the oversmoothing of features in graph neural networks when processing meteorological data while maintaining the convergence of deep models, preserving the observational statistical properties of the underlying layers.

[0037] Bidirectional Long Short-Term Memory (Bi-LSTM) networks utilize a bidirectional gating mechanism to extract long-range dependency features of meteorological sequences over historical year-on-year spans from both forward and backward dimensions. This design enables the model to extract the nonlinear evolution logic of meteorological features over time. By modeling historical year-on-year data, this module maintains the accuracy of forecast results and ensures the model's adaptability to scenarios with missing data.

[0038] In the feature weighting and output stage, this application introduces a multi-head attention mechanism based on learnable date query vectors. It uses 365 learnable date query vectors and the time-series features output by Bi-LSTM to perform cross-attention calculation, thereby achieving adaptive weighting of features for each date throughout the year. Finally, it outputs the daily cloud cover level probability through 12 independent monthly classifiers to better capture the seasonal features of different months.

[0039] Specifically, such as Figure 1 As shown, the annual cloud cover prediction method based on graph neural networks and monthly features includes:

[0040] Step 1: Obtain historical cloud cover data for geographical sampling points, and construct input features and output labels for each sampling point. The input features are 12-dimensional monthly feature vectors obtained based on daily cloud cover level statistics of historical years; the output labels are four-category daily cloud cover level vectors corresponding to the target prediction year, consisting of clear, cloudy, mixed, and unknown categories defined by the cloud mask status bits of the MODIS product; including:

[0041] The cloud mask state information (State QA) of the MOD09GA product is used as the raw data. The information of bits 0-1 in the State QA flag is decoded to construct a refined four-class discrete tag system. These correspond to: sunny, cloudy, mixed, and unknown, respectively.

[0042] For a given target prediction year t, construct the following monthly feature data for each geographic sampling point:

[0043] Input feature vector : is a 12-dimensional vector representing the average cloud cover level of the geographic sampling point in each month within the target prediction year t (obtained by rounding the arithmetic mean of all valid observations for that month; if there is no data for that month, it is marked as -1):

[0044] ,

[0045] Where: 1, 2, 3, 4 are the four-class cloud mask labels mapped by the MOD09GA official QA standard, t is the target prediction year, and the superscript T indicates transpose.

[0046] Output year tags A 365- or 366-dimensional vector representing the daily cloud cover level at that point within the target predicted year t:

[0047] ,

[0048] The model learns cross-scale mapping patterns from 12-dimensional monthly statistical features to 365 / 366-dimensional daily evolution details. This design enables the model to utilize macroscopic climatic background features to prospectively extrapolate the daily state of the entire year.

[0049] Step 2: Based on the geographical coordinates of each sampling point, construct an adjacency matrix representing spatial topological relationships using spherical distance and the K-Nearest Neighbors (KNN) algorithm;

[0050] Irregular geographic sampling points within the region are abstracted as graph nodes to construct a geospatial adjacency graph. The physical distance between any two points i and j is calculated using the Haversine Formula. :

[0051] ,

[0052] Where: Earth's radius Δlat and Δlon represent the latitude and longitude differences between points i and j, respectively. , Let i and j represent the latitudes of the two points i and j, respectively.

[0053] The adjacency matrix is ​​constructed using the K-nearest neighbor algorithm:

[0054] ,

[0055] The optimal number of spatial neighbor nodes K is 8, and KNN represents the K-nearest neighbor algorithm. In practical applications, it can be adjusted within the range of 4 to 20 according to the sampling point distribution density to ensure the local continuity of spatial feature aggregation.

[0056] Without loss of generality, those skilled in the art can employ spatial weight matrices based on radial basis functions (RBF), or Chebyshev networks and graph attention networks (GAT) as spatial extraction operators. Furthermore, for regularly distributed geographic sampling points, dilated convolution or deformable convolution can be used to capture cross-regional spatial dependencies. While these alternatives differ in computational efficiency or parameter count, their core logic remains the same: extracting geographic topological features through neighborhood aggregation.

[0057] Step 3: Input the input features and adjacency matrix into a pre-trained spatiotemporal graph neural network model, and output the daily cloud cover level prediction result for the target prediction year; wherein, the spatiotemporal graph neural network model consists of the following cascaded modules: a residual graph convolutional network layer, which aggregates spatial neighborhood features defined by the adjacency matrix based on the input features and outputs enhanced spatial features; a bidirectional long short-term memory network layer, which models the long-range dependency of the enhanced spatial features in the historical year-on-year time dimension and outputs latent features that fuse spatiotemporal dependencies; a multi-head self-attention weighted layer, which adaptively weights the time-series features of each date throughout the year based on the learnable date query vector and latent features; and multiple independent monthly classifiers, which output four-class daily cloud cover level probabilities.

[0058] like Figure 2 As shown, this invention constructs a deep learning framework composed of residual graph convolutional network layers, bidirectional long short-term memory network layers, and multi-head self-attention weighted layers. Its specific architecture and parameters are described below:

[0059] Residual GCN (Residual Generative Network) layer:

[0060] The model uses the input feature vector at each sampling point. Starting with 12-dimensional monthly historical features, the model first maps these features to a 128-dimensional high-dimensional space using the first layer of graph convolution (GCN Layer 1), achieving initial feature dimensionality upscaling and exchange. Subsequently, the features enter a cascaded aggregation stage consisting of three layers of graph convolution (GCN Layers 2-4). During this process, the model utilizes an adjacency matrix generated based on the K-nearest neighbor algorithm. By aggregating the feature information of neighboring nodes, each sampling point can perceive the environmental impact of the surrounding geographic space.

[0061] Specifically, a multi-layer graph convolution operator is used to aggregate features of neighboring nodes, where the first layer (see...) Figure 2 The GCN Layer 1 is a feature exchange layer responsible for mapping the input 12-dimensional monthly historical feature vector to a 128-dimensional high-dimensional feature space, achieving preliminary information exchange between nodes under the constraint of the adjacency matrix. GCN Layers 2-4 are feature aggregation layers, which further extract deep spatial topological features through three consecutive layers of graph convolution operations. Each layer integrates batch normalization (BN) to accelerate convergence, LeakyReLU nonlinear activation to extract nonlinear features, and regularized dropout layers to prevent model overfitting.

[0062] Residual connection:

[0063] To preserve the statistical significance of climate observations at the bottom layer and address the common "oversmoothing" problem in deep graph networks (i.e., node features tending to be consistent after multiple convolutions), the model introduces residual connection paths starting from the second layer: the original features from each layer's input are vector-superimposed with the output after graph convolution, batch normalization, and nonlinear activation. This design allows the original statistical information from the shallow layers to be passed across layers to the deep layer output, maintaining the expressive power of the deep network while preserving the historical climate characteristics of the nodes themselves. After iterative evolution through four layers of residual graph convolutions, the model finally outputs topological features that integrate multi-order spatial relationships. (dimension is) , (The number of sampling points) enables the feature dimensionality upgrade from isolated points to a spatially continuous field.

[0064] Specifically, from the second level (see...) Figure 2 The GCN Layer 2 introduces a residual connection mechanism to preserve the saliency of a node's own climate features. This mechanism establishes an identity mapping path, directly passing the features output from the previous layer to the output of the current convolutional layer for aggregation. This design ensures that when the model aggregates spatial neighborhood features through deep networks, it can preserve the original saliency of the node's own climate features, effectively alleviating the "feature oversmoothing" phenomenon commonly found in graph neural networks. Its forward propagation process is as follows:

[0065] ,

[0066] in:

[0067] For the convolution operation of the l-th layer graph,

[0068] A is the adjacency matrix.

[0069] BN stands for batch normalization.

[0070] This is a feature of the previous layer.

[0071] The number of graph convolutional layers can be set between 2 and 8 layers depending on the complexity of the geographical environment, and the recommended dimension for the hidden feature layers is between 64 and 512. Activation function. The activation function is non-linear, preferably LeakyReLU, with a preferred negative slope of 0.01. Specifically:

[0072] First, the initial input monthly historical feature vector is fed into the first graph convolutional layer for feature exchange, resulting in the initial spatial mapping features:

[0073] ,

[0074] A is the previously constructed adjacency matrix, X is the previously constructed monthly feature vector, and BN is the batch normalization process. It can be the LeakyReLU activation function.

[0075] Subsequently, the features enter the cascaded feature aggregation stage (see...). Figure 2 The GCN consists of Layers 2 through 4, and residual connection paths are introduced to preserve the statistical properties of the lower-level observations. For the l-th layer (layers 2, 3, and 4), its output features are... The calculation formula is:

[0076] ,

[0077] Finally, after the last feature aggregation layer, the output is a spatial topological feature that incorporates multi-level spatial neighborhood relationships:

[0078] ,

[0079] It is then passed as a temporal input to a bidirectional long short-term memory (Bi-LSTM) network for subsequent temporal evolution modeling.

[0080] Without loss of generality, those skilled in the art can employ spatial weight matrices based on radial basis functions (RBF), or Chebyshev networks and graph attention networks (GAT) as spatial extraction operators. Furthermore, for regularly distributed geographic sampling points, dilated convolution or deformable convolution can be used to capture cross-regional spatial dependencies. While these alternatives differ in computational efficiency or parameter count, their core logic remains the same: extracting geographic topological features through neighborhood aggregation.

[0081] Bidirectional Long Short-Term Memory (Bi-LSTM) network layer:

[0082] Spatial topology features of the preamp output The data is expanded along the historical year dimension to form a time series, which is then input into a three-layer bidirectional long short-term memory (Bi-LSTM) network. The Bi-LSTM utilizes a bidirectional gating mechanism to simultaneously extract long-range time dependencies from both the forward and reverse dimensions of historical year-on-year data, capturing the nonlinear evolution patterns of climate elements on multi-year scales. The output is a latent feature that integrates forward and reverse information. (Dimension Preservation) ).

[0083] Specifically, based on spatial feature extraction, a multi-layer bidirectional long short-term memory network is introduced to perform deep modeling of historical evolution trends:

[0084] ,

[0085] The output features after time series modeling This represents the spatial topological features output by the previous graph convolutional layer.

[0086] The two-way structure simultaneously captures the forward correlations and backward feedbacks of climate evolution along the time axis. The number of time series modeling layers is typically set to 2 to 4.

[0087] Without loss of generality, alternative solutions can employ gated recurrent units (GRUs) or bidirectional gated recurrent units (Bi-GRUs) to reduce parameter complexity and shorten training time. Furthermore, for long-sequence tasks, temporal convolutional networks (TCNs) can be used to capture long-range temporal dependencies through causal convolution and dilated convolution structures. These temporal processing units are essentially designed to extract nonlinear evolution patterns from historical year-on-year data, representing equivalent transformations of the temporal modeling module described in this application.

[0088] Multi-head Attention:

[0089] To further enhance feature selection and adapt to the need for daily prediction throughout the year, the model introduces an 8-head cross-multi-head attention mechanism. This mechanism uses 365 learnable date query vectors as the query baseline (Q), representing the prediction target for each day of the year; simultaneously... Two independent weight matrices are projected onto a key vector (K) and a value vector (V). The key vector is used to calculate the matching weight (i.e., attention score) between the target date and historical time, while the value vector carries the actual meteorological evolution content. After query-key dot product, scaling, and softmax normalization, the value vector is weighted and summed to adaptively aggregate information from the most relevant historical time for each date. This process outputs daily features (365×128 dimensions), achieving a cross-time mapping from historical time-series features to the daily scale of the target year.

[0090] Its specific parameterization and calculation sources are as follows:

[0091] 1. Query Vector (Q):

[0092] 365 learnable date lookup vectors are used. This vector sequence Q l ∈R 365×d Regardless of the input data, this represents the forecasting target for each day of the year, with the superscript d indicating the dimension. The mapping of its i-th head is represented as:

[0093] ,

[0094] 2. Key vector (K) and value vector (V):

[0095] Both K and V are temporal modeling features output by the preceding bidirectional long short-term memory network (Bi-LSTM). It is mapped from elsewhere. Although they originate from the same source, they are transformed through two sets of independent weight matrices, resulting in a physical difference:

[0096] Key vector K: Represents the "index" information of historical time-series features, used to calculate the relevance score with the query vector. Its mapping representation for the i-th head is:

[0097] ,

[0098] Value vector V: Represents the actual "meteorological evolution content" contained in the historical time series features, used for weighted fusion based on relevance scores. Its i-th head mapping is represented as:

[0099] ,

[0100] Through the above parameterized representation, the model reconstructs the latent features of the Bi-LSTM output using a cross-multi-head attention mechanism:

[0101] ,

[0102] in: , , This represents the learnable weight matrix for the i-th attention head. is the scaling factor, representing the dimension of the feature vector; Softmax is the normalization function, used to convert the calculated energy score into a weight coefficient between 0 and 1, representing the contribution of different time steps or spatial nodes to the forecast result.

[0103] This mechanism can adaptively allocate weights between historical years and spatial neighborhoods based on the dynamic changes in input features, thereby significantly improving the robustness of the forecast model to extreme weather fluctuations. The number of attention heads can be configured between 4 and 16 depending on the feature complexity.

[0104] Without loss of generality, as alternatives, squeeze-and-excitation networks (SENet) or convolutional block attention modules (CBAM) can be used to enhance the importance of features in the feature channels or spatial dimensions. Furthermore, weight allocation can be achieved through learnable scalar gating mechanisms or weighted mappings based on global average pooling. These methods all belong to the category of adaptive optimization of spatiotemporal features through mathematical transformations.

[0105] Classification output and loss function optimization:

[0106] Taking full account of the seasonal differences in meteorological elements, the model divides the daily features of the whole year into 12 groups by month, and inputs them into 12 independent monthly classifiers. Each classifier consists of a fully connected layer and is trained independently for the weight parameters of a specific month, which can more accurately depict the distribution pattern of cloud cover within that month. The classifier maps the 128-dimensional features of the corresponding date into 4-dimensional raw scores, which are then normalized by the Softmax function to predict the probabilities of four cloud cover levels (sunny, cloudy, mixed, and unknown).

[0107] Specifically, the model employs 12 independent classifier groups (one for each month, corresponding to January to December) at the end, and outputs the prediction confidence scores for each level using the Softmax function.

[0108] ,

[0109] The prediction confidence level is the number of samples taken on the target date that belong to the k-th cloud coverage level.

[0110] For the model output layer corresponding to the first Unnormalized score for cloud-like coverage level;

[0111] The result obtained after Softmax normalization belongs to the th The predicted probability of a class.

[0112] The model is trained end-to-end using the sum of the categorical cross-entropy between the predicted probabilities of all sampling points and all dates and the true labels as the loss function. Regularized dropout layers are embedded between layers, with a dropout rate preferably between 0.1 and 0.3. The AdamW optimizer, combined with cosine annealing learning rate scheduling, is employed to improve the model's generalization performance on unseen years. The loss function formula is as follows:

[0113] ,

[0114] M is the total number of sampling points. These are actual observed values. These are predicted values.

[0115] This function can effectively widen the probability gap between correct and incorrect grades, ensuring that the prediction results achieve high-precision classification and inference on an annual scale while adhering to the official MOD09GA standard.

[0116] Through the above hierarchical processing, the model achieves high-precision extrapolation from macro-level monthly statistical characteristics to micro-level daily cloud coverage details, making full use of the dual information of geospatial topological correlation and historical time evolution patterns.

[0117] Preferably, this application primarily uses historical cloud cover sequences as feature input. In practical applications, this technical solution can further integrate auxiliary meteorological factors such as temperature, humidity, and air pressure, or introduce spatial attributes such as land cover type and topographic elevation distribution (DEM) as multi-source input features. By increasing the feature dimension, the hierarchical processing architecture described in this application remains applicable; only the number of feature channels needs to be adjusted at the input layer.

[0118] Preferably, this application uses four-category cloud cover forecasting as the main embodiment. As an alternative, by modifying the activation function of the output layer (e.g., changing Softmax to Linear) and the loss function (e.g., changing cross-entropy loss to mean squared error loss), this architecture can be directly applied to regression prediction of cloud cover percentage, or extended to forecasting tasks of other continuous meteorological elements such as precipitation and aerosol optical depth (AOD).

[0119] Preferably, this invention uses 12 independent monthly classifiers corresponding to January to December. Alternatively, a unified fully connected classifier can be used, embedding month codes (such as sine / cosine positional codes or learnable month embeddings) into the input, or month-related feature modulation can be achieved through Conditional Batch Normalization. These schemes also achieve seasonal modeling and are equivalent transformations of the output module.

[0120] Example:

[0121] To verify the effectiveness and universality of the method of this invention, three typical climate zones were selected as experimental areas: the first experimental area (subtropical monsoon climate, with large interannual variability in cloud cover), the second experimental area (temperate continental climate, with stable climate), and the third experimental area (basin climate, with abundant clouds and fog, and moderate interannual variability). For each area, 500 fixed sampling points were randomly selected according to administrative boundaries (random seeds were 123, 456, and 789 respectively). Daily cloud cover categories (four categories in total: sunny, cloudy, mixed, and undetermined) from 2013 to 2024 were extracted from the MOD09GA product. Each sampling point includes latitude and longitude coordinates and a daily cloud cover label.

[0122] To simulate a forward-looking prediction scenario, the validation year was set to 2024, meaning data from years prior to 2024 was used for training to predict daily cloud cover in 2024. The training data was organized by year pairing: using the 12-month cloud cover categories of the input year as features, the daily categories for the 365 days of the output year were predicted. Two training periods were set up for the experiment:

[0123] (1) Short cycle: The training input years are from 2020 to 2022, and the output years are from 2021 to 2023, with a total of 3 year pairings and a training sample size of approximately 547,000 sample days;

[0124] (2) Long cycle: The training input years are from 2013 to 2022, and the output years are from 2014 to 2023, with a total of 10 annual pairings and a training sample size of approximately 1.825 million sample days.

[0125] Both periods ensure that the 2024 data is not used for training at all, but only for final validation.

[0126] The baseline method employs the historical mode method, which calculates the predicted cloud cover category as the most frequent cloud cover category for each calendar day at each sampling point, based on the highest frequency observed in all historical years prior to the validation year. To ensure fairness in the comparison, the historical year range used in the baseline method is consistent with the training years of the method in this invention (historical data from 2020 to 2023 is used for the short period, and historical data from 2013 to 2023 is used for the long period).

[0127] Evaluation metrics included overall accuracy, precision for each category, recall, F1 score, and macro-average F1. All experiments used the spatiotemporal graph neural network model proposed in this invention, with the following hyperparameters uniformly set: GCN hidden layer dimension 128, LSTM hidden layer dimension 256 (bidirectional), attention heads 8, Dropout coefficient 0.3, and the optimizer using AdamW with cosine annealing learning rate scheduling. The maximum number of training epochs was 200, and the early stopping mechanism's waiting epochs were set to 30.

[0128] The overall accuracy comparison results of the proposed method and the baseline method under long and short cycles in three regions are shown in Table 1 below.

[0129] Table 1

[0130]

[0131] As shown in Table 1, the method of this invention achieved better prediction accuracy than the baseline in all three different climate zones, verifying the universality of the method. It is particularly noteworthy that in the first experimental region with significant interannual cloud cover variability, the method of this invention achieved an accuracy of 76.06% in the short period, an improvement of 12.52 percentage points from the baseline of 63.54%, significantly exceeding the performance limit of traditional climatological statistical methods; in the long period, it also achieved 76.06%, an improvement of 7.32 percentage points from the baseline of 68.74%, demonstrating the model's strong ability to capture interannual fluctuations. In the second experimental region with stable climate, the method of this invention still achieved stable gains: a short-period accuracy of 53.14% (an improvement of 2.54 percentage points) and a long-period accuracy of 54.82% (an improvement of 1.20 percentage points), proving the robustness of the model under strong baseline conditions. In the third experimental region, the short-cycle accuracy of the method of this invention was 66.80%, an improvement of 6.96 percentage points from the baseline of 59.84%; the long-cycle accuracy was 67.77%, an improvement of 2.07 percentage points from the baseline of 65.70%, which is in line with the expected performance in the medium variability region.

[0132] Of particular note is that, in the three experimental regions, the accuracy of the method of this invention is essentially the same under both short-period (using only 4 years of data) and long-period (using 11 years of data) conditions. This result indicates that the method of this invention is insensitive to the length of historical data, requiring only a small amount of recent data to achieve superior performance, and exhibits extremely high data utilization efficiency. This characteristic makes this method particularly suitable for regions with scarce historical observational data.

[0133] To further analyze the model's ability to identify different cloud categories, Tables 2 to 4 list the F1 scores of each category in the first to third experimental regions under short-term experiments (the results for long-term experiments are similar and will not be repeated). It should be noted that the proportion of mixed-class samples in all three regions is less than 6%, and the number of undetermined samples is 0. The model has difficulty learning effective features from a limited number of samples, which is due to the inherent characteristics of the MOD09GA data.

[0134] Table 2

[0135]

[0136] Table 3

[0137]

[0138] Table 4

[0139]

[0140] This invention's method significantly improves the identification ability of the rare class (clear sky) while maintaining the advantage of the majority class (cloudy). In the first experimental region, the F1 score for clear sky jumped from 0.2566 to 0.4776, an increase of 86%, a breakthrough that is difficult to achieve with traditional methods; the F1 score for cloudy increased from 0.7674 to 0.8506, further consolidating the predictive advantage of the majority class. In the second and third experimental regions, the F1 score for clear sky remained basically unchanged, while the F1 score for cloudy increased by 0.0622 and 0.0780 respectively, indicating that the model can stably improve the predictive performance of the majority class under different climatic conditions.

[0141] Through systematic experiments in three different climate zones and two training cycles, the method of this invention demonstrates the following significant advantages:

[0142] First, in the first experimental region with large interannual variability of cloud cover, the short-period accuracy was improved by as much as 12.52 percentage points, and the F1 score for clear sky was improved by 86%, significantly exceeding the performance limit of traditional climatological methods, fully demonstrating the powerful predictive ability of the method of this invention in high-variability regions.

[0143] Second, the data utilization efficiency is extremely high. The same prediction accuracy as 11 years of historical data can be achieved with only 4 years of recent data, proving that the method of this invention is not sensitive to the length of historical data and is particularly suitable for rapid deployment in areas with scarce historical data.

[0144] Third, the method exhibits strong robustness. In the second experimental region with stable climate, the method of this invention can still achieve a stable gain of 1.2 to 2.5 percentage points. In the third experimental region with moderate variability, the improvement ranges from 2.1 to 7.0 percentage points, verifying the model's generalization ability under different climatic conditions.

[0145] Fourth, the method achieves balanced performance across categories. While maintaining accuracy for most classes, it effectively improves the recognition ability for rare classes, resulting in more balanced overall performance.

[0146] On the other hand, the present invention provides an annual cloud cover prediction device based on graph neural networks and monthly features, the various modules of which can implement the various steps of the aforementioned method, specifically including:

[0147] The acquisition module is used to acquire historical cloud cover data of geographical sampling points, and construct a 12-dimensional monthly input feature and an output label representing the daily cloud cover level vector for each sampling point;

[0148] The matrix construction module is used to construct an adjacency matrix representing spatial topological relationships based on the geographic coordinates of each sampling point and the spherical distance and K-nearest neighbor algorithm.

[0149] The prediction output module is used to input the 12-dimensional monthly input features and the adjacency matrix into a pre-trained spatiotemporal graph neural network model, and output the daily cloud cover level prediction results for the target prediction year. In the spatiotemporal graph neural network model, the residual graph convolutional network layer aggregates spatial neighborhood features defined by the adjacency matrix based on the input features and outputs enhanced spatial features; the bidirectional long short-term memory network layer models the long-range dependency of the enhanced spatial features in the historical year-on-year time dimension and outputs latent features that fuse spatiotemporal dependencies; and the multi-head self-attention weighted layer adaptively weights the temporal features of each date throughout the year based on the latent features and outputs the four-class daily cloud cover level probability.

[0150] Thirdly, the present invention provides an electronic device, comprising: one or more processors; and a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned annual cloud coverage prediction method based on graph neural networks and monthly features.

[0151] Fourthly, the present invention provides a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, enable the processor to implement the aforementioned annual cloud coverage prediction method based on graph neural networks and monthly features.

[0152] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting annual cloud cover based on graph neural networks and monthly features, characterized in that, The method includes: Step 1: Obtain historical cloud cover data for geographical sampling points, and construct a 12-dimensional monthly input feature and an output label representing the daily cloud cover level vector for each sampling point; Step 2: Based on the geographical coordinates of each sampling point, construct an adjacency matrix representing spatial topological relationships using spherical distance and the K-nearest neighbor algorithm; Step 3: Input the 12-dimensional monthly input features and adjacency matrix into the pre-trained spatiotemporal graph neural network model, and output the daily cloud cover level prediction results for the target prediction year; wherein, in the spatiotemporal graph neural network model, the residual graph convolutional network layer aggregates the spatial neighborhood features defined by the adjacency matrix based on the input features and outputs enhanced spatial features; the bidirectional long short-term memory network layer performs long-range dependency modeling on the enhanced spatial features in the historical year-on-year time dimension and outputs latent features that fuse spatiotemporal dependencies; the multi-head self-attention weighted layer adaptively weights the time-series features of each date throughout the year based on the latent features and outputs the four-class daily cloud cover level probability.

2. The annual cloud cover prediction method based on graph neural networks and monthly features according to claim 1, characterized in that, Step 1 includes: the input feature is a 12-dimensional monthly feature vector obtained based on the daily cloud coverage level statistics of historical years; the output label is a four-category daily cloud coverage level vector consisting of sunny, cloudy, mixed, and unknown, defined by the cloud mask status bits of MODIS products.

3. The annual cloud cover prediction method based on graph neural networks and monthly features according to claim 2, characterized in that, Each dimension of the 12-dimensional monthly feature vector is obtained by calculating the arithmetic mean of the cloud cover levels of all valid observation days in the corresponding month and rounding it to the nearest integer. If there is no valid observation data for the month, it is marked as a preset missing value.

4. The annual cloud cover prediction method based on graph neural networks and monthly features according to claim 1, characterized in that, Step 2 includes: calculating the physical distance between any two sampling points based on the spherical distance formula, using the K nearest neighbor algorithm, and constructing an adjacency matrix representing the spatial topological relationship based on the calculated physical distance. The rule for the value of the elements in the adjacency matrix is: if a sampling point belongs to the K nearest neighbor nodes of another sampling point, the corresponding element is set to one; otherwise, it is set to zero.

5. The annual cloud cover prediction method based on graph neural networks and monthly features according to claim 2, characterized in that, In step 3, the residual graph convolutional network layer includes a first graph convolutional layer to a fourth graph convolutional layer; the first graph convolutional layer is used to map the input 12-dimensional monthly feature vector to a 128-dimensional high-dimensional feature space; the second graph convolutional layer to the fourth graph convolutional layer are used to extract deep spatial topological features through three consecutive graph convolution operations; starting from the second graph convolutional layer, a residual connection mechanism is introduced to superimpose the features output from the previous layer with the output of the current layer after graph convolution, batch normalization and LeakyReLU nonlinear activation.

6. The annual cloud cover prediction method based on graph neural networks and monthly features according to claim 1, characterized in that, In step 3, the bidirectional long short-term memory network layer is composed of three cascaded bidirectional long short-term memory networks, which are used to perform bidirectional long-range time dependency modeling on the enhanced spatial features output by the residual graph convolutional network layer along the historical year dimension, and output implicit features that fuse forward and backward information.

7. The annual cloud cover prediction method based on graph neural networks and monthly features according to claim 1, characterized in that, The multi-head self-attention weighted layer adopts an eight-head cross-attention mechanism, using 365 learnable date query vectors as the query basis. The implicit features output by the bidirectional long short-term memory network layer are directly used as key vectors and value vectors. After attention calculation, daily features are output. The daily features are divided into 12 groups according to the month and input into 12 independent monthly classifiers. Each classifier consists of a fully connected layer and outputs the predicted probability of the four-class daily cloud coverage level through the Softmax function.

8. An annual cloud cover prediction device based on graph neural networks and monthly features, characterized in that, include: The acquisition module is used to acquire historical cloud cover data of geographical sampling points, and construct a 12-dimensional monthly input feature and an output label representing the daily cloud cover level vector for each sampling point; The matrix construction module is used to construct an adjacency matrix representing spatial topological relationships based on the geographic coordinates of each sampling point and the spherical distance and K-nearest neighbor algorithm. The prediction output module is used to input the 12-dimensional monthly input features and the adjacency matrix into a pre-trained spatiotemporal graph neural network model, and output the daily cloud cover level prediction results for the target prediction year. In the spatiotemporal graph neural network model, the residual graph convolutional network layer aggregates spatial neighborhood features defined by the adjacency matrix based on the input features and outputs enhanced spatial features; the bidirectional long short-term memory network layer models the long-range dependency of the enhanced spatial features in the historical year-on-year time dimension and outputs latent features that fuse spatiotemporal dependencies; and the multi-head self-attention weighted layer adaptively weights the temporal features of each date throughout the year based on the latent features and outputs the four-class daily cloud cover level probability.

9. An electronic device, characterized in that, include: One or more processors; Memory, used to store one or more programs; When one or more programs are executed by the one or more processors, the one or more processors implement the annual cloud coverage prediction method based on graph neural networks and monthly features as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, It stores executable instructions that, when executed by a processor, enable the processor to implement the annual cloud cover prediction method based on graph neural networks and monthly features as described in any one of claims 1-7.

Images