A multi-modal meteorological pollutant prediction method and system fusing optical aerosol remote sensing images and ground meteorological time series data

Abstract

The application belongs to the technical field of climate data analysis, and discloses a multi-modal meteorological pollutant prediction method and system fusing optical aerosol remote sensing images and ground meteorological time series data. Time series features representing the change of atmospheric aerosols in the region are obtained from the aerosol optical depth time series image provided by the remote sensing satellite, and time series data of ground meteorological observation are obtained. After STL decomposition and Transformer encoder, the deep representations of the aerosol mode and the meteorological mode are obtained respectively. A unified multi-modal representation is constructed to obtain the comprehensive feature representation of each climate variable. By maximizing the reward function, the adjacency matrix is optimized to establish the climate causal network. The long short-term memory network is used to predict the climate variable. The application has significant advantages in causal modeling and prediction tasks, and has effectiveness and practicality in atmospheric causal discovery and target pollutant variable prediction tasks.

Description

Technical Field

[0001] This invention belongs to the fields of climate data analysis, remote sensing and artificial intelligence technology, and in particular relates to a multimodal meteorological pollutant prediction method and system that integrates optical aerosol remote sensing images and ground meteorological time series data. Background Technology

[0002] Climate causality discovery aims to identify causal relationships between climate elements from observational data and reconstruct climate causal networks in the form of directed acyclic graphs (DAGs). This process is crucial for understanding the internal mechanisms of the climate system and supporting climate trend prediction and intervention decisions. For example, by constructing climate causal networks, the causal relationships between different climate variables can be visually presented, and the impact of changes in a particular factor on the entire climate system can be simulated. This allows for the assessment of the potential impacts of climate change and the development of corresponding countermeasures, while also providing a basis for climate variable prediction. In applications such as air pollution early warning and control, accurate prediction of pollutant concentrations is also essential for developing control measures.

[0003] Existing methods for discovering climate causality can be broadly categorized into two types: traditional statistical machine learning methods, including Granger causality tests, nonlinear state-space models, causal network learning algorithms, and structural equation modeling. These methods infer causal relationships based on statistical tests and mathematical models, offering advantages such as simple models and easily interpretable results, but also have limitations. For example, they are sensitive to noise and outliers, typically require large amounts of data, and perform poorly when dealing with high-dimensional and complex climate data. Deep learning methods, on the other hand, have shown remarkable performance in extracting complex patterns and nonlinear relationships from large-scale data in recent years. Some studies have begun to apply deep learning to the construction of climate causal networks and have made initial progress. Deep models can automatically learn hierarchical features and capture intricate dependencies in the climate system, gradually becoming a powerful tool for constructing and understanding climate causal networks. However, these studies primarily focus on identifying causal structures, lacking integrated schemes for applying the discovered causal structures to climate variable prediction. Meanwhile, climate systems often involve multi-source heterogeneous data: on the one hand, satellite and ground-based remote sensing provide large-scale imagery data showing the temporal changes in atmospheric composition; on the other hand, ground-based meteorological observations provide local time-series meteorological pollution data such as PM2.5, NO2, and SO2. These data modalities contain complementary information; for example, aerosol remote sensing images reflect the spatial distribution of the atmospheric environment, while meteorological time series records the evolution of local climate elements. Relying solely on a single data source may fail to fully capture the complete picture of climate phenomena, leading to incomplete or inaccurate causal relationship discovery. Based on this situation, there is an urgent need for a method that can effectively integrate multi-source climate observation data, extract cross-modal correlation features, and efficiently and robustly discover causal relationships and predict climate variables within a vast structural space. Summary of the Invention

[0004] To address the problems of existing climate causality discovery methods, which rely on single-modal data leading to insufficient information utilization, inaccurate causal identification, difficulty in structure search, and failure to apply causal structures to climate variable prediction, this invention provides a multimodal meteorological pollutant prediction method and system that integrates optical aerosol remote sensing images and ground meteorological time series data. By introducing a cross-modal attention mechanism, reinforcement learning strategy, and time series prediction model, it achieves unified representation, causal structure reconstruction, and variable prediction of multi-source climate data.

[0005] This invention provides the following technical solution:

[0006] A multimodal meteorological pollutant prediction method integrating optical aerosol remote sensing images and ground meteorological time series data includes the following steps:

[0007] We acquire time-series images of aerosol optical thickness provided by remote sensing satellites, and through ROI analysis, we obtain time-series characteristics of atmospheric aerosol changes in representative areas, as well as time-series data of ground meteorological observations.

[0008] Using the STL decomposition algorithm, the time series features of atmospheric aerosol changes and the multi-scale features mixed in the time series data of ground meteorological observations are separated. After passing through the Transformer encoder, the deep representations of aerosol modes and meteorological modes are obtained respectively.

[0009] The aerosol modes and the meteorological modes are aligned and fused to construct a unified multimodal representation, resulting in an updated aerosol mode representation that incorporates meteorological information, i.e., a comprehensive feature representation of each climate variable.

[0010] Based on the comprehensive feature representation of each climate variable, an adjacency matrix of a causal directed graph is generated. The adjacency matrix is ​​optimized by maximizing the reward function, thereby establishing a climate causal network.

[0011] Based on the aforementioned climate causal network, a long short-term memory network is used to predict climate variables, thereby enabling the prediction of future pollutant variables.

[0012] Preferably, methods for obtaining time-series characteristics of atmospheric aerosol changes in a representative region through ROI analysis include:

[0013] Let at discrete time Acquire images within the region of interest (ROI) ,in For the set of pixels within the ROI, For each time step, the number of pixels is [number]. Calculate the regional mean and regional variance of this region:

[0014] ;

[0015] ;

[0016] Based on this, the time-series characteristics of atmospheric aerosol changes in the representative region can be obtained. .

[0017] Preferably, the STL decomposition algorithm decomposes the observed sequence into three parts: trend, seasonality, and residuals.

[0018] ;

[0019] ;

[0020] ;

[0021] ;

[0022] ;

[0023] in, For the observation sequence, Indicates a smooth long-term trend. For seasonal items smoothed through the seasonal window, For the residual term, Represents a sequence The trend part is calculated using the moving average method. The seasonal portion is obtained through a periodic extraction method.

[0024] Preferred methods for constructing unified multimodal representations include:

[0025] Using the feature vector of aerosol mode as a query Using the characteristic matrix of meteorological modes as the key Sum Through calculation and The degree of matching Weighted summation yields the aerosol mode update representation incorporating meteorological information. ,Right now:

[0026] ;

[0027] in The latent space dimension used for scaling in the attention mechanism. This is the transpose symbol.

[0028] Preferably, the method for constructing the adjacency matrix includes:

[0029] By calculating the probability that there is a causal edge between each pair of variables This gives the candidate DAG structure under the current strategy, for any two variables. and , The calculation formula is:

[0030] ;

[0031] in, , Representing variables respectively and eigenvectors, , The weight matrix is ​​a learnable matrix. For learnable parameter vectors, This represents the Sigmoid activation function;

[0032] Then, the reward function for the environment is defined as the AIC score of the candidate graph and its sparsity penalty:

[0033] Let the candidate causal graph be There are a total of The nth variable node, the nth Nodes The parent set is Based on the idea of ​​local decomposability using log-likelihood, AIC is written as:

[0034] ;

[0035] in, To achieve maximum likelihood estimation The local log-likelihood of the following For the first The number of parameters in each local model;

[0036] To encourage sparse structures, we introduce Norm penalty, and thus reward function for:

[0037] ;

[0038] in It is an adjacency matrix. Represents the number of directed edges. Control sparsity.

[0039] Preferred methods for predicting climate variables using long short-term memory networks include:

[0040] For each target variable node in the climate causal network, an input sequence is constructed based on the historical time-series data of all its parent node variables, i.e., in time... At any given time, the target variable will be... All parent node variables The observations are combined into an input vector:

[0041] ;

[0042] LSTM networks recursively update their hidden states using the input sequence, thereby capturing the dynamic influence of parent node variables on the target variable. The LSTM state update is represented as:

[0043] ;

[0044] in, for The hidden state at all times The state transition function of the LSTM cell;

[0045] Subsequently, the predicted output is obtained by linearly mapping the hidden state:

[0046] ;

[0047] in and These are the weight matrix and bias vector of the output layer, respectively. For the target variable At any moment The predicted value.

[0048] This invention also provides a multimodal meteorological pollutant prediction system that integrates optical aerosol remote sensing images and ground meteorological time series data. Applying the aforementioned method, it includes a raw sequence acquisition unit, a feature representation unit, a multimodal fusion unit, a causal network unit, and a prediction unit.

[0049] The original sequence acquisition unit is used to acquire time-series images of aerosol optical thickness provided by remote sensing satellites, and through ROI analysis, obtain time-series characteristics of atmospheric aerosol changes in representative areas, as well as time-series data of ground meteorological observations.

[0050] The feature representation unit is used to separate the time series features of atmospheric aerosol changes and the multi-scale features mixed in the time series data of ground meteorological observations using the STL decomposition algorithm. After passing through the Transformer encoder, the deep representations of aerosol modes and meteorological modes are obtained respectively.

[0051] The multimodal fusion unit is used to align and fuse the aerosol mode and the meteorological mode to construct a unified multimodal representation and obtain an updated aerosol mode representation that incorporates meteorological information, i.e., a comprehensive feature representation of each climate variable.

[0052] The causal network unit is used to generate an adjacency matrix of a causal directed graph based on the comprehensive feature representation of each climate variable. The adjacency matrix is ​​optimized by maximizing the reward function, thereby establishing a climate causal network.

[0053] The prediction unit is used to predict climate variables based on the climate causal network and using a long short-term memory network, thereby enabling the prediction of future pollutant variables.

[0054] The beneficial effects of this invention are as follows:

[0055] This invention provides a multimodal meteorological pollutant prediction method and system that integrates optical aerosol remote sensing images and ground meteorological time series data. Compared with the prior art, this invention has the following advantages:

[0056] (1) Utilization of multimodal complementary information: By integrating remote sensing images and meteorological time series as two data sources, the complementary information of spatial distribution of atmospheric components and local meteorological changes is fully explored, overcoming the limitations of insufficient single-modal information and improving the accuracy and comprehensiveness of causal relationship identification.

[0057] (2) Cross-attention fusion mechanism: A feature alignment and fusion strategy based on cross-attention is adopted to effectively capture the correlation and mutual influence between different modal climate variables, and to achieve adaptive fusion at the feature level. This mechanism ensures that key causal signals are prominently represented in multimodal fused features, and enhances the model's sensitivity to cross-modal causal relationships.

[0058] (3) Temporal feature extraction from trend-seasonal decomposition: The STL decomposition method is introduced to decompose the time series data into two components: long-term trend and periodic fluctuation. This helps the model capture the slow evolution trend and periodic variation features respectively. Compared with using the original sequence directly, the features separated by trend-seasonal decomposition more clearly reflect the change patterns at different time scales, improving the robustness and effectiveness of downstream causal structure learning.

[0059] (4) Causal-guided predictive capability: After discovering the climate causal network, this invention also introduces an LSTM-based prediction model, which uses the information of the parent nodes in the causal graph to perform sequence prediction of the target variable, thus realizing the integration of causal discovery and climate variable prediction. This method makes full use of the driving relationship in the causal structure, improving the accuracy and reliability of pollutant concentration prediction. Attached Figure Description

[0060] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly described below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0061] Figure 1 This is a schematic diagram of the overall process of the present invention. Detailed Implementation

[0062] The core idea of ​​this invention lies in fusing feature representations of multimodal climate data and transforming causal structure learning into an optimizable policy search process, using the searched causal graph to predict pollutants. First, by performing trend and periodic decomposition and Transformer encoding on data from different modalities, this invention obtains two sets of feature representations that reflect the spatiotemporal patterns within each modality. Then, a cross-attention mechanism is used to align and fuse these two sets of features, extracting hidden correlation signals between climate variables from a cross-modal perspective, forming a unified representation to support causal relationship inference. Next, the causal network reconstruction problem is characterized as a sequence decision-making task in reinforcement learning: the model generates a causal graph structure based on the fused features and obtains feedback based on statistical scoring, continuously iterating and improving the policy. Furthermore, after modeling the causal relationships between climate variables, this invention further combines the causal graph with a time-series prediction task. By constructing an LSTM prediction model, using the causal parent node information of the target variable as input, the future trend of the target climate variable is predicted. This method fully utilizes the learned causal structure, improving prediction performance and enhancing the interpretability of the results.

[0063] In summary, this invention proposes a four-stage framework centered on "cross-modal Transformer fusion and reinforcement learning optimization," which fully integrates multi-source climate observation information and combines causal discovery with variable prediction, achieving efficient reconstruction of climate causal networks and accurate prediction of spatiotemporal changes in pollutants. While ensuring structure search efficiency, it enhances model robustness and interpretability; it is applicable to various scenarios such as air pollution causation analysis, extreme climate event tracing, and pollutant concentration prediction, possessing broad application prospects and practical value. Especially in climate system modeling and causal discovery, this invention unifies remote sensing imagery and meteorological time series for modeling, extracts key driving factors through spatiotemporal analysis, and uses causal structure learning to reveal the action chains and transmission paths between variables. The climate system model constructed accordingly can more accurately characterize the operational laws of mechanisms, providing verifiable technical support and decision-making basis for understanding climate processes and constructing internal system structure models.

[0064] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0065] Example 1

[0066] like Figure 1The image shows a multimodal meteorological pollutant prediction method that fuses optical aerosol remote sensing images and ground meteorological time series data, as described in Embodiment 1 of the present invention. The method mainly includes the following steps:

[0067] S1. Obtain time-series images of aerosol optical thickness provided by remote sensing satellites, and obtain time-series characteristics representing atmospheric aerosol changes in the region through ROI analysis, as well as time-series data of ground meteorological observations.

[0068] For time-series images of aerosol optical thickness provided by remote sensing satellites, image statistical operations can be performed on the area of ​​interest to obtain a time series representing the aerosol changes in that area. This embodiment uses the regional mean and mean-variance:

[0069] Let at discrete time Acquire images within the Region of Interest (ROI). ,in For the set of pixels within the ROI, For each time step, the number of pixels is [number]. Calculate the regional mean and regional variance of this region:

[0070] ;

[0071] ;

[0072] Based on this, time-series characteristics representing atmospheric aerosol changes in this region can be obtained. (or binary features) This operation converts spatial imagery data into a time series using mean / variance statistics and aligns it naturally with the surface meteorological series on the time axis.

[0073] S2. Using the STL decomposition algorithm, the time series features of atmospheric aerosol changes and the multi-scale features mixed in the time series data of ground meteorological observations are separated. After passing through the Transformer encoder, the deep representations of aerosol modes and meteorological modes are obtained respectively.

[0074] Based on the STL decomposition algorithm, the aerosol time series features in S1 are decomposed into components reflecting long-term change trends and components reflecting periodic seasonal fluctuations. In addition, the time series data of ground meteorological observations (such as variable series such as temperature, humidity, and precipitation) are decomposed to extract the trend and periodic components of each meteorological variable.

[0075] The STL decomposition algorithm breaks down the observed sequence into three parts: trend, seasonality, and residuals.

[0076] ;

[0077] ;

[0078] ;

[0079] ;

[0080] ;

[0081] in, For the observation sequence, Indicates a smooth long-term trend. For seasonal items smoothed through the seasonal window, For the residual term, Represents a sequence The trend part is calculated using the moving average method. The seasonal portion is obtained through a periodic extraction method.

[0082] This processing can separate potentially mixed multi-scale features in the original time series, helping subsequent models focus on patterns at different time scales. Next, within each data mode, the sequence signal after removing trend and seasonal components and its multi-scale combination are input into the Transformer encoder. Multi-head self-attention is used to capture the long- and short-term dependencies and intervariate correlations within a single mode, obtaining deep representations of the aerosol mode and the meteorological mode, respectively.

[0083] S3. Align and fuse aerosol modes and meteorological modes to construct a unified multimodal representation, and obtain an updated aerosol mode representation that incorporates meteorological information, i.e., a comprehensive feature representation of each climate variable.

[0084] Based on the remote sensing modal features and meteorological modal features obtained from S2, this step constructs a unified multimodal representation by aligning and fusing the two feature spaces. Specifically, the feature vector of the aerosol mode is used as the query. Using the characteristic matrix of meteorological modes as the key Sum Through calculation and The degree of matching Weighted summation yields the aerosol mode update representation incorporating meteorological information. ,Right now:

[0085] ;

[0086] in The latent space dimension used for scaling in the attention mechanism. This is the transpose symbol.

[0087] Similarly, using meteorological modal characteristics as the query... Remote sensing modal features are key Sum Performing a symmetric cross-attention operation yields meteorological mode representations incorporating aerosol remote sensing information. ,Right now

[0088] ;

[0089] in This refers to the latent space dimension used for scaling in the attention mechanism. Through the aforementioned bidirectional cross-attention computation, the model establishes semantic alignment between remote sensing image features and meteorological time-series features, thereby effectively representing the complementary relationships between heterogeneous modalities. Introducing multi-head attention allows for parallel modeling in different subspaces, enabling the network to focus on different types of cross-modal associations from multiple projection perspectives, thus more fully exploring the potential correlation structures between the two data modalities. Simultaneously, cross-attention allows the meteorological modality to adaptively focus on key factors closely related to its dynamic changes (such as wind speed, relative humidity, and temperature) at each time step; on the other hand, it prompts the meteorological modality to perceive the spatial distribution characteristics of atmospheric aerosol thickness in the remote sensing modality, thereby achieving bidirectional semantic mapping. This structure effectively alleviates information mismatch caused by inconsistencies in step size and scale between modalities.

[0090] The fusion module outputs a comprehensive representation of each meteorological variable, including its spatiotemporal pattern within the current mode and supplementary information from the other mode. Specifically, it first concatenates the features from the two modes along the channel dimension. Then, the system undergoes reweighting and nonlinear transformation sequentially through a multi-head self-attention layer and a feed-forward network (FFN): ;

[0091] in F For a unified representation of multimodal data, FFN stands for Feedforward Neural Network, and MultiHead stands for Multi-Head Self-Attention Layer.

[0092] Through the aforementioned cross-modal feature fusion, the complementary information from the two data sources is jointly expressed in a shared latent space, completing joint representation learning. Remote sensing images and meteorological time series are first projected onto a unified representation space and jointly optimized within an end-to-end framework, enabling the model to obtain a consistent cross-modal representation. This representation simultaneously integrates spatial distribution cues from remote sensing images and temporal evolution cues from meteorological series, and aligns corresponding patterns in different modalities through cross-attention, thereby achieving deep cross-modal fusion.

[0093] From a multi-scale perspective, the single-modal stage has decomposed and extracted components at different time scales, such as trends and seasonality. In the subsequent cross-modal fusion stage, this multi-scale information is further integrated, enabling the joint representation to simultaneously retain long-term trend influences and seasonal fluctuation characteristics. This synergy between cross-modal and multi-scale approaches makes the final feature representation more conducive to downstream causal structure learning: the representation of each variable contains more comprehensive and discriminative signals, enhancing the reliability and robustness of causal relationship discovery and estimation.

[0094] Regarding the fusion features of the output In summary, its encoded information can be summarized as follows: (i) spatial distribution and diffusion patterns provided by remote sensing imagery (e.g., regional morphology of aerosol optical thickness); (ii) temporal dynamics of ground meteorological observations (e.g., temporal dependence of wind direction, humidity, and temperature); and (iii) cross-modal high-order interaction relationships captured through cross-attention. This unified representation has advantages in both physical consistency and spatial resolution, providing more refined input for subsequent causal structure search and mechanism identification.

[0095] S4. Based on the comprehensive feature representation of each climate variable, generate the adjacency matrix of the causal directed graph. Optimize the adjacency matrix by maximizing the reward function, thereby establishing a climate causal network.

[0096] Based on feature fusion, this invention searches for the optimal structure of a causal network using a reinforcement learning algorithm. First, the combined representations of the variables obtained in step two are input into the model network, which generates an adjacency matrix of a causal directed graph based on the relationships between the representations. The model network calculates the probability that a causal edge exists between each pair of variables. This gives the candidate DAG structure under the current strategy, for any two variables. and , The calculation formula is:

[0097] ;

[0098] in, , Representing variables respectively and eigenvectors, , The weight matrix is ​​a learnable matrix. For learnable parameter vectors, This represents the Sigmoid activation function. (The rest of the text appears to be a series of seemingly unrelated phrases and sentences, possibly from different sources.) The sign of the value can be used to determine whether a path exists from... point to The causal edge tendency is obtained by mapping with the Sigmoid function, and the probability values ​​between 0 and 1 are used to construct the adjacency matrix of the causal graph.

[0099] Then, the reward function for the environment is defined as the AIC score of the candidate graph and its sparsity penalty:

[0100] Let the candidate causal graph be There are a total of The nth variable node, the nth Nodes The parent set is Based on the idea of ​​local decomposability using log-likelihood, AIC is written as:

[0101] ;

[0102] in, To achieve maximum likelihood estimation The local log-likelihood of the following For the first The number of parameters in a local model.

[0103] To encourage sparse structures, we introduce The penalty norm (number of edges) leads to the reward function. for:

[0104] ;

[0105] in It is an adjacency matrix. Represents the number of directed edges. Control sparsity.

[0106] By maximizing It can obtain sparse, interpretable, and well-generalized climate causal networks. AIC uses "goodness of fit plus parameter penalty" as a criterion to achieve a balance between accuracy and complexity. The smaller the value, the better the candidate map performs in interpreting observational data and suppressing overfitting.

[0107] S5. Based on the climate causal network, a long short-term memory network is used to predict climate variables, thereby enabling the prediction of future pollutant variables.

[0108] In this embodiment, based on the climate causal network constructed in S4, a long short-term memory network is used to predict climate variables. Specifically, for each target variable node to be predicted in the climate causal network, an input sequence is constructed based on the historical time-series data of all its parent node variables, i.e., in time... At any given time, the target variable will be... All parent node variables The observations are combined into an input vector:

[0109] ;

[0110] The input sequence is then fed into an LSTM network. The LSTM network uses this input sequence to recursively update its hidden state, thereby capturing the dynamic influence of parent node variables on the target variable. The LSTM state update is represented as:

[0111] ;

[0112] in, for The hidden state at all times This is the state transition function for the LSTM cell.

[0113] Subsequently, the predicted output is obtained by linearly mapping the hidden state:

[0114] ;

[0115] in and These are the weight matrix and bias vector of the output layer, respectively. For the target variable At any moment The predicted value is obtained. In this step, historical information of the parent node variables is used as input to the LSTM to establish a nonlinear time-series mapping relationship between the target variable and its causal parent nodes, thereby enabling the prediction of future pollutant variables.

[0116] Example 2

[0117] This embodiment verifies the effectiveness and practicality of the technical solution of the present invention through specific experimental data.

[0118] This experiment first discovers the causal structure of multimodal climate data, then constructs an LSTM prediction model based on the obtained causal graph for all parent nodes of the target variable, and finally evaluates the prediction performance. The specific process is as follows:

[0119] 1. Using the multimodal meteorological pollutant prediction method (Cross-modal Transformer with Reinforcement Learning for pollutant prediction, abbreviated as CTRL) that integrates optical aerosol remote sensing images and ground meteorological time series data of the present invention, and various baseline comparison methods (PC, LPCMCI, GRaSP, DiffAN, TS-ICD), the data are subjected to causal graph structure learning to obtain the set of parent nodes of the target variable;

[0120] 2. Using the time series data of the parent node variable as input, construct an LSTM model with the same configuration to predict the target climate variable;

[0121] 3. The mean absolute percentage error (MAPE) and root mean square error (RMSE) were used to evaluate the predictive performance of the LSTM model. All methods followed the same procedure described above (first discovering the causal graph, then performing LSTM predictions based on the parent nodes of the causal graph) to ensure the fairness of the comparison.

[0122] This experiment utilizes a multimodal atmospheric dataset constructed from near-real-time reanalysis data from the Copernicus Atmospheric Monitoring Service (CAMS). The dataset spans from January to July 2025, with a temporal resolution of 3 hours, and primarily includes particulate matter and gaseous pollutant indicators: PM2.5, PM10, SO2, NO2, CO, O3, and aerosol optical thickness. The multimodal characteristics and rich variable information of the data source provide a sufficient foundation for causal discovery and prediction.

[0123] The control experiment selected five typical temporal causal discovery methods: PC, LPCMCI, GRaSP, DiffAN, and TS-ICD. In the experiments, all methods first generated a causal graph, and then used the parent nodes of the target variable in the causal graph to train an LSTM model. For fairness, the LSTM network structure and hyperparameter settings were kept consistent for all methods; only the impact of differences in causal structure on prediction performance was compared.

[0124] As shown in Table 1, the comparative experimental results demonstrate that the CTRL method achieved the lowest MAPE and RMSE values ​​across all pollutant variable prediction tasks. This indicates that the causal structure recovered by CTRL is of higher quality, and the LSTM model trained from the causal parent node provides more accurate predictions. CTRL can fully utilize the complementary information between remote sensing optical aerosol imagery and surface meteorological sequences, exhibiting stronger robustness to weakly correlated signals and noise; for complex seasonal and slowly evolving signals, its side-direction determination is more stable. For example, in gaseous components dominated by particulate matter and background fields, the CTRL method captures cross-temporal trends and amplitude changes more fully than other baselines, demonstrating stronger generalization ability and noise resistance.

[0125] Table 1

[0126]

[0127] Furthermore, CTRL employs reinforcement learning strategies to efficiently search for optimal structures in a vast directed acyclic graph space, resulting in fast and accurate reconstructed causal networks that maintain stable performance even under scenarios involving cross-time periods, cross-regional distributions, and cross-modal shifts. The original CTRL paper also reports that using only the causal features discovered by CTRL as LSTM input for prediction yields better prediction results, further validating the beneficial effect of causal discovery on downstream prediction.

[0128] In summary, the multimodal climate causality discovery method (CTRL) proposed in this invention demonstrates significant advantages in both causal modeling and prediction tasks. CTRL not only recovers more accurate causal structures but also exhibits lower prediction errors in LSTM predictions based on causal parent nodes, showcasing the method's stability and superior generalization ability. These results fully demonstrate the effectiveness and practicality of the CTRL method in atmospheric causality discovery and target pollutant variable prediction tasks.

[0129] Example 3

[0130] This embodiment provides a multimodal meteorological pollutant prediction system that integrates optical aerosol remote sensing images and ground meteorological time series data. The system includes a raw sequence acquisition unit, a feature representation unit, a multimodal fusion unit, a causal network unit, and a prediction unit. Specifically:

[0131] The raw sequence acquisition unit is used to acquire time-series images of aerosol optical thickness provided by remote sensing satellites. Through ROI analysis, time-series characteristics representing atmospheric aerosol changes in the region are obtained, as well as time-series data of ground meteorological observations are acquired.

[0132] The feature representation unit is used to separate the time series features of atmospheric aerosol changes and the multi-scale features mixed in the time series data of ground meteorological observations using the STL decomposition algorithm. After passing through the Transformer encoder, the deep representations of aerosol modes and meteorological modes are obtained respectively.

[0133] The multimodal fusion unit is used to align and fuse aerosol modes and meteorological modes to construct a unified multimodal representation, thereby obtaining an updated aerosol mode representation that incorporates meteorological information, i.e., a comprehensive feature representation of each climate variable.

[0134] The causal network unit is used to generate an adjacency matrix of a causal directed graph based on the comprehensive feature representation of each climate variable. By maximizing the reward function, the adjacency matrix is ​​optimized, thereby establishing a climate causal network.

[0135] The prediction unit is used to predict climate variables based on climate causal networks and employs long short-term memory networks, thereby enabling the prediction of future pollutant variables.

[0136] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A multimodal meteorological pollutant prediction method integrating optical aerosol remote sensing images and ground meteorological time series data, characterized in that, Includes the following steps: We acquire time-series images of aerosol optical thickness provided by remote sensing satellites, and through ROI analysis, we obtain time-series characteristics of atmospheric aerosol changes in representative areas, as well as time-series data of ground meteorological observations. Using the STL decomposition algorithm, the time series features of atmospheric aerosol changes and the multi-scale features mixed in the time series data of ground meteorological observations are separated. After passing through the Transformer encoder, the deep representations of aerosol modes and meteorological modes are obtained respectively. The aerosol modes and the meteorological modes are aligned and fused to construct a unified multimodal representation, resulting in an updated aerosol mode representation that incorporates meteorological information, i.e., a comprehensive feature representation of each climate variable. Based on the comprehensive feature representation of each climate variable, an adjacency matrix of a causal directed graph is generated. This adjacency matrix is ​​then optimized by maximizing the reward function, thereby establishing a climate causal network. The method for constructing the adjacency matrix includes: By calculating the probability that there is a causal edge between each pair of variables. This gives the candidate DAG structure under the current strategy, for any two variables. and , The calculation formula is: ； in, , Representing variables respectively and eigenvectors, , The weight matrix is ​​a learnable matrix. For learnable parameter vectors, This represents the Sigmoid activation function; Then, the reward function for the environment is defined as the AIC score of the candidate graph and its sparsity penalty: Let the candidate causal graph be There are a total of The nth variable node, the nth Nodes The parent set is Based on the idea of ​​local decomposability using log-likelihood, AIC is written as: ； in, To achieve maximum likelihood estimation The local log-likelihood of the following For the first The number of parameters in each local model; To encourage sparse structures, we introduce Norm penalty, and thus reward function for: ； in It is an adjacency matrix. Represents the number of directed edges. Controlling sparsity; Based on the aforementioned climate causal network, a long short-term memory network is used to predict climate variables, thereby enabling the prediction of future pollutant variables.

2. The method according to claim 1, characterized in that: Methods for obtaining time-series characteristics of atmospheric aerosol changes in representative regions through ROI analysis include: Let at discrete time Acquire images within the region of interest (ROI) ,in For the set of pixels within the ROI, For each time step, the number of pixels is [number]. Calculate the regional mean and regional variance of this region: ； ； This allows us to obtain the time-series characteristics of atmospheric aerosol changes in the representative region. .

3. The method according to claim 1, characterized in that: The STL decomposition algorithm decomposes the observed sequence into three parts: trend, seasonality, and residuals. ； ； ； ； ； in, For the observation sequence, Indicates a smooth long-term trend. For seasonal items smoothed through the seasonal window, For the residual term, Represents a sequence The trend part is calculated using the moving average method. The seasonal portion is obtained through a periodic extraction method.

4. The method according to claim 1, characterized in that: Methods for constructing a unified multimodal representation include: Using the feature vector of aerosol mode as a query Using the characteristic matrix of meteorological modes as the key Sum Through calculation and The degree of matching Weighted summation yields the aerosol mode update representation incorporating meteorological information. ,Right now: ； in The latent space dimension used for scaling in the attention mechanism. This is the transpose symbol.

5. The method according to claim 1, characterized in that: Methods for predicting climate variables using long short-term memory networks include: For each target variable node in the climate causal network, an input sequence is constructed based on the historical time-series data of all its parent node variables, i.e., in time... At any given time, the target variable will be... All parent node variables The observations are combined into an input vector: ； LSTM networks recursively update their hidden states using the input sequence, thereby capturing the dynamic influence of parent node variables on the target variable. The LSTM state update is represented as: ； in, for The hidden state at all times The state transition function of the LSTM cell; Subsequently, the predicted output is obtained by linearly mapping the hidden state: ； in and These are the weight matrix and bias vector of the output layer, respectively. For the target variable At any moment The predicted value.

6. A multimodal meteorological pollutant prediction system integrating optical aerosol remote sensing images and ground meteorological time series data, employing the method described in any one of claims 1-5, characterized in that, It includes a raw sequence acquisition unit, a feature representation unit, a multimodal fusion unit, a causal network unit, and a prediction unit: The original sequence acquisition unit is used to acquire time-series images of aerosol optical thickness provided by remote sensing satellites, and through ROI analysis, obtain time-series characteristics of atmospheric aerosol changes in representative areas, as well as time-series data of ground meteorological observations. The feature representation unit is used to separate the time series features of atmospheric aerosol changes and the multi-scale features mixed in the time series data of ground meteorological observations using the STL decomposition algorithm. After passing through the Transformer encoder, the deep representations of aerosol modes and meteorological modes are obtained respectively. The multimodal fusion unit is used to align and fuse the aerosol mode and the meteorological mode to construct a unified multimodal representation and obtain an updated aerosol mode representation that incorporates meteorological information, i.e., a comprehensive feature representation of each climate variable. The causal network unit is used to generate an adjacency matrix of a causal directed graph based on the comprehensive feature representation of each climate variable. The adjacency matrix is ​​optimized by maximizing the reward function, thereby establishing a climate causal network. The prediction unit is used to predict climate variables based on the climate causal network and using a long short-term memory network, thereby enabling the prediction of future pollutant variables.

Images