Spatiotemporal dust storm event ai forecasting method fusing satellite remote sensing and meteorological data

Abstract

The application discloses a spatiotemporal dust storm event AI prediction method fusing satellite remote sensing and meteorological data, and relates to the technical fields of atmospheric pollution prediction and artificial intelligence. In view of the problem that existing methods are difficult to efficiently model multi-source spatiotemporal dust data, the method proposed in the application designs a dust storm prediction model DustMamba based on a multi-task architecture, aiming to jointly predict future PM10 concentration and dust storm occurrence through satellite remote sensing and meteorological data. The DustMamba is composed of a spatiotemporal encoder, a feature aggregation layer and a task-specific layer. The spatiotemporal encoder combines a visual Mamba and a three-dimensional convolutional network, and can efficiently extract spatiotemporal dependence information in multi-source data. The feature aggregation layer adopts a global attention module, and enhances the ability of the model in cross-dimension feature interaction. The task-specific layer is based on an independent two-dimensional convolutional predictor, so as to realize accurate prediction of different dust tasks.

Description

Technical Field

[0001] This invention relates to the fields of atmospheric pollution prediction and artificial intelligence technology, and more specifically, to an AI forecasting method for spatiotemporal dust storm events that integrates satellite remote sensing and meteorological data. Background Technology

[0002] Dust storms are natural phenomena caused by strong winds, resulting in the suspension and propagation of dust particles in the air. They commonly occur in arid and semi-arid regions. The mechanisms underlying dust storms involve multiple factors, including dry soil, increased wind speed, sparse vegetation, and unfavorable meteorological conditions. These factors work together to cause the rise, propagation, and settling of dust particles. Dust storms not only cause soil erosion and vegetation damage but also significantly reduce air quality and increase the concentration of suspended particulate matter (such as PM10), posing a serious threat to human health. Furthermore, dust storms can lead to economic impacts such as agricultural losses, increased traffic accidents, and infrastructure damage, and in severe cases, can even disrupt normal societal operations. Therefore, accurate prediction of dust storm events can not only allow for proactive preventative measures and reduce health risks but also provide a scientific basis for decision-making, thereby mitigating the ecological, social, and economic losses caused by dust storms.

[0003] Currently, numerical weather prediction (NMR) is the traditional method for predicting dust storms. This method primarily relies on meteorological models to analyze and simulate various factors that may influence dust storm occurrence, such as the selection of dust source areas, wind speed, soil moisture, vegetation cover, and atmospheric circulation, thereby quantitatively describing the generation, propagation path, and spatial distribution of dust storms. These numerical models can provide relatively intuitive dust storm prediction results, but they have several limitations. First, numerical weather prediction methods are extremely sensitive to the accuracy of initial conditions, and the high complexity and dynamic instability of atmospheric systems often lead to error accumulation, thus affecting the accuracy and reliability of predictions. Second, to achieve high spatiotemporal resolution predictions, numerical weather prediction models typically require substantial computational resources and high-performance computing infrastructure, making real-time, high-frequency dust storm predictions over large areas difficult and costly.

[0004] With the rapid development of artificial intelligence technologies, especially machine learning and deep learning, an increasing number of studies are applying these technologies to solve meteorological forecasting tasks. In recent years, machine learning methods have made significant progress in meteorological element forecasting and air quality forecasting, particularly demonstrating strong predictive capabilities for local dust storms and fine particulate matter (PM2.5, PM10, etc.). However, these machine learning-based studies typically focus on localized forecasts and largely rely on surface meteorological observation data or short-term local meteorological data. The effectiveness and adaptability of these methods still have certain limitations in modeling the global dynamics and complex spatiotemporal characteristics of dust storms.

[0005] The formation and propagation of dust storms involve multiple complex factors, including wind speed, air pressure, humidity, soil conditions, and atmospheric structure. These factors typically exhibit high nonlinearity and interdependence in both space and time, making it difficult for traditional machine learning methods to fully capture their spatiotemporal variations, thus limiting their predictive accuracy and applicability. Furthermore, existing machine learning models generally fail to adequately integrate multi-source data (such as satellite remote sensing data and meteorological data) and lack effective spatiotemporal modeling mechanisms. This makes it difficult for them to fully exploit the correlations and complementarities between different data sources when dealing with large-scale dust storm predictions, thereby affecting their predictive performance. Summary of the Invention

[0006] To address the aforementioned technical challenges, this invention provides an AI-based forecasting method for spatiotemporal dust storm events that integrates satellite remote sensing and meteorological data. Addressing the difficulty of existing methods in efficiently modeling multi-source spatiotemporal dust data, this invention proposes a multi-task architecture-based dust storm forecasting model, DustMamba, aiming to jointly predict future PM10 concentrations and dust storm occurrences using satellite remote sensing and meteorological data. DustMamba consists of three parts: a spatiotemporal encoder, a feature aggregation layer, and a task-specific layer. The spatiotemporal encoder combines visual Mamba and a 3D convolutional network to efficiently extract spatiotemporal dependency information from multi-source data. The feature aggregation layer employs a global attention module to enhance the model's ability to interact with features across dimensions. The task-specific layer is based on an independent 2D convolutional predictor to achieve accurate predictions for different dust storm tasks. This invention offers significant technical advantages, enabling efficient, multi-scale prediction of dust storm events and providing more accurate and timely data support for weather forecasting and emergency response. By accurately predicting the occurrence time, intensity, and impact range of dust storms, this invention provides a reliable scientific basis for meteorological disaster early warning, environmental protection, and public health prevention.

[0007] In a first aspect, the present invention provides an AI forecasting method for spatiotemporal dust storm events that integrates satellite remote sensing and meteorological data, the method comprising:

[0008] A dataset is acquired and preprocessed to obtain a preprocessed dataset; wherein the dataset includes satellite imagery data and meteorological reanalysis data, daily PM10 data and air quality data;

[0009] The preprocessed data is spatiotemporally aligned and interpolated to obtain unified input data;

[0010] A dust storm forecasting model is constructed, comprising a spatiotemporal encoder, a feature aggregation layer, and a task-specific layer. Based on unified input data, it is processed into an input tensor of dimension (B, C, T, H, W), where B is the batch size, C is the number of input channels, T is the historical time length, and H×W is the spatial resolution. After the input passes through the dust storm forecasting model, the outputs of four tasks are obtained: two regression prediction tasks and two classification prediction tasks. The output shape of each task is (B, H, W), representing the dust distribution at the next moment.

[0011] For various dust storm forecasting tasks, a multi-task loss function is constructed, and the dust storm forecasting model is trained based on the multi-task loss function;

[0012] During model training, the comprehensive loss is used as the optimization objective. The model parameters are continuously adjusted through the backpropagation algorithm to gradually minimize the multi-task loss and obtain the trained dust forecast model.

[0013] Further, a dataset is acquired, and the dataset is preprocessed to obtain a preprocessed dataset, including:

[0014] The data in the dataset is cleaned by using statistical methods to detect and remove data points with abnormal fluctuations, eliminating outliers and noisy data to ensure the accuracy and completeness of the data. For missing values ​​caused by equipment failure or data transmission problems, cubic spline interpolation is used to complete them to ensure the integrity of the data.

[0015] A normalization method is used to compress the cleaned data to a preset numerical range in order to eliminate the dimensional differences between different data dimensions.

[0016] Furthermore, the preprocessed data is spatiotemporally aligned and interpolated to obtain unified input data, including:

[0017] Project all types of data in the preprocessed data onto a unified spatial grid to ensure spatial alignment of all data. For PM10 concentration data with a resolution lower than a set threshold, use bilinear interpolation or spline interpolation to adjust it to the same high-resolution grid as the satellite remote sensing data.

[0018] All data were standardized to hourly time intervals, and missing data at different times was filled in using interpolation methods.

[0019] For meteorological data and satellite remote sensing data, linear interpolation is used to smooth the data for shorter time intervals to ensure the continuity and consistency of the time series.

[0020] For daily PM10 data with only day intervals, hourly air quality station data is used to correct it to obtain hourly gridded PM10 data.

[0021] Furthermore, the spatiotemporal encoder comprehensively extracts spatiotemporal features through a dual-channel architecture integrating 3D convolution and Mamba modules, and couples them through Hadamard product to ensure effective capture of local details and global context; wherein, the dual-channel architecture includes convolution channels and Mamba channels;

[0022] In the convolutional channels, the encoder employs two layers of 3D convolution to extract local spatiotemporal features, given a first input tensor x∈R. B×C×T×H×W R represents the set of real numbers, and the first output h of the convolution channel is obtained through the following convolution operation. c1 :

[0023] h c1 =Conv3D2(ReLU(Conv3D1(x)))

[0024] Where Conv3D1 and Conv3D2 both represent three-dimensional convolutional modules with a kernel size of 3×3×3, and ReLU is a linear rectified activation function;

[0025] In Mamba channels, the first input tensor x∈R is embedded using a patch-based embedding method. B×C×T×H×W Divide into non-overlapping patches of size P×P:

[0026] x′=FC(Rearrange(x,P)),

[0027] Where FC represents a fully connected layer, P is the patch size, Rearrange represents the patch embedding operation, and x′ is the reshaped tensor, x′∈R B×N×D ,in D = P, where P is the number of embedded patches. 2 • C represents the embedding dimension;

[0028] After dividing into non-overlapping patches, they are processed through multiple layers of Vim blocks;

[0029] In each Vim block, the reshaped tensor x′ generates a second output x″ by combining bidirectional sequence modeling and structured SSM. The reshaped tensor x′ is first normalized and linearly projected into two feature representations, a first feature u and a second feature v. Subsequently, a one-dimensional convolution operation is applied to the first feature u in the forward and backward directions to generate an intermediate feature u′. o intermediate feature u′ o It is converted into a set of learnable parameters A o B o Co and Δ o , where Δ o To ensure that Δ is positive, the softplus activation function is applied. o The specific calculation formula is as follows:

[0030]

[0031] Where, Δ o By Δ o After performing a linear transformation and adding a learnable bias parameter b, the softplus activation function is used to ensure that its value is positive, thereby adjusting the nonlinear transformation of the model.

[0032] potential state h o Recursive updates via SSM:

[0033]

[0034] in, This represents a matrix multiplication operation discretized by time step.

[0035] Output y in each direction o The calculation is performed using the following formula:

[0036]

[0037] Forward output y forward and backward output y backward Use v for gating and combine them as follows:

[0038] y combined =y forward ⊙SiLU(v)+y backward ⊙SiLU(v)

[0039] Where SiLU represents the Sigmoid linear unit activation function; ⊙ represents the Hadamard product operation;

[0040] Finally, the merged result is linearly transformed and added to x′ via a residual connection to generate the second output x″:

[0041] x″=Linear(y combined )+x′

[0042] Wherein, Linear represents a linear layer;

[0043] The output y of the Mamba channel at the last moment T ∈R B×N×D The feature map is reconstructed to the initial data dimension using an inverse embedding operation to obtain the final output h. c2 ∈RB×C×T×H×W It matches the data with the convolutional channels.

[0044] Furthermore, the feature aggregation layer employs a GAM module to enhance the feature representation provided by the spatiotemporal encoder; wherein, the GAM module integrates channel attention and spatial attention for weighted feature mapping, dynamically emphasizing channel dependencies and spatial correlations, and suppressing irrelevant noise information;

[0045] For a given second input tensor h c ∈R B×(C·T)×H×W Channel attention transforms it into a three-dimensional tensor h′ c ∈R B×(H·W)×(C·T) The importance of features is identified using a two-layer fully connected network, and then applied to the original input through element-wise multiplication.

[0046] h CA =h c ·σ(FC2(ReLU(FC1(h c ′))))

[0047] Where FC1 and FC2 represent fully connected layers used to adjust the channel dimension, ReLU represents the rectified linear activation function, and σ represents the sigmoid function; h CA Represents feature mapping;

[0048] The spatial attention module is used to capture the feature map h. CA Local spatial correlations are used to highlight key regions through convolution:

[0049] h SA =h CA ·σ(Conv2(ReLU(BN(Conv1(h CA )))))

[0050] Where Conv1 and Conv2 represent two 2D convolutional layers used to adjust the channel dimension, BN represents batch normalization, and h SA This represents the feature map ultimately generated by the GAM module.

[0051] Furthermore, the task-specific layer processes the aggregated features through a separate predictor or classifier to generate prediction results for various output tasks; for regression tasks that directly output numerical values, the predictor consists of two two-dimensional convolutional layers:

[0052]

[0053] in and This indicates a task-specific convolutional layer; This represents the model output result for a specific task.

[0054] The classifier adds a binary classification activation function to the predictor, expressed as:

[0055]

[0056] Furthermore, for various dust storm forecasting tasks, a multi-task loss function is constructed, and the dust storm forecasting model is trained based on the multi-task loss function, including:

[0057] In each training cycle, the input data and the true label y i Used to calculate the prediction for the i-th task The regression task uses mean squared error loss for calculation:

[0058]

[0059] Among them, L i This refers to the i-th task; This represents the prediction result for the j-th sample in the i-th task; N represents the total number of labeled data samples; and j represents the sample index.

[0060] The classification task uses binary cross-entropy loss for training:

[0061]

[0062] The total loss L of a batch is calculated using the following formula. batch :

[0063]

[0064] Where w i It is the weight of the i-th task, and Task is the total number of tasks;

[0065] To dynamically balance the task, w is dynamically adjusted during training based on the gradient norm. i The gradient norm g of the i-th task i for:

[0066]

[0067] in, This represents the gradient with respect to the parameter θ;

[0068] The average gradient norm for all tasks is: Update the task weights using the following formula to minimize g. i and Differences between them:

[0069]

[0070] Where η is the learning rate of the weights;

[0071] The weights are normalized to ensure that their sum is 1:

[0072]

[0073] Where w j This represents the j-th weight value;

[0074] In each training iteration, minimize the total loss L. batch By updating the model parameters through backpropagation, we can ensure that all tasks achieve balanced optimization.

[0075] Furthermore, after obtaining the trained dust forecasting model, the method further includes:

[0076] The trained dust storm forecasting model was evaluated using evaluation metrics.

[0077] For regression tasks (prediction of PM10 and BADI), mean squared error (MSE), root mean squared error (RMSE), mean absolute error (MAE), and / or coefficient of determination (R²) are used. 2 To evaluate the model's prediction accuracy;

[0078] For classification tasks, accuracy, recall, precision, and / or F1 score are used for evaluation.

[0079] Secondly, the present invention provides an AI forecasting device for spatiotemporal dust storm events that integrates satellite remote sensing and meteorological data, the device comprising:

[0080] The data preprocessing unit is configured to acquire a dataset and preprocess the dataset to obtain a preprocessed dataset; wherein the dataset includes satellite imagery data and meteorological reanalysis data, daily PM10 data and air quality data;

[0081] The alignment interpolation unit is configured to perform spatiotemporal alignment and interpolation on the preprocessed data to obtain unified input data;

[0082] The model building unit is configured to build a dust forecasting model. The dust forecasting model includes a spatiotemporal encoder, a feature aggregation layer, and a task-specific layer. Based on unified input data, it is processed into an input tensor of dimension (B, C, T, H, W), where B is the batch size, C is the number of input channels, T is the historical time length, and H×W is the spatial resolution. After the input passes through the dust forecasting model, it produces the outputs of four tasks: two regression prediction tasks and two classification prediction tasks. The output shape of each task is (B, H, W), representing the dust distribution at the next moment.

[0083] The loss function determination unit is configured to construct a multi-task loss function for various dust event forecasting tasks, and train the dust forecasting model based on the multi-task loss function;

[0084] The model training unit is configured to use the comprehensive loss as the optimization objective during the model training process, and continuously adjust the model parameters through the backpropagation algorithm to gradually minimize the multi-task loss and obtain the trained dust forecast model.

[0085] Thirdly, the present invention provides a readable storage medium storing one or more programs that can be executed by one or more processors to implement the method described above.

[0086] The present invention has at least the following beneficial effects:

[0087] This invention utilizes a multi-task learning framework in deep learning to successfully construct the DustMamba model, achieving high-precision prediction of dust storm events. It combines multi-source data, including remote sensing data from the FY-4A satellite, meteorological reanalysis data, and PM10 concentration data, and by fusing information from different time scales and spatial resolutions, comprehensively captures the spatiotemporal characteristics of dust storm occurrences.

[0088] First, in the data preprocessing stage, the differences and inconsistencies between data sources are eliminated through cleaning, standardization, and spatiotemporal alignment, ensuring the high quality and structure of the input data. This stage includes imputing missing values, removing outliers, and spatiotemporal alignment between different data sources, ensuring data consistency and usability, and providing a reliable data foundation for subsequent modeling.

[0089] Next, the DustMamba model employs a multi-task architecture consisting of a spatiotemporal encoder, a feature aggregation layer, and a task-specific layer. By optimizing model parameters using a multi-task loss function, it achieves accurate predictions for multiple tasks, such as PM10 concentration prediction and dust storm probability prediction. In the spatiotemporal encoder, the model introduces a visual Mamba structure and a 3D convolutional network, efficiently extracting spatiotemporal dependencies from satellite remote sensing and meteorological data, fully utilizing the spatiotemporal information of the data. The feature aggregation layer utilizes a global attention mechanism, further enhancing the model's ability to interact with features across dimensions. The task-specific layer is designed independently for each prediction task to ensure prediction accuracy across different tasks.

[0090] Through joint optimization of multi-task losses, the DustMamba model can not only accurately predict PM10 concentrations, but also effectively predict the probability of dust storms and other related meteorological indicators, such as BADI, DRBTD, and DST, thus providing a scientific basis for dust storm early warning and emergency response.

[0091] In summary, this invention offers significant advantages. First, by utilizing a framework of multi-source data fusion and multi-task learning, it fully explores the potential correlations between data points, enhancing the generalization ability and accuracy of the prediction model. Second, the DustMamba model, while ensuring efficient computation, effectively handles large-scale spatiotemporal data, achieving superior performance in the spatiotemporal prediction of dust storms, demonstrating higher accuracy and practicality compared to traditional methods. Finally, this invention can provide more scientific and accurate technical support for the prediction of extreme weather events such as dust storms, reducing disaster losses and protecting life and property, thus possessing broad application prospects and significant social value. Attached Figure Description

[0092] Figure 1 An overall flowchart of an AI forecasting method for spatiotemporal dust storm events that integrates satellite remote sensing and meteorological data according to an embodiment of the present invention is shown.

[0093] Figure 2 A schematic diagram of the model structure according to an embodiment of the present invention is shown;

[0094] Figure 3 A structural diagram of a spatiotemporal dust storm event AI forecasting device that integrates satellite remote sensing and meteorological data according to an embodiment of the present invention is shown. Detailed Implementation

[0095] To enable those skilled in the art to better understand the technical solutions of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. The embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and specific examples, but this is not intended to limit the present invention. If there is no necessary sequential relationship between the various steps described herein, the order in which they are described as examples should not be considered a limitation. Those skilled in the art should understand that the order can be adjusted, as long as it does not disrupt the logical consistency between them and render the entire process impossible.

[0096] In the field of atmospheric science and meteorological forecasting, dust storms, as extreme weather events with severe environmental and social impacts, have received widespread attention. Dust storms not only deteriorate air quality and increase the concentration of suspended particulate matter (such as PM10), seriously threatening human health, but also cause significant damage to agricultural production, transportation, and infrastructure. Therefore, accurate and timely prediction of the occurrence and intensity of dust storms is of great significance for reducing disaster losses and protecting people's lives and property. However, due to the suddenness, localization, and complexity of dust storm events, existing prediction methods still have certain limitations.

[0097] While existing numerical weather prediction methods can simulate atmospheric processes, they require significant computational resources, and the parameter tuning process for these models is complex. They often struggle to achieve high-precision predictions for highly localized and sudden weather events like dust storms. Furthermore, numerical methods typically rely on accurate knowledge of initial conditions, while dust storms are influenced by a variety of complex factors, leading to considerable uncertainty in practical applications. On the other hand, while statistical machine learning methods have made some progress in predicting pollutant concentrations, they struggle to provide comprehensive spatiotemporal dust distribution information and are ill-suited to handling complex dust storm events. Therefore, existing prediction methods still suffer from insufficient accuracy and inefficiency when dealing with large-scale, high spatiotemporal resolution dust storm events.

[0098] In view of the above problems, this invention provides an AI-based forecasting method for spatiotemporal dust storm events that integrates satellite remote sensing and meteorological data. By effectively fusing multi-source data, it improves the spatiotemporal accuracy of dust storm prediction, overcoming the problems of data isolation and insufficient spatiotemporal information extraction in traditional methods. Secondly, it proposes a deep learning-based multi-task modeling framework. Through the design of a spatiotemporal encoder, feature aggregation layer, and task-specific layer, it can accurately predict multiple dust storm-related tasks (such as dust storm probability, PM10 concentration, and dust storm intensity) while capturing the spatiotemporal characteristics and nonlinear relationships of dust storms. Extreme weather events such as dust storms have a serious impact on human production and life. Accurate prediction can provide a scientific basis for disaster prevention and mitigation, emergency response, and socio-economic activities. The implementation of this invention will make an important contribution to reducing disaster losses caused by dust storms and protecting people's lives and property, further improving the intelligence level of meteorological forecasting and environmental management, and promoting scientific response to meteorological disasters and social security.

[0099] It's important to note that the Mamba model is a deep learning architecture based on a Structured State-Space Model (SSM), specifically designed for efficiently processing long sequences of data. By introducing a selection mechanism to reparameterize the SSM, Mamba can filter out irrelevant information while preserving key features, exhibiting excellent linear scalability. Unlike traditional convolutional operations, Mamba employs a hardware-aware algorithm, improving computational efficiency through a scanning mechanism, significantly accelerating computation, especially on large-scale GPUs. When processing long sequences of data, its computational cost increases linearly with sequence length, offering a clear advantage over models like the Transformer, making it particularly suitable for fields such as natural language processing, computer vision, and medicine. These advantages make it an ideal choice for handling spatiotemporal modeling tasks related to dust storms.

[0100] Therefore, the core of this embodiment lies in proposing an innovative multi-task dust storm prediction model, DustMamba, based on deep learning hybrid modeling. This model, through its efficient Mamba architecture, fully mines the potential dust features in satellite remote sensing and meteorological data to achieve accurate dust storm prediction. By designing a multi-task architecture consisting of a spatiotemporal encoder, a feature aggregation layer, and a task-specific layer, the proposed model effectively captures the spatiotemporal characteristics of dust storms while achieving accurate predictions of different dust detection indicators, thus providing support for accurate forecasting and emergency response to dust storm events.

[0101] Specifically, such as Figure 1 As shown, the spatiotemporal dust storm event AI forecasting method that integrates satellite remote sensing and meteorological data includes the following steps S10 to S50.

[0102] S10. Obtain the dataset and preprocess it to obtain the preprocessed dataset; the dataset includes satellite imagery data and meteorological reanalysis data, daily PM10 data and air quality data.

[0103] In one exemplary embodiment, during the data collection and preprocessing phase, high-quality datasets from multiple sources are used to construct a dust storm prediction model. These specifically include FY-4A satellite imagery data and meteorological reanalysis data, daily PM10 data at 1km resolution in China, and air quality data from cities across the country. By integrating these multi-source data, a high-quality input set with spatiotemporal consistency is constructed.

[0104] First, the data is cleaned by using statistical methods to detect and remove data points with abnormal fluctuations, eliminating outliers and noisy data to ensure accuracy and completeness. For missing values ​​caused by equipment failure or data transmission problems, cubic spline interpolation is used to complete the data and ensure its integrity.

[0105] Subsequently, due to the inconsistent dimensions and significant differences in numerical ranges between data from different sources (such as the order-of-magnitude difference between wind speed and PM10 concentration), a normalization (Min-MaxScaling) method was employed to compress the data into a uniform numerical range, thereby eliminating the dimensional differences between different data dimensions. This approach ensures that the impact of various input features on the model is balanced during subsequent model training.

[0106] The preprocessed dataset not only improves data quality but also ensures the effective fusion of multi-source data in dust storm prediction, providing strong support for accurately capturing the spatiotemporal characteristics and meteorological impacts of dust storms.

[0107] S20. Perform spatiotemporal alignment and interpolation on the preprocessed data to obtain unified input data.

[0108] In an exemplary embodiment, the temporal frequency and spatial resolution of these data are further unified in the spatiotemporal alignment and interpolation step (i.e., step S20). First, due to the differences in spatial resolution between different data sources, various types of data need to be projected onto a unified spatial grid to ensure spatial alignment of all data. For PM10 concentration data with lower resolution, interpolation methods (such as bilinear interpolation or spline interpolation) are used to adjust it to the same high-resolution grid as the satellite remote sensing data. Second, in order to handle the time mismatch problem between different data sources, all data are unified into hourly time intervals, and missing data at different times are filled in using interpolation methods. For meteorological data and satellite remote sensing data, linear interpolation is used to smooth the data in shorter time intervals to ensure the continuity and consistency of the time series. For high-resolution daily PM10 data of 1km in China with only daily intervals, hourly granular data of national urban air quality stations are used to correct it to obtain hourly granular gridded PM10 data.

[0109] These spatiotemporal alignment and interpolation operations successfully eliminated the spatiotemporal differences between different data sources, providing high-quality and unified input data for subsequent feature fusion and model training, ensuring that the model can fully learn the spatiotemporal characteristics of dust storm events and their complex relationship with meteorological elements.

[0110] S30. Construct a dust forecasting model; the dust forecasting model includes a spatiotemporal encoder, a feature aggregation layer, and a task-specific layer. Based on unified input data, it is processed into an input tensor of dimension (B,C,T,H,W), where B is the batch size, C is the number of input channels, T is the historical time length, and H×W is the spatial resolution. After the input passes through the dust forecasting model, the outputs of four tasks are obtained, namely two regression prediction tasks and two classification prediction tasks. The output shape of each task is (B,H,W), which represents the dust distribution at the next moment.

[0111] It should be noted that this paper names the "dust forecasting model" as "DustMamba." This model is a multi-task spatiotemporal sequence prediction framework based on Mamba and convolution, used to simultaneously handle multiple prediction tasks for dust events. The model structure is as follows: Figure 2 As shown, DustMamba consists of three core components: a spatiotemporal encoder, a feature aggregation layer, and a task-specific layer. For a given input data X containing meteorological, satellite remote sensing, and dust index data, it is processed into an input tensor of dimension (B, C, T, H, W), where B is the batch size, C is the number of input channels, T is the historical time length, and H×W is the spatial resolution. After passing through DustMamba, the input yields outputs for four tasks: two regression prediction tasks (PM10 and BADI) and two classification prediction tasks (DRBTD and DST). Each task output has a shape of (B, H, W), representing the dust distribution at the next time step.

[0112] In one exemplary embodiment, the spatiotemporal encoder comprehensively extracts spatiotemporal features through a dual-channel architecture integrating 3D convolution and Mamba modules, coupled via Hadamard products to ensure efficient capture of local details and global context. In the convolutional channels, the encoder employs two layers of 3D convolution to extract local spatiotemporal features. Given a first input tensor x∈R... B×C×T×H×W R represents the set of real numbers, and the first output h of this channel can be obtained through the following convolution operation. c1 :

[0113] h c1 =Conv3D2(ReLU(Conv3D1(x)))

[0114] Conv3D1 and Conv3D2 represent three-dimensional convolutional modules with a kernel size of (3×3×3), and ReLU is a linear rectified activation function. This channel excels at encoding local spatiotemporal relationships and is used to capture abrupt changes in meteorological elements or satellite remote sensing images.

[0115] In a Mamba channel, the input tensor x∈R is first processed using a patch-based embedding method. B×C×T×H×W Divide into non-overlapping patches of size P×P:

[0116] x′=FC(Rearrange(x,P))

[0117] Where FC represents a fully connected layer, P is the patch size, and Rearrange represents the patch embedding operation. These patches are flattened and linearly projected into a high-dimensional latent space. The reshaped tensor is x′∈R. B×N×D ,in D = P, where P is the number of embedded patches. 2 • C represents the embedding dimension. This is then processed through multiple layers of Vim blocks.

[0118] In each Vim block, the input tensor x′ is used to generate the output x″ by combining bidirectional sequence modeling and structured SSM, thereby achieving effective modeling of spatial and temporal features in the vision task. The input tensor x′ is first normalized and linearly projected into two feature representations: a first feature u and a second feature v. Subsequently, a one-dimensional convolution operation is applied to the first feature u in the forward and backward directions, respectively, to generate the intermediate feature u′. o , where o∈{forward,backward}. Then, the intermediate feature u′ o It is converted into a set of learnable parameters A o B o C o and Δ o , where Δ o This ensures that the value is positive by applying the softplus activation function. Δ o The specific calculation formula is as follows:

[0119]

[0120] Where, Δ o By Δ o After performing a linear transformation and adding a learnable bias parameter b, the softplus activation function is used to ensure that its value is positive, thereby adjusting the nonlinear transformation of the model.

[0121] potential state h o Recursive updates via SSM:

[0122]

[0123] in, This represents a matrix multiplication operation discretized by time step.

[0124] The output in each direction can be calculated using the following formula:

[0125]

[0126] Forward and backward output y forward and y backward Use v for gating and combine them as follows:

[0127] y combined =y forward ⊙SiLU(v)+y backward ⊙SiLU(v)

[0128] Where SiLU represents the Sigmoid linear unit activation function.

[0129] Finally, the merged result is linearly transformed and added to x′ via a residual connection to generate the second output x″:

[0130] x″=Linear(y combined )+x′

[0131] Here, Linear represents a linear layer.

[0132] This Vim structure effectively integrates temporal dynamics and spatial modeling, ensuring that the model captures long-range dependencies and contextual relationships. Subsequently, the output y of Mamba at the last moment is... T ∈R B×N×D The feature map is reconstructed to the initial data dimension using an inverse embedding operation to obtain the final output h. c2 ∈R B×C×T×H×W It matches the data with the convolutional channels.

[0133] The first output h of the convolutional channels is integrated. c1 The final output h of the Mamba channel c2 Obtain the fusion tensor h c :

[0134] h c =h c1 ⊙h c2

[0135] In the formula, ⊙ represents the Hadamard product operation.

[0136] This dual-channel fusion strategy leverages the advantages of two spatiotemporal modeling methods. The convolutional channel features provide high-resolution, local spatiotemporal information, while the Mamba channel features introduce large-scale contextual understanding, ensuring the effective encoding of local and global spatiotemporal patterns and providing rich feature representations for the accurate identification and prediction of dust storms.

[0137] In one exemplary embodiment, the feature aggregation layer employs GAM to enhance the feature representation provided by the spatiotemporal encoder. GAM integrates channel attention and spatial attention for weighted feature mapping, dynamically emphasizing channel dependencies and spatial correlations while suppressing irrelevant noise information. First, for a given second input tensor h... c ∈R B×(C·T)×H×W Channel attention transforms it into a three-dimensional tensor h′ c ∈R B×(H·W)×(C·T) The importance of features is identified using a two-layer fully connected network, and then applied to the original input through element-wise multiplication.

[0138] h CA=h c ·σ(FC2(ReLU(FC1(h c ′))))

[0139] The fully connected layers FC1 and FC2 are used to adjust the channel dimension, ReLU represents the linear rectified activation function, and σ represents the sigmoid function.

[0140] Furthermore, a spatial attention module is used to capture the feature map h. CA Local spatial correlations are used to highlight key regions through convolution:

[0141] h SA =h CA ·σ(Conv2(ReLU(BN(Conv1(h CA )))))

[0142] Two 2D convolutional layers, Conv1 and Conv2, are used to adjust the channel dimension. BN represents batch normalization, ReLU represents the rectified linear activation function, and σ represents the sigmoid function. The integration of channel and spatial attention ensures that the network can capture global and local dependencies. By combining these two complementary attention mechanisms, the GAM module enables the model to prioritize key information while maintaining computational efficiency, achieving effective aggregation of the spatiotemporal features of dust storms.

[0143] In one exemplary embodiment, the task-specific layer processes the aggregated features through a separate predictor or classifier to generate prediction results for various output tasks. For regression tasks that directly output numerical values, the predictor consists of two two-dimensional convolutional layers with a kernel size of (1×1):

[0144]

[0145] Conv1 and Conv2 are task-specific convolutional layers, while the classifier adds a binary classification activation function to the predictor.

[0146]

[0147] In the formula, σ represents the sigmoid function; This represents the model output for a specific task. Input features h are processed independently by multiple task-specific modules. SA The task-specific layer stacks the outputs of all tasks along a new dimension to produce the final prediction tensor.

[0148] S40. For various sandstorm event forecasting tasks, construct a multi-task loss function and train the sandstorm forecasting model based on the multi-task loss function.

[0149] In one exemplary embodiment, a multi-task loss function is constructed for various sandstorm event forecasting tasks to ensure that all tasks are fully optimized without interference. In each training cycle, the input data x and the true label y are... i Used to calculate the prediction for the i-th task The regression task then uses the mean square error (MSE) loss for calculation:

[0150]

[0151] Among them, L i This refers to the i-th task; This represents the prediction result for the j-th sample in the i-th task; N represents the total number of labeled data samples; and j represents the sample index.

[0152] For classification tasks, binary cross-entropy (BCE) loss is used for training.

[0153]

[0154] Where N represents the total number of labeled data samples. The total loss of a batch is the weighted sum of the losses from all tasks:

[0155]

[0156] Where w i is the weight of task i, and Task is the total number of tasks.

[0157] To dynamically balance the task, the GradNorm method is introduced, which dynamically adjusts w during training based on the gradient norm. i The gradient norm of task i is:

[0158]

[0159] in, This represents the gradient with respect to the parameter θ.

[0160] The average gradient norm for all tasks is: Update task weights to minimize g i and Differences between them:

[0161]

[0162] Where η is the learning rate of the weights. The weights are normalized to ensure that their sum is 1:

[0163]

[0164] Where T represents the total training duration, j is the training time index; w j Let L represent the j-th weight value. In each training iteration, minimize the total loss L. batch By updating the model parameters through backpropagation, balanced optimization is ensured for all tasks.

[0165] S50. During the model training process, the comprehensive loss is used as the optimization objective. The model parameters are continuously adjusted through the backpropagation algorithm to gradually minimize the multi-task loss and obtain the trained dust forecast model.

[0166] In one exemplary embodiment, during model training, a comprehensive loss function L is used. batch As an optimization objective, the model parameters are continuously adjusted through backpropagation to gradually minimize multi-task losses, thereby improving the model's prediction accuracy for each task. This comprehensive loss function integrates the losses from multiple tasks, ensuring that regression tasks (such as predicting PM10 and BADI concentrations) and classification tasks (such as classifying DRBTD and DST values) are considered simultaneously during the optimization process. In regression tasks, by minimizing loss functions such as mean squared error (MSE), the model can accurately fit the concentration values ​​of PM10 and BADI; while in classification tasks, by minimizing cross-entropy loss, the model's classification performance is optimized, ensuring accurate classification of DRBTD and DST values. Through this multi-task learning, the model can not only accurately predict dust storm-related indicators but also effectively share information across different tasks, improving overall prediction accuracy. During each training iteration, the optimization of the loss function allows the model to adaptively adjust parameters to minimize prediction errors and gradually approach actual observations. Through continuous iteration, the model can accurately predict key indicators related to dust storm events (such as PM10 and BADI concentrations and DRBTD and DST values), thus providing a more accurate scientific basis for dust storm prediction.

[0167] In one exemplary embodiment, after model training is complete, the evaluation phase comprehensively measures the model's predictive performance using a series of standardized metrics. For regression tasks (prediction of PM10 and BADI), the main metrics used are mean squared error (MSE), root mean squared error (RMSE), mean absolute error (MAE), and coefficient of determination (R²). 2 The accuracy of a model's predictions is evaluated using metrics such as MSE and RMSE. MSE and RMSE measure the difference between the model's predicted values ​​and the actual values, while MAE assesses the mean absolute value of the prediction error. 2These metrics are used to measure the model's goodness of fit and prediction accuracy. For classification tasks (DRBTD and DST value prediction), accuracy, recall, precision, and F1 score are used for evaluation. Accuracy measures the proportion of correctly classified samples in the prediction results; recall measures the model's ability to identify positive samples; precision assesses the proportion of truly positive samples among those predicted as positive; and the F1 score is the harmonic mean of precision and recall, comprehensively considering the performance balance of the classification model across different tasks. These evaluation metrics comprehensively reflect the model's performance in dust storm prediction, aiding in further optimization and improvement. Through this multi-dimensional evaluation method, the dust storm prediction model of this invention can fully verify its application effectiveness on different tasks, providing scientific support for actual dust storm event early warning.

[0168] In summary, this invention innovatively designs the DustMamba model for dust storm events. The key technical points of this solution will be elaborated in detail below:

[0169] Multi-task learning model design: The innovation of this invention lies in constructing the DustMamba dust storm prediction model based on multi-task learning and introducing a multi-task loss optimization strategy to simultaneously predict multiple dust storm-related indicators, such as PM10 concentration and dust storm occurrence probability (BADI, DRBTD, DST), within the same model. By fusing satellite remote sensing data, meteorological reanalysis data, and high-resolution PM10 data, the DustMamba model can fully utilize the spatiotemporal information in these multi-source data to accurately capture the dynamic changes and spatial distribution characteristics of dust storms, thereby achieving high-precision dust storm prediction.

[0170] Spatiotemporal Encoder and Global Attention Mechanism: This technique effectively handles the complex spatiotemporal dependencies of dust storms through a spatiotemporal encoder and a global attention mechanism (GAM). The spatiotemporal encoder combines 3D convolution and Vim blocks to extract spatiotemporal features at both local and global scales from the input data, ensuring comprehensive feature representation. The global attention mechanism enhances the model's learning ability during the feature aggregation stage, integrating channel attention and spatial attention to capture key spatiotemporal information, highlighting crucial features influencing the occurrence and propagation of dust storms, and improving the model's prediction accuracy.

[0171] Task-Specific Prediction Layer: The DustMamba model employs a task-specific prediction layer, with each task (such as PM10 concentration prediction or dust storm probability prediction) having its own independent predictor. This design allows each task to be independently optimized based on shared spatiotemporal features, avoiding interference between tasks and improving task robustness and prediction accuracy.

[0172] Through these technological innovations, the DustMamba model has not only improved the accuracy of dust storm prediction, but also significantly enhanced its ability to process large-scale, highly complex data, providing strong technical support for accurate forecasting and timely emergency response.

[0173] This invention also provides an AI forecasting device for spatiotemporal dust storm events that integrates satellite remote sensing and meteorological data, such as... Figure 3 As shown, the device includes:

[0174] The data preprocessing unit 301 is configured to acquire a dataset and preprocess the dataset to obtain a preprocessed dataset; wherein the dataset includes satellite imagery data and meteorological reanalysis data, daily PM10 data and air quality data;

[0175] Alignment interpolation unit 302 is configured to perform spatiotemporal alignment and interpolation on the preprocessed data to obtain unified input data;

[0176] Model building unit 303 is configured to build a dust forecasting model. The dust forecasting model includes a spatiotemporal encoder, a feature aggregation layer, and a task-specific layer. Based on unified input data, it is processed into an input tensor of dimension (B, C, T, H, W), where B is the batch size, C is the number of input channels, T is the historical time length, and H×W is the spatial resolution. After the input passes through the dust forecasting model, it produces the outputs of four tasks: two regression prediction tasks and two classification prediction tasks. The output shape of each task is (B, H, W), representing the dust distribution at the next moment.

[0177] The loss function determination unit 304 is configured to construct a multi-task loss function for various dust event forecasting tasks, and train the dust forecasting model based on the multi-task loss function;

[0178] Model training unit 305 is configured to use the comprehensive loss as the optimization objective during model training, continuously adjust the model parameters through the backpropagation algorithm, and gradually minimize the multi-task loss to obtain the trained dust forecast model.

[0179] In some embodiments, the data preprocessing unit is further configured to:

[0180] The data in the dataset is cleaned by using statistical methods to detect and remove data points with abnormal fluctuations, eliminating outliers and noisy data to ensure the accuracy and completeness of the data. For missing values ​​caused by equipment failure or data transmission problems, cubic spline interpolation is used to complete them to ensure the integrity of the data.

[0181] A normalization method is used to compress the cleaned data to a preset numerical range in order to eliminate the dimensional differences between different data dimensions.

[0182] In some embodiments, the alignment interpolation unit is further configured to:

[0183] Project all types of data in the preprocessed data onto a unified spatial grid to ensure spatial alignment of all data. For PM10 concentration data with a resolution lower than a set threshold, use bilinear interpolation or spline interpolation to adjust it to the same high-resolution grid as the satellite remote sensing data.

[0184] All data were standardized to hourly time intervals, and missing data at different times was filled in using interpolation methods.

[0185] For meteorological data and satellite remote sensing data, linear interpolation is used to smooth the data for shorter time intervals to ensure the continuity and consistency of the time series.

[0186] For daily PM10 data with only day intervals, hourly air quality station data is used to correct it to obtain hourly gridded PM10 data.

[0187] In some embodiments, the spatiotemporal encoder comprehensively extracts spatiotemporal features through a dual-channel architecture integrating 3D convolution and Mamba modules, and couples them through Hadamard product to ensure effective capture of local details and global context; wherein, the dual-channel architecture includes convolution channels and Mamba channels;

[0188] In the convolutional channels, the encoder employs two layers of 3D convolution to extract local spatiotemporal features, given a first input tensor x∈R. B×C×T×H×W R represents the set of real numbers, and the first output h of the convolution channel is obtained through the following convolution operation. c1 :

[0189] h c1 =Conv3D2(ReLU(Conv3D1(x)))

[0190] Where Conv3D1 and Conv3D2 both represent three-dimensional convolutional modules with a kernel size of 3×3×3, and ReLU is a linear rectified activation function;

[0191] In Mamba channels, the first input tensor x∈R is embedded using a patch-based embedding method. B×C×T×H×W Divide into non-overlapping patches of size P×P:

[0192] x′=FC(Rearrange(x,P)),

[0193] Where FC represents a fully connected layer, P is the patch size, Rearrange represents the patch embedding operation, and x′ is the reshaped tensor, x′∈R B×N×D ,in D = P, where P is the number of embedded patches. 2 • C represents the embedding dimension;

[0194] After dividing into non-overlapping patches, they are processed through multiple layers of Vim blocks;

[0195] In each Vim block, the reshaped tensor x′ generates a second output x″ by combining bidirectional sequence modeling and structured SSM. The reshaped tensor x′ is first normalized and linearly projected into two feature representations, a first feature u and a second feature v. Subsequently, a one-dimensional convolution operation is applied to the first feature u in the forward and backward directions to generate an intermediate feature u′. o intermediate feature u′ o It is converted into a set of learnable parameters A o B o C o and Δ o , where Δ o To ensure that Δ is positive, the softplus activation function is applied. o The specific calculation formula is as follows:

[0196]

[0197] Where, Δ o By Δ o After performing a linear transformation and adding a learnable bias parameter b, the softplus activation function is used to ensure that its value is positive, thereby adjusting the nonlinear transformation of the model.

[0198] potential state h o Recursive updates via SSM:

[0199]

[0200] in, This represents a matrix multiplication operation discretized by time step.

[0201] Output y in each direction o The calculation is performed using the following formula:

[0202]

[0203] Forward output y forward and backward output y backward Use v for gating and combine them as follows:

[0204] y combined =yforward ⊙SiLU(v)+y backward ⊙SiLU(v)

[0205] Where SiLU represents the Sigmoid linear unit activation function; ⊙ represents the Hadamard product operation;

[0206] Finally, the merged result is linearly transformed and added to x′ via a residual connection to generate the second output x″:

[0207] x″=Linear(y combined )+x′

[0208] Wherein, Linear represents a linear layer;

[0209] The output y of the Mamba channel at the last moment T ∈R B×N×D The feature map is reconstructed to the initial data dimension using an inverse embedding operation to obtain the final output h. c2 ∈R B×C×T×H×W It matches the data with the convolutional channels.

[0210] In some embodiments, the feature aggregation layer employs a GAM module to enhance the feature representation provided by the spatiotemporal encoder; wherein the GAM module integrates channel attention and spatial attention for weighted feature mapping, dynamically emphasizing channel dependencies and spatial correlations, and suppressing irrelevant noise information;

[0211] For a given second input tensor h c ∈R B×(C·T)×H×W Channel attention transforms it into a three-dimensional tensor h′ c ∈R B×(H·W)×(C·T) The importance of features is identified using a two-layer fully connected network, and then applied to the original input through element-wise multiplication.

[0212] h CA =h c ·σ(FC2(ReLU(FC1(h c ′))))

[0213] Where FC1 and FC2 represent fully connected layers used to adjust the channel dimension, ReLU represents the rectified linear activation function, and σ represents the sigmoid function; h CA Represents feature mapping;

[0214] The spatial attention module is used to capture the feature map h. CA Local spatial correlations are used to highlight key regions through convolution:

[0215] h SA =hCA ·σ(Conv2(ReLU(BN(Conv1(h CA )))))

[0216] Where Conv1 and Conv2 represent two 2D convolutional layers used to adjust the channel dimension, BN represents batch normalization, and h SA This represents the feature map ultimately generated by the GAM module.

[0217] In some embodiments, the task-specific layer processes the aggregated features through a separate predictor or classifier to generate prediction results for various output tasks; for regression tasks that directly output numerical values, the predictor consists of two two-dimensional convolutional layers:

[0218]

[0219] in and This indicates a task-specific convolutional layer; This represents the model output result for a specific task.

[0220] The classifier adds a binary classification activation function to the predictor, expressed as:

[0221]

[0222] In some embodiments, the model training unit is configured as follows:

[0223] In each training cycle, the input data and the true label y i Used to calculate the prediction for the i-th task The regression task uses mean squared error loss for calculation:

[0224]

[0225] Among them, L i This represents the loss of the i-th task; This represents the prediction result for the j-th sample in the i-th task; N represents the total number of labeled data samples; and j represents the sample index.

[0226] The classification task uses binary cross-entropy loss for training:

[0227]

[0228] The total loss L of a batch is calculated using the following formula. batch :

[0229]

[0230] Where w iIt is the weight of the i-th task, and Task is the total number of tasks;

[0231] To dynamically balance the task, w is dynamically adjusted during training based on the gradient norm. i The gradient norm g of the i-th task i for:

[0232]

[0233] in, This represents the gradient with respect to the parameter θ;

[0234] The average gradient norm for all tasks is: Update the task weights using the following formula to minimize g. i and Differences between them:

[0235]

[0236] Where η is the learning rate of the weights;

[0237] The weights are normalized to ensure that their sum is 1:

[0238]

[0239] Where w j This represents the j-th weight value;

[0240] In each training iteration, minimize the total loss L. batch By updating the model parameters through backpropagation, we can ensure that all tasks achieve balanced optimization.

[0241] In some embodiments, the apparatus further includes a model evaluation unit, the model evaluation unit being configured to:

[0242] The trained dust storm forecasting model was evaluated using evaluation metrics.

[0243] For regression tasks (prediction of PM10 and BADI), mean squared error (MSE), root mean squared error (RMSE), mean absolute error (MAE), and / or coefficient of determination (R²) are used. 2 To evaluate the model's prediction accuracy;

[0244] For classification tasks, accuracy, recall, precision, and / or F1 score are used for evaluation.

[0245] It should be noted that the structures of the various spatiotemporal dust storm event AI forecasting devices that fuse satellite remote sensing and meteorological data described in this embodiment belong to the same technical concept as the previously described spatiotemporal dust storm event AI forecasting method that fuses satellite remote sensing and meteorological data. They achieve the same beneficial effects through the same principle, and will not be elaborated here.

[0246] This invention also provides a readable storage medium storing one or more programs that can be executed by one or more processors to implement the methods described in any of the above embodiments.

[0247] The above description is intended to be illustrative and not restrictive. For example, the above examples (or one or more of them) can be used in combination with each other. Other embodiments can be used by those skilled in the art when reading the above description. Furthermore, in the above detailed description, various features may be grouped together to simplify the invention. This should not be construed as an intention that a feature of an unclaimed invention is necessary for any claim. Rather, the subject matter of the invention may be less than all the features of a particular embodiment of the invention. Thus, the following claims are incorporated herein by reference as examples or embodiments, wherein each claim is an independent, separate embodiment, and these embodiments are contemplated to be combined with each other in various combinations or arrangements. The scope of the invention should be determined by reference to the appended claims and the full scope of their equivalents.

Claims

1. An AI-based forecasting method for spatiotemporal dust storm events that integrates satellite remote sensing and meteorological data, characterized in that, The method includes: Acquire the dataset and preprocess it to obtain the preprocessed dataset; the dataset includes satellite imagery data and meteorological reanalysis data, daily PM10 data and air quality data; Spatiotemporal alignment and interpolation are performed on the preprocessed data to obtain unified input data; A dust forecasting model is constructed, including a spatiotemporal encoder, a feature aggregation layer, and a task-specific layer. Based on unified input data, it is processed into an input tensor. After the input passes through the dust forecasting model, two regression prediction tasks and two classification prediction tasks are obtained. The output of each task represents the dust distribution at the next moment. For various dust storm forecasting tasks, a multi-task loss function is constructed to train the dust storm forecasting model; The spatiotemporal encoder comprehensively extracts spatiotemporal features through a dual-channel architecture that integrates 3D convolution and Mamba modules, and couples them through Hadamard product; the dual-channel architecture includes convolution channels and Mamba channels; In the convolutional channel, the encoder uses two layers of 3D convolution to extract local spatiotemporal features. Given a first input tensor x, the first output h of the convolutional channel is obtained through the convolution operation. c1 ; In the Mamba channel, the first input tensor x is divided into non-overlapping patches of size P×P by a patch embedding method to obtain the reshaped tensor x′. After dividing into non-overlapping patches, they are processed through multiple layers of Vim blocks; In each Vim block, the reshaped tensor x′ generates a second output x″ by combining bidirectional sequence modeling and structured SSM; the reshaped tensor x′ is normalized and linearly projected into a first feature u and a second feature v, and a one-dimensional convolution operation is applied to the first feature u in the forward and backward directions, respectively, to generate an intermediate feature u′. o intermediate feature u′ o It is converted into a set of learnable parameters A o B o C o and Δ o ; potential state h o Recursive updates are performed using SSM; Output y in each direction o Through the potential state h o and learnable parameter C o calculate; Forward output y forward and backward output y backward The second feature v is used for gating, and the results are combined to obtain the merged result; The merged result is linearly transformed and added to x′ through residual connection to generate the second output x″. The output y of the Mamba channel at the last moment T The feature map is reconstructed to the initial data dimension using an inverse embedding operation to obtain the final output h. c2 Matching the data with the convolutional channels; The feature aggregation layer uses the GAM module to enhance the feature representation provided by the spatiotemporal encoder; the GAM module integrates channel attention and spatial attention for weighted feature mapping; For a given second input tensor h c Channel attention transforms it into a three-dimensional tensor h′ c The importance of features is identified using a two-layer fully connected network, and the feature map h is obtained by applying element-wise multiplication to the original input. CA : The spatial attention module is used to capture the feature map h. CA The local spatial correlations in the model are used to obtain the feature maps finally generated by the GAM module. Task-specific layers process aggregated features through separate predictors or classifiers to generate prediction results for various output tasks.

2. The spatiotemporal dust storm event AI forecasting method fusing satellite remote sensing and meteorological data according to claim 1, characterized in that, The first output h of the convolution channel is obtained through the following convolution operation. c1 : h c1 =Conv3D2(ReLU(Conv3D1(x))) Where Conv3D1 and Conv3D2 both represent three-dimensional convolutional modules with a kernel size of 3×3×3, and ReLU is a linear rectified activation function; In Mamba channels, the first input tensor x∈R is embedded using a patch-based embedding method. B×C×T×H×W Divide into non-overlapping patches of size P×P: x′=FC(Rearrange(x,P)), Where FC represents a fully connected layer, P is the patch size, Rearrange represents the patch embedding operation, and x′ is the reshaped tensor, x′∈R B×N×D Where N = T·H·W, P 2 D = P, where P is the number of embedded patches. 2 • C is the embedding dimension, B is the batch size, C is the number of input channels, T is the historical time length, and H×W is the spatial resolution.

3. The spatiotemporal dust storm event AI forecasting method fusing satellite remote sensing and meteorological data according to claim 1, characterized in that, Where Δ o By applying the softplus activation function, we ensure that it is a positive value, Δ. o The specific calculation formula is as follows: Where, Δ o By using intermediate features u′ o After performing a linear transformation and adding a learnable bias parameter b, the softplus activation function is used to ensure that its value is positive, thereby adjusting the nonlinear transformation of the model. potential state h o Recursive updates via SSM: in, This represents a matrix multiplication operation discretized by time step. Output y in each direction o The calculation is performed using the following formula: Forward output y forward and backward output y backward Use v for gating and combine them as follows: and combined / and forward ⊙SiLU(v)+y backward ⊙SiLU(v) Where SiLU represents the Sigmoid linear unit activation function; ⊙ represents the Hadamard product operation; The merged result is linearly transformed and added to x′ via residual concatenation to generate the second output x″: x″=Linear(y combined )+x′ Here, Linear represents a linear layer.

4. The spatiotemporal dust storm event AI forecasting method fusing satellite remote sensing and meteorological data according to claim 1, characterized in that, The feature map h is calculated using the following formula. CA : h CA =h c ·σ(FC2(ReLU(FC1(h c ′)))) Where FC1 and FC2 represent fully connected layers used to adjust the channel dimension, ReLU represents the rectified linear activation function, and σ represents the sigmoid function; h CA Represents feature mapping; The spatial attention module is used to capture the feature map h. CA The local spatial correlations in the model are used to emphasize key regions through convolution, resulting in the following formula for calculating the final feature map generated by the GAM module: h SA =h CA ·σ(Conv2(ReLU(BN(Conv1(h CA ))))) Where Conv1 and Conv2 represent two 2D convolutional layers used to adjust the channel dimension, BN represents batch normalization, and h SA This represents the feature map ultimately generated by the GAM module; In the task-specific layer, for regression tasks that directly output numerical values, the predictor consists of two two-dimensional convolutional layers: in and This indicates a task-specific convolutional layer; This represents the model output result for a specific task. The classifier adds a binary classification activation function to the predictor, expressed as:

5. The spatiotemporal dust storm event AI forecasting method fusing satellite remote sensing and meteorological data according to claim 1, characterized in that, Obtain the dataset and preprocess it to obtain a preprocessed dataset, including: The data in the dataset is cleaned by using statistical methods to detect and remove data points with abnormal fluctuations, eliminating outliers and noisy data to ensure the accuracy and completeness of the data. For missing values ​​caused by equipment failure or data transmission problems, cubic spline interpolation is used to complete them to ensure the integrity of the data. A normalization method is used to compress the cleaned data to a preset numerical range in order to eliminate the dimensional differences between different data dimensions.

6. The spatiotemporal dust storm event AI forecasting method fusing satellite remote sensing and meteorological data according to claim 1, characterized in that, The preprocessed data is spatiotemporally aligned and interpolated to obtain unified input data, including: Project all types of data in the preprocessed data onto a unified spatial grid to ensure spatial alignment of all data. For PM10 concentration data with a resolution lower than a set threshold, use bilinear interpolation or spline interpolation to adjust it to the same high-resolution grid as the satellite remote sensing data. All data were standardized to hourly time intervals, and missing data at different times was filled in using interpolation methods. For meteorological data and satellite remote sensing data, linear interpolation is used to smooth the data for shorter time intervals to ensure the continuity and consistency of the time series. For daily PM10 data with only day intervals, hourly air quality station data is used to correct it to obtain hourly gridded PM10 data.

7. The spatiotemporal dust storm event AI forecasting method fusing satellite remote sensing and meteorological data according to claim 1, characterized in that, For various dust storm forecasting tasks, a multi-task loss function is constructed, and the dust storm forecasting model is trained based on the multi-task loss function, including: In each training cycle, the input data and the true label y i Used to calculate the prediction for the i-th task The regression task uses mean squared error loss for calculation: Among them, L i This represents the loss of the i-th task; This represents the prediction result for the j-th sample in the i-th task; N represents the total number of labeled data samples; and j represents the sample index. The classification task uses binary cross-entropy loss for training: The total loss L of a batch is calculated using the following formula. batch : Where w i It is the weight of the i-th task, and Task is the total number of tasks; To dynamically balance the task, w is dynamically adjusted during training based on the gradient norm. i The gradient norm g of the i-th task i for: in, This represents the gradient with respect to the parameter θ; The average gradient norm for all tasks is: Update the task weights using the following formula to minimize g. i and Differences between them: Where η is the learning rate of the weights; The weights are normalized to ensure that their sum is 1: Where w j This represents the j-th weight value; In each training iteration, minimize the total loss L. batch By updating the model parameters through backpropagation, we can ensure that all tasks achieve balanced optimization.

8. The spatiotemporal dust storm event AI forecasting method fusing satellite remote sensing and meteorological data according to claim 1, characterized in that, After obtaining the trained dust forecasting model, the method further includes: The trained dust storm forecasting model was evaluated using evaluation metrics. For regression tasks, mean squared error (MSE), root mean squared error (RMSE), mean absolute error (MAE), and / or coefficient of determination (R²) are used. 2 To evaluate the model's prediction accuracy; For classification tasks, accuracy, recall, precision, and / or F1 score are used for evaluation.

9. An AI forecasting device for spatiotemporal dust storm events integrating satellite remote sensing and meteorological data, based on the method as described in any one of claims 1 to 8, characterized in that, The device includes: The data preprocessing unit is configured to acquire a dataset and preprocess the dataset to obtain a preprocessed dataset; wherein the dataset includes satellite imagery data and meteorological reanalysis data, daily PM10 data and air quality data; The alignment interpolation unit is configured to perform spatiotemporal alignment and interpolation on the preprocessed data to obtain unified input data; The model building unit is configured to build a dust forecasting model. The dust forecasting model includes a spatiotemporal encoder, a feature aggregation layer, and a task-specific layer. Based on unified input data, it is processed into an input tensor of dimension (B, C, T, H, W), where B is the batch size, C is the number of input channels, T is the historical time length, and H×W is the spatial resolution. After the input passes through the dust forecasting model, it produces the outputs of four tasks: two regression prediction tasks and two classification prediction tasks. The output shape of each task is (B, H, W), representing the dust distribution at the next moment. The loss function determination unit is configured to construct a multi-task loss function for various dust event forecasting tasks, and train the dust forecasting model based on the multi-task loss function; The model training unit is configured to use the comprehensive loss as the optimization objective during the model training process, and continuously adjust the model parameters through the backpropagation algorithm to gradually minimize the multi-task loss and obtain the trained dust forecast model.

10. A non-transitory computer-readable storage medium storing instructions that, when executed by a processor, perform the method according to any one of claims 1 to 8.

Images