A sandstorm prediction method based on multi-source data fusion and conditional diffusion generation
By using multi-source data fusion and conditional diffusion generation, the dust storm prediction task is clearly defined as a spatiotemporal sequence problem. The conditional diffusion model is trained, which solves the problem of insufficient data utilization in existing technologies and achieves high-precision dust storm prediction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- ZHEJIANG UNIV OF TECH
- Filing Date
- 2025-02-14
- Publication Date
- 2026-05-15
AI Technical Summary
Existing technologies for dust storm prediction lack standardized datasets and clear prediction definitions, rely on traditional machine learning, do not fully utilize multi-source data, and do not comprehensively consider influencing factors.
By employing multi-source data fusion and conditional diffusion generation methods, the dust storm prediction task is clearly defined as a spatiotemporal sequence problem. The conditional diffusion model is trained through a multi-source data noise prediction network, and predictions are made using satellite cloud images, meteorological reanalysis data, and dust storm data.
It has achieved accurate sandstorm prediction, improving the accuracy and comprehensiveness of predictions, and is superior to traditional methods.
Smart Images
Figure CN120198792B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of dust storm forecasting technology, specifically relating to a dust storm prediction method based on multi-source data fusion and conditional diffusion generation. Background Technology
[0002] Dust storms, as a natural disaster in arid and semi-arid regions, are characterized by their rapid occurrence, high intensity, and significant impact. The occurrence of dust storms requires three basic conditions: strong winds, drought, and a source of dust. Dust storm forecasting utilizes historical meteorological data, including satellite data, to predict the occurrence of dust storms in the future. China is one of the countries most severely affected by dust storms globally. Large-scale dust storms cause enormous economic losses to northern China every year. To prevent such weather disasters, predicting the range and intensity of dust storms is essential.
[0003] Artificial intelligence technology is still in its infancy in the field of dust storm prediction. Although many studies have explored the application capabilities of artificial intelligence and deep learning technologies in dust storm prediction from multiple perspectives and achieved optimistic results, there are still several problems that need to be solved: (1) There are differences in data input and output among existing studies, and there is a lack of a standardized dust storm dataset and a clear definition of the dust storm prediction problem. (2) Most studies still rely on traditional machine learning techniques and have not yet incorporated emerging deep learning technologies into dust storm prediction research. (3) Existing studies do not consider the influencing factors of dust storms comprehensively enough, fail to make full use of multi-source observation data such as meteorological data and satellite remote sensing products, and have insufficient exploration of methods for multi-source data fusion. Summary of the Invention
[0004] To overcome the shortcomings of existing technologies, this invention provides a dust storm prediction method based on multi-source data fusion and conditional diffusion. By using multi-source data fusion and conditional diffusion, this method addresses the problems of existing technologies relying on traditional machine learning techniques and insufficient exploration of multi-source data, based on a clear definition of the dust storm prediction task and data.
[0005] The technical solution adopted by this invention to solve its technical problem is:
[0006] A dust storm prediction method based on multi-source data fusion and conditional diffusion includes the following steps:
[0007] S1. The dust storm prediction task is clearly defined as a spatiotemporal sequence prediction problem. The input is meteorological data at multiple times, and the output is dust storm change information at multiple times.
[0008] S2. Add noise to the dust storm data according to the noise addition formula to obtain noisy dust storm data; train the multi-source data noise prediction network by using historical satellite cloud images, meteorological reanalysis data, dust storm data, and time step t, and then complete the training of the conditional diffusion model.
[0009] S3. The image will contain a completely Gaussian noise map, and the current time step t will be displayed. x The network inputs historical satellite cloud images, meteorological reanalysis data, and dust storm data into a multi-source data noise prediction network to predict the noise in the current noise map. Then, it removes the noise from the current time frame using a denoising formula to obtain t. x-1 Gaussian noise plot at time t x-1 The dust storm data is always noisy; this operation is repeated until time t0 to obtain the denoised dust storm data, which is the prediction result.
[0010] Furthermore, in S1, the spatiotemporal sequence prediction problem includes the following parts:
[0011] The input data consists of meteorological data collected over the previous N time points, including satellite cloud imagery and meteorological reanalysis data. The output data, i.e., the target, is the change in dust storms over the next M time points, including changes in the range and intensity of dust storms. The prediction model used needs to accept data input from multiple time points and output data from multiple time points.
[0012] Preferably, the meteorological reanalysis data is subdivided into (1) atmospheric motion data and (2) temperature and pressure data. The atmospheric motion data includes data such as wind speed at two meters altitude, wind direction at two meters altitude, wind speed at ten meters altitude, and wind direction at ten meters altitude. The temperature and pressure data includes data such as air temperature, dew point temperature, surface air pressure, and average sea level air pressure.
[0013] Furthermore, in S2, the multi-source data noise prediction network integrates satellite cloud images, atmospheric motion data, temperature and air pressure data, and dust storm data, thereby promoting noise prediction through multi-source data fusion.
[0014] First, in the downsampling stage, satellite cloud images, atmospheric motion data, temperature and pressure data, and dust storm data each have their own processing channels. Features of each type of data are extracted through an encoder, a time-step embedding module, a residual linking module, and downsampling operations, with at least four scales of features extracted for each type of data. Then, the meteorological data fusion module connects the various types of data and extracts fused data features through 3D convolution, group normalization, and the SILU activation function, extracting fused data features at least four scales. Next, in the upsampling stage, the feature sizes are restored through the time-step embedding module and the residual linking module. Finally, the prediction head, i.e., the fully connected layer, outputs the noise prediction results.
[0015] Preferably, the downsampling stage is divided into four different processing channels to process satellite cloud images, atmospheric motion data, temperature and air pressure data, and sandstorm data, respectively.
[0016] First, satellite cloud images, atmospheric motion data, temperature and pressure data, and dust storm data differ in the number of data channels, requiring feature size alignment during encoding before subsequent downsampling. Second, conditional diffusion generates predicted dust storm data, which needs to be generated from a specific time step t. x Therefore, time information needs to be encoded and time step embedding is required when extracting features of dust storm data at various scales. However, satellite cloud images, atmospheric motion data, and temperature and air pressure data serve as auxiliary conditions to help the diffusion process. Therefore, time step embedding is not required when extracting features at various scales.
[0017] More preferably, the time step embedding module can embed time step information into sandstorm data information, and use time information to guide the implementation of conditional diffusion generation;
[0018] First, the time-step information is encoded using the SILU activation function and a fully connected layer. The dust storm data at time step t, including historical and predicted data, is used for feature extraction through group normalization, SILU activation, and 3D convolution. The two data points are then summed bit-by-bit. Next, the sum is further processed using group normalization, SILU activation, Dropout layer, and 3D convolution to extract more features. Finally, the extracted features are residually concatenated with the initial dust storm data to obtain the final time-step embedded features.
[0019] Furthermore, the encoder performs the following operations:
[0020] The input data passes through the input layer and the output layer in sequence. The input layer performs initial processing through group normalization, SILU activation function, and 3D convolution. The output layer obtains the encoding result through group normalization, SILU activation function, Dropout layer, and 3D convolution, thus achieving data size alignment.
[0021] Furthermore, the residual connection module performs the following operations:
[0022] The input data is processed through group normalization, SILU activation function, Dropout layer, and 3D convolution to extract features. The extracted features are then added bit-by-bit to the initial output data to obtain the residual connection result.
[0023] Furthermore, the downsampling operation performs the following operations:
[0024] The input features are first reduced in size through max pooling, then normalized through group normalization, and finally extracted through two-dimensional convolution.
[0025] Furthermore, the meteorological data fusion module can fuse and extract features from multi-source data, extracting features from multi-source data in at least four dimensions;
[0026] First, the multi-source data features extracted at the same scale during the downsampling stage are concatenated. Then, 3D convolution is used for feature extraction, and the extracted features are subjected to group normalization and SILU activation. Finally, the 3D convolution, group normalization, and SILU activation are repeated once more to obtain the fused features at that scale.
[0027] Furthermore, the upsampling stage can perform upsampling operations on the fusion features extracted by the meteorological data fusion module to gradually restore the feature size.
[0028] For each scale of fused features, the feature size is enlarged through time-step embedding module, residual module, bilinear interpolation, and two-dimensional convolution, and then connected with the fused features of the previous scale; the time-step embedding module, residual module, bilinear interpolation, and feature connection are continued; the operation is repeated until all fused data features are processed to restore the feature size.
[0029] In this invention, the forecast data includes satellite data and meteorological reanalysis data related to dust storms. The satellite data includes R0.47 and R0.65 satellite band data from Fengyun-4A; the meteorological reanalysis data includes wind speed and direction at two meters altitude from ERA5 and MERRA-2 reanalysis data. This application explicitly defines dust storm prediction as a spatiotemporal sequence prediction problem, specifying that meteorological data collected in the previous N time periods are the input, and dust storm changes in the next M time periods are the output. It employs a conditional diffusion generation model for progressive denoising to predict dust storms, and uses multi-source data fusion to predict noise during the diffusion process, finally obtaining information on future dust storm changes. This application extracts valuable information from complex meteorological data, enabling accurate dust storm prediction.
[0030] The main advantages of this invention are: it can accurately predict sandstorms. Attached Figure Description
[0031] Figure 1 This is a flowchart of a dust storm prediction method based on multi-source data fusion and conditional diffusion.
[0032] Figure 2 Inputs and outputs for dust storm prediction tasks.
[0033] Figure 3This is a schematic diagram of the conditional diffusion model.
[0034] Figure 4 This is a schematic diagram of the structure of a multi-source data noise prediction network.
[0035] Figure 5 This is a schematic diagram of the time step embedding module. Detailed Implementation
[0036] The present invention will now be further described with reference to the accompanying drawings.
[0037] Reference Figures 1-5 A dust storm prediction method based on multi-source data fusion and conditional diffusion generation is proposed. This method involves training and constructing a prediction network model based on a clear dust storm prediction task, and then using the trained prediction network model to process multi-source data to obtain information on dust storm changes in the future.
[0038] The dust storm prediction method based on multi-source data fusion and conditional diffusion includes the following steps:
[0039] Step S1: Define the dust storm prediction task as a spatiotemporal sequence prediction problem. The input is meteorological data at multiple times, and the output is dust storm change information at multiple times.
[0040] Specifically, such as Figure 2 As shown, the data input for dust storm prediction is meteorological data from N historical moments, including satellite cloud images, meteorological reanalysis data, and dust storm data; the output data is information on dust storm changes over the next M moments.
[0041] In one specific embodiment, for the input data, the size of the satellite cloud image is 4×5×160×480×7, the size of the meteorological reanalysis data is 4×5×160×480×12, and the size of the dust storm data is 4×5×160×480×1. Here, 4 represents the batch size, 5 represents the number of historical and future time points, the time interval is 15 minutes, 160 and 480 represent the spatial dimension of each channel, and 7, 12, and 1 represent the number of channels for each data point.
[0042] It should be noted that this application defines dust storm prediction as a spatiotemporal prediction problem. Different models can be used to solve spatiotemporal prediction problems. This application uses an advanced conditional diffusion model to implement dust storm prediction.
[0043] Step S2: Add noise to the dust storm data according to the noise addition formula to obtain noisy dust storm data; train the multi-source data noise prediction network by using historical satellite cloud images, meteorological reanalysis data, dust storm data, and time step t, and then complete the training of the conditional diffusion model.
[0044] This embodiment adds noise to the dust storm data and trains a multi-source data noise prediction network using time steps and multi-source data from historical moments to complete the training of the conditional diffusion model.
[0045] In a specific embodiment, such as Figure 3 As shown, this implementation requires training a noise prediction network for subsequent dust storm data generation. First, noise is added to the dust storm data, with the added noise determined by the time step t, resulting in noisy dust storm data. Then, the noisy dust storm data, the time step t, and multi-source data from historical moments are input into the multi-source data noise prediction network to minimize the difference between the noise predicted by the network and the added noise. Among them, the multi-source data from historical moments serve as a condition to promote the prediction of the noise network. The multi-source data includes satellite cloud images, meteorological reanalysis data, and dust storm data. The satellite cloud images mainly include bands such as R0.47 and R0.65. The meteorological reanalysis data is subdivided into (1) atmospheric motion data and (2) temperature and pressure data. The atmospheric motion data mainly includes wind speed at two meters, wind direction at two meters, wind speed at ten meters, and wind direction at ten meters. The temperature and pressure data mainly include air temperature, dew point temperature, surface air pressure, and mean sea level air pressure. The dust storm data mainly includes the location and intensity information of the dust storm.
[0046] Specifically, multi-source data noise prediction networks, such as Figure 4 As shown, in the downsampling stage, multi-source data undergoes data size alignment via corresponding encoders, followed by time-step embedding, residual connection, and downsampling operations to extract features from various data types, extracting features at least four scales. Dust storm data is constrained by time step t during the noise addition stage; therefore, a time-step embedding module is needed to encode temporal information, followed by feature extraction via the residual module and downsampling operations. The time-step embedding module is as follows: Figure 5 As shown, the time step information is encoded through SILU activation and a fully connected layer, and then added bitwise to the dust storm data features extracted through group normalization, SILU activation, and 3D convolution. Further feature extraction is then performed through group normalization, SILU activation, Dropout layer, and 3D convolution. Finally, residual concatenation is performed with the initial dust storm data to obtain the final time step embedding features. Satellite cloud images, atmospheric motion data, and temperature and pressure data, as conditions, do not require time information encoding and can be directly extracted using residual modules and downsampling operations.
[0047] In one specific embodiment, the number of input channels, output channels, kernel size, stride, and fill width of the 3D convolution in the satellite cloud image encoder are 7, 32, 3, 1, 1 and 32, 32, 3, 1, 1, respectively; the number of input channels, output channels, kernel size, stride, and fill width of the 3D convolution in the atmospheric motion data encoder are 6, 32, 3, 1, 1 and 32, 32, 3, 1, 1, respectively; the number of input channels, output channels, kernel size, stride, and fill width of the 3D convolution in the temperature and barometric pressure data encoder are 6, 32, 3, 1, 1 and 32, 32, 3, 1, 1, respectively; and the number of input channels, output channels, kernel size, stride, and fill width of the 3D convolution in the dust storm data encoder are 1, 32, 3, 1, 1 and 32, 32, 3, 1, 1, respectively. Therefore, after passing through the encoder, all data dimensions are expanded to 32 channels, thereby achieving data size alignment.
[0048] Satellite cloud images, atmospheric motion data, and temperature and pressure data have the same data size after being encoded, and the subsequent processing is similar. Taking satellite cloud images as an example, when extracting features of the first size, the number of input channels, number of output channels, kernel size, stride, and padding width of the 3D convolution in the residual module are 32, 32, 3, 1, and 1, respectively; when extracting features of the second size, the number of input channels, number of output channels, kernel size, stride, and padding width of the 2D convolution and the 3D convolution in the residual module during downsampling are 32, 64, 3, 1, and 1, respectively, and 64, 64, 3, 1, and 1, respectively; when extracting features of the third size, ... In the downsampling operation, the number of input channels, number of output channels, kernel size, stride, and padding width of the 2D convolution and the 3D convolution in the residual module are 64, 128, 3, 1, 1 and 128, 128, 3, 1, 1, respectively. When extracting features of the fourth size, the number of input channels, number of output channels, kernel size, stride, and padding width of the 2D convolution and the 3D convolution in the residual module are 128, 256, 3, 1, 1 and 256, 256, 3, 1, 1, respectively. For dust storm data, when extracting the first scale feature, the input channels, output channels, kernel size, stride, and padding width of the 3D convolution in the time-step embedding module are 32, 32, 3, 1, and 1, respectively. When extracting the second scale feature, the input channels, output channels, kernel size, stride, and padding width are 64, 64, 3, 1, and 1, respectively. When extracting the third scale feature, the input channels, output channels, kernel size, stride, and padding width are 128, 128, 3, 1, and 1, respectively. When extracting the fourth scale feature, the input channels, output channels, kernel size, stride, and padding width are 256, 256, 3, 1, and 1, respectively. All pooling operations used are max pooling, with kernel size, stride, and padding width of 2, 2, and 0, respectively.
[0049] This embodiment obtains data features at one scale through an encoder, downsampling operation, time step embedding module, and residual module. When multiple scale data features are required, multiple feature extraction units are needed, which will not be elaborated here.
[0050] During the data fusion phase, the meteorological data fusion module connects various data features at the same scale and extracts fused data features through three-dimensional convolution, group normalization, and SILU activation function, extracting fused features at least four scales.
[0051] In one specific embodiment, the data features at the first scale have 128 channels after concatenation, and the number of input channels, output channels, kernel size, stride, and padding width used in the 3D convolution are 128, 32, 3, 1, 1 and 32, 32, 3, 1, 1, respectively; the data features at the second scale have 256 channels after concatenation, and the number of input channels, output channels, kernel size, stride, and padding width used in the 3D convolution are 256, 64, 3, 1, 1 and 64, 64, 3, 1, 1, respectively; the data features at the third scale... The number of channels after concatenation of the first scale data features is 512, and the number of channels after concatenation of the fourth scale data features is 1024. The number of input channels, output channels, kernel size, stride, and padding width of the 3D convolution used are 512, 128, 3, 1, 1 and 128, 128, 3, 1, 1, respectively. The number of channels after concatenation of the fourth scale data features is 1024, and the number of input channels, output channels, kernel size, stride, and padding width of the 3D convolution used are 1024, 512, 3, 1, 1 and 512, 512, 3, 1, 1, respectively.
[0052] In the upsampling phase, starting with the fused feature with the smallest data size, the upsampling operation restores the data size through a time-step embedding module and a residual connection module, and then connects it with the fused feature from the previous scale. This process is repeated until the data size is restored. Specifically, the upsampling operation uses bilinear interpolation to expand the data size and two-dimensional convolution to enhance the expressive power of the feature map.
[0053] In one specific embodiment, for features at the fourth scale, the number of input channels, number of output channels, kernel size, stride, and padding width of the 3D convolution in the upsampling time-step embedding module and the residual connection module are 512, 512, 3, 1, 1 and 512, 256, 3, 1, 1, respectively. The number of input channels, number of output channels, kernel size, stride, and padding width of the 2D convolution in the upsampling operation are 256, 128, 3, 1, 1, respectively. For features at the third scale, the number of input channels, number of output channels, kernel size, stride, and padding width of the 3D convolution in the upsampling time-step embedding module and the residual connection module are 256, 256, 3, 1, 1 and 256, 128, 3, 1, 1, respectively. The input channels, output channels, kernel size, stride, and padding width of the 2D convolution in the upsampling operation are 128, 64, 3, 1, and 1, respectively. For the features at the second scale, the input channels, output channels, kernel size, stride, and padding width of the 3D convolution in the upsampling time-step embedding module and the residual connection module are 128, 128, 3, 1, 1 and 128, 64, 3, 1, 1, respectively. The input channels, output channels, kernel size, stride, and padding width of the 2D convolution in the upsampling operation are 64, 32, 3, 1, 1, respectively. For the features at the first scale, the input channels, output channels, kernel size, stride, and padding width of the 3D convolution in the upsampling time-step embedding module and the residual connection module are 64, 64, 3, 1, 1 and 64, 32, 3, 1, 1, respectively.
[0054] During the prediction phase, the features are passed through the prediction head, i.e., the fully connected layer, to output the noisy prediction results.
[0055] In one specific embodiment, there are two fully connected layers: the first fully connected layer, i.e., the time-series fully connected layer, has an input dimension of 10 and an output dimension of 5; the second fully connected layer, i.e., the channel fully connected layer, has an input dimension of 32 and an output dimension of 1.
[0056] It should be noted that the goal of the conditional diffusion model generation model in the training phase is to train the noise prediction network. The multi-source data noise prediction network designed in this application uses satellite cloud images and meteorological reanalysis data as conditions to achieve accurate noise prediction.
[0057] Step S3: The image contains a completely Gaussian noise map, and the current time step t is used to... x The network inputs historical satellite cloud images, meteorological reanalysis data, and dust storm data into a multi-source data noise prediction network to predict the noise in the current noise map. Then, it removes the noise from the current time frame using a denoising formula to obtain t. x-1 Gaussian noise plot at time t x-1The dust storm data is always noisy. This operation is repeated until time t0 to obtain the denoised dust storm data, i.e., the prediction result.
[0058] This embodiment uses a multi-source data noise prediction network trained by application. By utilizing time step information and multi-source data from historical moments, it gradually removes noise from completely Gaussian noise images to obtain clear dust storm data, which is the final dust storm prediction result.
[0059] In a specific embodiment, such as Figure 3 As shown, after the conditional diffusion model, i.e., the multi-source data noise prediction network, is trained, it needs to be used for dust storm prediction. In a specific embodiment, the maximum time step is 1000. During the prediction phase, from t... 1000 Starting with a completely Gaussian noise map at time step t, the completely Gaussian noise map and time step t are... 1000 And the multi-source data from historical moments are input into the multi-source noise prediction network to obtain t 1000 The noise prediction result at time t is obtained, and then t is obtained according to the denoising formula. 999 Dust storm data with Gaussian noise at all times. Then, t... 999 Dust storm data with Gaussian noise at any given time, t 999 Multi-source data from time steps and historical moments are input into a multi-source noise prediction network to obtain t. 999 The noise prediction result at time t is obtained, and then t is obtained according to the denoising formula. 998 The dust storm data contains Gaussian noise at each time step. The noise prediction network and denoising formula are repeatedly applied until the dust storm data at time t0 is obtained, which is the final prediction result.
[0060] The training process for the conditional diffusion generative model in this application is as follows:
[0061] First, the dataset was preprocessed by dividing it into three parts in a 7:1:2 ratio: training set, validation set, and test set. The training set included satellite cloud images, meteorological reanalysis data, and dust storm data from March to May 2020-2021, covering latitude and longitude ranges of 32.4N-52.7N and 64.3E-129.2E. The validation and test sets included satellite cloud images, meteorological reanalysis data, and dust storm data from March and April-May 2022, respectively. The satellite cloud images and meteorological reanalysis data were normalized to make the data more conducive to model training. To accommodate the model, all data were set to a size of 160×480.
[0062] During training:
[0063] The parameters of the detection network model, such as the convolutional layers, were initialized with random weights. The model was then optimized using the Adam model optimizer with a learning rate of 0.0002 and trained for 200 cycles.
[0064] The training set sequence data is input into the model, and noise is added to the dust storm data using randomly generated time steps t. Then, the noise-added dust storm data, time step t, and multi-source data from historical moments are input into a multi-source data prediction network to obtain the noise prediction results. The network's prediction results are compared with the added noise to obtain the loss L. The parameters of the noise prediction network are adjusted by minimizing the loss function using AdamW to optimize the noise prediction performance.
[0065] Train the multi-source data noise prediction network model until the loss converges, and obtain the best-performing multi-source data noise prediction network for noise prediction.
[0066] This application also provides experimental data on the dataset used for training the conditional generative model, comparing the proposed solution with other deep learning prediction methods (U-net, 3D U-net, ConvLSTM, Earthform). Performance metrics used for comparison include the Dice coefficient, Key Success Index (SCI), Heidrick Skill Score (HSS), Kappa coefficient, and Bias score. A Bias score closer to 1 indicates better performance, and higher values for Dice, SCI, HSS, and Kappa also indicate better performance. The comparison results are shown in Table 1.
[0067] Methods Dice CSI HSS Kappa Bias U-net 0.2659 0.1720 0.2623 0.4012 10.2083 3D U-net 0.4371 0.3107 0.4335 0.4800 1.7637 ConvLSTM 0.4518 0.3191 0.4485 0.4999 2.0152 Earthfomer 0.4481 0.3253 0.4450 0.5817 7.5836 Method of the present invention 0.5017 0.3726 0.4971 0.5041 0.6307
[0068] Table 1
[0069] Table 1 above shows the comparison results of different deep learning models in the dust storm prediction task. The method of this application outperforms all previous methods on Dice, CSI, HSS, and Bias, and is second only to the Earthform model on Kappa. It can be seen that the performance of the method of this invention is better than other deep learning methods.
[0070] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0071] The embodiments described in this specification are merely examples of implementations of the inventive concept and are for illustrative purposes only. The scope of protection of this invention should not be considered limited to the specific forms described in these embodiments; rather, it extends to equivalent technical means conceived by those skilled in the art based on the inventive concept.
Claims
1. A method for predicting dust storms based on multi-source data fusion and conditional diffusion, characterized in that, The method includes the following steps: S1. The dust storm prediction task is clearly defined as a spatiotemporal sequence prediction problem. The input is meteorological data at multiple times, and the output is dust storm change information at multiple times. S2. Add noise to the dust storm data according to the noise addition formula to obtain noisy dust storm data; train the multi-source data noise prediction network by using historical satellite cloud images, meteorological reanalysis data, dust storm data, and time step t, and then complete the training of the conditional diffusion model. The multi-source data noise prediction network integrates satellite cloud images, atmospheric motion data, temperature and air pressure data, and dust storm data to improve noise prediction through multi-source data fusion. First, in the downsampling stage, satellite cloud images, atmospheric motion data, temperature and pressure data, and dust storm data each have their own processing channels. Features of each type of data are extracted through an encoder, a time-step embedding module, a residual linking module, and downsampling operations, with at least four scales of features extracted for each data type. Next, the meteorological data fusion module connects the various data types and extracts fused data features through 3D convolution, group normalization, and the SILU activation function, extracting fused data features at least four scales. Following this, in the upsampling stage, the feature sizes are restored through the time-step embedding module and the residual linking module. Finally, the prediction head, i.e., the fully connected layer, outputs the noisy prediction results. The time step embedding module can embed time step information into sandstorm data information and use time information to guide the implementation of conditional diffusion generation; First, the time-step information is encoded using the SILU activation function and a fully connected layer. The dust storm data at time step t, including historical and predicted dust storm data, is used for feature extraction through group normalization, SILU activation function, and 3D convolution. The two data are then summed bit by bit. Next, the summed result is further processed through group normalization, SILU activation function, Dropout layer, and 3D convolution to extract more features. Finally, the extracted features are residually linked with the initial dust storm data to obtain the final time-step embedded features. The encoder performs the following operations: the input data passes through the input layer and the output layer in sequence. The input layer performs initial processing through group normalization, SILU activation function, and 3D convolution. The output layer obtains the encoding result through group normalization, SILU activation function, Dropout layer, and 3D convolution, thereby achieving data size alignment. The residual linking module performs the following operations: the input data is processed through group normalization, SILU activation function, Dropout layer, and 3D convolution to extract features; the extracted features are then added bit-by-bit to the initial output data to obtain the residual linking result. The downsampling operation performs the following steps: the input features are first reduced in size through max pooling, then normalized through group normalization, and finally the features are extracted through two-dimensional convolution. The meteorological data fusion module can fuse and extract features from multi-source data, and extract features from multi-source data in at least four dimensions. First, the multi-source data features of the same scale extracted in the downsampling stage are concatenated. Then, 3D convolution is used for feature extraction, and the extracted features are subjected to group normalization and SILU activation. Finally, 3D convolution, group normalization, and SILU activation are repeated once to obtain the fused features at that scale. The upsampling stage can perform upsampling operations on the fusion features extracted by the meteorological data fusion module to gradually restore the feature size; S3. The image will contain a completely Gaussian noise map, and the current time step t will be displayed. x The network inputs historical satellite cloud images, meteorological reanalysis data, and dust storm data into a multi-source data noise prediction network to predict the noise in the current noise map. Then, it removes the noise from the current time frame using a denoising formula to obtain t. x-1 Gaussian noise plot at time t x-1 The dust storm data is always noisy; this operation is repeated until time t0 to obtain the denoised dust storm data, which is the prediction result.
2. The dust storm prediction method based on multi-source data fusion and conditional diffusion as described in claim 1, characterized in that, In step S1 of the method, the spatiotemporal sequence prediction problem includes the following parts: The input data consists of meteorological data collected over the previous N time points, including satellite cloud imagery and meteorological reanalysis data. The output data, i.e., the target is the changes in dust storms over the next M time points, including changes in the range and intensity of dust storms; the prediction model used needs to accept data input at multiple time points and output data at multiple time points.
3. The dust storm prediction method based on multi-source data fusion and conditional diffusion as described in claim 2, characterized in that, The meteorological reanalysis data is further subdivided into atmospheric motion data and temperature and pressure data. The atmospheric motion data includes wind speed at two meters, wind direction at two meters, wind speed at ten meters, and wind direction at ten meters. The temperature and pressure data includes air temperature, dew point temperature, surface pressure, and mean sea level pressure.
4. The dust storm prediction method based on multi-source data fusion and conditional diffusion as described in claim 1, characterized in that, The downsampling stage is divided into four different processing channels, which respectively process satellite cloud images, atmospheric motion data, temperature and air pressure data, and dust storm data. First, satellite cloud images, atmospheric motion data, temperature and pressure data, and dust storm data differ in the number of data channels, requiring feature size alignment during encoding before subsequent downsampling. Second, conditional diffusion generates predicted dust storm data, which needs to be calculated from time step t. x Therefore, time information needs to be encoded and time step embedding is required when extracting features of dust storm data at various scales. However, satellite cloud images, atmospheric motion data, and temperature and air pressure data serve as auxiliary conditions to help the diffusion process. Therefore, time step embedding is not required when extracting features at various scales.
5. The dust storm prediction method based on multi-source data fusion and conditional diffusion as described in claim 1, characterized in that, For each scale of fused features, the feature size is enlarged through time-step embedding module, residual module, bilinear interpolation, and two-dimensional convolution, and then connected with the fused features of the previous scale; the time-step embedding module, residual module, bilinear interpolation, and feature connection are continued; the operation is repeated until all fused data features are processed to restore the feature size.