A method, apparatus, device, and storage medium for predicting the movement path of a dust storm.

Abstract

This invention relates to a method, apparatus, device, and storage medium for predicting the movement path of dust storms. The method involves: acquiring an aerosol optical thickness dataset and geographic background information of a dust storm event; generating a binarized labeled dust storm image sequence based on the aerosol optical thickness dataset; selecting key features from the geographic background information using a random forest feature importance method; concatenating the dust storm image sequence with the key features to obtain an input dataset; constructing an initial mixture model for predicting the dust storm's movement path; training the initial mixture model using the input dataset to obtain a target mixture model; and using the target mixture model to predict the future movement path of the dust storm. This approach achieves efficient and accurate prediction of dust storm movement paths.

Description

Technical Field

[0001] This invention relates to the field of natural disaster prevention, and more specifically, to a method, apparatus, equipment, and storage medium for predicting the movement path of sandstorms. Background Technology

[0002] With the gradual change in global climate, dust storms have become increasingly frequent in recent years, and their impact range has expanded to thousands of kilometers. According to the World Meteorological Organization, a dust storm is defined as a strong wind blowing large amounts of dust from bare, dry soil into the atmosphere and carrying it hundreds to thousands of kilometers away. Dust storms mainly occur in arid and semi-arid regions, and their destructive power is enormous, potentially causing direct disasters such as collapsed buildings and injuries or deaths to people and livestock. Furthermore, dust storms absorb and scatter solar radiation, altering the microphysical properties of clouds, reducing precipitation, further exacerbating drought, lowering air quality, and triggering various diseases. The short-term and long-term health effects of dust storms on urban residents are significant, particularly increasing the risk of respiratory and cardiovascular diseases. Therefore, spatiotemporal monitoring, modeling, forecasting, and developing early warning systems for dust storm movement paths are of great importance for mitigating and preventing the environmental, health, and socioeconomic impacts of dust storms on urban areas.

[0003] Existing methods for predicting dust storm paths mainly fall into three categories: The first category simulates the propagation process of dust storms based on atmospheric dynamic equations combined with meteorological data. These models rely on high-precision meteorological data input, are sensitive to initial conditions, and are suitable for long-term, large-scale predictions. However, they struggle to meet the short-term prediction needs under complex dynamic environments and have limited ability to model the spatial distribution characteristics of dust storms. The second category utilizes satellite remote sensing data (such as aerosol optical depth, AOD), ground-based lidar, and other sensors to monitor the concentration and spatial distribution of dust in the atmosphere, and then analyzes this data to predict the movement path of dust storms. Remote sensing technology can provide high temporal and spatial resolution data, facilitating real-time monitoring of dust storm dynamics and improving prediction accuracy when combined with ground observations. However, this method is overly reliant on the speed of remote sensing data acquisition and processing, making it difficult to perform complex dynamic simulations independently. It typically requires combination with other models (such as numerical models or deep learning models). The third category comprises deep learning methods, which can integrate multi-dimensional environmental factors to improve prediction accuracy. Models such as Convolutional Neural Networks (CNNs) and Long Short-Term Memory Networks (LSTMs) demonstrate superior performance in spatiotemporal feature extraction. However, CNNs lack temporal memory capabilities, and LSTMs alone struggle to capture spatial features, which limits their effectiveness in predicting dust storm paths.

[0004] Research has revealed several key shortcomings in existing dust storm path prediction technologies: First, they struggle to handle complex spatiotemporal dependencies. Specifically, traditional models, such as numerical models or remote sensing data methods, are ill-suited to addressing the intricate spatiotemporal dependencies in dust storm paths. Numerical models are highly dependent on initial conditions, making them ill-suited for short-term, high-spatiotemporal resolution predictions; remote sensing technologies rely heavily on data acquisition and processing speed, hindering complex dynamic simulations. Second, they lack long-term memory capabilities. Traditional CNN models, in particular, lack long-term memory and cannot utilize historical data for prediction, resulting in high short-term accuracy but declining long-term accuracy. Third, they are highly dependent on meteorological data. Specifically, numerical models and some remote sensing methods are heavily reliant on meteorological data, with prediction accuracy significantly affected by the quality of that data. Summary of the Invention

[0005] In view of this, the purpose of the present invention is to provide a method, apparatus, device and storage medium for predicting the movement path of a sandstorm, so as to achieve efficient and accurate prediction of the movement path of a sandstorm.

[0006] In a first aspect, embodiments of this application provide a method for predicting the movement path of a sandstorm, the method comprising:

[0007] Obtain aerosol optical thickness datasets and geographical background information for dust storm events;

[0008] Based on the aerosol optical thickness dataset, a sequence of dust storm images with binary labels is generated, and key features are selected from the geographic background information using the random forest feature importance method.

[0009] The input dataset is obtained by concatenating the sandstorm image sequence with the key features.

[0010] An initial mixture model is constructed for predicting the movement path of a sandstorm, and a target mixture model is obtained by training the initial mixture model using the input dataset;

[0011] The target hybrid model is used to predict the future movement path of dust storms.

[0012] Optionally, the aerosol optical thickness dataset includes several aerosol optical thickness distribution maps, and the generation of a binarized labeled dust storm image sequence based on the aerosol optical thickness dataset includes:

[0013] The Otsu thresholding algorithm was used to binarize the optical thickness distribution maps of aerosols to obtain several images to be labeled.

[0014] Pixels in each image to be labeled that have a pixel value higher than a preset threshold are labeled as dusty areas, and pixels in the image to be labeled that have a pixel value lower than the preset threshold are labeled as non-dusty areas, thus obtaining several labeled images;

[0015] The dust storm image sequence is constructed based on each labeled image.

[0016] Optionally, the step of filtering key features from the geographic background information using the random forest feature importance method includes:

[0017] The average reduced impurities (MDI) score for each feature in the geographic background information was calculated using a random forest model.

[0018] The key features are determined based on the MDI score ranking of each feature.

[0019] Optionally, the initial hybrid model includes three convolutional layers, one bidirectional long short-term memory network layer, two max pooling layers, two fully connected layers, and a dropout layer.

[0020] Optionally, in the three convolutional layers, the first convolutional layer uses 16 3*1 filters, the second convolutional layer uses 16 3*16 filters, and the third convolutional layer uses 16 3*16 filters; the activation function of each convolutional layer is ReLU, and the model weights are initialized using the GlorotUniform initialization method.

[0021] Optionally, the bidirectional long short-term memory network layer contains 32 units, which can simultaneously capture both the forward and reverse information of the sequence.

[0022] Optionally, after predicting the future movement path of the dust storm using the target mixture model, the method further includes:

[0023] Obtain several mobile path prediction results, and count the number of true positive results, true negative results, false positive results, and false negative results in the mobile path prediction results;

[0024] The overall accuracy, F1 score, and Kappa coefficient are determined based on the number of true positive results, true negative results, false positive results, and false negative results in the predicted movement path.

[0025] The performance of the target mixture model is evaluated based on the overall accuracy, F1 score, and Kappa coefficient.

[0026] Secondly, embodiments of this application provide a dust storm movement path prediction device, the device comprising:

[0027] The data acquisition module is used to acquire aerosol optical thickness datasets and geographical background information for dust storm events;

[0028] The data processing module is used to generate a sequence of binarized dust storm images based on the aerosol optical thickness dataset, and to filter out key features from the geographic background information using the random forest feature importance method.

[0029] The dataset construction module is used to perform feature concatenation between the sandstorm image sequence and the key features to obtain the input dataset;

[0030] The model training module is used to construct an initial hybrid model for predicting the movement path of sandstorms, and to train the initial hybrid model using the input dataset to obtain a target hybrid model.

[0031] The path prediction module is used to predict the future movement path of dust storms using the target hybrid model.

[0032] Optionally, the aerosol optical thickness dataset includes several aerosol optical thickness distribution maps, and the generation of a binarized labeled dust storm image sequence based on the aerosol optical thickness dataset includes:

[0033] The Otsu thresholding algorithm was used to binarize the optical thickness distribution maps of aerosols to obtain several images to be labeled.

[0034] Pixels in each image to be labeled that have a pixel value higher than a preset threshold are labeled as dusty areas, and pixels in the image to be labeled that have a pixel value lower than the preset threshold are labeled as non-dusty areas, thus obtaining several labeled images;

[0035] The dust storm image sequence is constructed based on each labeled image.

[0036] Optionally, the step of filtering key features from the geographic background information using the random forest feature importance method includes:

[0037] The average reduced impurities (MDI) score for each feature in the geographic background information was calculated using a random forest model.

[0038] The key features are determined based on the MDI score ranking of each feature.

[0039] Optionally, the initial hybrid model includes three convolutional layers, one bidirectional long short-term memory network layer, two max pooling layers, two fully connected layers, and a dropout layer.

[0040] Optionally, in the three convolutional layers, the first convolutional layer uses 16 3*1 filters, the second convolutional layer uses 16 3*16 filters, and the third convolutional layer uses 16 3*16 filters; the activation function of each convolutional layer is ReLU, and the model weights are initialized using the GlorotUniform initialization method.

[0041] Optionally, the bidirectional long short-term memory network layer contains 32 units, which can simultaneously capture both the forward and reverse information of the sequence.

[0042] Optionally, after predicting the future movement path of the dust storm using the target mixture model, the method further includes:

[0043] Obtain several mobile path prediction results, and count the number of true positive results, true negative results, false positive results, and false negative results in the mobile path prediction results;

[0044] The overall accuracy, F1 score, and Kappa coefficient are determined based on the number of true positive results, true negative results, false positive results, and false negative results in the predicted movement path.

[0045] The performance of the target mixture model is evaluated based on the overall accuracy, F1 score, and Kappa coefficient.

[0046] Thirdly, embodiments of this application provide a computer device, including: a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the computer device is running, the processor communicates with the memory via the bus. When the machine-readable instructions are executed by the processor, the steps of the sandstorm movement path prediction method described in any of the optional embodiments of the first aspect are performed.

[0047] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program that, when executed by a processor, performs the steps of the sandstorm movement path prediction method described in any of the optional embodiments of the first aspect.

[0048] The technical solution provided in this application includes, but is not limited to, the following beneficial effects:

[0049] This application first collects two key datasets related to dust storms: aerosol optical depth (AOD) datasets and geographic background information. The AOD dataset, acquired through satellite remote sensing, reflects the concentration and spatial distribution of dust in the atmosphere. Geographic background information includes environmental characteristics such as relative humidity, surface air temperature, surface skin temperature, wind speed and direction at different altitudes. These data provide a rich source of information for the model, contributing to a more comprehensive understanding of the formation and movement mechanisms of dust storms.

[0050] Then, the AOD dataset was binarized using the Otsu thresholding algorithm to generate a sequence of binarized dust storm images. Simultaneously, the random forest feature importance method was employed to select the key features with the greatest impact on the dust storm's movement path from the geographic background information. This step, by simplifying the data and selecting key features, improved data processing efficiency, reduced model complexity and computational cost, while avoiding overfitting and improving the model's generalization ability.

[0051] Next, the processed dust storm image sequences are concatenated with the selected key geographic background features to form a complete input dataset. This integrated dataset provides the model with richer information, helping it to more accurately capture the spatiotemporal characteristics of dust storms, thereby improving the accuracy and reliability of predictions.

[0052] An initial hybrid model is then constructed, combining several convolutional layers and bidirectional long short-term memory network layers. The initial hybrid model is trained using the concatenated input dataset to obtain the target hybrid model. This innovative model structure, combining 1DCNN and BiLSTM, effectively handles the complex spatiotemporal dependencies in the movement path of dust storms, significantly improving the model's predictive ability.

[0053] Finally, the trained objective mixture model is used to predict the future path of the dust storm. After obtaining the prediction results, the model's performance is evaluated, the number of true positives, true negatives, false positives, and false negatives in the prediction results are counted, and indicators such as overall accuracy, F1 score, and Kappa coefficient are calculated. This step not only provides scientific support for actual disaster prevention and mitigation work, but also ensures the reliability of the model's prediction results through multi-dimensional performance evaluation, providing a basis for further model optimization.

[0054] Through the above steps, from data acquisition to model training, and finally to prediction and evaluation, each step has a clear beneficial effect, and together they construct a complete and efficient method for predicting sandstorm paths. This method can not only achieve efficient and accurate prediction of sandstorm movement paths, but also provide scientific support for actual disaster prevention and mitigation work, and has important application value.

[0055] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0056] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0057] Figure 1 The flowchart of a method for predicting the movement path of a sandstorm provided in Embodiment 1 of the present invention is shown;

[0058] Figure 2 A flowchart of a method for generating sandstorm image sequences provided in Embodiment 1 of the present invention is shown;

[0059] Figure 3 The flowchart of a key feature screening method provided in Embodiment 1 of the present invention is shown;

[0060] Figure 4 A flowchart of a model performance evaluation method provided in Embodiment 1 of the present invention is shown;

[0061] Figure 5 This diagram illustrates the structure of a sandstorm movement path prediction device provided in Embodiment 2 of the present invention.

[0062] Figure 6 A schematic diagram of the structure of a computer device provided in Embodiment 3 of the present invention is shown. Detailed Implementation

[0063] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0064] Example 1

[0065] To facilitate understanding of this application, the following is combined with... Figure 1 The flowchart of the method for predicting the movement path of a sandstorm provided in Embodiment 1 of the present invention is shown below, which will be described in detail for Embodiment 1 of this application.

[0066] See Figure 1 As shown, Figure 1 The flowchart of a method for predicting the movement path of a sandstorm according to Embodiment 1 of the present invention is shown, wherein the method includes steps S101 to S105:

[0067] S101: Obtain aerosol optical thickness dataset and geographic background information for dust storm events.

[0068] Specifically, this study collected two key datasets related to dust storms: aerosol optical thickness (AOD) datasets and geographic background information. AOD datasets, acquired through satellite remote sensing, reflect the concentration and spatial distribution of dust in the atmosphere. NASA's MERRA-2 dataset is a crucial source for obtaining high temporal resolution (hourly) and spatial resolution (0.5° × 0.625°) AOD data, essential for accurately capturing the dynamic changes of dust storms. Generated by NASA's Office of Global Simulation and Assimilation (GMAO), the MERRA-2 dataset is based on satellite reanalysis models, combining advanced Very High Resolution Radiometer (AVHRR), Multi-Angle Imaging Spectroradiometer (MISR), MODISAOD datasets, and AOD datasets from ground-based observations of the Aerosol Robotics Network (AERONET). This study used the "dustextinction aerosol optical thickness (AOT) 550nm" data from MERRA-2, covering 84 dust events and 2016 storm hours from 2000 to 2024.

[0069] Geographic background information includes environmental characteristics such as relative humidity, surface air temperature, surface skin temperature, wind speed and direction at different altitudes. This data can typically be obtained from meteorological observation stations or reanalysis datasets, providing the model with the environmental context for dust storm occurrence and movement. By combining this geographic background information with the AOD dataset, the model can more comprehensively understand and predict the movement path of dust storms, providing strong support for disaster prevention and mitigation and environmental protection.

[0070] S102: Generate a sequence of binarized dust storm images based on the aerosol optical thickness dataset, and select key features from the geographic background information using the random forest feature importance method.

[0071] Specifically, the aerosol optical thickness dataset contains remote sensing images of dust storm events, with AOD as the target variable. Each remote sensing image is processed using the Otsu thresholding algorithm, dividing pixels into "dust pixels" and "non-dust pixels," resulting in a series of labeled images that clearly show the extent and location of the dust storm. Secondly, to improve the model's efficiency and accuracy, it is necessary to filter out the key features that have the greatest impact on the dust storm's movement path from a wealth of geographical background information. This is achieved using the random forest feature importance method, which evaluates the importance of each feature in the prediction task and ranks the features according to the average reduced impurities (MDI) score, ultimately identifying the key features.

[0072] S103: The sandstorm image sequence and the key features are concatenated to obtain the input dataset.

[0073] Specifically, the purpose of this step is to combine spatial and temporal features to form a complete input dataset. Specifically, the processed dust storm image sequence (containing spatial information) is concatenated with selected key geographical background features (containing temporal information) to provide the model with comprehensive input. This allows the model to consider both the spatial distribution and temporal changes of dust storms during the learning process, thereby better capturing the spatiotemporal characteristics of dust storm movement paths.

[0074] S104: Construct an initial hybrid model for predicting the movement path of a sandstorm, and train the initial hybrid model using the input dataset to obtain a target hybrid model.

[0075] Specifically, the initial hybrid model combines convolutional layers (1DCNN) and bidirectional long short-term memory (BiLSTM) layers. 1DCNN is primarily used to extract spatial features from dust storm image sequences, identifying fundamental features and reducing data dimensionality through convolutional and pooling operations. BiLSTM captures long-term dependencies in time series data, simultaneously considering both positive and negative contextual information, thus improving the model's understanding and predictive capabilities for time series data. During model construction, the model weights are initialized using the GlorotUniform initialization method. Then, the initial hybrid model is trained using a concatenated input dataset. By adjusting the model's parameters, it accurately learns the characteristics and patterns of dust storm movement paths, ultimately yielding the target hybrid model.

[0076] During model training, the deep learning framework consists of three main parts: model input, prediction model, and model output. The data (1992 hours out of 2016 storm hours) was divided into 1592 training samples (80% of the data), 200 test samples (10% of the data), and 200 validation samples (10% of the data). The input to the deep learning model included raw MERRA-2 AOD data (images) at time step t-1 and geographic information at time step t-1. The output included predicted dust storm images for 24 hours (t to t+24). The labeled AOD data layer was used as the target layer during the deep learning model training phase. Dust storm movement paths are greatly influenced by geographic background; therefore, the AOD dataset and geographic background information (i.e., context information) were overlaid to construct the input layer. Due to the time series nature of the problem, the input consisted of raw AOD images of 71×81 pixels corresponding to each hour of storm movement, along with contextual geographic information. The model's predictions were achieved by using a neural network to extract spatial and temporal features from the input data.

[0077] S105: Use the target hybrid model to predict the future movement path of the dust storm.

[0078] Specifically, a trained objective mixture model is used to predict the movement path of future dust storms. After obtaining the prediction results, statistical analysis is needed to evaluate the model's performance. Specifically, the number of true positives, true negatives, false positives, and false negatives in the prediction results is counted. Then, based on these statistical data, indicators such as overall accuracy, F1 score, and Kappa coefficient are calculated to quantify the model's predictive accuracy and reliability. This step is crucial for verifying the model's effectiveness and practicality and also provides a basis for further model optimization.

[0079] In an optional implementation, see Figure 2 As shown, Figure 2 The flowchart illustrates a method for generating a sandstorm image sequence according to Embodiment 1 of the present invention. The aerosol optical thickness dataset contains several aerosol optical thickness distribution maps. The step of generating a binarized sandstorm image sequence based on the aerosol optical thickness dataset includes steps S201 to S203:

[0080] S201: The Otsu thresholding algorithm is used to binarize the optical thickness distribution maps of each aerosol to obtain several images to be labeled.

[0081] Specifically, the Otsu thresholding algorithm is applied to each acquired AOD distribution map. This algorithm automatically determines the optimal threshold for image segmentation, maximizing the inter-class variance between the foreground (dust-covered areas) and the background (non-dust-covered areas). Through this process, each AOD distribution map is converted into a binary image, where pixel values ​​above the threshold are labeled as "dust pixels," and those below the threshold are labeled as "non-dust pixels." The key to this step is that the Otsu algorithm adaptively determines the optimal threshold without manual intervention, thereby improving processing efficiency and accuracy.

[0082] S202: Mark the pixels in each image to be marked that are higher than a preset threshold as dusty areas, and mark the pixels in the images to be marked that are lower than the preset threshold as non-dusty areas, thus obtaining several marked images.

[0083] Specifically, pixels with values ​​higher than a preset threshold in each image to be labeled are marked as dusty areas. These pixels typically appear as high-brightness areas in the image, indicating a high concentration of dust. Pixels with values ​​lower than the threshold are marked as non-dusty areas. These pixels appear as low-brightness areas in the image, indicating little or no dust. This labeling process yields several labeled images that clearly distinguish between dusty and non-dusty areas. These images will serve as the basis for subsequently constructing a dust storm image sequence.

[0084] S203: Construct the sandstorm image sequence based on each labeled image.

[0085] Specifically, hourly images are processed into sequential data, fused, and data outside the study area are removed to obtain all pixels at time t-1, where t is the current time. This sequence contains the distribution of the dust storm at different time points, visually demonstrating the movement and changes of the dust storm. This image sequence will serve as one of the model's inputs to capture the spatial characteristics and temporal dynamics of the dust storm, providing crucial data support for subsequent model training and prediction.

[0086] In an optional implementation, see Figure 3 As shown, Figure 3 The flowchart of a key feature filtering method provided in Embodiment 1 of the present invention is shown, wherein the key features are filtered from the geographic background information using the random forest feature importance method, including steps S301 to S303:

[0087] S301: The average reduced impurity (MDI) score of each feature in the geographic background information is calculated using a random forest model.

[0088] Specifically, the MERRA-2 dataset provides a variety of environmental features, including 18 features such as relative humidity, surface air temperature, surface skin temperature, surface wind speed, surface wind direction, 10-meter wind direction, 50-meter wind direction, 500-meter wind direction and their corresponding temperature and humidity, and pressure and its corresponding temperature and humidity. To select the most effective parameters from these features for predicting dust storm paths, this paper employs the Random Forest Feature Importance (RFFI) technique. RFFI identifies the most important features or variables in the dataset for a specific prediction task. It trains a random forest model on the dataset and then evaluates the contribution of each feature to the model's prediction accuracy.

[0089] Random Forest (RFFI) is an ensemble learning method that improves the accuracy and stability of a model by constructing multiple decision trees and aggregating their results. Each decision tree is split based on feature importance during construction, and the Mean Decrease in Impurity (MDI) score measures the importance of each feature in the model. Specifically, the MDI score quantifies the reduction in impurities (or increase in purity) that a feature brings when used in a decision tree split. A higher MDI score indicates a greater impact of the feature on the model. RFFI ultimately obtains the final importance score for each feature by averaging the MDI scores across all decision trees in the model.

[0090] This method allows for the selection of the most valuable features for predicting the movement path of dust storms from a wealth of geographical background information, thereby reducing model complexity and computational costs while improving model efficiency and accuracy.

[0091] S302: Determine the key features based on the MDI score ranking of each feature.

[0092] Specifically, the features are first ranked according to their importance scores in the model. Features with higher scores play a more important role in the model and can better explain and predict the movement path of dust storms. By setting a threshold or selecting the top N features, the key features with the greatest impact on the dust storm's movement path can be identified. The purpose of this step is to reduce the model's complexity and computational cost, while improving its efficiency and accuracy. By retaining only the key features, we can build a simpler and more effective model, avoiding overfitting and improving generalization ability.

[0093] Furthermore, data standardization is a crucial step in ensuring model performance. Using a Z-score scaler to standardize features transforms the original data into a standard normal distribution by setting the mean to 0 and the standard deviation to 1. This standardization method helps ensure that all features are distributed in the same proportions, allowing the model to learn them equitably. Simultaneously, standardization helps reduce the impact of outliers, which can significantly affect model performance. The formula below shows the Z-score scaler normalization result X. scaled The calculation process:

[0094]

[0095] Where X represents the original data, μ is the mean of the original data, and σ is the standard deviation of the original data. This standardization process can further improve the stability and predictive ability of the model.

[0096] In an optional implementation, the initial hybrid model includes three convolutional layers, one bidirectional long short-term memory network layer, two max pooling layers, two fully connected layers, and a dropout layer.

[0097] Specifically, convolutional layers excel at processing sequential data, automatically extracting spatial features through convolution operations, thus reducing reliance on data preprocessing and feature engineering. They utilize multiple filters sliding across the input data to capture local high-dimensional features and further reduce data dimensionality through pooling operations, improving the model's computational efficiency and robustness.

[0098] Bidirectional Long Short-Term Memory (LSTM) networks focus on processing time-series data. Their unique bidirectional structure allows them to model sequence data simultaneously from both forward and backward directions. This mechanism enables LSM networks to capture long-term dependencies between time steps in a time series, effectively utilizing historical information to improve prediction accuracy. For the task of predicting dust storm paths, long-term dependencies in the time series are crucial because dust storm movement is often influenced by a combination of previous meteorological conditions and geographical factors.

[0099] In the hybrid model, convolutional layers first extract spatial features from the input dust storm image sequence, capturing the spatial patterns and structure of dust distribution in the images. Subsequently, a bidirectional long short-term memory network layer takes over the processing, modeling these spatial features in the temporal dimension to analyze the movement trends and changes of the dust storm at different time points. Through this deep fusion of spatial and temporal features, the hybrid model can comprehensively understand and predict the movement path of dust storms, overcoming the limitations of single models when processing complex spatiotemporal data.

[0100] In an optional implementation, in the three convolutional layers, the first convolutional layer uses 16 3*1 filters, the second convolutional layer uses 16 3*16 filters, and the third convolutional layer uses 16 3*16 filters; the activation function of each convolutional layer is ReLU, and the model weights are initialized using the GlorotUniform initialization method.

[0101] Specifically, the convolutional layer structure can be summarized as follows: In the hidden layers of a 1DCNN (One-Dimensional Convolutional Neural Network), the network involves two special matrix operations: convolutional layers and pooling layers. Convolutional layers act as local feature extractors, scanning the input data and extracting local high-dimensional features through multiple different filters. This is achieved by sliding the filter with a specific stride (ω). i ) in input data (x t The data is convolved with the input data, and neurons in each convolutional layer perform non-linear computations, dot products, and can selectively connect to neurons in the next layer to generate multiple features (c t，i Each neuron achieves local connectivity and shares weights, thereby reducing model complexity and accelerating training efficiency. The convolution equation is shown below:

[0102]

[0103] Where * represents the convolution operation, ω i b represents the filter vector. i c represents the bias vector. t，i This represents the output data features of a 1D CNN. One-dimensional convolutions still contain redundant information. To reduce redundancy and improve the robustness of feature extraction, pooling layers (downsampling) are added. Local averaging or max pooling reduces computational cost and improves learning efficiency. Convolutions and pooling can be repeated multiple times until the feature map is reduced to 1×1. Finally, the fully connected layer flattens the features, and the output layer uses the optimal loss function to calculate the optimal parameters. CNNs often use stochastic gradient descent to optimize training and improve model performance.

[0104] In an optional implementation, the bidirectional long short-term memory network layer contains 32 units and is capable of simultaneously capturing both forward and reverse information of the sequence.

[0105] Specifically, compared to traditional recurrent neural networks (RNNs), the Bidirectional Long Short-Term Memory (BiLSTM) network layer can simultaneously capture both forward and backward contextual information. This not only effectively handles long-term dependencies but also mitigates the possibility of gradient vanishing, thus better predicting current information. The core concepts of BiLSTM include cell states and gate structures. Cell states can transmit important information, overcoming the limitations of short-term memory and ensuring the effective transmission of long-term dependent information. BiLSTM includes three gate structures: input gate, forget gate, and output gate, each with its specific function, used for selectively remembering, forgetting, or outputting information, respectively. For forward propagation, the input sequence is set to X = [x1, x2, ..., xn], where xn is the input at time step n. The computation process of the forward LSTM is as follows:

[0106] 1. Formula for calculating the forgetting gate:

[0107] Among them, f t The current output of the forget gate; σ is the sigmoid activation function; W f It is a learnable parameter matrix, the dimensionality of which depends on the input features and the number of hidden units; h t-1 and These represent the hidden state at the previous time step and the input at the current time step, respectively; b f This is the bias vector for the forget gate, used to adjust the output of the forget gate, and is related to W. f Together, they ensure the flexibility of the model in learning forgetting rules.

[0108] 2. Calculation formula for input gate:

[0109]

[0110] Among them, i t The current output of the input gate is σ; σ is the Sigmoid activation function; W I The weight matrix of the input gate is used to control the current input. and the previous hidden state h t-1 The degree of contribution to the new cell state; h t-1 and These represent the hidden state at the previous time step and the input at the current time step, respectively; b i This is the bias vector of the input gate, used to adjust the input gate output; As candidate cell states, the values ​​are restricted to the [-1, 1] interval by the tanh activation function; W C The weight matrix for candidate cell states is used to calculate... The weighting parameter; b cThis is the bias vector for the candidate cell state, used to adjust the candidate state output.

[0111] 3. Unit Update:

[0112] Among them, c t c represents the current cell state. t-1 This represents the cell state at the previous time step, a long-term memory transmitted in the BiLSTM network at time t-1, containing information from all past time steps; f t This is the output of the forget gate at the current moment, used to control c. t-1 Forgotten information; Candidate cell state; i t This is the current output of the input gate.

[0113] 4. Update output gates and hidden status:

[0114] h t =o t ⊙tanh(c t );

[0115] Among them, o t The current output of the output gate; σ is the Sigmoid activation function; W o h is the output gate weight matrix. t-1 and These represent the hidden state at the previous time step and the input at the current time step, respectively; b o This is the bias vector of the output gate, used to adjust the output of the output gate, and is related to W. o To ensure the model's learning ability; h t The current hidden state; c t This represents the current state of the cell.

[0116] This application presents a hybrid model based on convolutional layers and bidirectional long short-term memory (LSTM) networks. This model comprises 13 different layers, including two convolutional layers, one bidirectional LSTM layer, one flattened layer, three fully connected layers, multiple batch normalization layers, and dropout layers. These layers are organized into several main modules: a feature extraction module and a prediction module. The feature extraction module contains two convolutional layers and one bidirectional LSTM layer: the first convolutional layer uses 16 3x1 filters, the second uses 16 3x16 filters, and the third uses 16 3x16 filters. The activation function for the convolutional layers is ReLU, and the model weights are initialized using GlorotUniform to improve the stability and convergence efficiency of model training. The bidirectional LSTM layer, containing 32 units, can simultaneously capture both forward and backward information of the sequence, enhancing the model's ability to handle time-dependent features. Dimensionality reduction is achieved between the convolutional layers using two max-pooling layers with a pooling window size of 2, reducing feature dimensions, improving computational efficiency, and mitigating the risk of overfitting.

[0117] The prediction module consists of two fully connected layers: one fully connected layer with 64 neurons using the ReLU activation function, and the other output layer using the softmax activation function based on the number of classes (softmax activation is commonly used in classification output layers, interpreting the output as a probability distribution, i.e., the predicted probability for each class). A Dropout layer is added between these two fully connected layers to prevent overfitting. A total of 23,346 trainable parameters were used during training.

[0118] In an optional implementation, see Figure 4 As shown, Figure 4 The flowchart of a model performance evaluation method provided in Embodiment 1 of the present invention is shown. After predicting the future movement path of a sandstorm using the target hybrid model, the method further includes steps S401 to S403:

[0119] S401: Obtain several movement path prediction results, and count the number of true positive results, true negative results, false positive results, and false negative results in the movement path prediction results.

[0120] Specifically, the trained objective mixture model is first used to predict the future path of dust storms, yielding a series of predictions. These results are presented as images, showing the possible locations and extents of dust storms at different time points. Next, a detailed statistical analysis of these predictions is performed. Specifically, the predicted results are compared with the actual dust storm paths, and the number of the following four outcomes is statistically analyzed:

[0121] True positive (TP): The number of results that the model correctly predicted as dusty areas.

[0122] True negative (TN): The number of results where the model correctly predicts a non-dust-affected area.

[0123] False positives (FP): The number of results in which the model incorrectly predicts a non-dust area as a dusty area.

[0124] False negative (FN): The number of times the model incorrectly predicts a dusty area as a non-dusty area.

[0125] These statistics form the basis for evaluating model performance, and they can intuitively reflect the accuracy and reliability of the model in the prediction process.

[0126] S402: Determine the overall accuracy, F1 score, and Kappa coefficient based on the number of true positive results, true negative results, false positive results, and false negative results in the predicted movement path.

[0127] Specifically, this data is used to calculate three key performance metrics: overall accuracy, F1 score, and Kappa coefficient. Overall accuracy, obtained by dividing the number of correctly predicted outcomes (TP and TN) by the sum of all predictions, reflects the proportion of the model's correct predictions. The F1 score, a useful quantitative measure of the balance between precision and completeness, is calculated by their harmonic mean and measures the model's balance between accuracy and completeness. The Kappa coefficient indicates the degree of agreement between the predicted and reference data. A Kappa value of 100% indicates 100% agreement, while a value of 0 indicates disagreement. These three metrics comprehensively reflect the model's performance from different perspectives.

[0128] Furthermore, the formula for calculating Overall Accuracy is as follows:

[0129]

[0130] The formula for calculating the F1 value is as follows:

[0131]

[0132] The formula for calculating the Kappa coefficient k is as follows:

[0133]

[0134] N = TP + TN + FP + FN;

[0135] Where TP is the number of results correctly predicted as dusty areas; TN is the number of results correctly predicted as non-dusty areas; FP is the number of results incorrectly predicted as dusty areas; FN is the number of results incorrectly predicted as non-dusty areas; Precision is the accuracy rate; Recall is the recall rate; ρ o ρ represents the observational consistency ratio, that is, the proportion of the actual classification result that matches the true label; e This represents the expected consistency ratio, which is the expected consistency ratio assuming the classification is random, where N is the total number of samples.

[0136] S403: Evaluate the performance of the target mixture model based on the overall accuracy, F1 score, and Kappa coefficient.

[0137] Specifically, the performance of the target mixture model is comprehensively evaluated based on the calculated overall accuracy, F1 score, and Kappa coefficient. Analyzing these metrics allows us to assess the model's accuracy and reliability in predicting dust storm paths. For example, a high overall accuracy indicates that the model can correctly predict most situations; a high F1 score suggests a good balance between precision and recall; and a high Kappa coefficient indicates that the model's predictions closely match reality, exceeding the expectation of mere coincidence. This multi-dimensional evaluation method ensures a comprehensive and objective understanding of the model's performance, providing crucial reference for further optimization and practical application.

[0138] Example 2

[0139] Embodiment 2 of the present invention provides a sandstorm movement path prediction device, see [link to related document]. Figure 5 As shown, Figure 5 A schematic diagram of a sandstorm movement path prediction device provided in Embodiment 2 of the present invention is shown, wherein the device includes:

[0140] The data acquisition module 501 is used to acquire aerosol optical thickness datasets and geographical background information for dust storm events;

[0141] Data processing module 502 is used to generate a sequence of binarized dust storm images based on the aerosol optical thickness dataset, and to filter out key features from the geographic background information using the random forest feature importance method;

[0142] The dataset construction module 503 is used to perform feature concatenation between the sandstorm image sequence and the key features to obtain the input dataset;

[0143] The model training module 504 is used to construct an initial hybrid model for predicting the movement path of sandstorms, and to train the initial hybrid model using the input dataset to obtain a target hybrid model.

[0144] The path prediction module 505 is used to predict the future movement path of the dust storm using the target hybrid model.

[0145] In an optional implementation, the aerosol optical thickness dataset includes several aerosol optical thickness distribution maps, and the generation of a binarized labeled dust storm image sequence based on the aerosol optical thickness dataset includes:

[0146] The Otsu thresholding algorithm was used to binarize the optical thickness distribution maps of aerosols to obtain several images to be labeled.

[0147] Pixels in each image to be labeled that have a pixel value higher than a preset threshold are labeled as dusty areas, and pixels in the image to be labeled that have a pixel value lower than the preset threshold are labeled as non-dusty areas, thus obtaining several labeled images;

[0148] The dust storm image sequence is constructed based on each labeled image.

[0149] In an optional implementation, the step of filtering key features from the geographic background information using a random forest feature importance method includes:

[0150] The average reduced impurities (MDI) score for each feature in the geographic background information was calculated using a random forest model.

[0151] The key features are determined based on the MDI score ranking of each feature.

[0152] In an optional implementation, the initial hybrid model includes three convolutional layers, one bidirectional long short-term memory network layer, two max pooling layers, two fully connected layers, and a dropout layer.

[0153] In an optional implementation, in the three convolutional layers, the first convolutional layer uses 16 3*1 filters, the second convolutional layer uses 16 3*16 filters, and the third convolutional layer uses 16 3*16 filters; the activation function of each convolutional layer is ReLU, and the model weights are initialized using the GlorotUniform initialization method.

[0154] In an optional implementation, the bidirectional long short-term memory network layer contains 32 units and is capable of simultaneously capturing both forward and reverse information of the sequence.

[0155] In an optional implementation, after predicting the future path of a dust storm using the target hybrid model, the method further includes:

[0156] Obtain several mobile path prediction results, and count the number of true positive results, true negative results, false positive results, and false negative results in the mobile path prediction results;

[0157] The overall accuracy, F1 score, and Kappa coefficient are determined based on the number of true positive results, true negative results, false positive results, and false negative results in the predicted movement path.

[0158] The performance of the target mixture model is evaluated based on the overall accuracy, F1 score, and Kappa coefficient.

[0159] Example 3

[0160] Based on the same application concept, see [link / reference] Figure 6 As shown, Figure 6 A schematic diagram of the structure of a computer device provided in Embodiment 3 of the present invention is shown, wherein, as Figure 6 As shown, the computer device 600 provided in Embodiment 3 of this application includes:

[0161] The computer device 600 includes a processor 601, a memory 602, and a bus 603. The memory 602 stores machine-readable instructions that can be executed by the processor 601. When the computer device 600 is running, the processor 601 communicates with the memory 602 through the bus 603. When the machine-readable instructions are executed by the processor 601, the steps of the sandstorm movement path prediction method shown in Embodiment 1 are performed.

[0162] Example 4

[0163] Based on the same concept, this application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the steps of the dust storm movement path prediction method described in any of the above embodiments.

[0164] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the system and apparatus described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0165] The computer program product for predicting the movement path of sandstorms provided in this embodiment of the invention includes a computer-readable storage medium storing program code. The instructions included in the program code can be used to execute the methods described in the preceding method embodiments. For specific implementation details, please refer to the method embodiments, which will not be repeated here.

[0166] The sandstorm movement path prediction device provided in this embodiment of the invention can be specific hardware on a device or software or firmware installed on the device. The implementation principle and technical effects of the device provided in this embodiment of the invention are the same as those in the foregoing method embodiments. For the sake of brevity, any parts not mentioned in the device embodiments can be referred to the corresponding content in the foregoing method embodiments. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can all be referred to the corresponding processes in the above method embodiments, and will not be repeated here.

[0167] In the embodiments provided by this invention, it should be understood that the disclosed apparatus and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the coupling or direct coupling or communication connection shown or discussed may be through some communication interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0168] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0169] In addition, the functional units in the embodiments provided by the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0170] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0171] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. In addition, the terms "first", "second", "third", etc. are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0172] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention. All should be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method of predicting a moving path of a sandstorm, characterized by, The method includes: Obtain aerosol optical thickness dataset and geographic background information for dust storm events; Based on the aerosol optical thickness dataset, a sequence of dust storm images with binary labels is generated, and key features are selected from the geographic background information using the random forest feature importance method. The input dataset is obtained by concatenating the sandstorm image sequence with the key features. An initial mixture model is constructed for predicting the movement path of a sandstorm, and a target mixture model is obtained by training the initial mixture model using the input dataset; The target hybrid model is used to predict the future movement path of dust storms; The initial hybrid model includes three convolutional layers, one bidirectional long short-term memory network layer, two max pooling layers, two fully connected layers, and a dropout layer. In the 3 convolutional layers, the first convolutional layer uses 16 3 x 3 filters, the second convolutional layer uses 16 3 x 3 filters, and the third convolutional layer uses 16 3 x 3 filters; the activation function of each convolutional layer is ReLU, and the model weight is initialized by GlorotUniform initialization method; The bidirectional long short-term memory network layer contains 32 units and can simultaneously capture both forward and reverse information of a sequence.

2. The method according to claim 1, characterized in that, The aerosol optical thickness dataset contains several aerosol optical thickness distribution maps. The generation of a binarized, labeled dust storm image sequence based on the aerosol optical thickness dataset includes: The Otsu thresholding algorithm was used to binarize the optical thickness distribution maps of aerosols to obtain several images to be labeled. Pixels in each image to be labeled that have a pixel value higher than a preset threshold are labeled as dusty areas, and pixels in the image to be labeled that have a pixel value lower than the preset threshold are labeled as non-dusty areas, thus obtaining several labeled images; The dust storm image sequence is constructed based on each labeled image.

3. The method according to claim 1, characterized in that, The process of selecting key features from the geographic background information using the random forest feature importance method includes: The average reduced impurities (MDI) score for each feature in the geographic background information was calculated using a random forest model. The key features are determined based on the MDI score ranking of each feature.

4. The method according to claim 1, characterized in that, After predicting the future path of a dust storm using the target fusion model, the method further includes: Obtain several mobile path prediction results, and count the number of true positive results, true negative results, false positive results, and false negative results in the mobile path prediction results; The overall accuracy, F1 score, and Kappa coefficient are determined based on the number of true positive results, true negative results, false positive results, and false negative results in the predicted movement path. The performance of the target mixture model is evaluated based on the overall accuracy, F1 score, and Kappa coefficient.

5. A dust storm movement path prediction device, characterized in that, The device includes: The data acquisition module is used to acquire aerosol optical thickness datasets and geographical background information for dust storm events; The data processing module is used to generate a sequence of binarized dust storm images based on the aerosol optical thickness dataset, and to filter out key features from the geographic background information using the random forest feature importance method. The dataset construction module is used to perform feature concatenation between the sandstorm image sequence and the key features to obtain the input dataset; The model training module is used to construct an initial hybrid model for predicting the movement path of sandstorms, and to train the initial hybrid model using the input dataset to obtain a target hybrid model. The path prediction module is used to predict the future movement path of the dust storm using the target hybrid model; The initial hybrid model includes three convolutional layers, one bidirectional long short-term memory network layer, two max pooling layers, two fully connected layers, and a dropout layer. Of the three convolutional layers, the first convolutional layer uses 16 3x3 convolutions. The filter is 1, and the second convolutional layer uses 16 3 filters. The third convolutional layer uses 16 filters and 16 3-bit filters. A 16-bit filter; the activation function of each convolutional layer is ReLU, and the model weights are initialized using the GlorotUniform initialization method; The bidirectional long short-term memory network layer contains 32 units and can simultaneously capture both forward and reverse information of a sequence.

6. A computer device, characterized in that, include: The computer device includes a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the computer device is running, the processor communicates with the memory via the bus. When the machine-readable instructions are executed by the processor, they perform the steps of the dust storm movement path prediction method as described in any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, performs the steps of the dust storm movement path prediction method as described in any one of claims 1 to 4.

Images