A method for predicting dust storm trajectories based on SSA-EDLSTM fusion network

Abstract

This invention discloses a method for predicting dust storm trajectories based on an SSA-EDLSTM fusion network, relating to the field of dust storm trajectory prediction. The method includes the following steps: acquiring historical wind field data, topographic elevation data, and historical aerosol optical thickness data of the target area; calculating slope and aspect based on the topographic elevation data, and constructing topographic and wind field interaction features with wind speed and wind direction respectively; inputting the topographic and wind field interaction features and historical wind field data into an SSA-MSED network to predict future wind field data; inputting future wind field data and historical aerosol optical thickness data into an ED-ConvLSTM network to generate future aerosol trajectories; and using a dust storm area detection algorithm to detect dust storm areas, generating dust storm trajectories and a visualization report. This method integrates SSA-MSED and ED-ConvLSTM networks, accurately capturing the influence of topography and the temporal characteristics of dust storms, significantly improving the long-term prediction accuracy and regional detection accuracy of dust storm trajectories.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of sandstorm trajectory prediction, and more particularly to a sandstorm trajectory prediction method based on a SSA-EDLSTM fusion network. BACKGROUND

[0002] Sandstorms, as a type of severe natural disaster, have a significant impact on the environment, human health, and social and economic development. Accurate prediction of sandstorm trajectories is crucial for reducing the damage caused by sandstorms. In recent years, with the development of deep learning technology, especially the application of convolutional neural networks (CNN) and long short-term memory networks (LSTM) in spatiotemporal data processing, the accuracy of sandstorm prediction has been significantly improved.

[0003] However, existing sandstorm trajectory prediction methods based on image processing have obvious shortcomings in handling complex terrain conditions. Traditional models often fail to fully consider the key physical process that topography factors cause the sandstorm trajectory to deviate by changing the local wind field structure, resulting in deviations between the predicted results and the actual situation. In addition, due to the lack of effective modeling of the interaction mechanism between topography, wind field, and sand transport, these models are limited in their application in complex terrain areas. Especially in mountainous or valley areas with complex topographic changes, the changes in wind speed and direction are difficult to accurately capture, further affecting the accuracy of sandstorm trajectory prediction.

[0004] On the other hand, existing deep learning models such as ConvLSTM, while effectively simulating the evolution trend of sandstorms in the short term, perform poorly in long-time step prediction. The error accumulation problem leads to a rapid decline in prediction accuracy as the prediction time step increases. For example, when predicting several hours or even longer periods in the future, the model's determination coefficient R 2 will decrease from a higher level at the beginning to a lower value, indicating its insufficient long-term dependency modeling capability. In addition, other networks such as 3D-CNN and PredRNN also face similar problems, and their error propagation and information forgetting phenomena are particularly prominent when processing long time series data.

[0005] At the same time, current methods based on remote sensing image segmentation of sandstorm areas also face a series of challenges. The problem that weak sand or sandy land surface reflectivity is too close to sandstorm reflectivity will cause confusion in the segmented area, affecting the accuracy of segmentation. Moreover, remote sensing image processing requires high-performance computing resources, and processing large-scale datasets is particularly time-consuming and laborious. Coupled with the influence of meteorological conditions such as cloud cover and precipitation, the resolution and quality of remote sensing images are difficult to meet the needs of practical applications.

[0006] Therefore, how to design a sandstorm movement trajectory prediction method based on a SSA-EDLSTM fusion network can fully consider the complex interaction mechanism among the terrain, wind field and sandstorm, and realize high-precision prediction of the sandstorm movement trajectory under complex terrain conditions is a problem that those skilled in the art need to solve. SUMMARY

[0007] Therefore, the present application provides a sandstorm movement trajectory prediction method based on a SSA-EDLSTM fusion network, which accurately captures the terrain influence and sandstorm time series characteristics by fusing SSA-MSED and ED-ConvLSTM networks, significantly improves the long-term prediction accuracy and regional detection accuracy of the sandstorm movement trajectory, and provides effective technical support for sandstorm early warning and disaster prevention and mitigation decision-making under complex terrain.

[0008] In order to achieve the above purpose, the present application adopts the following technical scheme:

[0009] A sandstorm movement trajectory prediction method based on a SSA-EDLSTM fusion network comprises the following steps:

[0010] S1, obtaining historical wind field data, terrain elevation data and historical aerosol optical thickness data of a target area;

[0011] S2, calculating the slope and slope direction based on the terrain elevation data, and constructing the terrain and wind field interaction features with the wind speed and wind direction, respectively;

[0012] S3, inputting the terrain and wind field interaction features and the historical wind field data into the SSA-MSED network to predict future wind field data;

[0013] S4, inputting the future wind field data and the historical aerosol optical thickness data into the ED-ConvLSTM network to generate future aerosol movement trajectory;

[0014] S5, based on the future aerosol movement trajectory, using a sandstorm region detection algorithm to detect the sandstorm region, and generating a sandstorm movement trajectory and a visual report.

[0015] Preferably, in the S2, the terrain and wind field interaction features are constructed by:

[0016] defining the slope terrain slope and the interaction data Ts Wv of the wind speed wind velocity :

[0017] Ts Wv = terrain slope * wind velocity

[0018]

[0019] Define slope aspect of terrain aspect With wind direction direction Interaction data Ta Wd :

[0020] Ta Wd =|wind direction -terrain aspect |

[0021]

[0022] Among them, terrain h For the sea slope height of the target area, wind U With wind V These represent the east-west and north-south components of the wind field in the target area, respectively.

[0023] Preferably, the SSA-MSED network includes an encoder and a decoder; wherein,

[0024] The encoder is composed of N series-connected modules, each module containing:

[0025] Convolutional layers are used to perform convolution operations through multiple convolutional kernels to extract spatial features at different scales.

[0026] The spatial self-attention mechanism layer is used to calculate the correlation between different locations in the input feature map and capture the long-distance dependence of key terrain features on the wind field.

[0027] The ConvLSTM layer is used to build a time-dynamic model of the wind field and capture the trend of wind field evolution over time.

[0028] The decoder is symmetrical to the encoder structure, receives the encoder's hidden state, and outputs the predicted wind field.

[0029] Preferably, the computation process of the spatial self-attention mechanism layer includes:

[0030] The input feature map is projected using convolution weights WQ, WK, and WV to generate a query matrix Q, a key matrix K, and a value matrix V.

[0031] Based on the query matrix Q and the key matrix K, calculate the attention weight matrix A:

[0032] A = softmax(QK) T )

[0033] Weighted and residual concatenated output Y from value matrix V:

[0034] Y = γ(A·V) + X

[0035] Where γ represents a learnable parameter.

[0036] Preferably, the ConvLSTM layer is represented as follows:

[0037] ConvLSTM(X t ,h t-1 ,c t-1 )=(h t ,c t )

[0038] Among them, X t h represents the input data at time t. t-1 and c t-1 h represents the hidden state and the cell state at the previous time step, respectively. t and c t These represent the hidden state and the cell state at the current moment, respectively.

[0039] Preferably, the ED-ConvLSTM network adopts a cascaded encoder-decoder structure, including M cascaded modules, each module containing an encoder and a decoder;

[0040] The encoder output of the m-th module is:

[0041]

[0042] Among them, X t Indicates input data, These represent the hidden state and cell state of the m-th module encoder at time t, respectively. Let represent the hidden state and cell state of the m-th module encoder at time t-1, respectively.

[0043] Preferably, in the ED-ConvLSTM network,

[0044] The hidden state of the m-th module encoder at time t Passed to the decoder in the same module:

[0045]

[0046] Cell state of the m-th module encoder at time t The encoder passed to the (m+1)th module:

[0047]

[0048] in, Let these represent the hidden state and cell state of the (m+1)th module encoder at time t, respectively. This represents the hidden state of the (m+1)th module encoder at time t-1.

[0049] Preferably, in step S5, detecting a sandstorm area using a sandstorm area detection algorithm includes:

[0050] Calculate the instantaneous rate of change of AOD, Rate(t):

[0051]

[0052] Where Δt is the time interval, C(t) is the aerosol concentration at time t, and C(t-1) is the aerosol concentration at the previous moment.

[0053] Define a rate of change threshold λ. When Rate(t) > the rate of change threshold λ, it is marked as a dust storm area.

[0054] Preferably, the defined rate of change threshold λ includes:

[0055] Time series data of aerosol concentration from typical dust storm events were selected and preprocessed.

[0056] Calculate the rate of change of aerosol concentration during each dust storm event, statistically analyze its distribution characteristics, and select the top 5% quantiles of the distribution as the initial threshold.

[0057] By combining cross-validation and adjusting the threshold through multiple experiments, its performance in different sandstorm events was evaluated, and the optimal threshold was finally determined.

[0058] Preferably, in step S5, the visualization report includes: a spatiotemporal comparison map of the predicted trajectory and the actual trajectory, an error heat map, and a coverage map of the dynamic detection results of the sandstorm area.

[0059] As can be seen from the above technical solutions, compared with the prior art, the technical solutions of the present invention have the following advantages:

[0060] Beneficial effects:

[0061] 1. This method significantly enhances the model's ability to model wind field changes under complex terrain conditions by introducing the Spatial Self-Attention (SSA) mechanism into the Multi-Scale Encoder-Decoder Network (SSA-MSED). The terrain slope and aspect are constructed through the interaction features with wind speed and wind direction, enabling the model to accurately capture the local influence of terrain on the wind field, thereby achieving high-precision prediction of the wind field.

[0062] 2. The ED-ConvLSTM network employs a cascaded encoder-decoder structure, enabling in-depth mining of the spatiotemporal features in dust storm time series and, combined with predicted wind field data, achieving long-term predictions of dust storm trajectories. Through the design of multiple cascaded modules, the model maintains information continuity, effectively captures long-term temporal dependencies, and improves the stability and accuracy of long-term predictions.

[0063] 3. The SSA-EDLSTM fusion network combines the advantages of the SSA-MSED and ED-ConvLSTM networks, considering both the influence of terrain on the wind field and utilizing historical dust storm evolution data for temporal feature modeling. This dual-network collaborative architecture fully leverages the high-precision wind field feature extraction capability of SSA-MSED and the powerful temporal data modeling capability of ED-ConvLSTM, thereby achieving more accurate predictions of dust storm trajectories, especially demonstrating excellent prediction performance over long time spans.

[0064] 4. The dust storm area detection algorithm in this method can dynamically respond to instantaneous changes in concentration, accurately capturing the sharp fluctuations in aerosol concentration during dust storms, and significantly improving the detection accuracy of dust storm areas. By statistically analyzing historical dust storm event data to determine the threshold and combining it with cross-validation optimization, the algorithm exhibits excellent detection performance under different meteorological conditions, providing more accurate spatial distribution information for dust storm early warning. Attached Figure Description

[0065] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0066] Figure 1 A flowchart of a dust storm trajectory prediction method based on an SSA-EDLSTM fusion network is provided for an embodiment of the present invention.

[0067] Figure 2 This is a schematic diagram of the SSA-MSED network structure provided in an embodiment of the present invention;

[0068] Figure 3 This is a schematic diagram of the ED-ConvLSTM network structure provided in an embodiment of the present invention. Detailed Implementation

[0069] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0070] like Figure 1 As shown, this embodiment provides a method for predicting the trajectory of sandstorms based on an SSA-EDLSTM fusion network. The SSA-EDLSTM fusion network includes an SSA-MSED network and an ED-ConvLSTM network, and includes the following steps:

[0071] S1. Obtain historical wind field data, topographic elevation data, and historical aerosol optical thickness data for the target area;

[0072] S2. Calculate the slope and aspect based on the terrain elevation data, and construct the interaction features between terrain and wind field with wind speed and wind direction respectively;

[0073] S3. Input the interaction features between terrain and wind field and historical wind field data into the SSA-MSED network to predict future wind field data;

[0074] S4. Input the future wind field data and historical aerosol optical thickness data into the ED-ConvLSTM network to generate the future aerosol motion trajectory.

[0075] S5. Based on the future aerosol movement trajectory, use the dust storm area detection algorithm to detect the dust storm area and generate the dust storm movement trajectory and visualization report.

[0076] This method achieves accurate prediction of dust storm trajectories under complex terrain conditions by fusing a spatial self-attention mechanism (SSA) and a multi-scale encoder-decoder network (SSA-MSED). It not only accurately captures the impact of terrain on wind fields, enabling high-precision wind field prediction, but also, by incorporating an ED-ConvLSTM network, deeply mines the spatiotemporal characteristics of dust storm time series for long-term prediction. Simultaneously, the innovative dust storm region detection algorithm improves the detection accuracy of dust storm regions, providing more precise spatial distribution information for dust storm early warning and more reliable data support for disaster prevention and mitigation decisions.

[0077] The following provides a further detailed explanation of each step in the above method;

[0078] In this embodiment S1, historical wind field data, topographic elevation data, and historical aerosol optical thickness data of the target area are acquired.

[0079] This step is the foundational data collection phase for predicting dust storm trajectories. First, historical wind field data for the target area needs to be acquired. This data typically includes key meteorological elements such as wind speed and direction, which directly determine the propagation path and speed of the dust storm. Historical wind field data can be obtained through various means, including meteorological observation stations, satellite remote sensing, or reanalysis datasets (such as MERRA-2), ensuring that the spatiotemporal coverage and accuracy of the data meet prediction requirements.

[0080] Meanwhile, topographic elevation data is also indispensable, as changes in terrain significantly affect local wind field structure, thus influencing the trajectory of dust storms. Topographic elevation data can be obtained through digital elevation models (DEMs), which provide detailed height information of the target area's surface. Furthermore, historical aerosol optical depth (AOD) data is another key input. AOD data reflects the concentration distribution of dust particles in the atmosphere and is crucial for identifying dust storm regions and tracking their trajectories. This historical AOD data can also be acquired through satellite remote sensing or ground-based observation stations, providing necessary spatiotemporal characteristics for subsequent dust storm trajectory prediction.

[0081] In this embodiment S2, the slope and aspect are calculated based on the terrain elevation data, and the terrain and wind field interaction features are constructed with wind speed and wind direction respectively.

[0082] The construction of terrain and wind field interaction features includes:

[0083] Define slope terrain slope With wind speed velocity Interactive data Ts Wv :

[0084] Ts Wv =terrain slope ×wind velocity

[0085]

[0086] Define slope aspect of terrain aspect With wind direction direction Interaction data Ta Wd :

[0087] Ta Wd =|wind direction -terrain aspect |

[0088]

[0089] Among them, terrain h For the sea slope height of the target area, wind UWith wind V These represent the east-west and north-south components of the wind field in the target area, respectively.

[0090] This step optimizes the simulation capability of local wind fields by introducing the interactive features of slope-wind speed and slope-wind direction, thereby improving the accuracy of dust storm trajectory prediction.

[0091] In this embodiment S3, the interaction features between terrain and wind field and historical wind field data are input into the SSA-MSED network to predict future wind field data.

[0092] In regions with complex topography, wind fields are highly variable, primarily manifested in significant fluctuations in wind speed and direction. These variations are influenced by a combination of factors, including topographic relief, mountain obstruction, and valley ventilation effects. Therefore, to accurately predict future wind fields, deep learning models not only need powerful spatiotemporal modeling capabilities but also must delve into the characteristics of topography's influence on wind fields to effectively characterize their complex nonlinear relationships.

[0093] like Figure 2 As shown, the SSA-MSED network proposed in this embodiment includes an encoder and a decoder;

[0094] The encoder consists of multiple modules, each containing a convolutional layer, a spatial self-attention layer, and a ConvLSTM layer. Assuming there are N such stages, the input sequence contains S frames (temporal length), with each frame having a size of (B, C, H, W), where B represents the batch size, C is the number of channels, and H and W are the image height and width, respectively. In the encoder, the input sequence is sequentially passed through the convolutional layers of each stage. The output of these convolutional layers is then passed to the ConvLSTM layer of each stage for temporal modeling.

[0095] Each module contains, in sequence: a convolutional layer, used to perform convolution operations through multiple convolutional kernels to extract spatial features at different scales; a spatial self-attention mechanism layer, used to calculate the correlation between different locations in the input feature map and capture the long-distance dependence of key terrain features on the wind field; and a ConvLSTM layer, used to establish a temporal dynamic model of the wind field and capture the trend of wind field evolution over time.

[0096] The decoder is symmetrical to the encoder structure, receives the encoder's hidden state, and outputs the predicted wind field.

[0097] Furthermore, the computational process of the spatial self-attention mechanism layer includes:

[0098] The input feature map is projected using convolution weights WQ, WK, and WV to generate a query matrix Q, a key matrix K, and a value matrix V.

[0099] Based on the query matrix Q and the key matrix K, calculate the attention weight matrix A:

[0100] A = softmax(QK) T )

[0101] Weighted and residual concatenated output Y from value matrix V:

[0102] Y = γ(A·V) + X

[0103] Where γ represents a learnable parameter.

[0104] Furthermore, the ConvLSTM layer is represented as:

[0105] ConvLSTM(X t ,h t-1 ,c t-1 )=(h t ,c t )

[0106] Among them, X t h represents the input data at time t. t-1 and c t-1 h represents the hidden state and the cell state at the previous time step, respectively. t and c t These represent the hidden state and the cell state at the current moment, respectively.

[0107] ConvLSTM combines the spatial feature extraction capabilities of CNNs with the time series modeling capabilities of LSTMs, enabling it to better capture the evolutionary trends of wind fields over time. Through ConvLSTM computation at multiple time steps, it can effectively extract the evolutionary patterns of wind fields, enhancing the predictive ability for changes in wind speed and direction.

[0108] The input for the i-th stage first passes through a convolutional layer, and then the convolutional output is fed into a ConvLSTM layer for processing via a spatial self-attention mechanism module, as shown in the following formula:

[0109] ConvLSTM(X t ,h t-1 ,c t-1 )=(h t ,c t )

[0110] Among them, X t It is the input time series data, representing the input at time t, h t-1 and c t-1 These are the hidden state and the unit state of the previous moment, h. t and c tThese represent the hidden state and the unit state at the current time step, respectively. In each stage, the convolutional layer extracts spatial features, while ConvLSTM handles temporal relationships. For the entire encoder, assuming there are N stages, its output hidden state is:

[0111] H encoder =(h1,h2,...,h N )

[0112] Furthermore, the decoder works similarly to the encoder, but it starts from the encoder's hidden states and progressively decodes them back to the original output sequence. The decoder operates by starting from the state of the last stage and progressively passing information forward to generate predictions. The decoder's main task is to recover the spatial information of the original data from the encoder's feature representations and generate the predicted results, which are the future wind field data.

[0113] The output of each stage of the decoder is:

[0114] H decoder =(h′1,h′2,...,h′) N )

[0115] The entire input to the encoder-decoder network is a time-series data X = {X1, X2, ..., X}. S}, where each X t The input sequence is a grid of size (B, C, H, W). The encoder and decoder process this input sequence using their respective ConvLSTM and convolutional layers. The encoder process is as follows:

[0116] H encoder =Encoder(X,H) prev )

[0117] The decoder process is as follows:

[0118] H decoder =Decoder(H encoder )

[0119] The final output is the decoder's output, represented as:

[0120] Y = H decoder

[0121] Regarding data processing, the input data for training the model is mainly divided into dimensions (200, 18, 10, 4, 64, 64), where 200 indicates that the dataset contains 200 samples; 18 indicates that each sample contains 18 time-series data points; 10 indicates that each time-series data point contains data from 10 time steps; and 4 indicates that the wind field data at each time step contains 4 feature channels, including wind field data (U, V) and two interaction data (Ts). Wv Ta Wd The 64 indicates that each wind field data point has a spatial resolution of 64×64 grids, reflecting the spatial distribution of each time series. The output data is divided into dimensions (200, 18, 1, 2, 64, 64). Here, 1 represents the data for the next time step, and 2 represents the wind field (U, V) data for the next time step.

[0122] The SSA-MSED network in this step combines a spatial self-attention mechanism and an encoder-decoder structure to accurately capture the influence of terrain on wind fields and achieve high-precision prediction of wind fields in dust storm areas. The core of the SSA-MSED network enhances its ability to model the complex nonlinear relationship between terrain and wind fields through a spatial self-attention mechanism, while utilizing the encoder-decoder structure to effectively reduce dimensionality and extract features from high-dimensional data, thereby improving the accuracy and robustness of predictions.

[0123] In this embodiment S4, future wind field data and historical aerosol optical thickness data are input into the ED-ConvLSTM network to generate future aerosol motion trajectories.

[0124] In predicting the trajectory of dust storms, the extraction of time-series features is a key issue affecting prediction accuracy. While accurate wind field prediction is important, relying solely on wind field information is insufficient to fully characterize the movement patterns of dust storms. The formation and evolution of dust storms are influenced by multiple factors, among which the evolution patterns and spatiotemporal correlations in historical dust storm time series are crucial for predicting their future trajectories. Due to the highly nonlinear and complex spatiotemporal evolution of dust storms, traditional prediction methods struggle to effectively model their dynamic characteristics.

[0125] The proposed ED-ConvLSTM network in this step aims to delve deeper into the time-series features of dust storms and improve the accuracy of trajectory prediction. Its core is to utilize the powerful spatiotemporal feature extraction capability of the ConvLSTM network combined with the Encoder-Decoder structure, enabling the model to effectively capture the historical evolution patterns of dust storms and infer future trajectories based on these patterns.

[0126] like Figure 3As shown, the ED-ConvLSTM network employs a cascaded encoder-decoder structure, comprising M cascaded modules. Within each module, the hidden states and cell states in the encoder block are passed to the decoder block to support the entire prediction process. For long-term prediction tasks, this state information is not limited to transmission within a module but can be carried over across multiple modules. In this structure, the hidden states and cell states of the encoder block are passed from the encoder block of one module to the encoder block of the next module, thereby achieving cross-module temporal information transmission. This design enables the model to capture long-term temporal dependencies and maintain information continuity across multiple modules, further enhancing its ability to model long-step sequences. For long-step prediction tasks, cross-module temporal information transmission is crucial.

[0127] Furthermore, each module contains an encoder and a decoder;

[0128] The encoder output of the m-th module is:

[0129]

[0130] Among them, X t Indicates input data, These represent the hidden state and cell state of the m-th module encoder at time t, respectively. Let represent the hidden state and cell state of the m-th module encoder at time t-1, respectively.

[0131] Furthermore, the hidden state of the m-th module encoder at time t. Passed to the decoder in the same module:

[0132]

[0133] Cell state of the m-th module encoder at time t The encoder passed to the (m+1)th module:

[0134]

[0135] in, Let these represent the hidden state and cell state of the (m+1)th module encoder at time t, respectively. This represents the hidden state of the (m+1)th module encoder at time t-1;

[0136] Combining the encoder and decoder processes, the model's prediction result can be expressed as:

[0137]

[0138] The final Y tThis is the predicted AOD data for times {t+1,t+2,t+3,…,t+10}.

[0139] In this embodiment, the fusion network SSA-EDLSTM integrates the wind field data predicted by SSA-MSED. The dust storm data from the past 10 time steps are input into the ED-ConvLSTM network to predict the dust storm trajectory for the next 10 time steps. This fusion method can fully utilize the accurate wind field modeling capability of SSA-MSED and the temporal feature capture capability of ED-ConvLSTM to achieve more accurate prediction of dust storm trajectories, while effectively considering the influence of topographic factors on wind field and dust storm evolution.

[0140] The following provides further explanation of the process of generating the future dust storm trajectory:

[0141] There is a close relationship between aerosols and dust storms, with dust storms being a significant component of aerosols. Dust storms are primarily formed when strong winds lift and transport surface dust particles into the atmosphere, where these particles suspend and form aerosols. Therefore, the distribution and trajectory of aerosols can reflect the propagation path and impact range of dust storms. During a dust storm, a large number of dust particles are lifted into the atmosphere by strong winds, forming aerosols. These particles have a wide range of sizes, from a few micrometers to tens of micrometers, and can remain suspended in the atmosphere for extended periods and be transported long distances by airflow. The scattering and absorption characteristics of dust aerosols significantly affect atmospheric visibility and radiation balance; therefore, monitoring aerosol optical depth (AOD) can indirectly identify the presence and intensity of dust storms. To predict aerosol trajectories over long time steps, a sequence-to-sequence method is used, where T... i The AOD data at times {t-9, t-8, t-7, ..., t} are used as the input pair of the ED-ConvLSTM network. o Predict AOD data at times {t+1,t+2,t+3,…,t+10}.

[0142] In dust storm trajectory prediction based on ED-ConvLSTM, only the influence of wind field is considered, while the potential effect of topographic features on wind field is ignored. However, the undulation of topography has a significant impact on the structure and direction of wind field, especially in complex terrain areas such as mountains, canyons, and coastlines, where wind field changes significantly due to topography. Therefore, incorporating topographic factors into the aerosol trajectory prediction model can more accurately reflect the dynamic changes of the actual wind field, thereby improving the accuracy of the prediction results. The main steps of the dust storm trajectory prediction method based on SSA-EDLSTM are as follows: First, data processing is performed. Aerosol optical thickness data is extracted from the MERRA-2 remote sensing dataset, and all data are processed according to time series. In the fusion network SSA-EDLSTM, the input data of the SSA-MSED network are wind field data and topographic data from time t-9 to time t, and the wind field data at time t+1 is predicted. The input data of the ED-ConvLSTM network are wind field data and aerosol optical thickness data from time t-9 to time t, and the wind field data at time t+1 is predicted. The outputs of the two networks are combined and then input into the ED-ConvLSTM network to obtain the aerosol optical thickness data at time t+2. This process is repeated to obtain the aerosol optical thickness data from time t+1 to t+10.

[0143] In this embodiment S5, based on the future aerosol movement trajectory, a dust storm area detection algorithm is used to detect the dust storm area and generate a dust storm movement trajectory and visualization report.

[0144] Among them, the dust storm area detection algorithm includes the following:

[0145] Calculate the instantaneous rate of change of AOD, Rate(t):

[0146]

[0147] Δt is the time interval, C(t) is the aerosol concentration at time t, and C(t-1) is the aerosol concentration at the previous time.

[0148] Define a rate of change threshold λ. When Rate(t) > the rate of change threshold λ, it is marked as a dust storm area.

[0149] Furthermore, the definition of the rate of change threshold λ includes:

[0150] Time series data of aerosol concentration from typical dust storm events were selected and preprocessed.

[0151] Calculate the rate of change of aerosol concentration during each dust storm event, statistically analyze its distribution characteristics, and select the top 5% quantiles of the distribution as the initial threshold.

[0152] By combining cross-validation and adjusting the threshold through multiple experiments, its performance in different sandstorm events was evaluated, and the optimal threshold was finally determined.

[0153] Furthermore, the visualization report includes: a spatiotemporal comparison map of the predicted trajectory and the actual trajectory, an error heat map, and a coverage map of the dynamic detection results of the dust storm area.

[0154] Among them, the spatiotemporal comparison map of predicted trajectory and actual trajectory can intuitively show the difference between the model prediction results and the actual situation, which helps to evaluate the prediction accuracy of the model; the error heat map reflects the magnitude of prediction error in different areas through color depth, helping researchers to quickly locate areas with large prediction deviations; the dust storm area dynamic detection result coverage map combines the output of the dust storm detection algorithm to show the spatiotemporal distribution of the dust storm area.

[0155] This embodiment proposes a dust storm trajectory prediction method based on an SSA-EDLSTM fusion network. By combining the advantages of Spatial Self-Attention (SSA), Multi-Scale Encoding / Decoding Network (SSA-MSED), and ED-ConvLSTM network, it achieves high-precision prediction of wind fields and aerosol trajectories under complex terrain conditions. This method not only accurately captures the impact of terrain on wind fields but also improves the accuracy of long-term predictions by deeply mining the spatiotemporal characteristics of dust storm time series. Simultaneously, an innovative dust storm region detection algorithm further enhances the accuracy of dust storm region identification, and combined with visualization reports, provides intuitive and reliable data support for disaster prevention and mitigation decision-making.

[0156] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0157] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A sandstorm movement trajectory prediction method based on an SSA-EDLSTM fusion network, characterized in that, The method comprises the following steps: S1, obtaining historical wind field data, terrain elevation data and historical aerosol optical depth data of a target area; S2, calculating slope and aspect based on the terrain elevation data, and constructing terrain and wind field interaction features with wind speed and direction respectively; S3, inputting the terrain and wind field interaction features and the historical wind field data into the SSA-MSED network to predict future wind field data; the SSA-MSED network comprises an encoder and a decoder; the encoder is composed of N modules connected in series, and each module comprises, in sequence, a convolution layer for performing convolution operation through multiple convolution kernels to extract spatial features of different scales, a spatial self-attention mechanism layer for calculating the correlation of different positions in the input feature map to capture the long-distance dependence of key terrain features on the wind field, and a ConvLSTM layer for establishing a time dynamic model of the wind field to capture the trend of the wind field evolution over time; the decoder is symmetrical to the structure of the encoder, receives the hidden state of the encoder and outputs the predicted wind field; S4, inputting the future wind field data and the historical aerosol optical depth data into the ED-ConvLSTM network to generate future aerosol movement trajectories; the ED-ConvLSTM network adopts a cascaded encoding and decoding structure and comprises M cascaded modules, each module comprising an encoder and a decoder; the output of the encoder of the mth module is: wherein X t represents input data, respectively represent the hidden state, the cell state of the mth module encoder at time t, respectively represent the hidden state, the cell state of the mth module encoder at time t-1. In the ED-ConvLSTM network, the hidden state of the mth module encoder at time t is passed to the same module decoder: Cell state of the mth module encoder at time t Encoder passed to the m+1th module wherein, respectively denote the hidden state and the cell state of the m+1th module encoder at time t, denotes the hidden state of the m+1th module encoder at time t-1; S5, detecting a sandstorm area based on the future aerosol movement trajectories by using a sandstorm area detection algorithm to generate sandstorm movement trajectories and a visual report.

2. The sandstorm movement trajectory prediction method based on the SSA-EDLSTM fusion network according to claim 1, characterized in that, In S2, the construction of the terrain and wind field interaction features comprises: Definition of the slope terrain slope Interaction with the wind speed wind velocity The interaction data Ts Wv : Ts Wv = terrain slope x wind velocity Definition of slope orientation terrain aspect Interaction with wind direction wind direction Interaction data Ta Wd : Ta Wd = |wind direction -terrain aspect | where terrain h is the height of the sea floor in the target area, wind U is the wind speed, and wind V are the east and north components of the wind field in the target area, respectively.

3. The sandstorm movement trajectory prediction method based on the SSA-EDLSTM fusion network according to claim 1, characterized in that, The calculation process of the spatial self-attention mechanism layer comprises: Projecting the input feature map through convolution weights WQ, WK and WV to generate a query matrix Q, a key matrix K and a value matrix V; Calculating an attention weight matrix A based on the query matrix Q and the key matrix K: A = softmax(QK T ) Weighting the value matrix V and performing residual connection to output Y: Y = γ (A·V) + X Wherein, γ represents a learnable parameter.

4. The sandstorm movement trajectory prediction method based on the SSA-EDLSTM fusion network according to claim 1, characterized in that, The ConvLSTM layer is represented as: ConvLSTM(X t ,h t-1 ,c t-1 ) = (h t ,c t ) wherein X t represents the input data at the tth moment, h t-1 and c t-1 respectively represent the hidden state and the cell state at the previous moment, h t and c t respectively represent the hidden state and the cell state at the current moment.

5. The sandstorm movement trajectory prediction method based on the SSA-EDLSTM fusion network according to claim 1, characterized in that, In S5, the sandstorm area detection algorithm comprises: Calculating the AOD instantaneous rate of change Rate(t): Wherein, Δt is the time interval, C(t) is the aerosol concentration at time t, and C(t-1) is the aerosol concentration at the previous moment; Defining a change rate threshold λ, and marking as a sandstorm area when Rate(t) > the change rate threshold λ.

6. The sandstorm movement trajectory prediction method based on the SSA-EDLSTM fusion network according to claim 5, characterized in that, The definition of the change rate threshold λ comprises: Selecting aerosol concentration time series data of typical sandstorm events and performing data preprocessing; Calculating the change rate of the aerosol concentration during each sandstorm event, and selecting the top 5% quantile of the distribution as the initial threshold; Combining cross-validation, adjusting the threshold through multiple experiments, evaluating its performance in different sandstorm events, and finally determining the optimal threshold.

7. The sandstorm movement trajectory prediction method based on the SSA-EDLSTM fusion network according to claim 1, characterized in that, In S5, the visual report comprises a spatiotemporal comparison chart of the predicted trajectory and the real trajectory, an error heat map, and a sandstorm area dynamic detection result overlay chart.

Images