Sand storm motion trail prediction method based on SSA-EDLSTM fusion network

Abstract

The invention discloses a sand storm motion trail prediction method based on an SSA-EDLSTM fusion network, and relates to the field of sand storm motion trail prediction, and the method comprises the following steps: obtaining historical wind field data, terrain elevation data and historical aerosol optical thickness data of a target region; calculating a gradient and a slope direction based on terrain elevation data, and respectively constructing terrain and wind field interaction characteristics with a wind speed and a wind direction; inputting the terrain and wind field interaction characteristics and historical wind field data into an SSA-MSED network, and predicting future wind field data; inputting the future wind field data and the historical aerosol optical thickness data into the ED-ConvLSTM network to generate a future aerosol movement track; and carrying out sandstorm region detection by using a sandstorm region detection algorithm, and generating a sandstorm movement track and a visual report. According to the method, SSA-MSED and ED-ConvLSTM networks are fused, the terrain influence and sand storm time sequence characteristics are accurately captured, and the long-term prediction precision and the region detection accuracy of the sand storm motion trail are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of sandstorm trajectory prediction, and more specifically to a sandstorm trajectory prediction method based on an SSA-EDLSTM fusion network. Background Art

[0002] Dust storms, as a serious natural disaster, have significant impacts on the environment, human health, and socioeconomic development. Accurately predicting the trajectory of sandstorms is crucial for minimizing the damage they cause. In recent years, the advancement of deep learning technology, particularly the application of convolutional neural networks (CNNs) and long short-term memory networks (LSTMs) to spatiotemporal data processing, has significantly improved the accuracy of sandstorm forecasts.

[0003] However, existing image processing-based sandstorm trajectory prediction methods have significant shortcomings when dealing with complex terrain conditions. Traditional models often fail to fully account for the critical physical process by which terrain factors shift the sandstorm trajectory by altering the local wind field structure, resulting in deviations between predictions and actual conditions. Furthermore, the lack of effective modeling of the interaction between terrain, wind field, and sand and dust transport limits the application of these models in complex terrain areas. This is especially true in areas with complex topography, such as mountainous areas or river valleys, where variations in wind speed and direction are difficult to accurately capture, further impacting the accuracy of sandstorm trajectories.

[0004] On the other hand, although existing deep learning models such as ConvLSTM can effectively simulate the evolution trend of sandstorms in the short term, they perform poorly in long-term predictions. The error accumulation problem causes the prediction accuracy to drop rapidly as the prediction time step increases. For example, when making predictions for the next few hours or even longer time periods, the model's coefficient of determination R 2 The initial high level will drop to a lower value, indicating that its ability to model long-term dependencies is insufficient. In addition, other networks such as 3D-CNN and PredRNN also face similar problems. They are particularly prone to error propagation and information forgetting when processing long time series data.

[0005] At the same time, current methods for segmenting sandstorm areas based on remote sensing imagery face a number of challenges. The reflectivity of surfaces like thin sand or sandy ground that are too close to that of a sandstorm can cause confusion in the segmented areas, affecting segmentation accuracy. Furthermore, remote sensing image processing requires high-performance computing resources, and processing large datasets is particularly time-consuming and labor-intensive. Furthermore, meteorological conditions such as cloud cover and precipitation can hinder the resolution and quality of remote sensing imagery from meeting the requirements of practical applications.

[0006] Therefore, how to design a sandstorm trajectory prediction method based on the SSA-EDLSTM fusion network that can fully consider the complex interaction mechanism between terrain, wind field and sandstorms, and achieve high-precision prediction of sandstorm trajectories under complex terrain conditions is an urgent problem that technicians in this field need to solve. Summary of the Invention

[0007] In view of this, the present invention provides a sandstorm motion trajectory prediction method based on the SSA-EDLSTM fusion network. By fusing the SSA-MSED and ED-ConvLSTM networks, it accurately captures the terrain influence and sandstorm temporal characteristics, significantly improves the long-term prediction accuracy of sandstorm motion trajectories and the regional detection accuracy, and provides effective technical support for sandstorm early warning and disaster prevention and mitigation decision-making in complex terrain.

[0008] In order to achieve the above object, the present invention adopts the following technical solutions:

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

[0010] S1. Obtain historical wind field data, terrain elevation data, and historical aerosol optical depth data for the target area;

[0011] S2. Calculate the slope and slope direction based on the terrain elevation data, and construct the terrain and wind field interaction characteristics with the wind speed and wind direction respectively;

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

[0013] S4. Input the future wind field data and historical aerosol optical depth data into the ED-ConvLSTM network to generate the future aerosol motion trajectory;

[0014] S5. Based on the future aerosol motion trajectory, a sandstorm area detection algorithm is used to detect the sandstorm area, and a sandstorm motion trajectory and a visualization report are generated.

[0015] Preferably, in S2, constructing the terrain and wind field interaction characteristics includes:

[0016] Define slope terrain slope and wind speed velocity Interaction data Ts Wv :

[0017] Ts Wv =terrain slope ×wind velocity

[0018]

[0019] Define aspect terrain aspect and wind direction direction Interaction data Ta Wd :

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

[0021]

[0022] Among them, terrain h is the sea slope height of the target area, wind U with wind V They are the east-west component and north-south component 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-level modules connected in series, and each level of modules sequentially includes:

[0025] Convolutional layer, used to perform convolution operations through multiple convolution kernels to extract spatial features of different scales;

[0026] A spatial self-attention mechanism layer is used to calculate the correlation between different positions in the input feature map and capture the long-range dependency of key terrain features on the wind field;

[0027] The ConvLSTM layer is used to build a temporal dynamic model of the wind farm and capture the temporal evolution trend of the wind farm.

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

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

[0030] Project the input feature map through the convolution weights WQ, WK, and WV to generate the query matrix Q, key matrix K, and 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] Weight the value matrix V and residually connect the output Y:

[0034] Y=γ(A·V)+X

[0035] Among them, γ represents a learnable parameter.

[0036] Preferably, the ConvLSTM layer is expressed as:

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

[0038] Among them, X t represents the input data at time t, h t-1 and c t-1 Represent the hidden state and unit state of the previous moment, h t and c t Represent the hidden state and unit state at the current moment respectively.

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

[0040] The encoder output of the mth module is:

[0041]

[0042] Among them, X t Represents input data, Represent the hidden state and cell state of the m-th module encoder at time t, They 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 mth module encoder at time t Passed to the decoder of the same module:

[0045]

[0046] The cell state of the mth module encoder at time t Passed to the encoder of the m+1th module:

[0047]

[0048] in, denote the hidden state and cell state of the m+1th module encoder at time t, respectively. represents the hidden state of the m+1th module encoder at time t-1.

[0049] Preferably, in S5, detecting the 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 change rate threshold λ, and mark the area as a sandstorm area when Rate(t)>change rate threshold λ.

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

[0055] Select the aerosol concentration time series data of typical sandstorm events and perform data preprocessing;

[0056] The rate of change of aerosol concentration during each sandstorm event was calculated, its distribution characteristics were statistically analyzed, and the top 5% quantile of the distribution was selected as the initial threshold;

[0057] Combined with cross-validation, the threshold was adjusted through multiple experiments to evaluate its performance in different sandstorm events, and finally the optimal threshold was determined.

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

[0059] It can be seen from the above technical solution that compared with the prior art, the technical solution of the present invention has 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 mechanism (SSA) into the multi-scale encoder-decoder network (SSA-MSED). The terrain slope and aspect are constructed by interacting with wind speed and direction, enabling the model to accurately capture the local impact of terrain on the wind field, thereby achieving high-precision prediction of the wind field.

[0062] 2. The ED-ConvLSTM network utilizes a cascaded encoder-decoder structure, enabling in-depth exploration of the spatiotemporal characteristics of sandstorm time series. Combined with predicted wind data, it enables long-term prediction of sandstorm trajectories. Through the design of multiple cascaded modules, the model maintains information continuity, effectively capturing long-term temporal dependencies and improving the stability and accuracy of long-step predictions.

[0063] 3. The SSA-EDLSTM fusion network combines the advantages of the SSA-MSED network and the ED-ConvLSTM network, taking into account the impact of terrain factors on wind fields and utilizing historical sandstorm evolution data for time series feature modeling. This dual-network collaborative architecture fully utilizes the high-precision wind field feature extraction capabilities of SSA-MSED in the network and the powerful ability of ED-ConvLSTM in time series data modeling, thereby achieving more accurate predictions of sandstorm movement trajectories, especially excellent prediction performance over long time spans.

[0064] 4. The sandstorm area detection algorithm in this method dynamically responds to instantaneous changes in concentration, accurately capturing the dramatic fluctuations in aerosol concentration during a sandstorm, significantly improving the accuracy of sandstorm area detection. By statistically analyzing historical sandstorm event data to determine the threshold and combining it with cross-validation optimization, the algorithm demonstrates excellent detection performance under diverse meteorological conditions, providing more accurate spatial distribution information for sandstorm warnings. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0066] Figure 1 A flow chart of a sandstorm trajectory prediction method based on an SSA-EDLSTM fusion network provided in an embodiment of the present invention;

[0067] Figure 2 Schematic diagram of the SSA-MSED network structure provided by an embodiment of the present invention;

[0068] Figure 3 Schematic diagram of the ED-ConvLSTM network structure provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0069] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

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

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

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

[0073] S3, input the terrain and wind field interaction characteristics 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 depth data into the ED-ConvLSTM network to generate the future aerosol motion trajectory;

[0075] S5. Based on the future aerosol motion trajectory, a sandstorm area detection algorithm is used to detect the sandstorm area, and a sandstorm motion trajectory and a visualization report are generated.

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

[0077] The following further describes each step in the above method in detail;

[0078] In this embodiment, S1, historical wind field data, terrain elevation data, and historical aerosol optical depth data of the target area are obtained;

[0079] This step is the basic data collection stage for sandstorm trajectory prediction. First, historical wind data for the target area must be obtained. This data typically includes key meteorological factors such as wind speed and direction, which directly determine the propagation path and speed of the sandstorm. Historical wind data can be obtained through various channels, such as meteorological observation stations, satellite remote sensing, or reanalysis datasets (such as MERRA-2), ensuring that the temporal and spatial coverage and accuracy of the data meet the prediction requirements.

[0080] At the same time, terrain elevation data is also indispensable. The undulation of terrain will significantly affect the local wind field structure, and thus affect the movement trajectory of sandstorms. Terrain elevation data can be obtained through digital elevation models (DEMs), which provide detailed height information of the surface of the target area. In addition, historical aerosol optical depth (AOD) data is also one of the key inputs. AOD data reflects the concentration distribution of sand and dust particles in the atmosphere and is an important basis for identifying sandstorm areas and tracking their movement trajectories. These historical AOD data can also be obtained through satellite remote sensing or ground observation stations, providing the necessary spatiotemporal characteristic information for subsequent sandstorm trajectory predictions.

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

[0082] Among them, constructing the interaction characteristics of terrain and wind field includes:

[0083] Define slope terrain slope and wind speed velocity Interaction data Ts Wv :

[0084] Ts Wv =terrain slope ×wind velocity

[0085]

[0086] Define aspect terrain aspect and wind direction direction Interaction data Ta Wd :

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

[0088]

[0089] Among them, terrain h is the sea slope height of the target area, wind Uwith wind V They are the east-west component and north-south component of the wind field in the target area, respectively.

[0090] In this step, the interactive characteristics of slope-wind speed and slope-wind direction are introduced to optimize the simulation capability of the local wind field, thereby improving the accuracy of sandstorm trajectory prediction.

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

[0092] Wind fields in areas with complex terrain are highly variable, primarily manifesting in significant fluctuations in wind speed and direction. These variations are influenced by a variety of factors, including topographical undulations, mountain obstructions, and valley ventilation. Therefore, to accurately predict future wind fields, deep learning models must not only possess robust spatiotemporal modeling capabilities but also deeply explore the impact of terrain 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 of which contains a convolutional layer, a spatial self-attention layer, and a ConvLSTM layer. Assume there are N such stages in total, and the input sequence consists of S frames (time series length), with each frame size being (B, C, H, W), where B represents the batch size, C is the number of channels, and H and W are the height and width of the image, respectively. In the encoder, the input sequence passes through the convolutional layers of each stage in turn for convolution operations. The output of the convolutional layers is then passed to the ConvLSTM layers of each stage for time series modeling.

[0095] Each level of the module consists of: a convolutional layer, which performs convolution operations using multiple convolution kernels to extract spatial features at different scales; a spatial self-attention mechanism layer, which calculates the correlation between different positions in the input feature map and captures the long-range dependency of key terrain features on the wind field; and a ConvLSTM layer, which builds a temporal dynamic model of the wind field and captures the temporal evolution of the wind field.

[0096] The decoder is symmetrical to the encoder structure, receiving the encoder hidden state and outputting the predicted wind field.

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

[0098] Project the input feature map through the convolution weights WQ, WK, and WV to generate the query matrix Q, key matrix K, and 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] Weight the value matrix V and residually connect the output Y:

[0102] Y=γ(A·V)+X

[0103] Among them, γ represents a learnable parameter.

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

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

[0106] Among them, X t represents the input data at time t, h t-1 and c t-1 Represent the hidden state and unit state of the previous moment, h t and c t Represent the hidden state and unit state at the current moment respectively.

[0107] ConvLSTM combines the spatial feature extraction capabilities of CNN with the time series modeling capabilities of LSTM, enabling better capture of wind field evolution trends over time. ConvLSTM calculations across multiple time steps effectively extract the evolutionary patterns of wind fields, enhancing the ability to predict changes in wind speed and direction.

[0108] For the i-th stage, the input first passes through the convolution layer, and then the convolution output is sent to the ConvLSTM layer for processing through the spatial self-attention mechanism module. The specific formula is as follows:

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

[0110] Among them, X t is the input time series data, indicating the input at time t, h t-1 and c t-1 They are the hidden state and unit state of the previous moment, h t and c tThey are the hidden state and unit state at the current moment, respectively. At each stage, the convolution layer is used to extract spatial features, while ConvLSTM processes temporal relationships. For the entire encoder, assuming there are N stages, the hidden state of its output is:

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

[0112] The decoder operates similarly to the encoder, but starts with the encoder's hidden state and gradually decodes it back to the original output sequence. The decoder operates by starting with the state of the last stage, passing information forward, and generating predictions. The decoder's primary task is to recover the spatial information of the original data from the encoder's feature representation and generate predictions, 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 of the encoder-decoder network is a time series data X = {X1, X2, ..., X S}}, where each X t Is a grid data of size (B, C, H, W). The encoder and decoder process this input sequence through their respective ConvLSTM and convolution layers. The encoder process is:

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

[0117] The decoder process is:

[0118] H decoder =Decoder(H encoder )

[0119] The final output is the output of the decoder, expressed as:

[0120] Y=H decoder

[0121] In terms of data processing, the input data for model training is mainly divided into the dimensions of (200, 18, 10, 4, 64, 64), where 200 means that the data set contains 200 sample sets; 18 means that each sample set contains 18 time series data points; 10 means that each time series data point contains 10 time steps of data; 4 means that the wind field data of each time step contains 4 feature channels, including wind field data (U, V) and two interaction data (Ts Wv 、Ta Wd ); 64 indicates that the spatial resolution of each wind field data point is a 64×64 grid, reflecting the spatial distribution of each time series. The output data is divided into dimensions of (200, 18, 1, 2, 64, 64). 1 represents the data for one time step in the future, and 2 represents the wind field (U, V) data for one time step in the future.

[0122] The SSA-MSED network in this step combines a spatial self-attention mechanism with an encoder-decoder architecture to accurately capture the impact of terrain on wind fields and achieve high-precision wind forecasts in sandstorm areas. The core of the SSA-MSED network uses spatial self-attention to enhance the network's ability to model the complex nonlinear relationship between terrain and wind fields. The encoder-decoder architecture also effectively reduces the dimensionality of high-dimensional data and extracts features, thereby improving the accuracy and robustness of predictions.

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

[0124] When predicting the trajectory of a sandstorm, extracting time series features is crucial for accurate prediction. While accurate wind prediction is crucial, relying solely on wind field information cannot fully capture the dynamics of a sandstorm. The formation and evolution of sandstorms are influenced by multiple factors, particularly the evolution patterns and spatiotemporal correlations within historical sandstorm time series. Due to the highly nonlinear and complex spatiotemporal evolution of sandstorms, traditional prediction methods struggle to effectively model their dynamic characteristics.

[0125] The ED-ConvLSTM network proposed in this step is used to further explore the time series characteristics of sandstorms and improve the accuracy of trajectory prediction. Its core is to use the powerful spatiotemporal feature extraction capabilities of the ConvLSTM network combined with the Encoder-Decoder structure, enabling the model to effectively capture the historical evolution patterns of sandstorms and infer future trajectories based on these patterns.

[0126] like Figure 3As shown in the figure, the ED-ConvLSTM network adopts a cascaded encoder-decoder structure, which includes M cascaded modules. Within each module, the hidden state and cell state in the encoder block are passed to the decoder block to support the entire prediction process. For long-term prediction tasks, these state information is not limited to the transmission within the module, but can be continued between multiple modules. Under this structure, the hidden state and cell state of the encoder block are passed from the encoder block of one module to the encoder block of the next module, thereby realizing the temporal information transmission across modules. This design enables the model to capture long-term temporal dependencies and maintain the continuity of information between multiple modules, further enhancing the modeling ability of long time step sequences. For long time step prediction tasks, the temporal information transmission across modules is crucial;

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

[0128] The encoder output of the mth module is:

[0129]

[0130] Among them, X t Represents input data, Represent the hidden state and cell state of the m-th module encoder at time t, They 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 mth module encoder at time t is Passed to the decoder of the same module:

[0132]

[0133] The cell state of the mth module encoder at time t Passed to the encoder of the m+1th module:

[0134]

[0135] in, denote the hidden state and cell state of the m+1th module encoder at time t, respectively. represents the hidden state of the m+1th module encoder at time t-1;

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

[0137]

[0138] The final Y tis the predicted AOD data at {t+1, t+2, t+3,…, t+10}.

[0139] In this embodiment, the fusion network SSA-EDLSTM combines the wind field data predicted by SSA-MSED This data, along with the sandstorm data from the past 10 time steps, is fed into the ED-ConvLSTM network to predict the sandstorm's trajectory for the next 10 time steps. This fusion approach leverages the precise wind field modeling capabilities of SSA-MSED and the temporal feature capture capabilities of ED-ConvLSTM to achieve more accurate predictions of sandstorm trajectories, while also effectively accounting for the impact of terrain on wind fields and sandstorm evolution.

[0140] The following further explains the process of generating future sandstorm movement trajectories:

[0141] There is a close relationship between aerosols and dust storms, and dust storms are an important component of aerosols. Dust storms are mainly formed by strong winds that roll up dust particles on the surface and transport them into the atmosphere. These particles are suspended in the atmosphere to form aerosols. Therefore, the distribution and movement trajectory of aerosols can reflect the propagation path and impact range of sandstorms. When a sandstorm occurs, a large number of sand and dust particles are rolled up by strong winds into the atmosphere to form aerosols. The particle size range of these particles is wide, ranging from a few microns to tens of microns. They can be suspended in the atmosphere for a long time and transmitted over long distances with air currents. The scattering and absorption characteristics of dust aerosols for light can significantly affect atmospheric visibility and radiation balance. Therefore, by monitoring the aerosol optical depth (AOD), the existence and intensity of sandstorms can be indirectly identified. In order to make long-step predictions of the trajectory of aerosols, a sequence-to-sequence method is used to convert T i = The AOD data at time {t-9, t-8, t-7, ..., t} is used as the input of the ED-ConvLSTM network T o =Prediction is performed based on the AOD data at time {t+1, t+2, t+3,…, t+10}.

[0142] The ED-ConvLSTM-based sandstorm trajectory prediction method only considers the impact of the wind field, ignoring the potential influence of terrain features on the wind field. However, topographic fluctuations significantly affect the structure and direction of the wind field. This is particularly true in complex terrain areas such as mountainous areas, canyons, and coastal areas, where the wind field can vary significantly due to topography. Therefore, incorporating topographic factors into the aerosol trajectory prediction model can more accurately reflect the dynamic changes in the actual wind field, thereby improving the accuracy of the prediction results. The main step in the SSA-EDLSTM-based sandstorm trajectory prediction method is data processing. Aerosol optical depth data is extracted from the MERRA-2 remote sensing dataset. All data is processed as a time series. In the fusion network SSA-EDLSTM, the SSA-MSED network takes as input wind field data and terrain data from time t-9 to time t, and predicts the wind field data at time t+1. The ED-ConvLSTM network takes as input wind field data and aerosol optical depth data from time t-9 to time t, and predicts the wind field data at time t+1. The outputs of the two networks are combined and fed into the ED-ConvLSTM network to obtain the aerosol optical density data at time t+2. This cycle repeats to obtain the aerosol optical density data from time t+1 to time t+10.

[0143] In this embodiment S5, based on the future aerosol motion trajectory, a sandstorm area detection algorithm is used to perform sandstorm area detection, and a sandstorm motion trajectory and a visualization report are generated.

[0144] Among them, the detection of sandstorm areas in combination with the sandstorm area detection algorithm includes:

[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 moment;

[0148] Define a change rate threshold λ, and mark the area as a sandstorm area when Rate(t)>change rate threshold λ.

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

[0150] Select the aerosol concentration time series data of typical sandstorm events and perform data preprocessing;

[0151] The rate of change of aerosol concentration during each sandstorm event was calculated, its distribution characteristics were statistically analyzed, and the top 5% quantile of the distribution was selected as the initial threshold;

[0152] Combined with cross-validation, the threshold was adjusted through multiple experiments to evaluate its performance in different sandstorm events, and finally the optimal threshold was determined.

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

[0154] Among them, the spatiotemporal comparison map of predicted trajectories and actual trajectories 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 size of the prediction error in different regions through the depth of color, helping researchers to quickly locate areas with large prediction deviations; the dynamic detection result coverage map of the sandstorm area combines the output of the sandstorm detection algorithm to show the spatiotemporal distribution of the sandstorm area.

[0155] This example proposes a sandstorm trajectory prediction method based on the SSA-EDLSTM fusion network. By combining the advantages of the spatial self-attention mechanism (SSA), the multi-scale encoder-decoder network (SSA-MSED), and the 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 exploring the spatiotemporal characteristics of sandstorm time series. At the same time, the innovative sandstorm area detection algorithm further improves the identification accuracy of sandstorm areas, and combined with visual 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 the differences from other embodiments. References to the same or similar parts between the various embodiments are sufficient. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple. For relevant parts, refer to the method description.

[0157] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to one 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 present invention. Therefore, the present invention is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A sandstorm trajectory prediction method based on SSA-EDLSTM fusion network, characterized by: The following steps are involved: S1. Obtain historical wind field data, terrain elevation data, and historical aerosol optical depth data for the target area; S2. Calculate the slope and slope direction based on the terrain elevation data, and construct the terrain and wind field interaction characteristics with the wind speed and wind direction respectively; S3, input the terrain and wind field interaction characteristics and historical wind field data into the SSA-MSED network to predict future wind field data; S4. Input the future wind field data and historical aerosol optical depth data into the ED-ConvLSTM network to generate the future aerosol motion trajectory; S5. Based on the future aerosol motion trajectory, a sandstorm area detection algorithm is used to detect the sandstorm area, and a sandstorm motion trajectory and a visualization report are generated.

2. The sandstorm trajectory prediction method based on the SSA-EDLSTM fusion network according to claim 1 is characterized in that: In S2, constructing the terrain and wind field interaction features includes: Define slope terrain slope and wind speed velocity Interaction data Ts Wv : Ts Wv =terrain slope ×wind velocity Define aspect terrain aspect and wind direction direction Interaction data Ta Wd : Are Wd |Wind direction -field aspect | Among them, terrain h is the sea slope height of the target area, wind U with wind V They are the east-west component and north-south component of the wind field in the target area, respectively.

3. The sandstorm trajectory prediction method based on the SSA-EDLSTM fusion network according to claim 1 is characterized in that: The SSA-MSED network includes an encoder and a decoder; wherein, The encoder is composed of N-level modules connected in series, and each level of modules sequentially includes: Convolutional layer, used to perform convolution operations through multiple convolution kernels to extract spatial features of different scales; A spatial self-attention mechanism layer is used to calculate the correlation between different positions in the input feature map and capture the long-range dependency of key terrain features on the wind field; The ConvLSTM layer is used to build a temporal dynamic model of the wind farm and capture the temporal evolution trend of the wind farm. The decoder is symmetrical to the encoder structure, receives the encoder hidden state and outputs the predicted wind field.

4. The sandstorm trajectory prediction method based on the SSA-EDLSTM fusion network according to claim 3 is characterized in that: The computation process of the spatial self-attention mechanism layer includes: Project the input feature map through the convolution weights WQ, WK, and WV to generate the query matrix Q, key matrix K, and value matrix V; Based on the query matrix Q and the key matrix K, calculate the attention weight matrix A: A=softmax(QK T ) Weight the value matrix V and residually connect the output Y: Y=γ(A·V)+X Among them, γ represents a learnable parameter.

5. The sandstorm trajectory prediction method based on the SSA-EDLSTM fusion network according to claim 3 is characterized in that: The ConvLSTM layer is represented as: ConvLSTM(X t ,h t-1 ,c t-1 )=(h t ,c t ) Among them, X t represents the input data at time t, h t-1 and c t-1 Represent the hidden state and unit state of the previous moment, h t and c t Represent the hidden state and unit state at the current moment respectively.

6. The sandstorm trajectory prediction method based on the SSA-EDLSTM fusion network according to claim 1 is characterized in that: The ED-ConvLSTM network adopts a cascaded encoding and decoding structure, including M cascaded modules, each module containing an encoder and a decoder; The encoder output of the mth module is: Among them, X t Represents input data, Represent the hidden state and cell state of the m-th module encoder at time t, They represent the hidden state and cell state of the m-th module encoder at time t-1 respectively.

7. The sandstorm trajectory prediction method based on the SSA-EDLSTM fusion network according to claim 6 is characterized in that: In the ED-ConvLSTM network, The hidden state of the mth module encoder at time t Passed to the same module decoder: The cell state of the mth module encoder at time t Passed to the encoder of the m+1th module: in, denote the hidden state and cell state of the m+1th module encoder at time t, respectively. represents the hidden state of the m+1th module encoder at time t-1.

8. The sandstorm trajectory prediction method based on the SSA-EDLSTM fusion network according to claim 1 is characterized in that: In S5, detecting the sandstorm area using the sandstorm area detection algorithm includes: Calculate the instantaneous rate of change of AOD Rate(t): 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; Define a change rate threshold λ, and mark the area as a sandstorm area when Rate(t)>change rate threshold λ.

9. The sandstorm trajectory prediction method based on the SSA-EDLSTM fusion network according to claim 8 is characterized in that: Defining the change rate threshold λ includes: Select the aerosol concentration time series data of typical sandstorm events and perform data preprocessing; The rate of change of aerosol concentration during each sandstorm event was calculated, its distribution characteristics were statistically analyzed, and the top 5% quantile of the distribution was selected as the initial threshold; Combined with cross-validation, the threshold was adjusted through multiple experiments to evaluate its performance in different sandstorm events, and finally the optimal threshold was determined.

10. The sandstorm trajectory prediction method based on the SSA-EDLSTM fusion network according to claim 1 is characterized in that: In S5, the visualization report includes: a spatiotemporal comparison diagram of the predicted trajectory and the actual trajectory, an error heat map, and a dynamic detection result coverage map of the sandstorm area.

Images