Tropical cyclone abnormal path forecasting method based on space-time multi-scale fusion and evidence learning

Abstract

The invention discloses a tropical cyclone abnormal path forecasting method based on spatio-temporal multi-scale fusion and evidence learning. The tropical cyclone abnormal path forecasting method comprises the steps of preprocessing collected satellite cloud picture data; marking the preprocessed satellite cloud picture on the basis of the IBTrACS optimal path set; constructing a tropical cyclone abnormal path forecasting model based on space-time multi-scale fusion and evidence learning; training the constructed tropical cyclone abnormal path forecasting model; and forecasting the abnormal path of the tropical cyclone at the future moment by using the trained abnormal path forecasting model of the tropical cyclone. The tropical cyclone abnormal path forecasting model provided by the invention not only can provide a high-precision path forecasting result, but also can accurately identify and forecast an abnormal path type, and provides reliable decision support and reference basis for a service forecasting department.

Description

Technical Field

[0001] This invention belongs to the field of satellite remote sensing monitoring, specifically involving a method for predicting anomalous tropical cyclone tracks based on spatiotemporal multi-scale fusion and evidence learning. Background Technology

[0002] Accurate tropical cyclone track forecasting is a crucial technical support for ensuring the safety of people and property in coastal areas. Tropical cyclone track forecasting predicts the area a tropical cyclone will pass through every 6 or 12 hours using specific methods, typically providing the center location of the cyclone. Despite significant progress in tropical cyclone track forecasting, many unpredictable track scenarios remain, potentially leading to substantial errors and uncertainties in the forecast. The most challenging cases are anomalous tracks (UTs), such as sudden changes in direction, rotation, or stagnation, characterized by rapid changes in the cyclone's motion within a short period.

[0003] Most studies define abrupt changes in tropical cyclone movement as an angle exceeding a certain threshold within a specified time period. Thanks to improvements in numerical models, particularly the development of ensemble forecasting systems, the prediction of anomalous tropical cyclone tracks has made some progress, but the error remains significant. Therefore, operational typhoon track forecasting and warnings still face considerable challenges. In recent years, artificial intelligence (AI) methods such as machine learning and deep learning have been widely applied in tropical cyclone forecasting. For anomalous tropical cyclone track forecasting, combining satellite cloud imagery data and fully exploring the potential of AI offers possibilities for improving anomalous tropical cyclone track prediction. Summary of the Invention

[0004] Purpose of the invention: In order to overcome the shortcomings of the existing technology, a method for predicting anomalous tracks of tropical cyclones based on spatiotemporal multi-scale fusion and evidence learning is provided.

[0005] Technical Solution: To achieve the above objectives, this invention provides a method for predicting anomalous tropical cyclone tracks based on spatiotemporal multi-scale fusion and evidence learning, comprising the following steps:

[0006] S1: Preprocess the collected satellite cloud image data;

[0007] S2: Annotate the preprocessed satellite cloud images based on the IBTRAS optimal path set;

[0008] S3: Construct a tropical cyclone anomaly track forecasting model based on spatiotemporal multi-scale fusion and evidence learning;

[0009] S4: Train the constructed tropical cyclone anomalous track forecasting model based on the data from steps S1 and S2;

[0010] S5: Use the trained tropical cyclone anomaly track prediction model to predict the anomaly track of tropical cyclones at future times.

[0011] Furthermore, the preprocessing procedure in step S1 includes:

[0012] First, a heat map of the satellite cloud image is calculated. Then, the infrared channel of the satellite cloud image is stitched with the heat map and the geographic coordinates are aligned. Finally, the time difference and vorticity are calculated.

[0013] Furthermore, in the preprocessing procedure of step S1:

[0014] The calculation method for the heat map of satellite cloud imagery is as follows: A two-dimensional Gaussian distribution is generated using the image center (256, 256) as the center, and the intensity adaptive standard deviation σ = 50 × (1 + I), where I is the intensity at the current time; the calculation formula is: exp(-dist 2 / (2σ 2 )) 2 , where: dist 2 = (h-256) 2 +(w-256) 2 h and w are the pixel values ​​in the height and width directions of the image, respectively. The generated heatmap is normalized to [0,1] after the center weight is enhanced by squaring, forming a physical constraint feature map with the center of the tropical cyclone as the focus and the greater the intensity, the wider the coverage.

[0015] The calculation method for geographic coordinate alignment is as follows: calculate the geographic range boundary of the center location of all tropical cyclones in the entire time series, and then calculate the cropping range of each frame image in a fixed grid based on the position of the cyclone center relative to the geographic boundary of the entire series at the current moment, and output a uniform standard size.

[0016] Time Difference (Diff) = SH i+N -SH i , among which, SH i The composite value of the satellite cloud image and heat map represents the moment of acquisition, and N represents the acquisition interval;

[0017] Among them, the wind direction Zonal wind Diff s This represents the contour difference portion of the time difference.

[0018] Furthermore, the labeling type in step S2 includes the unique thermal coding of the anomalous path of the tropical cyclone, which includes: [0,0,0] - no deflection, [0,1,0] - northward deflection and [0,0,1] - westward deflection; the calculation method is: the average angular difference of the tropical cyclone's movement is calculated based on the longitude and latitude of the IBTRAS optimal path set 12 hours before and after (t-6,t,t+6), with a leftward deflection greater than or equal to 30 degrees or a rightward deflection greater than or equal to 45 degrees.

[0019] Furthermore, the tropical cyclone anomaly track prediction model in step S3 includes a satellite cloud image physical feature extraction module, a spatiotemporal wavelet transform module, a time-domain dimensionality reduction module, a fusion pooling module, and an evidence learning module.

[0020] Furthermore, the satellite cloud image physical feature extraction module includes a 7×7 convolutional layer, a max pooling layer, and a ResNet18 backbone network connected in sequence, wherein the fully connected layer at the end of the ResNet18 backbone network is replaced by a 1×1 convolutional layer; the input of the satellite cloud image physical feature extraction module is the spliced ​​value of time difference and vorticity, wherein the time difference includes cloud image difference and heat map difference, and the input dimension is [T,C,H,W].

[0021] Furthermore, the spatiotemporal wavelet transform module includes a multi-scale spatial wavelet transform unit and a multi-scale temporal wavelet transform unit, and the specific operation process includes:

[0022] The input features of the spatiotemporal transformation module are the output features of stage 2 of the ResNet18 network in the satellite cloud image physical feature extraction module, and the size of the output feature dimension is [missing information]. Then, the frequency domain features are sequentially passed through a series of multi-scale spatial wavelet transform units and multi-scale time wavelet transform units to output the frequency domain features, the size of which is...

[0023] The input dimension of the multi-scale spatial wavelet transform unit is: The input features are decomposed into multiple scales using trainable Haar wavelet filters with different kernel sizes (kernel_size, 2×2 and 4×4). First, the last two dimensions of the features are decomposed using zero values: high... Hekuan The kernel is padded so that its width and height are divisible by the kernel size. The padded width and height are denoted as h and w, respectively, resulting in a feature dimension of [T,C,h,w]. Then, an unfold operation is used to transform the spatial dimension into a sliding window format of [T,C,h,w,kernel_size,kernel_size]. Each window is convolved with four Haar wavelet bases (LL, LH, HL, HH) to obtain wavelet coefficients [T,4C,h,w]. Frequency components are weighted using a frequency attention mechanism (adaptive average pooling to [T,4C,1,1], followed by a convolutional layer and Sigmoid activation to generate attention weights). The weighted wavelet features at each scale are then uniformly pooled to [T,4C,1,1]. Then, the layers are stitched together along the channel dimension, and the dimensionality is reduced to [value] using a 1×1 convolutional layer. Output;

[0024] The input of the multi-scale time wavelet transform unit is the output of the multi-scale spatial wavelet transform unit, with dimension [missing information]. Using predefined low-pass and high-pass filters (coefficients [0.7071, 0.7071] and [0.7071, -0.7071]), adjacent time steps are grouped together using a grouping operation. Then, the low-frequency and high-frequency wavelet coefficients are calculated, and the dimension becomes... The coefficients were then reshaped back to spatial dimensions. And convert it to a 3D convolution input format. Feature extraction is performed by applying a temporal convolution kernel (3,1,1) to the low-frequency and high-frequency components respectively;

[0025] Finally, it is obtained through channel splicing. Dimensionality reduced to 1×1×1 using 3D convolution with a kernel size of 1×1×1 Finally, reshape the dimension. Output.

[0026] Furthermore, the input features of the time-domain dimensionality reduction module are the output features of stage 4 of the ResNet18 network in the satellite cloud image physical feature extraction module, and the size of the output feature dimension is [missing value]. Then, dimensionality reduction is performed using a 3D convolution with a kernel size of 1×1×1. Output time-domain features.

[0027] Furthermore, the fusion pooling module includes a feature fusion unit and an attention pooling unit, and its specific operation process includes:

[0028] The input to the fusion pooling module is the frequency domain feature output from the spatiotemporal wavelet transform module and the time domain feature output from the time domain dimensionality reduction module, both with a feature size of [missing value]. Then, the fused features are output sequentially through a series of feature fusion units and attention pooling units, and the size of the output fused features is [C].

[0029] The feature fusion unit first concatenates the frequency domain features and time domain features along the channel dimension to obtain... Then, a 1×1 convolutional layer is used to reduce the dimensionality to... Output;

[0030] The input to the attention pooling unit is the output of the feature fusion unit, and the dimension is... First, reshape into The spatiotemporal structure of this unit employs a hierarchical pooling strategy that prioritizes spatial information over temporal information: the spatial attention layer calculates the importance scores of all spatial locations within each time step using two layers of multilayer perceptrons. After softmax normalization, the spatial dimensions are weighted and aggregated to obtain The temporal features are then calculated; subsequently, the temporal attention layer also uses a two-layer multilayer perceptron structure to calculate the attention weights [T,1] at each time step. After softmax normalization, the temporal dimension is weighted and pooled to finally output the global feature representation [C].

[0031] Furthermore, the evidence learning module includes an evidence feature extraction unit, an evidence regression unit, and an uncertainty calculation unit, and its specific operation process includes:

[0032] The evidence learning module takes the output features of the fusion pooling module as input, with a feature size of [C]. Then, it passes through the evidence feature extraction unit, the evidence regression unit, and the uncertainty calculation unit in sequence to output the Dirichlet parameter α and the abnormal path prediction probability P.

[0033] Among them, the evidence feature extraction unit extracts evidence features through a multilayer perceptron and a random deactivation layer [C]. The evidence regression unit regresses the evidence value [N] through a linear layer; the uncertainty calculation unit calculates the Dirichlet parameter α and the anomaly path prediction probability P, where the Dirichlet parameter α = evidence regression value + 1, and the anomaly path prediction probability P = α. i / ∑α i .

[0034] Furthermore, in step S4, the loss function used during model training is Loss. total The formula is as follows:

[0035] Loss total =w class Loss class +w ev Loss ev

[0036] Among them, w class and w ev The dynamic weights are given by the following formula:

[0037] w class =max(0.6, 0.8 - epoch * 0.05)

[0038] w ev =1-w class

[0039] Where epoch represents the number of iterations;

[0040] Loss class The classification loss is expressed by the following formula:

[0041] Loss class = -log(P+ε)

[0042] Where ε=1×10 -6 ;

[0043] Loss ev The formula for representing evidence loss is as follows:

[0044]

[0045] Among them, Loss nll = -log(P+ε), ε = 1×10 -6 ;

[0046] Dir(α) represents the Dirichlet distribution of the Dirichlet parameter α, and Β(α) is the Beta function, Β(α) = ∏(Γ(α) i )) / Γ(∑α i ), where Γ(·) is the gamma function; Ψ(α) = log(Γ(α)), where Ψ(α) measures the "sharpness" of the distribution. The parameter difference is represented by (Ψ(α)-Ψ(∑α)), which represents the relative importance of each category in the target distribution; α KL =max(0.1*(0.95**epoch),0.01) is the dynamic weight, initially: α KL ≈0.1, emphasizing learning under uncertainty; later stage: α KL →0.01, focusing on accuracy. All specific values ​​in the loss function are experimental values.

[0047] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0048] 1. Because the spatiotemporal multi-scale fusion method proposed in this invention has a better modeling effect on the multi-scale spatiotemporal characteristics and non-stationary properties of typhoons, the prediction accuracy of abnormal paths is higher than that of not using the spatiotemporal multi-scale fusion method.

[0049] 2. Since this invention utilizes the evidence learning method to output uncertainty and participate in the final prediction probability decision, it can reduce the error of anomaly probability prediction;

[0050] 3. Since the forecasting model of this invention can provide the probability of abnormal paths, it can provide a reference for business path forecasting, further improving the accuracy of business path forecasting. Attached Figure Description

[0051] Figure 1 This is a schematic flowchart of the method of the present invention;

[0052] Figure 2 This is a structural diagram of a tropical cyclone anomalous track forecasting model. Detailed Implementation

[0053] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading this invention, any modifications of the invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0054] Example 1:

[0055] like Figure 1 As shown, this embodiment provides a method for predicting anomalous tropical cyclone tracks based on spatiotemporal multi-scale fusion and evidence learning, including the following steps:

[0056] S1: Acquisition and preprocessing of Himawari-8 satellite data

[0057] The preprocessing process includes:

[0058] First, a heat map of the satellite cloud image is calculated. Then, the infrared channel of the satellite cloud image is stitched with the heat map and the geographic coordinates are aligned. Finally, the time difference and vorticity are calculated.

[0059] The calculation method for the heat map of satellite cloud imagery is as follows: A two-dimensional Gaussian distribution is generated using the image center (256, 256) as the center, and the intensity adaptive standard deviation σ = 50 × (1 + I), where I is the intensity at the current time; the calculation formula is: exp(-dist 2 / (2σ 2 )) 2 , where: dist 2 = (h-256) 2 +(w-256) 2h and w are the pixel values ​​in the height and width directions of the image, respectively. The generated heatmap is normalized to [0,1] after the center weight is enhanced by squaring, forming a physical constraint feature map with the center of the tropical cyclone as the focus and the greater the intensity, the wider the coverage.

[0060] The calculation method for geographic coordinate alignment is as follows: calculate the geographic range boundary of the center location of all tropical cyclones in the entire time series, and then calculate the cropping range of each frame image in a fixed grid based on the position of the cyclone center relative to the geographic boundary of the entire series at the current moment, and output it uniformly as a standard size of 350×350.

[0061] Time Difference (Diff) = SH i+N -SH i , among which, SH i The composite value of the satellite cloud image and heat map represents the time of acquisition, and N represents the acquisition interval;

[0062] Among them, the wind direction Zonal wind Diff s This represents the contour difference portion of the time difference.

[0063] S2: Based on the IBTRAS optimal path set, the preprocessed satellite cloud images are labeled to generate labels, which are the ground truth values, for subsequent loss function calculation and model evaluation;

[0064] The annotation types include the unique thermal coding of anomalous tracks of tropical cyclones, including: [0,0,0] - no deflection, [0,1,0] - northward deflection and [0,0,1] - westward deflection; the calculation method is: according to the standard of the 85-906 key project, the average difference in the directional angle of tropical cyclones is calculated based on the longitude and latitude of the IBTRAS optimal track set 12 hours before and after (t-6,t,t+6), with left deflection greater than or equal to 30 degrees or right deflection greater than or equal to 45 degrees.

[0065] S3: Construct a tropical cyclone anomaly track forecasting model based on spatiotemporal multi-scale fusion and evidence learning;

[0066] like Figure 2 As shown, the tropical cyclone anomalous track prediction model includes a satellite cloud image physical feature extraction module, a spatiotemporal wavelet transform module, a time-domain dimensionality reduction module, a fusion pooling module, and an evidence learning module.

[0067] First, the basic features of physical quantities are extracted using the satellite cloud image physical feature extraction module. Then, the basic features are processed separately using a frequency domain processing branch (mainly composed of a spatiotemporal wavelet transform module) and a time domain processing branch (mainly composed of a time domain dimensionality reduction module) to generate frequency domain features and time domain features. Next, the frequency domain features and time domain features generated by the frequency domain processing branch and the time domain processing branch are fused using a fusion pooling module to generate fused features. Finally, based on the fused features, the evidence learning module outputs the probability of an anomalous tropical cyclone path and the Dirichlet parameter α over the next 6 hours.

[0068] The satellite cloud image physical feature extraction module consists of a 7×7 convolutional layer, a max pooling layer, and a ResNet18 backbone network connected in sequence. The fully connected layer at the end of the ResNet18 backbone network is replaced by a 1×1 convolutional layer. The input of the satellite cloud image physical feature extraction module is the concatenated value of temporal difference and vorticity. The temporal difference includes cloud image difference and heat map difference. The input dimension is [T,C,H,W].

[0069] The satellite cloud image physical feature extraction module takes physical features of dimension [T, C, H, W] as input, where T represents the time series length, C represents the channel dimension after concatenating the time difference and vorticity, and the time difference includes cloud image difference and heat map difference, and H×W represents the height and width of the image. In the first two layers, 7×7 convolution is selected, followed by max pooling to effectively reduce dimensionality and extract features from the original image.

[0070] The spatiotemporal wavelet transform module includes a multi-scale spatial wavelet transform unit and a multi-scale temporal wavelet transform unit. The specific operation process includes:

[0071] The input features of the spatiotemporal transformation module are the output features of stage 2 of the ResNet18 network in the satellite cloud image physical feature extraction module, and the size of the output feature dimension is [missing information]. Then, the frequency domain features are sequentially passed through a series of multi-scale spatial wavelet transform units and multi-scale time wavelet transform units to output the frequency domain features, the size of which is...

[0072] The input dimension of the multi-scale spatial wavelet transform unit is: The input features are decomposed into multiple scales using trainable Haar wavelet filters with different kernel sizes (kernel_size, 2×2 and 4×4). First, the last two dimensions of the features are decomposed using zero values: high... Hekuan The kernel is padded so that its width and height are divisible by the kernel size. The padded width and height are denoted as h and w, respectively, resulting in a feature dimension of [T,C,h,w]. Then, an unfold operation is used to transform the spatial dimension into a sliding window format of [T,C,h,w,kernel_size,kernel_size]. Each window is convolved with four Haar wavelet bases (LL, LH, HL, HH) to obtain wavelet coefficients [T,4C,h,w]. Frequency components are weighted using a frequency attention mechanism (adaptive average pooling to [T,4C,1,1], followed by a convolutional layer and Sigmoid activation to generate attention weights). The weighted wavelet features at each scale are then uniformly pooled to [T,4C,1,1]. Then, the layers are stitched together along the channel dimension, and the dimensionality is reduced to [value] using a 1×1 convolutional layer. Output;

[0073] The input of the multi-scale time wavelet transform unit is the output of the multi-scale spatial wavelet transform unit, with dimension [missing information]. Using predefined low-pass and high-pass filters (coefficients [0.7071, 0.7071] and [0.7071, -0.7071]), adjacent time steps are grouped together using a grouping operation. Then, the low-frequency and high-frequency wavelet coefficients are calculated, and the dimension becomes... The coefficients were then reshaped back to spatial dimensions. And convert it to a 3D convolution input format Feature extraction is performed by applying a temporal convolution kernel (3,1,1) to the low-frequency and high-frequency components respectively;

[0074] Finally, it is obtained through channel splicing. Dimensionality reduced to 1×1×1 using 3D convolution with a kernel size of 1×1×1 Finally, reshape the dimension back. Output.

[0075] The input features of the time-domain dimensionality reduction module are the output features of stage 4 of the ResNet18 network in the satellite cloud image physical feature extraction module, and the size of the output feature dimension is [missing value]. Then, dimensionality reduction is performed using a 3D convolution with a kernel size of 1×1×1. Output time-domain features.

[0076] The fusion pooling module includes a feature fusion unit and an attention pooling unit, and its specific operation process includes:

[0077] The input to the fusion pooling module is the frequency domain feature output from the spatiotemporal wavelet transform module and the time domain feature output from the time domain dimensionality reduction module, both with a feature size of [missing value]. Then, the fused features are sequentially passed through a series of feature fusion units and attention pooling units to output fused features of size [C].

[0078] The feature fusion unit first concatenates the frequency domain features and time domain features along the channel dimension to obtain... Then, a 1×1 convolutional layer is used to reduce the dimensionality to... Output;

[0079] The input to the attention pooling unit is the output of the feature fusion unit, and the dimension is... First, reshape into The spatiotemporal structure of this unit employs a hierarchical pooling strategy that prioritizes spatial information over temporal information: the spatial attention layer calculates the importance scores of all spatial locations within each time step using two layers of multilayer perceptrons. After softmax normalization, the spatial dimensions are weighted and aggregated to obtain The temporal features are then calculated; subsequently, the temporal attention layer also uses a two-layer multilayer perceptron structure to calculate the attention weights [T,1] at each time step. After softmax normalization, the temporal dimension is weighted and pooled to finally output the global feature representation [C].

[0080] The evidence learning module includes an evidence feature extraction unit, an evidence regression unit, and an uncertainty calculation unit. The specific operation process includes:

[0081] The evidence learning module takes the output features of the fusion pooling module as input, with a feature size of [C]. Then, it passes through the evidence feature extraction unit, the evidence regression unit, and the uncertainty calculation unit in sequence to output the Dirichlet parameter α and the abnormal path prediction probability P.

[0082] Among them, the evidence feature extraction unit extracts evidence features through a multilayer perceptron and a random deactivation layer [C]. The evidence regression unit regresses the evidence value [N] through a linear layer; the uncertainty calculation unit calculates the Dirichlet parameter α and the anomaly path prediction probability P, where the Dirichlet parameter α = evidence regression value + 1, and the anomaly path prediction probability P = α. i / ∑α i .

[0083] The effect of the Dirichlet parameter α is reflected in the following two aspects:

[0084] 1. The probability of predicting an abnormal path is P = α i / ∑α i Therefore, it is used to calculate the probability P of abnormal path prediction;

[0085] 2. Used to measure the uncertainty of the model and participate in the calculation of the uncertain part in the subsequent loss function.

[0086] S4: Train the constructed tropical cyclone anomalous track forecasting model based on the data from steps S1 and S2;

[0087] The loss function used during model training is Loss total The formula is as follows:

[0088] Loss total =w class Loss class +w ev Loss ev

[0089] Among them, w class and w ev The dynamic weights are given by the following formula:

[0090] w class =max(0.6, 0.8 - epoch * 0.05)

[0091] w ev =1-w class

[0092] Where epoch represents the number of iterations;

[0093] Loss class The classification loss is expressed by the following formula:

[0094] Loss class = -log(P+ε)

[0095] Where ε=1×10 -6 ;

[0096] Loss ev The formula for representing evidence loss is as follows:

[0097]

[0098] Among them, Loss nll = -log(P+ε), ε = 1×10 -6 ;

[0099] Dir(α) represents the Dirichlet distribution of α, and Β(α) is the Beta function. Β(α)=∏(Γ(α) i )) / Γ(∑α i ), where Γ(·) is the gamma function. Ψ(α) = log(Γ(α)). It can handle large values ​​without encountering numerical problems. Ψ(α) measures the "sharpness" of the distribution. This represents the parameter differences. (Ψ(α)-Ψ(∑α)) represents the relative importance of each category in the target distribution. α KL =max(0.1*(0.95**epoch),0.01) is the dynamic weight, initially: αKL ≈0.1, emphasizing learning under uncertainty; later stage: α KL →0.01, primarily focusing on accuracy. All specific values ​​in the loss function are experimental values.

[0100] S5: Use the trained tropical cyclone anomalous track forecasting model to forecast the anomalous track of tropical cyclones in the future.

[0101] In this embodiment, the continuous time-lapse images of Himawari8 data are input, and the probability of an abnormal tropical cyclone path occurring in the next 6 hours is output for testing.

[0102] Example 2:

[0103] To verify the effectiveness and efficacy of the method of the present invention, the following experiments and analyses were conducted in this embodiment:

[0104] The study area for this embodiment is the Northwest Pacific region, 0-50°N latitude and 100-180°E longitude. All experiments were performed on an NVIDIA GeForce RTX 4090 GPU equipped with 24GB of video memory, using a batch size of 16, and model training was conducted using the Adam optimizer within the PyTorch framework under Python 3.8.

[0105] To achieve effective model training and evaluation, the dataset was divided into two subsets: 2000-2019 as the training dataset and 2018-2022 as the test dataset. Considering that anomalous tropical cyclone paths are rare events, and the number of non-anomalous path samples far exceeds the number of suddenly enhanced samples, an oversampling strategy was employed during data loading to address class imbalance. To ensure the reliability of the evaluation results, an uncertainty-aware evaluation framework was established in this embodiment: calculated using the Dirichlet parameter α... The model confidence is quantified using (where N is the number of classes), and samples with total uncertainty below a preset threshold are selected as high-confidence predictions for performance evaluation. Model performance is primarily evaluated using the F1 score, supplemented by the accuracy standard classification metric. The specific calculation formula is as follows:

[0106]

[0107] Where L is the set of category labels, |L| is the number of categories, N is the total number of samples, and y i and Let be the true label and the predicted label of sample i, respectively, and l(·) be the indicator function. For accuracy, For recall rate, TP l ,FP l ,FN lThese represent the number of true positives, false positives, and false negatives, respectively. This uncertainty-aware assessment method effectively improves the reliability and robustness of model performance evaluation by focusing on high-confidence predictions.

[0108] The model was tested by taking continuous time-series images from the Himawari8 data set as input and outputting the probability of anomalies in the tropical cyclone track over the next 6 hours. The test results are shown in Tables 1, 2, and 3.

[0109] Table 1: Data Contribution Analysis Results, 2018-2022

[0110]

[0111] Table 1 presents the data contribution analysis results from 2018 to 2022. The role of various satellite-derived features was evaluated using three different input configurations: configuration without heatmaps, configuration without eddy covariance, and configuration including the complete input.

[0112] Table 2: Module Contribution Analysis Results, 2018-2022

[0113]

[0114] Table 2 presents the module contribution analysis results from 2018 to 2022. Three sets of experiments were constructed to evaluate the role of different modules within the model, particularly the influence of wavelet transform. Here, w / o SWT, w / o TWT, and w / o SWT&TWT represent model variants tested without the SWT module, without the TWT module, and without both modules, respectively.

[0115] Table 3: Prediction Results of the Model for Various Anomaly Path Types, 2018-2022

[0116]

[0117] Evaluation results show that, based on the data contribution analysis in Table 1, the model achieves optimal performance when the input includes both heatmaps and vorticity. This demonstrates that enhancing the expression of typhoon cloud system features through manual feature enhancement and incorporating physical features such as vorticity effectively improves the model's learning ability and prediction accuracy. Regarding the module contribution analysis in Table 2, the model performs best when it integrates the spatiotemporal wavelet transform module, confirming the crucial role of capturing multi-scale spatiotemporal features of typhoons in forecasting anomalous tropical cyclone tracks. As for the practical application effects in Table 3, the tropical cyclone anomalous track forecasting model based on spatiotemporal multi-scale fusion and evidence learning not only provides high-precision track forecasting results but also accurately identifies and forecasts anomalous track types, providing reliable decision support and reference for operational forecasting departments. In summary, the ablation experiment fully validates the effectiveness of each component of the model, while the practical application results demonstrate the practical value and application prospects of this method in tropical cyclone anomalous track forecasting tasks.

Claims

1. A method for predicting anomalous tropical cyclone tracks based on spatiotemporal multi-scale fusion and evidence learning, characterized in that, Includes the following steps: S1: Preprocess the collected satellite cloud image data; S2: Annotate the preprocessed satellite cloud images based on the IBTRAS optimal path set; S3: Construct a tropical cyclone anomaly track forecasting model based on spatiotemporal multi-scale fusion and evidence learning; S4: Train the constructed tropical cyclone anomalous track forecasting model based on the data from steps S1 and S2; S5: Use the trained tropical cyclone anomaly track prediction model to predict the anomaly track of tropical cyclones at future times.

2. The method for predicting anomalous tropical cyclone tracks based on spatiotemporal multi-scale fusion and evidence learning according to claim 1, characterized in that, The preprocessing procedure in step S1 includes: First, a heat map of the satellite cloud image is calculated. Then, the infrared channel of the satellite cloud image is stitched with the heat map and the geographic coordinates are aligned. Finally, the time difference and vorticity are calculated.

3. The method for predicting anomalous tropical cyclone tracks based on spatiotemporal multi-scale fusion and evidence learning according to claim 2, characterized in that, In the preprocessing process of step S1: The calculation method for the heat map of satellite cloud imagery is as follows: A two-dimensional Gaussian distribution is generated using the image center (256, 256) as the center, and the intensity adaptive standard deviation σ = 50 × (1 + I), where I is the intensity at the current time; the calculation formula is: exp(-dist 2 / (2σ 2 )) 2 , where: dist 2 = (h-256) 2 +(w-256) 2 h and w are the pixel values ​​in the height and width directions of the image, respectively. The generated heatmap is normalized to [0,1] after the center weight is enhanced by squaring, forming a physical constraint feature map with the center of the tropical cyclone as the focus and the greater the intensity, the wider the coverage. The calculation method for geographic coordinate alignment is as follows: calculate the geographic range boundary of the center location of all tropical cyclones in the entire time series, and then calculate the cropping range of each frame image in a fixed grid based on the position of the cyclone center relative to the geographic boundary of the entire series at the current moment, and output a uniform standard size. Time Difference (Diff) = SH i+N -SH i , among which, SH i The composite value of the satellite cloud image and heat map represents the moment of acquisition, and N represents the acquisition interval; Among them, the wind direction Zonal wind Diff s This represents the contour difference portion of the time difference.

4. The method for predicting anomalous tropical cyclone tracks based on spatiotemporal multi-scale fusion and evidence learning according to claim 3, characterized in that, The tropical cyclone anomaly track prediction model in step S3 includes a satellite cloud image physical feature extraction module, a spatiotemporal wavelet transform module, a time-domain dimensionality reduction module, a fusion pooling module, and an evidence learning module.

5. The method for predicting anomalous tropical cyclone tracks based on spatiotemporal multi-scale fusion and evidence learning according to claim 4, characterized in that, The satellite cloud image physical feature extraction module includes a 7×7 convolutional layer, a max pooling layer, and a ResNet18 backbone network connected in sequence. The fully connected layer at the end of the ResNet18 backbone network is replaced by a 1×1 convolutional layer. The input of the satellite cloud image physical feature extraction module is the spliced ​​value of time difference and vorticity. The time difference includes cloud image difference and heat map difference. The input dimension is [T,C,H,W].

6. The method for predicting anomalous tropical cyclone tracks based on spatiotemporal multi-scale fusion and evidence learning according to claim 5, characterized in that, The spatiotemporal wavelet transform module includes a multi-scale spatial wavelet transform unit and a multi-scale temporal wavelet transform unit, and its specific operation process includes: The input features of the spatiotemporal transformation module are the output features of stage 2 of the ResNet18 network in the satellite cloud image physical feature extraction module, and the size of the output feature dimension is [missing information]. Then, the frequency domain features are sequentially passed through a series of multi-scale spatial wavelet transform units and multi-scale time wavelet transform units to output the frequency domain features, the size of which is... The input dimension of the multi-scale spatial wavelet transform unit is: The input features are decomposed into multiple scales using trainable Haar wavelet filters with different kernel sizes. First, the last two dimensions of the features are decomposed using zero values: high... Hekuan The kernel is padded so that its width and height are divisible by the kernel size. The padded width and height are denoted as h and w, respectively, resulting in a feature dimension of [T, C, h, w]. Then, an unfold operation is used to transform the spatial dimension into a sliding window format of [T, C, h, w, kernel_size, kernel_size]. Each window is convolved with four Haar wavelet bases to obtain wavelet coefficients [T, 4C, h, w]. Frequency components are weighted using a frequency attention mechanism, and the weighted wavelet features at each scale are then uniformly pooled. Then, the layers are stitched together along the channel dimension, and the dimensionality is reduced to [value] using a 1×1 convolutional layer. Output; The input of the multi-scale time wavelet transform unit is the output of the multi-scale spatial wavelet transform unit, with dimension [missing information]. Using predefined low-pass and high-pass filters, adjacent time steps are grouped together through a grouping operation, and their low-frequency and high-frequency wavelet coefficients are calculated, resulting in a dimension of [missing information]. The coefficients were then reshaped back to spatial dimensions. And convert it to a 3D convolution input format. Temporal convolution kernels are applied to extract features from low-frequency and high-frequency components respectively; Finally, it is obtained through channel splicing. Dimensionality reduced to 1×1×1 using 3D convolution with a kernel size of 1×1×1 Finally, reshape the dimension. Output.

7. The method for predicting anomalous tropical cyclone tracks based on spatiotemporal multi-scale fusion and evidence learning according to claim 6, characterized in that, The input features of the time-domain dimensionality reduction module are the output features of stage 4 of the ResNet18 network in the satellite cloud image physical feature extraction module, and the size of the output feature dimension is [missing information]. Then, dimensionality reduction is performed using a 3D convolution with a kernel size of 1×1×1. Output time-domain features.

8. The method for predicting anomalous tropical cyclone tracks based on spatiotemporal multi-scale fusion and evidence learning according to claim 7, characterized in that, The fusion pooling module includes a feature fusion unit and an attention pooling unit, and its specific operation process includes: The input to the fusion pooling module is the frequency domain feature output from the spatiotemporal wavelet transform module and the time domain feature output from the time domain dimensionality reduction module, both with a feature size of [missing value]. Then, the fused features are output sequentially through a series of feature fusion units and attention pooling units, and the size of the output fused features is [C]. The feature fusion unit first concatenates the frequency domain features and time domain features along the channel dimension to obtain... Then, a 1×1 convolutional layer is used to reduce the dimensionality to... Output; The input to the attention pooling unit is the output of the feature fusion unit, and the dimension is... First, reshape into The spatiotemporal structure of this unit employs a hierarchical pooling strategy that prioritizes spatial information over temporal information: the spatial attention layer calculates the importance scores of all spatial locations within each time step using two layers of multilayer perceptrons. After softmax normalization, the spatial dimensions are weighted and aggregated to obtain The temporal features are then calculated; subsequently, the temporal attention layer also uses a two-layer multilayer perceptron structure to calculate the attention weights [T,1] at each time step. After softmax normalization, the temporal dimension is weighted and pooled to finally output the global feature representation [C].

9. The method for predicting anomalous tropical cyclone tracks based on spatiotemporal multi-scale fusion and evidence learning according to claim 8, characterized in that, The evidence learning module includes an evidence feature extraction unit, an evidence regression unit, and an uncertainty calculation unit. Its specific operation process includes: The evidence learning module takes the output features of the fusion pooling module as input, with a feature size of [C]. Then, it passes through the evidence feature extraction unit, the evidence regression unit, and the uncertainty calculation unit in sequence to output the Dirichlet parameter α and the abnormal path prediction probability P. Among them, the evidence feature extraction unit extracts evidence features through a multilayer perceptron and a random deactivation layer [C]. The evidence regression unit regresses the evidence value [N] through a linear layer; the uncertainty calculation unit calculates the Dirichlet parameter α and the anomaly path prediction probability P, where the Dirichlet parameter α = evidence regression value + 1, and the anomaly path prediction probability P = α. i / ∑α i .

10. The method for predicting anomalous tropical cyclone tracks based on spatiotemporal multi-scale fusion and evidence learning according to claim 9, characterized in that, The training of the tropical cyclone anomalous track prediction model in step S4 includes: Loss function used total The formula is as follows: Loss total =w class ·Loss class +w ev ·Loss ev Among them, w class and w ev The dynamic weights are given by the following formula: w class =max(0.6,0.8-epoch*0.05) In ev =1-in class Where epoch represents the number of iterations; Loss class The classification loss is expressed by the following formula: Loss class =-log(P+ε) Where ε=1×10 -6 ; Loss ev The formula for representing evidence loss is as follows: Among them, Loss nll =-log(P+ε),ε=1×10 -6 ; Dir(α) represents the Dirichlet distribution of the Dirichlet parameter α, and Β(α) is the Beta function, Β(α) = ∏(Γ(α) i )) / Γ(∑α i ), where Γ(·) is the gamma function; If α > 0, then Ψ(α) = log(Γ(α)), where Ψ(α) measures the "sharpness" of the distribution. The parameter difference is represented by (Ψ(α)-Ψ(∑α)), which represents the relative importance of each category in the target distribution; α KL =max(0.1*(0.95**epoch),0.01) is the dynamic weight, initially: α KL ≈0.1, emphasizing learning under uncertainty; later stage: α KL →0.01, focus on accuracy.

Images