An interpretable typhoon track and intensity prediction method based on task-sensitive learning

Abstract

The present application belongs to the technical field of meteorological disaster prediction, and particularly relates to an interpretable typhoon path and intensity prediction method based on task-sensitive learning. The present application comprises constructing and training a prediction network model, extracting terrain features based on complete terrain elevation data through a dynamic terrain updating unit; obtaining cause-and-effect direction guided features based on task-customized terrain features; obtaining preliminary task-customized meteorological features based on typhoon historical attribute time series features, task identification one-hot encoding and original meteorological features; obtaining task-customized attribute meteorological features based on preliminary task-customized meteorological features and original meteorological features; fusing the task-customized terrain features, cause-and-effect direction guided features and task-customized attribute meteorological features with the typhoon historical attribute time series features, and inputting them into several prediction sub-tasks respectively to obtain prediction results of different tasks. The present application improves the accuracy, interpretability and optimization efficiency of typhoon path and intensity prediction.

Description

Technical Field

[0001] This invention belongs to the field of meteorological disaster forecasting technology, specifically involving an interpretable typhoon path and intensity prediction method based on task-sensitive learning. Background Technology

[0002] Currently, mainstream methods for predicting typhoon tracks and intensities mainly include numerical weather prediction (NMR) methods and deep learning-based methods. NMR models (such as ECMWF-IFS and CMA-GFS) simulate atmospheric physical processes numerically, exhibiting high physical consistency, but they require extremely high computational resources, making it difficult to meet real-time and efficient operational needs. With the development of artificial intelligence, deep learning-based typhoon track and intensity prediction methods have gradually gained attention due to their lower computational cost and considerable predictive performance. Early deep learning methods often used historical attribute sequences as input, failing to fully utilize environmental meteorological factors. Subsequent research has gradually incorporated reanalysis of meteorological data, significantly improving prediction accuracy.

[0003] However, existing deep learning methods still have significant shortcomings in typhoon track and intensity prediction. Firstly, most methods primarily focus on atmospheric environmental variables, neglecting the role of key environmental factors such as topography, especially during typhoon landfall and coastal evolution phases, where the impact of complex terrain on typhoon track and intensity is not effectively modeled. Secondly, existing methods typically employ a uniform environmental feature sharing mechanism, ignoring the heterogeneous environmental factor requirements of different typhoon attribute prediction tasks, thus limiting model interpretability and prediction accuracy. Furthermore, while sharing parameters (e.g., parameters of the environmental meteorological variable encoder) across multiple tasks in the typhoon track and intensity prediction sub-tasks facilitates information transfer between tasks, it also introduces the risk of gradient conflicts, which can slow model convergence, weaken the performance of some tasks, and affect the overall optimization effect of the model. To address these issues, there is an urgent need to propose new typhoon track and intensity prediction methods to effectively model environmental and topographical factors, thereby improving model interpretability and prediction accuracy. Summary of the Invention

[0004] The purpose of this invention is to address the problems in existing technical solutions, such as insufficient modeling of key environmental factors like topography in typhoon track and intensity prediction, failure to effectively distinguish the heterogeneous requirements of different prediction tasks for environmental factors, and gradient conflicts caused by parameter sharing across multiple tasks. This invention proposes an interpretable typhoon track and intensity prediction method based on task-sensitive learning, thereby improving the accuracy, interpretability, and optimization efficiency of typhoon track and intensity prediction.

[0005] To achieve the above objectives, the technical solution provided by this invention is as follows:

[0006] An interpretable typhoon track and intensity prediction method based on task-sensitive learning includes:

[0007] A prediction network model is constructed, which includes a dynamic terrain perception module and a task-sensitive environment routing module. The dynamic terrain perception module includes a dynamic terrain update unit and a terrain causal direction guidance unit; the task-sensitive environment routing module includes a differentiable channel selection unit and a cross-attention feature fusion unit.

[0008] Acquire complete terrain elevation data, and extract task-customized terrain features based on the complete terrain elevation data through dynamic terrain update units;

[0009] The task-customized terrain features are input into the terrain causal direction guidance unit to obtain the causal direction guidance features;

[0010] Acquire typhoon attribute data and environmental meteorological variables, input the environmental meteorological variables into the environmental meteorological variable encoder to obtain the raw meteorological features, preprocess the typhoon attribute data and input it into the typhoon historical attribute sequence encoder to extract the typhoon historical attribute time series features of different prediction sub-tasks respectively.

[0011] Based on the historical attributes and temporal characteristics of typhoons, the unique thermal coding of task identifiers, and the original meteorological characteristics, preliminary task-customized meteorological characteristics are obtained by selecting units through differentiable channels.

[0012] Based on the preliminary task-customized meteorological features and the original meteorological features, the task-customized attribute meteorological features are obtained through the cross-attention feature fusion unit;

[0013] Based on the task-customized meteorological features, task-customized terrain features, causal direction guidance features, and typhoon historical attribute time series features, the typhoon attribute prediction results for different prediction sub-tasks are obtained.

[0014] Furthermore, the acquisition of complete terrain elevation data, and the extraction of task-customized terrain features based on the complete terrain elevation data through a dynamic terrain update unit, includes:

[0015] At each forecast time, a local topographic elevation patch of a predetermined size is obtained by cropping from the complete topographic elevation data, centered on the latitude and longitude of the forecast typhoon center at the previous forecast time.

[0016] Local topographic elevation images along the horizontal and vertical directions respectively Applying the gradient operator yields the first spatial derivative. and ;

[0017] Calculate slope and dominant slope aspect based on the first-order spatial derivative;

[0018] Topographic roughness is calculated using the standard deviation of elevations within a local topographic elevation image;

[0019] Elevation, slope, dominant aspect, and terrain roughness are input into the terrain encoder and converted into elevation features, slope features, aspect features, and roughness features. Different combinations of elevation features, slope features, aspect features, and roughness features are used as task-customized terrain features.

[0020] Furthermore, the step of inputting task-customized terrain features into the terrain causal direction guidance unit to obtain causal direction guidance features includes:

[0021] Based on the task-customized terrain features, a terrain query vector is obtained, a terrain prototype dictionary is constructed, and the key vector and value vector of each prototype are obtained using the terrain prototype dictionary.

[0022] Calculate the correlation between the terrain query vector and the prototype key vector, determine the degree of matching between the task-customized terrain features and the prototype, select the prototype most relevant to the current terrain, and generate the corresponding causal mask.

[0023] Extract the corresponding terrain semantic information from the numerical vector of the selected prototype, fuse the terrain semantic information with the typhoon historical attribute time series features, task-customized attribute meteorological features and task-customized terrain features, and perform tensor multiplication with the causal mask to obtain causal branch features.

[0024] Causal branch features are used as causal direction guiding features.

[0025] Furthermore, the following constraints are imposed on the terrain causal direction guidance unit, including:

[0026] The complementary part of the causal mask is defined as a promiscuous mask;

[0027] The historical time-series features of typhoon attributes, the meteorological features of task-customized attributes, and the terrain features of task-customized attributes are fused together and tensor multiplication is performed with the hybrid mask to obtain hybrid branch features.

[0028] The causal branch features and the mixed branch features are respectively input into two directional classifiers with the same structure to obtain the probability prediction values ​​of each directional category;

[0029] The causal branch features are randomly shuffled and fused with the mixed branch features element by element, and then input into a random fusion multilayer perceptron to output the probability prediction value of the intervention branch.

[0030] A uniform distribution loss constraint is imposed on the probability predictions of the heterogeneous branches, and a loss constraint based on classification differences is introduced between the probability predictions of the causal and intervention branches and the true directional labels.

[0031] Furthermore, based on the historical time-series characteristics of typhoon attributes, the unique thermal encoding of the task identifier, and the original meteorological characteristics, preliminary task-customized meteorological characteristics are obtained through a differentiable channel selection unit, including:

[0032] Map the original meteorological features with C channels to a feature tensor with double the number of channels, 2C.

[0033] The task identifier features are obtained by mapping the one-hot encoding of the task identifier to a high-dimensional feature space using a structured task identifier encoder.

[0034] The historical typhoon attributes and time-series features of the prediction sub-tasks are concatenated with the corresponding task identification features and input into the channel correlation prompter to generate the channel correlation score vector for each prediction sub-task.

[0035] The top C channels with the highest scores in the channel correlation score vector of each prediction subtask are selected to obtain the preliminary task-customized meteorological features for each prediction subtask.

[0036] Furthermore, the process of selecting the top C channels with the highest scores from the channel correlation score vectors of each prediction sub-task yields the preliminary task-customized meteorological features for each prediction sub-task, including:

[0037] Gumbel noise is introduced into the channel correlation score vector, and the noise intensity is adjusted by scaling parameters to obtain the perturbed score vector;

[0038] The perturbated score vector is used to construct a cost matrix, and the cost matrix is ​​then sorted according to the temperature parameter. Scale the matrix and take the negative exponent to obtain the initial similarity matrix;

[0039] The Sinkhorn algorithm is used to normalize and project the initial similarity matrix to obtain a soft assignment matrix, and the soft selection probability is obtained based on the soft assignment matrix.

[0040] The soft selection probability is multiplied element-wise with the expanded feature tensor along the channel dimension, and then the number of channels is increased by a convolutional operation. Down to Preliminary customized meteorological characteristics for the task were obtained.

[0041] Furthermore, the task-customized attribute meteorological features obtained by the cross-attention feature fusion unit based on the preliminary task-customized meteorological features and the original meteorological features include:

[0042] Generate query vectors based on customized meteorological features from the initial task.

[0043] The original meteorological features are projected into key vectors and value vectors;

[0044] The correlation between the query vector and the key vector is calculated by the cross-attention feature fusion unit, and the value vector is weighted by attention weights to obtain the fused meteorological features.

[0045] The fused meteorological features are nonlinearly mapped through a feedforward network to output task-customized meteorological features.

[0046] Furthermore, based on task-customized meteorological features, task-customized terrain features, causal direction guidance features, and typhoon historical attribute time series features, the typhoon attribute prediction results for different prediction sub-tasks are obtained, including:

[0047] The customized meteorological features of different prediction subtasks are used as the initial state inputs of the corresponding prediction subtasks. The customized terrain features, causal direction guidance features and typhoon historical attribute time series features of the current moment are summed and fused element by element to obtain the fused features of different prediction subtasks. The fused features of different prediction subtasks and the prediction results of the previous moment are input into the corresponding prediction sub-branch. The hidden state of each prediction sub-branch is updated together with the network state of the previous moment to obtain the typhoon attribute prediction results of different prediction subtasks.

[0048] Furthermore, it also includes joint optimization of the prediction network model, including:

[0049] Before each backpropagation during the training process of the prediction network model, the loss of each prediction subtask is obtained, and the task affinity between each prediction subtask is calculated based on the gradient information of each task with respect to the shared parameters.

[0050] The loss weight for each task is dynamically calculated, and the loss weight is adaptively allocated based on task affinity.

[0051] Based on the aforementioned loss weights, the losses of each prediction sub-task are weighted and combined to form the prediction loss.

[0052] The parameters of the prediction network model are optimized based on the prediction loss and the constraints imposed on the terrain causal direction guidance unit.

[0053] Compared with the prior art, the significant advantages of this invention are: it can not only effectively model the impact of terrain factors on typhoon path and intensity, but also perform task-sensitive modeling for the heterogeneous requirements of different prediction tasks, and alleviate conflicts between tasks through joint optimization, thereby improving the interpretability and optimization efficiency of the model while improving prediction accuracy. Attached Figure Description

[0054] Figure 1 This is a flowchart of an interpretable typhoon track and intensity prediction method based on task-sensitive learning according to the present invention.

[0055] Figure 2 This is a schematic diagram of the predictive network model of the present invention. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0057] like Figure 1 As shown, an interpretable typhoon track and intensity prediction method based on task-sensitive learning is provided, including:

[0058] Step S1: Construct and train a prediction network model, which includes a dynamic terrain perception module, a task-sensitive environment routing module, and a joint optimization module based on task affinity.

[0059] The prediction network model constructed in this embodiment is as follows: Figure 2 As shown, it includes a dynamic terrain perception module, a task-sensitive environment routing module, and a task affinity-based joint optimization module. The dynamic terrain perception module includes a dynamic terrain update unit and a terrain causal direction guidance unit; the task-sensitive environment routing module includes a differentiable channel selection unit and a cross-attention feature fusion unit.

[0060] Training network models is a relatively mature technique in this field, and will not be elaborated upon here.

[0061] Step S2: Dynamically incorporate topographic factors of the typhoon's location into the typhoon path and intensity prediction process, including S2.1 dynamic topographic update unit and S2.2 topographic causal direction guidance unit.

[0062] Step S2.1: Preload complete topographic elevation data of the target area, the Northwest Pacific, into memory to efficiently support subsequent dynamic cropping and feature extraction operations. At each prediction time, based on the typhoon's current location (i.e., the latitude and longitude of the typhoon center at the previous prediction time; at the initial time, this location corresponds to the last time in the typhoon path observation sequence), a local topographic elevation patch of a fixed size is cropped from the complete topographic data, centered on this location. This ensures that the cut-off area fully covers the typhoon's activity range and its potential path.

[0063] In one specific embodiment, the geographic extent of the cropping window is set to a region of 5.12° × 5.12°.

[0064] After obtaining local terrain elevation images, the following additional terrain attribute extraction operations are performed:

[0065] Local topographic elevation images along the horizontal and vertical directions respectively Applying the gradient operator yields the first spatial derivative. and In one specific embodiment, the gradient operator is the Sobel operator.

[0066] The slope, which characterizes the steepness of the terrain, is calculated based on the above derivative. Slope direction, which characterizes topographic relief :

[0067]

[0068] in, It is the arctangent function in the four quadrants.

[0069] Using local topographic elevation images Calculating terrain roughness using the standard deviation of internal elevation .

[0070] The terrain attributes such as elevation, slope, aspect, and roughness are input into the terrain encoder, and each terrain attribute is converted into a unified high-dimensional feature representation.

[0071] In one specific embodiment, the terrain encoder consists of four convolutional modules (each module consisting of a two-dimensional convolutional layer, a batch normalization layer, and a LeakyReLU activation function, respectively), an adaptive average pooling layer, and a fully connected layer. The convolutional layers are used to progressively extract the spatial texture and shape features of the input terrain raster; the batch normalization layer stabilizes the distribution and accelerates convergence; and LeakyReLU enhances nonlinear representation and alleviates the "neuron death" phenomenon. The four-level convolutional modules achieve multi-scale compression and feature convergence through stride downsampling. The adaptive average pooling layer aggregates the spatial dimensions into a global semantic description, and finally, a fixed-length terrain encoding feature vector is output through the fully connected layer.

[0072] The set of trainable parameters for the entire terrain encoder is denoted as . The parameters include the weights and biases of each convolutional layer, the scaling and translation parameters of the batch normalization layer, and the weights and biases of the fully connected layer, all of which are floating-point learnable variables. The input channel count, output channel count, kernel size, stride, and padding width of the four 2D convolutional layers are set to (1, 16, 5, 2, 2), (16, 32, 5, 2, 2), (32, 64, 5, 2, 2), and (64, 128, 5, 2, 2), respectively. Each convolutional layer is followed by a batch normalization layer with matched channel counts (16, 32, 64, and 128 channels respectively) and a LeakyReLU activation function. An adaptive average pooling layer compresses feature maps of arbitrary spatial dimensions into a (1×1) global representation, and then a fully connected layer transforms the 128-dimensional channel features into 32-dimensional encoded features at output.

[0073] This terrain encoder structure ensures good adaptability to different input sizes while realizing joint modeling of global and local terrain information, and finally outputs a fixed-length representation vector of the terrain data.

[0074] Differentiated combinations of terrain features are employed to address different objectives in typhoon track and intensity prediction. In one specific embodiment, the prediction of the typhoon center's longitude and latitude utilizes a combination of elevation, slope, and aspect features; the prediction of the minimum central pressure uses only elevation features; and the prediction of the maximum wind speed near the center combines elevation and roughness features. The resulting task-customized terrain features are then used as input to subsequent network modules to support the prediction of various attributes.

[0075] Step S2.2: Project the task-customized terrain features of the longitude and latitude prediction task in Step S2.1 onto a high-dimensional feature space using a linear mapping method to obtain a terrain query vector. Simultaneously, construct and train a learnable terrain prototype dictionary. This dictionary contains multiple prototypes representing different terrain types. Each prototype obtains its key vector and numerical vector in the high-dimensional feature space through linear mapping to characterize its features and its impact on typhoon path changes.

[0076] In one specific embodiment, a discrete sampling cross-attention mechanism based on Gumbel-Softmax is employed to calculate the correlation between the terrain query vector and the key vector of the prototype, determine the matching degree between the task-customized terrain features and the prototype, select the prototype most relevant to the current terrain, and construct a corresponding causal mask based on the similarity weight distribution. Specifically, based on the calculated similarity weights between the terrain features and each prototype, Softmax normalization is used to obtain the attention distribution of each prototype; then, based on the index positions of the prototypes with the highest weights in this distribution, a weighted selection matrix, i.e., the causal mask, is calculated. This mask corresponds to the terrain features in the tensor dimension and is used to perform weighted filtering of terrain features in the subsequent feature fusion process, selecting terrain feature components that have a significant causal correlation with the typhoon path change. The complementary part of this causal mask is defined as a confounding mask, used to reflect non-causal confounding components in the terrain that may introduce errors.

[0077] Furthermore, based on the correlation results, the corresponding terrain semantic information is extracted from the numerical vector of the selected prototype and fused with the typhoon historical attribute temporal features, task-customized attribute meteorological features, and task-customized terrain features. The comprehensive features that integrate the typhoon historical attribute temporal features, task-customized attribute meteorological features, and task-customized terrain features are then multiplied with the causal mask and the hybrid mask respectively by tensor multiplication, specifically element-wise multiplication, to obtain the features of the causal branch and the hybrid branch.

[0078] In one specific embodiment, the features of specific causal branches and hybrid branches are input into two structurally identical direction classifiers, essentially multilayer perceptrons, composed of two cascaded linear transformation modules. Each linear transformation module sequentially includes a normalization layer, a fully connected linear layer, and a nonlinear activation layer. Specifically, the input features first undergo the first normalization and fully connected linear transformation, then are processed by the ReLU activation function, followed by the second normalization and fully connected linear transformation. Finally, the output is normalized using logarithmic softmax to obtain the probability prediction value for each direction category. This structure can achieve classification and prediction of the future development direction of typhoons.

[0079] To enhance the causal discrimination capability of this unit, a stochastic fusion multilayer perceptron is implemented. During forward inference, this multilayer perceptron randomly shuffles the causal branch features, fusing them element-wise with the mixed branch features. The resulting data is then processed sequentially through normalization, fully connected layers, ReLU activation, normalization, another fully connected layer, and logarithmic softmax normalization; this represents the intervention branch. This process, by introducing a stochastic feature fusion mechanism, effectively simulates the prediction results under causal intervention, improving the model's ability to model the causality of directional predictions.

[0080] The trainable parameter set of this multilayer perceptron includes the weights and biases of each layer, all of which are floating-point learnable variables. The input and output dimensions of the two fully connected layers are (64, 64) and (64, 8), respectively, where 8 represents the number of typhoon future development direction classifications, ranging from 0 to 7 representing east, southeast, south, southwest, west, northwest, north, and northeast. The input dimensions of the normalization layers are consistent with the output of the previous layer.

[0081] To improve the causal accuracy of direction prediction, a uniform distribution constraint is imposed on the probability predictions of the confounding branches during model training. This constraint is mitigated by minimizing the KL divergence with the uniform distribution. Simultaneously, cross-entropy loss is applied between the probability predictions of the causal and intervention branches and the true direction labels, guiding the model to focus more on the causal role of terrain. Finally, the prediction results of the causal direction branches are extracted as causal direction guidance features to guide the specific coordinate prediction process for typhoon longitude and latitude, thereby improving the accuracy of typhoon path prediction.

[0082] Step S3: Incorporate environmental meteorological variables into the typhoon track and intensity prediction process in a task-adaptive selection manner, including S3.1 Differentiable channel selection unit and S3.2 Cross-attention feature fusion unit.

[0083] Step S3.1: Input the environmental meteorological variables into the environmental meteorological variable encoder, extract the features of each meteorological variable, and convert them into a unified high-dimensional feature representation.

[0084] In one specific embodiment, the environmental meteorological variable encoder consists of six independent convolutional branches (corresponding to geopotential height, meridional wind components, zonal wind components, temperature, sea surface temperature, and total column water vapor, respectively). Each branch is composed of a two-dimensional convolutional layer, a batch normalization layer, a LeakyReLU activation function, and a two-dimensional pooling layer in sequence, followed by a flattening and fully connected layer to achieve vectorization. The convolutional layers are used to extract the spatial features of the input meteorological raster step by step, the batch normalization layer is used to stabilize the feature distribution and accelerate convergence, and the LeakyReLU activation function is used to enhance the nonlinear expression. The pooling layer achieves multi-scale compression and feature aggregation through spatial downsampling, and the flattening and fully connected layers further transform the local features of each meteorological variable into fixed-length embedding vectors.

[0085] The set of trainable parameters for the entire environmental meteorological variable encoder is denoted as . The convolutional layer includes weights and biases for each convolutional layer, scaling and translation parameters for batch normalization layers, and weights and biases for fully connected layers, all of which are floating-point learnable variables. For the four types of multi-channel input variables across four pressure layers—potential height, meridional wind component, zonal wind component, and temperature—the convolutional layer parameters (number of input channels, number of output channels, kernel size, stride, and padding width) are set to (4, 16, 3, 1, 1), respectively. For the two types of single-channel input variables—sea surface temperature and total column water vapor—the convolutional layer parameters (number of input channels, number of output channels, kernel size, stride, and padding width) are set to (1, 16, 3, 1, 1), respectively. Each convolutional layer is followed by a batch normalization layer with matched channel numbers and a LeakyReLU activation function, where the LeakyReLU slope parameter is set to 0.01. For pooling operations, an average pooling layer is used to compress the spatial dimension to an 8×8 feature representation. Subsequently, the output of each branch is flattened to obtain a vector of length 1024, which is then input into a fully connected layer and mapped to an embedding feature vector of dimension 16. Finally, the embedding feature vectors obtained from each branch are concatenated to obtain the original meteorological features.

[0086] This design ensures the structural uniformity and comparability of different meteorological variables, while maintaining the actual physical meaning of meteorological characteristic channels.

[0087] Next, by expanding the number of channels through grouped convolution, the original meteorological features with C channels obtained from the environmental meteorological variable encoder are mapped to a feature tensor with double the number of channels, 2C. This allows for greater flexibility in the adaptive selection of subsequent meteorological variables. In a specific embodiment, the grouped convolution uses 1×1 grouped convolution to achieve channel expansion. Its convolutional layer parameters (number of input channels, number of output channels, kernel size, stride, and padding width) are set to (96, 192, 1, 1, 1), with 6 groups, corresponding to the parallel processing of 6 types of meteorological physical quantities.

[0088] To guide the channel selection process, task-sensitive guidance information needs to be constructed. Typhoon attribute data is preprocessed and input into a typhoon historical attribute sequence encoder to extract the temporal features of typhoon historical attributes for different prediction sub-tasks. A structured task identifier encoder is used to map the task identifier one-hot into a high-dimensional feature space, obtaining the task identifier features. The temporal features of the typhoon historical attributes of the prediction sub-task are concatenated with the corresponding task identifier features and input into a channel correlation indicator to generate a channel correlation score vector. This score reflects the correlation between the current prediction sub-task and different meteorological channels.

[0089] In one specific embodiment, the typhoon historical attribute sequence encoder consists of four single-layer long short-term memory recurrent neural networks, each processing historical attribute sequences for different subtasks. The structured task identifier encoder consists of a single fully connected layer with parameters (input dimension, output dimension) set to (4, 32). The channel correlation cue consists of a single fully connected layer and a LeakyReLU activation function, with the fully connected layer parameters (input dimension, output dimension) set to (96, 96) and the LeakyReLU slope parameter set to 0.01.

[0090] Based on the channel relevance score vector obtained from the channel relevance cue, the top C channels with the highest scores are considered most relevant to the subtask. Therefore, it is necessary to select the top C channels with the highest scores. However, the ordinary Top-C operation is a hard selection, which leads to non-differentiability and hinders end-to-end training. This unit models the channel selection problem as an optimal transport problem and uses a differentiable approximation method to achieve channel selection. Random perturbations are added to the channel scores to construct the cost matrix between the channel and the "selected / unselected" state. The cost matrix is ​​then iteratively normalized using the Sinkhorn algorithm to obtain a soft assignment matrix that satisfies the marginal constraints, thereby achieving an approximate Top-C selection.

[0091] In one specific embodiment, in the channel correlation score vector Introducing Gumbel noise By scaling parameters Adjust the noise intensity to enhance the randomness of the selection process and the diversity of training:

[0092]

[0093] in, Indicates the first The perturbed score vector of each prediction subtask Indicates the first The channel correlation score vectors for each prediction subtask are as follows: typhoon center longitude prediction, typhoon center latitude prediction, typhoon center minimum pressure prediction, and typhoon near-center maximum wind speed prediction. Each subtask corresponds to an independent set of channel correlation score vectors and its subsequent perturbation and soft selection processes, thereby ensuring that each prediction subtask can obtain differentiated, task-customized meteorological characteristics.

[0094] Then, the perturbed score vector Constructing a cost matrix The two columns represent the cost of assigning each channel to the "unselected" and "selected" states, respectively:

[0095]

[0096] in, Indicates the first The cost matrix of each prediction subtask.

[0097] To facilitate optimal transport modeling, the cost matrix is ​​adjusted according to the temperature parameter. Scaling and taking the negative exponent yields the initial similarity matrix:

[0098]

[0099] in, Indicates the first The initial similarity matrix of each prediction subtask.

[0100] Using the Sinkhorn algorithm Performing normalized projection yields a soft allocation matrix that satisfies the edge constraints:

[0101]

[0102] in, Indicates the first The soft assignment matrix for each prediction subtask. The target column edge constraint represents two states (not selected, selected).

[0103] The Sinkhorn algorithm approximates the optimal solution through iterative row and column normalization. For ease of representation, the following details the specific iterative process. Corresponding to the above This refers to taking the soft allocation matrix corresponding to a specific prediction subtask as an example. The specific iterative process is as follows:

[0104] Repeat execution step:

[0105]

[0106]

[0107] in, This represents element-wise division. It is the numerical stability constant. Representation matrix The Column vectors are used for summing column by column. Representation matrix The Row vectors are used for summing row by row.

[0108] After iteration, the soft allocation matrix is ​​obtained. The second column This refers to the soft-select probability of each channel, reflecting the degree of preference of the current forecasting subtask for each meteorological channel.

[0109] Ultimately, the soft choice probability With the expanded feature tensor Element-wise multiplication in the channel dimension:

[0110]

[0111] in, Indicates the first The environmental and meteorological characteristics of each prediction subtask after soft selection Indicates the first The soft-choice probability of each prediction subtask Indicates the first The expanded feature tensor of each prediction subtask.

[0112] Then, another convolution operation is used to increase the number of channels from... Down to Preliminary customized meteorological characteristics for the task were obtained. This serves as the input for subsequent feature fusion steps. In a specific embodiment, the parameters of this convolutional layer (number of input channels, number of output channels, kernel size, stride, and padding width) are set to (192, 96, 1, 1, 1), respectively.

[0113] This unit can automatically capture meteorological factors that are highly relevant to different forecasting tasks, significantly enhancing the flexibility and adaptability of multi-task forecasting while preserving interpretability.

[0114] Step S3.2: The preliminary task-customized meteorological features obtained through the differentiable channel selection unit are projected through a fully connected layer to generate a query vector. Simultaneously, the unselected original meteorological features are projected into key vectors and value vectors, respectively. Subsequently, the correlation between the query vector and the key vector is calculated using a cross-attention mechanism to obtain an unnormalized attention similarity matrix. Attention weights are then used to weight the value vectors, thereby achieving the fusion of the original meteorological information and mitigating the risk of key information loss and overfitting that may result from relying solely on adaptively selected meteorological variables.

[0115] To further enhance the model's generalization ability and avoid over-reliance on a few variables, this unit introduces a channel discarding mechanism in cross-attention. A binary mask is applied to the attention similarity matrix, retaining only the attention scores corresponding to some key channels, while setting the scores of other channels to zero.

[0116] In one specific embodiment, a top-retention operation is performed row-by-row on the calculated attention similarity matrix, that is, only the top-ranked items in each row are retained. The position corresponding to the maximum similarity score The discard ratio is set to 0.1, resulting in a binary mask matrix where the retained positions are 1 and the rest are 0. This binary mask matrix is ​​then applied to the unnormalized attention matrix, maintaining the original values ​​for the retained positions and assigning values ​​to the discard positions. The attention matrix after this processing is applied... After normalization, a sparse attention weight matrix is ​​obtained. The weights at the discard positions will be... Normalization forces compression to 0. Finally, the value vectors are weighted and summed using this weight matrix to obtain the fused meteorological features.

[0117] The fused meteorological features are nonlinearly mapped through a feedforward network to output task-customized attribute meteorological features, enhancing feature representation capabilities. In a specific embodiment, the feedforward network consists of: one layer... Convolution is used for channel expansion, one layer Depth convolution is used for local spatial modeling, one layer of GELU activation function is used for nonlinear mapping and recalibration, and one layer of... Convolution is used for channel reconstruction. Specifically, the output of depthwise convolution is divided into two branches along each channel. and Through gating Nonlinear mapping and recalibration are implemented. The parameters of the feedforward network convolutional layers (number of input channels, number of output channels, kernel size, stride, padding width) are set as follows: the first convolutional layer is (96, 384, 1, 1, 0); the second convolutional layer is (384, 384, 3, 1, 1), with channel-wise grouping (groups=384); the third convolutional layer is (192, 96, 1, 1, 0). All three convolutional layers use a learnable bias term (bias=bias).

[0118] The customized meteorological features obtained in step S3 will be used as input for subsequent branches of typhoon track and intensity prediction.

[0119] Step S4: Dynamically fuse the task-customized terrain features, causal direction guidance features, task-customized meteorological attribute features, and typhoon historical attribute time series features, and input them into multiple parallel sub-branches used to predict the typhoon center longitude, center latitude, center minimum pressure, and near-center maximum wind speed, thereby achieving synchronous prediction of typhoon path and intensity.

[0120] Specifically, the task-customized meteorological features are used as the initial state input for each corresponding sub-task. In the subsequent prediction process, at each prediction time, the task-customized terrain features, causal direction guidance features, and typhoon historical attribute time-series features corresponding to the current time are summed and fused element by element to obtain the fused features of different tasks. These features are then combined with the network state of the previous time step (the hidden state and internal memory information of the long short-term memory recurrent neural network in the previous time step) to update the hidden state of the prediction unit. This achieves the temporal interaction and joint modeling of terrain, meteorological, and historical information, and finally outputs the typhoon attribute prediction results for different prediction sub-tasks.

[0121] Each target attribute prediction subtask employs a Long Short-Term Memory (LSTM) recurrent neural network (RNN) structure, using an input sequence composed of fused features from each time step as input to recursively model the temporal variation characteristics of the target attribute. At each prediction time step, the output of the previous time step and the fused features of the current position are input into the corresponding LSM RNN. After updating the hidden state, a fully connected layer completes the numerical mapping, outputting the predicted value of the target attribute at that time step. Repeating this process chronologically yields the attribute prediction sequence for the entire prediction time window (i.e., the typhoon attribute prediction results for different prediction subtasks). Finally, the outputs of the four sub-branches together constitute the typhoon track and intensity prediction results.

[0122] In one specific embodiment, the typhoon path and intensity prediction branch includes four parallel Long Short-Term Memory (LSTM) recurrent neural network (RNN) decoders, corresponding to the typhoon center longitude, center latitude, minimum central pressure, and maximum near-center wind speed, respectively. Each LSN decoder has an input dimension of 32, a hidden layer dimension of 64, and a network layer count of 1. Each fully connected layer has an input dimension of 64 and an output dimension of 1, used to map the hidden state output of the LSN decoder to a single attribute value.

[0123] Step S5: Calculate the error between the predicted longitude, latitude, minimum central pressure, and maximum near-center wind speed of the typhoon center output in Step S4 and the corresponding actual values ​​to obtain the loss value for each task. Calculate the gradient of each task's loss in the shared parameter subspace, and construct a task affinity matrix based on the gradient information to measure the cooperative and conflicting relationships between different tasks in the shared parameter space. Then, by inductively calculating the affinity matrix, obtain the weight score for each task, and dynamically normalize and adjust the loss weights in each training batch. Finally, combine the losses of each task according to the weights obtained through adaptive learning to form the overall prediction loss, thereby achieving joint optimization of the loss function. This effectively suppresses gradient conflicts and enhances the cooperative effect between tasks while ensuring the interaction of shared parameter information.

[0124] In one specific embodiment, firstly for each prediction task Calculate its value in the shared parameter subspace. loss gradient on: Mission losses The L2 norm is used for calculation.

[0125] To quantify the interaction relationships between different tasks in the shared space, an asymmetric task affinity matrix is ​​constructed. ( This indicates the number of prediction tasks (specifically 4 in this application), and its elements are defined as:

[0126]

[0127] All prediction subtasks constitute a set. ={Typhoon center longitude prediction, typhoon center latitude prediction, typhoon center minimum pressure prediction, typhoon near-center maximum wind speed prediction} Used to identify different prediction subtasks, each corresponding to a set. The index of each prediction subtask in the [reference]. Indicates the first Each prediction subtask shares a parameter subspace. The loss gradient on the , where the numerator represents the prediction subtask. With prediction subtask Gradient alignment in the shared parameter space, the denominator is used for normalization to eliminate the influence of gradient magnitude, and a constant. Numerical stability is guaranteed. Because the dot product form is asymmetric, this matrix can characterize the unidirectional influence relationship between tasks.

[0128] Next, the matrix Sum each row to get the task. Original influence score:

[0129]

[0130] in, Indicates the first The raw influence score for each prediction subtask.

[0131] Subsequently, a learnable direction vector is introduced. ,right Scaling is performed, and the softplus function is used to ensure that the result is non-negative and robust to outliers:

[0132]

[0133] in, Indicates the first The task weights are obtained by scaling the original influence scores of each prediction subtask.

[0134] After obtaining the task weight Then, it is used for standardization to obtain the normalized weights of each task. :

[0135]

[0136] in, Indicates the first Normalized weights for each prediction subtask.

[0137] Finally, the multi-tasking loss can be expressed as:

[0138]

[0139] in, This represents the multi-task loss for each prediction subtask.

[0140] In each backpropagation step of model training, the weights The optimization priority of multiple tasks is dynamically adjusted based on the task affinity of the current batch. This mechanism prioritizes tasks with strong synergy and suppresses tasks with severe gradient conflicts, thereby achieving effective collaboration between tasks during the optimization process of shared parameters and significantly improving the overall performance and convergence efficiency of typhoon track and intensity prediction.

[0141] The training process for the prediction network model in this application is as follows:

[0142] First, the dataset was preprocessed. Typhoon attribute data came from the best track dataset (CMA-BST, China Meteorological Administration – Best Track Dataset), environmental meteorological variables came from ERA5 (ECMWF Reanalysis 5th Generation) reanalysis data, and topographic factors came from high-resolution complete topographic elevation data of the Northwest Pacific region. The complete dataset containing these three data types covers 1861 typhoon cases in the Northwest Pacific Ocean from 1950 to 2024. Among them, the complete dataset from 1950 to 2016 was divided into training and validation sets, with 80% used for training and 20% for validation; the complete typhoon data from 2023 to 2024 served as an independent test set. The study area covers longitudes from 100°E to 179.75°E and latitudes from 0° to 59.75°N. Environmental meteorological variables include geopotential height, temperature, zonal wind, and meridional wind at four pressure levels (200 hPa, 500 hPa, 700 hPa, and 850 hPa), as well as sea surface temperature and total column water vapor, covering a 25°×25° area around the typhoon center with a spatial resolution of 0.25°. Topographic factors are dynamically trimmed by dynamic topographic update units, with a trimmed block size of 5.12°×5.12° around the typhoon center and a resolution of 0.01° to account for the influence of topography on the typhoon's track and intensity evolution. All input data are normalized before training to improve the convergence efficiency of the prediction network model.

[0143] During the training phase, the prediction network model was implemented using the PyTorch framework, and experiments were conducted using an NVIDIA RTX A6000 GPU. The optimizer was Adam, with an initial learning rate of 0.0001, a batch size of 64, and 100 training epochs. The historical input phase was set to 48 hours (i.e., 8 historical moments with a time resolution of 6 hours), and the prediction phase was set to 72 hours (i.e., 12 future moments).

[0144] The historical typhoon attribute sequence, obtained after preprocessing the typhoon attribute data, is input into the typhoon historical attribute sequence encoder to obtain the temporal features of the historical typhoon attributes. Topographic factors are input into the dynamic terrain perception module, and after passing through the dynamic terrain update unit and the terrain causal direction guidance unit, task-related terrain features and causal direction guidance features are extracted. Environmental meteorological variables are input into the environmental meteorological variable encoder, and after feature extraction and channel expansion, they enter the task-sensitive environment routing module. After passing through the differentiable channel selection unit and the cross-attention feature fusion unit, task-customized meteorological features are obtained. Subsequently, the terrain features, causal direction guidance features, task-customized meteorological features, and the temporal features of the historical typhoon attributes are concatenated and fused as the input for the typhoon path and intensity prediction branch.

[0145] In the prediction phase, the prediction network model comprises four parallel prediction subtasks, corresponding to the typhoon center longitude, center latitude, minimum central pressure, and maximum near-center wind speed, respectively. Each sub-branch employs a long short-term memory recurrent neural network structure to perform temporal modeling on the input task-customized features, outputting a prediction sequence for multiple future time points. The mean squared error is calculated between the prediction results and the actual observations to obtain the loss function for each task.

[0146] During the joint optimization phase, the task affinity-based joint optimization module calculates the gradient of each task's loss in the shared parameter subspace, constructs a task affinity matrix, dynamically measures the cooperation and conflict relationships between tasks, and then adjusts the loss weights in real time to obtain the weighted overall prediction loss, which serves as the optimization objective for updating network parameters. Furthermore, by incorporating constraints imposed on the terrain causal direction guidance unit, the prediction network model is jointly optimized, and the terrain prototype dictionary is updated.

[0147] Through the above process, the prediction network model can simultaneously consider the combined effects of historical attributes, environmental meteorological factors, and topographical factors, and achieve efficient prediction of typhoon track and intensity under the constraints of a joint optimization mechanism. After training until convergence, the prediction network model with the best performance on the validation set is selected for testing and actual prediction.

[0148] This application also provides experimental data, comparing the proposed technical solution with traditional numerical weather prediction methods (IFS used by the European Centre for Medium-Range Weather Forecasts (ECMWF), large meteorological models (Pangu model), and other deep learning methods (GRU, GBRNN, MMSTN, MGTCF, TC-Diffuser, and TCN_M). The performance metrics used for comparison are the absolute errors between the predictions and actual data for each task, including trajectory (km), pressure (hPa), and wind speed (m / s).

[0149] The definition of a coastal typhoon is as follows: if the land coverage of the area surrounding the typhoon (within a 4° × 4° range) exceeds 50% within any consecutive 48-hour period of a 72-hour window, then the typhoon is considered a coastal (nearshore or landfall) typhoon.

[0150] Table 1 Comparison of prediction performance under coastal typhoon scenarios

[0151]

[0152] Table 2 Comparison of prediction performance under all typhoon scenarios

[0153]

[0154] The comparison results are shown in Table 1 and Table 2:

[0155] As can be seen from the comparison results in Table 1, in the scenario of coastal typhoons, the technical solution of this application outperforms the traditional technical solution in terms of forecast performance of trajectory, pressure and wind speed.

[0156] As can be seen from the comparison results in Table 2, under all typhoon scenarios, the technical solution of this application is superior to the traditional technical solution in terms of short-term (within 24 hours) track forecast performance and intensity forecast performance for all forecast lead times.

[0157] Among them, IFS (Integrated Forecasting System) is the core of the numerical weather prediction system used by the European Centre for Medium-Range Weather Forecasts.

[0158] GRU (Gated Recurrent Unit): It is a type of RNN (Recurrent Neural Network), a variant of LSTM, and has a simpler structure.

[0159] GBRNN (Graphical Bidirectional Recurrent Neural Network): It is good at processing non-Euclidean data (such as weather station networks and irregular grids) and can learn the contextual information of sequences from both the "past to future" and "future to past" directions at the same time.

[0160] MMSTN (Multimodal Spatiotemporal Network): Typically, the spatial features of satellite images and the temporal series features of the atmospheric environment are extracted separately through designed network branches, and then they are fused for prediction.

[0161] MGTCF (Multimodal Spatiotemporal Fusion Network): By fusing multi-source, multimodal data (e.g., atmospheric environmental fields in reanalysis data) and utilizing a spatiotemporal attention mechanism, it jointly learns the complex relationship between this information and typhoon track and intensity, thereby achieving more accurate typhoon track and intensity prediction.

[0162] TC-Diffuser (Tropical Cyclone Diffuser): A method for predicting typhoon tracks and intensities based on a diffusion model. It generates multiple possible and reasonable future typhoon tracks and intensities from random noise through a denoising process.

[0163] TCN_M (Temporal Convolutional Networks - Meteorological Applications): is an architecture that uses temporal convolutional networks to process time series data.

[0164] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A method for interpretable typhoon track and intensity prediction based on task-sensitive learning, characterized in that, The method comprises the following steps: constructing a prediction network model, the prediction network model comprising a dynamic terrain perception module and a task-sensitive environment routing module, wherein the dynamic terrain perception module comprises a dynamic terrain updating unit and a terrain causal direction guiding unit; the task-sensitive environment routing module comprises a differentiable channel selection unit and a cross-attention feature fusion unit; obtaining complete terrain elevation data, and extracting task-customized terrain features based on the complete terrain elevation data through the dynamic terrain updating unit; inputting the task-customized terrain features into the terrain causal direction guiding unit to obtain causal direction guiding features; obtaining typhoon attribute data and environmental meteorological variables, inputting the environmental meteorological variables into an environmental meteorological variable encoder to obtain original meteorological features, and preprocessing the typhoon attribute data and inputting the typhoon historical attribute sequence encoder to extract typhoon historical attribute time series features of different prediction sub-tasks; based on the typhoon historical attribute time series features, task identification one-hot encoding and original meteorological features, obtaining preliminary task-customized meteorological features through the differentiable channel selection unit; based on the preliminary task-customized meteorological features and the original meteorological features, obtaining task-customized attribute meteorological features through the cross-attention feature fusion unit; based on the task-customized attribute meteorological features, the task-customized terrain features, the causal direction guiding features and the typhoon historical attribute time series features, obtaining typhoon attribute prediction results of different prediction sub-tasks; wherein, the inputting of the task-customized terrain features into the terrain causal direction guiding unit to obtain the causal direction guiding features comprises: obtaining a terrain query vector based on the task-customized terrain features, constructing a terrain prototype dictionary, and obtaining a key vector and a numerical vector of each prototype using the terrain prototype dictionary; calculating the correlation between the terrain query vector and the key vector of the prototype, determining the matching degree of the task-customized terrain features and the prototype, and selecting the most relevant prototype to the current terrain to generate a corresponding causal mask; extracting corresponding terrain semantic information from the numerical vector of the selected prototype, fusing the terrain semantic information with the typhoon historical attribute time series features, the task-customized attribute meteorological features and the task-customized terrain features, and performing tensor multiplication with the causal mask to obtain causal branch features; the causal branch features are used as the causal direction guiding features; the following constraints are imposed on the terrain causal direction guiding unit, including: the complementary part of the causal mask is defined as a mixed mask; fuse the typhoon historical attribute time series features, the task-customized attribute meteorological features and the task-customized terrain features, and perform tensor multiplication with the mixed mask to obtain mixed branch features; inputting the causal branch features and the mixed branch features into two direction classifiers with the same structure to obtain probability prediction values of each direction class; fuse the causal branch features and the mixed branch features by randomly shuffling and element-wise addition, and input them into a random fusion multi-layer perception machine to output probability prediction values of an intervention branch; the probability prediction values of the mixed branch are subjected to a uniform distribution loss constraint, and the probability prediction values of the causal branch and the intervention branch are subjected to a loss constraint based on the classification difference between the real direction labels. 2.The task-sensitive learning based interpretable typhoon track and intensity prediction method according to claim 1, wherein, The complete terrain elevation data is acquired, and task customized terrain features are extracted from the complete terrain elevation data by a dynamic terrain updating unit, including: At each prediction time, a local terrain elevation patch of a predetermined size is cropped from the complete terrain elevation data with the predicted typhoon center longitude and latitude of the last prediction time as the center; respectively along a horizontal direction and a vertical direction applying a gradient operator to obtain first order spatial derivatives and ; The slope and dominant direction slope are calculated based on the first-order spatial derivative; The terrain roughness is calculated using the standard deviation of the elevations in the local terrain elevation patch; The elevation, slope, dominant direction slope and terrain roughness are input into a terrain encoder to be converted into elevation features, slope features, slope direction features and roughness features, and different combinations of the elevation features, slope features, slope direction features and roughness features are used as task customized terrain features. 3.The task-sensitive learning based interpretable typhoon track and intensity prediction method according to claim 1, wherein, The preliminary task customized meteorological features are obtained from the typhoon historical attribute time series features, task identification one-hot encoding and original meteorological features by a differentiable channel selection unit, including: The original meteorological features with a channel number of C are mapped to a feature tensor with a doubled channel number of 2C; The task identification one-hot encoding is mapped to a high-dimensional feature space using a structured task identification encoder to obtain task identification features; The typhoon historical attribute time series features of the prediction subtasks are spliced with the corresponding task identification features, and input into a channel correlation prompter to generate a channel correlation score vector for each prediction subtask; The top C channels with the highest scores in the channel correlation score vector of each prediction subtask are selected to obtain the preliminary task customized meteorological features of each prediction subtask.

4. The interpretable typhoon track and intensity prediction method based on task-sensitive learning according to claim 3, wherein, The preliminary task customized meteorological features of each prediction subtask are obtained by selecting the top C channels with the highest scores in the channel correlation score vector of each prediction subtask, including: Gumbel noise is introduced into the channel correlation score vector, and the noise intensity is adjusted by a scaling parameter to obtain a perturbed score vector; The perturbed score vector is structured into a cost matrix, which is scaled and exponentiated negative to obtain an initial similarity matrix The perturbed score vector is structured into a cost matrix, which is scaled and exponentiated negative to obtain an initial similarity matrix The Sinkhorn algorithm is used to normalize and project the initial similarity matrix to obtain a soft assignment matrix, and the soft selection probability is obtained based on the soft assignment matrix; The soft selection probability is multiplied with the extended feature tensor element by element in the channel dimension, and a convolution operation is performed to reduce the number of channels from to , to obtain preliminary task customized weather features.

5. The interpretable typhoon track and intensity prediction method based on task-sensitive learning of claim 1, wherein, The task customized attribute meteorological features are obtained from the preliminary task customized meteorological features and the original meteorological features by a cross-attention feature fusion unit, including: A query vector is generated based on the preliminary task customized meteorological features; The original meteorological features are projected into a key vector and a value vector; The correlation between the query vector and the key vector is calculated by the cross-attention feature fusion unit, and the value vector is weighted using attention weights to obtain a fused meteorological feature; The fused meteorological feature is nonlinearly mapped by a feedforward network to output the task customized attribute meteorological feature.

6. The interpretable typhoon track and intensity prediction method based on task-sensitive learning of claim 1, wherein, The typhoon attribute prediction results of different prediction subtasks are obtained based on the task customized attribute meteorological features, the task customized terrain features, the causal direction guide features and the typhoon historical attribute time series features, including: The task customization attribute meteorological features of different prediction subtasks are respectively taken as initial state inputs of the corresponding prediction subtasks, the task customization topographic features, the causal direction guide features and the typhoon historical attribute time series features at the current time are element-wise added and fused to obtain the fusion features of different prediction subtasks, the fusion features of different prediction subtasks and the prediction results at the previous time are input into the corresponding prediction sub-branches, and the hidden states of the prediction sub-branches are updated together based on the network state at the previous time to obtain the typhoon attribute prediction results of different prediction subtasks.

7. The interpretable typhoon track and intensity prediction method based on task-sensitive learning of claim 1, wherein, Further comprising jointly optimizing the prediction network model, including: Before each back propagation in the prediction network model training process, the loss of each prediction subtask is obtained respectively, and the task affinity between the prediction subtasks is calculated based on the gradient information of the shared parameters of each task; The loss weight of each task is dynamically calculated, and the loss weight is adaptively allocated according to the task affinity; Based on the loss weight, the losses of the prediction subtasks are weighted and combined to form a prediction loss; The parameters of the prediction network model are optimized based on the prediction loss and the constraint applied to the topographic causal direction guide unit.

Images