Tropical Cyclone Scale Monitoring Method and Device Based on Deep Learning and Spatiotemporal Coding

Abstract

The present application relates to a tropical cyclone scale monitoring method and device based on deep learning and spatio-temporal coding. The method includes: acquiring multi-source satellite image data and the international best track dataset of tropical cyclones, and performing preprocessing; performing spatio-temporal independent field coding based on the international best track dataset to generate three-channel time scale field data reflecting the interannual variation characteristics, seasonal variation rules, and evolution time record characteristics of tropical cyclones, and two-channel spatial scale field data characterizing the spatial position information of tropical cyclones, and splicing them with the three-channel satellite image data obtained by preprocessing to obtain a tropical cyclone multi-scale comprehensive feature dataset and inputting it into a trained TC-Resnet model for tropical cyclone scale monitoring, and predicting the multi-scale monitoring results of tropical cyclones at the target moment. Using this method can make full use of the spatio-temporal information of tropical cyclones to achieve high-precision monitoring of tropical cyclone scales.

Description

Technical Field

[0001] This application relates to the field of deep learning technology, and in particular, to a tropical cyclone scale monitoring method and device based on deep learning and spatio-temporal coding. Background Art

[0002] The tropical cyclone (TC) scale is an important indicator for researchers to measure the influence range of a tropical cyclone system, and it is of decisive significance for revealing the potential harm degree that a tropical cyclone will bring. However, due to the physical processes of the interactions between multiple atmospheres and oceans, and the interactions between tropical cyclones and climate phenomena at multiple scales, these complex non-linear interaction effects make the monitoring research on the tropical cyclone scale full of challenges.

[0003] With the development of remote sensing technology, the objective quantitative monitoring of the tropical cyclone scale has gradually become possible. Some researchers use geostationary satellite infrared data to estimate the scale of tropical cyclones. They use the radius of maximum wind speed (RMAX) and the tangential wind speed at 182 km (V182) as the main parameters, construct a symmetric tangential wind field through an improved Rankine vortex model, and add the storm movement vector to obtain a complete two-dimensional wind field. Some researchers have also constructed a TC surface wind field analysis system (MTCSWA) based on multi-platform satellite data. This system integrates multi-source observation data such as scatterometers, microwave sounders, and cloud wind guidance, and can provide estimates of the radii of wind speeds of 34 knots, 50 knots, and 64 knots. Some researchers have also proposed a relatively simple method to estimate the TC wind field structure. This method is based on the commonly available information, including storm data (position, movement, and intensity) and TC scale parameters, and uses a modified Rankine vortex to fit the wind field to estimate the radii of wind speeds of 34 knots, 50 knots, and 64 knots. The verification results show that the mean absolute errors are 54 n mi (nautical miles), 32 n mi, and 20 n mi respectively. In addition, some researchers have proposed an objective technique based on geostationary satellite infrared images to characterize the structural and intensity changes of tropical cyclones. This method calculates the deviation angle variance (DAV) of the brightness temperature gradient to measure the degree of symmetry of the cloud system.

[0004] These remote sensing and objective estimation technologies provide important technical support for the TC scale monitoring in sea areas lacking conventional observation data. However, due to the spatio-temporal resolution limitations of the observation means and the complexity of the physical relationships, there are still certain uncertainties in these methods for TC scale estimation. Summary of the Invention

[0005] Based on this, it is necessary to provide a tropical cyclone scale monitoring method and device based on deep learning and spatio-temporal coding for the above technical problems, which can make full use of the spatio-temporal information of tropical cyclones to achieve high-precision monitoring of tropical cyclone scales, and support the exploration of the change mechanism of tropical cyclone scales based on the visualization of model parameters.

[0006] A tropical cyclone scale monitoring method based on deep learning and spatio-temporal coding, the method comprising:

[0007] Obtain multi-source satellite image data and the international best track dataset of tropical cyclones, and preprocess the multi-source satellite image data;

[0008] Perform spatio-temporal independent field coding based on the international best track dataset to generate three-channel time-scale field data reflecting the interannual change characteristics, seasonal change rules, and evolution time record characteristics of tropical cyclones, and two-channel space-scale field data representing the spatial position information of tropical cyclones. Then, by splicing with the three-channel satellite image data obtained from preprocessing, a tropical cyclone multi-scale comprehensive feature dataset is obtained;

[0009] Divide the tropical cyclone multi-scale comprehensive feature dataset into a training set, a validation set, and a test set according to a preset ratio. Input the training set and the validation set into a pre-constructed TC-Resnet model for optimized training to obtain a trained TC-Resnet model; wherein, the model consists of a feature extraction layer and a fully connected layer; the feature extraction layer extracts the tropical cyclone scale features of the input image through a convolutional module, a max pooling module, and four groups of pyramid convolutional residual modules connected in sequence; the fully connected layer obtains the feature map output by the feature extraction layer, performs average pooling, flattening, and linear layer classification processing in sequence, and outputs the estimation of multi-type scale parameters of the tropical cyclone; wherein, the size parameters include the radius of maximum wind speed, the radius of 34-knot wind speed, the radius of 50-knot wind speed, and the radius of 64-knot wind speed;

[0010] Input the test set into the trained TC-Resnet model for tropical cyclone scale monitoring, and predict the multi-scale monitoring results of the tropical cyclone at the target time.

[0011] In one embodiment, obtaining the multi-source satellite image data and the international best track dataset of tropical cyclones, and preprocessing the multi-source satellite image data includes:

[0012] Obtain multi-source satellite image data of tropical cyclones and the international best track dataset; among them, the multi-source satellite image data includes complementary GridSat-B1 satellite image data, CMORPH satellite-derived precipitation image data, and FY series meteorological satellite image data. The GridSat-B1 satellite image data includes infrared channel image data and water vapor channel image data. In the time dimension, all satellite image data are aligned to be consistent with the international best track dataset. In the space dimension, all satellite image data are interpolated to the same spatial resolution as the CMORPH satellite-derived precipitation image data. After time alignment and spatial interpolation, the multi-source satellite image data are complemented, and the complemented multi-source satellite image data are dimensionally stitched to construct a satellite image dataset. Further, data cleaning and spatial position screening are performed on the satellite image dataset to obtain the final satellite image dataset.

[0013] In one embodiment, data cleaning and spatial position screening are performed on the satellite image dataset to obtain the final satellite image dataset, including: screening and removing satellite image data in the satellite image dataset that do not meet the preset scale parameter constraint conditions according to the physical characteristics of the tropical cyclone scale parameters; the scale parameter constraint conditions include R34 > R50 > R64, R34 ≥ RMW, R50 ≥ RMW, and R64 ≥ RMW, where RMW, R34, R50, and R64 are the radius of maximum wind speed, the radius of 34-knot wind speed, the radius of 50-knot wind speed, and the radius of 64-knot wind speed of the tropical cyclone respectively; extracting the satellite image data in the satellite image dataset after data cleaning that are located in a set area around the longitude and latitude position of the tropical cyclone path center recorded in the international best track dataset to obtain the final satellite image dataset.

[0014] In one embodiment, spatio-temporal independent field encoding is performed based on the international best track dataset to generate three-channel time-scale field data reflecting the interannual variation characteristics, seasonal variation laws, and evolution time record characteristics of tropical cyclones and two-channel spatial-scale field data characterizing the spatial position information of tropical cyclones, and by stitching with the three-channel satellite image data obtained by preprocessing, a tropical cyclone multi-scale comprehensive feature dataset is obtained, including:

[0015] Perform spatio-temporal independent field encoding based on the international best path dataset. In the time dimension, generate three-channel time-scale field data, including a long-term time-scale field reflecting the interannual variation characteristics of tropical cyclones, a medium-term time-scale field reflecting the seasonal variation law of tropical cyclones, and a short-term time-scale field capturing the evolution time record characteristics of the synoptic scale of tropical cyclones; in the space dimension, generate two-channel space-scale field data, including a latitude field and a longitude field characterizing the spatial position information of tropical cyclones; among them, all the data obtained by spatio-temporal independent field encoding have the same spatial resolution as the preprocessed satellite image data, and the value of each grid point obtained by encoding represents the spatio-temporal information corresponding to the grid point position.

[0016] Stitch together the three-channel time-scale field data, the two-channel space-scale field data, and the infrared channel image data, water vapor channel image data, and CMORPH satellite-inverted precipitation image data in the final satellite image dataset to obtain a tropical cyclone multi-scale comprehensive feature dataset containing 8 channels.

[0017] In one embodiment, the four groups of pyramid convolutional residual modules in the TC-Resnet model respectively contain 3 layers, 8 layers, 36 layers, and 3 layers of pyramid convolutional residual blocks; among them, each pyramid convolutional residual block is composed of a first convolutional layer, a second pyramid convolutional layer, a third convolutional layer, and a fourth SE attention mechanism module. The data input to the pyramid convolutional residual block is divided into a main path and a shortcut path. After the data on the main path is processed by three layers of convolution and input into the SE attention mechanism module for attention weighting, it is stitched together with the data on the shortcut path and input into the next layer for processing.

[0018] In one embodiment, after obtaining the input feature map, the pyramid convolutional layer first divides the input feature map into different branches. For the data of each branch, apply the feature extraction ability branches with different convolutional kernel sizes in parallel for convolution, and stitch together the feature maps obtained by convolution of each feature extraction ability branch in the data dimension to obtain the final output feature map; among them, each feature extraction ability branch is assigned a different number of channels, and for the feature extraction ability branch with a larger convolutional kernel, the number of channels assigned to it is less, and the corresponding number of groups is more;

[0019] After obtaining the input feature map, the SE attention mechanism module first performs feature compression transformation, then performs feature recalibration on each channel in the compressed and transformed feature map to obtain channel attention information, and finally multiplies the channel attention information with the original input feature map channel by channel with a weight coefficient to finally output a feature map with channel attention.

[0020] In one embodiment, the data in the TC-Resnet model is processed by a batch normalization layer and a Mish activation function after each convolution.

[0021] In one embodiment, an additional SE attention mechanism module or attention module is further added to the first layer of the TC-Resnet model; wherein, the additional SE attention mechanism module is used to assign weights to each type of data input to the model and perform visual analysis of the data weights, and analyze and obtain the importance of each type of data for the model to monitor the multi-scale results of tropical cyclones; the additional attention module is used to perform visual analysis of the model weights after normalizing various types of satellite image data input to the model, and analyze and obtain the key attention areas of the model for various types of satellite image data.

[0022] In one embodiment, optimizing the training of the TC-Resnet model includes: calculating a loss function according to the estimated multi-type scale parameters of the tropical cyclone predicted and output by the TC-Resnet model and the true values of the scale parameters of the corresponding types, and optimizing the training of the TC-Resnet model according to the loss function and the Adam optimizer; wherein, the loss function is expressed as

[0023] ;

[0024] Wherein, is the number of input feature maps, is the feature map serial number and ; is the estimated scale parameter of the tropical cyclone predicted and output by the model; is the true value of the scale parameter.

[0025] A tropical cyclone scale monitoring device based on deep learning and spatio-temporal coding, the device includes:

[0026] A preprocessing module, configured to obtain multi-source satellite image data of tropical cyclones and an international best track dataset, and preprocess the multi-source satellite image data;

[0027] A spatio-temporal coding module, configured to perform spatio-temporal independent field coding based on the international best track dataset, generate three-channel time scale field data reflecting the interannual change characteristics, seasonal change rules, and evolution time record characteristics of tropical cyclones and two-channel spatial scale field data characterizing the spatial position information of tropical cyclones, and splice them with the three-channel satellite image data obtained by preprocessing to obtain a tropical cyclone multi-scale comprehensive feature dataset;

[0028] A model training module is used to divide the tropical cyclone multi-scale comprehensive feature dataset into a training set, a validation set, and a test set according to a preset ratio, and input the training set and the validation set into a pre-constructed TC-Resnet model for optimized training to obtain a trained TC-Resnet model. Among them, the model consists of a feature extraction layer and a fully connected layer. The feature extraction layer extracts the tropical cyclone scale features of the input image through a convolutional module, a max pooling module, and four groups of pyramid convolutional residual modules connected in sequence. The fully connected layer obtains the feature map output by the feature extraction layer, performs average pooling, flattening, and linear layer classification processing in sequence, and outputs the multi-type scale parameter estimation of the tropical cyclone. Among them, the size parameters include the radius of maximum wind speed, the radius of 34-knot wind speed, the radius of 50-knot wind speed, and the radius of 64-knot wind speed.

[0029] A scale monitoring module is used to input the test set into the trained TC-Resnet model for tropical cyclone scale monitoring, and predict the multi-scale monitoring results of the tropical cyclone at the target time.

[0030] The above-mentioned tropical cyclone scale monitoring method and device based on deep learning and spatio-temporal coding have the following beneficial effects:

[0031] 1. On the basis of satellite image data, three-channel time scale field data reflecting the interannual variation characteristics, seasonal variation laws, and evolution time record characteristics of tropical cyclones and two-channel spatial scale field data representing spatial position information generated by spatio-temporal coding based on the international best track dataset are further spliced to construct an 8-channel tropical cyclone multi-scale comprehensive feature dataset, which makes up for the limitations of a single data source, effectively integrates the historical spatio-temporal evolution information and real-time satellite observation characteristics of tropical cyclones, and significantly improves the multi-scale monitoring performance of the model.

[0032] 2. For the non-linear data situation where the tropical cyclone wind field often presents an asymmetric structure and there are large differences in different data samples, in the constructed TC-Resnet model, based on pyramid convolutional residual modules including feature extraction ability branches with different convolutional receptive field sizes and SE attention mechanism modules that can perform feature recalibration on each channel of the input data, etc., so as to be able to fully focus on the tropical cyclone multi-scale comprehensive feature data input into the model and realize the multi-scale intelligent collaborative monitoring of tropical cyclones.

[0033] 3. By adding an SE attention mechanism module or an attention module to the first layer of the model, it supports visual analysis of the input data weights and visual analysis of the model weights, which can reveal the data types and regions that the model focuses on, so as to be able to further improve the monitoring performance of the model and optimize the model training. Description of the Drawings

[0034] Figure 1Schematic flow chart of a tropical cyclone scale monitoring method based on deep learning and spatio-temporal coding in an embodiment;

[0035] Figure 2 Schematic diagram of the composition of a tropical cyclone multi-scale comprehensive feature dataset in an embodiment;

[0036] Figure 3 Schematic diagram of the architecture of the TC-Resnet model in an embodiment;

[0037] Figure 4 Schematic diagram of the comparison between the fitting results of the TC-Resnet model and the fitting results of other existing models in an embodiment; where, Figure 4 (a) Schematic diagram of the comparison of fitting results for the RMW scale parameter, Figure 4 (b) Schematic diagram of the comparison of fitting results for the R64 scale parameter, Figure 4 (c) Schematic diagram of the comparison of fitting results for the R50 scale parameter, Figure 4 (d) Schematic diagram of the comparison of fitting results for the R34 scale parameter;

[0038] Figure 5 Schematic diagram of the distribution of the mean error (ME) and the mean absolute error weighted by the true value (ME-weight) of the TC-Resnet model in an embodiment; where, Figure 5 (a) Schematic diagram of the error distribution for the RMW scale parameter, Figure 5 (b) Schematic diagram of the error distribution for the R64 scale parameter, Figure 5 (c) Schematic diagram of the error distribution for the R50 scale parameter, Figure 5 (d) Schematic diagram of the error distribution for the R34 scale parameter. Detailed implementation manners

[0039] In order to make the objectives, technical solutions and advantages of the present application clearer and more understandable, the present application 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 only used to explain the present application and are not used to limit the present application.

[0040] In one embodiment, as Figure 1 shown, a tropical cyclone scale monitoring method based on deep learning and spatio-temporal coding is provided, including the following steps:

[0041] Step S1, obtaining multi-source satellite image data of tropical cyclones and the international best track dataset, and preprocessing the multi-source satellite image data.

[0042] Specifically, this application selects GridSat-B1 satellite image data, CMORPH satellite-derived precipitation image data, and FY series meteorological satellite image data in the Northwest Pacific Ocean region from 90°E to 160°E and 5°S to 55°N. The three types of satellite image data complement each other. Among them, the GridSat-B1 satellite image data includes infrared channel image data (IR, approximately 11 μm) and water vapor channel image data (WV, approximately 6.7 μm).

[0043] The International Best Track Archive for Climate Stewardship (IBTrACS) is the most authoritative and complete database of historical tropical cyclone records in the world currently, compiled in cooperation with the World Meteorological Organization (WMO) and other organizations and individuals from around the world. These complete records with extremely high temporal resolution can provide important record materials for researchers to analyze tropical cyclones.

[0044] Preprocess the multi-source satellite image data, including the following steps:

[0045] First, in order to properly put the data into model training, in the time dimension, align all satellite image data to a 3-hourly record consistent with the International Best Track Archive for Climate Stewardship (IBTrACS). Further, in order to enable the data to be put into model training and for the multi-channel data effects to be integrated with each other, in the space dimension, interpolate all satellite image data to the same spatial resolution as the CMORPH satellite-derived precipitation image data, with a grid longitude and latitude size of 0.25°×0.25°.

[0046] After time alignment and spatial interpolation, complete the multi-source satellite image data, and splice the dimensions of the completed multi-source satellite image data to construct a satellite image dataset. Among them, completion includes: for some satellite image data with missing values, use methods such as cubic spline to reasonably interpolate and complete. And due to task adjustments such as satellite observations, for some satellite image data with completely missing values, use FY satellite data to complete, so as to ensure the integrity and continuity of the data.

[0047] Further clean the data and screen the spatial positions of the satellite image dataset to obtain the final satellite image dataset.

[0048] Among them, data cleaning is to screen and remove satellite image data in the satellite image dataset that does not meet the preset scale parameter constraint conditions according to the physical characteristics of tropical cyclone scale parameters; the scale parameter constraint conditions include R34 > R50 > R64, R34 ≥ RMW, R50 ≥ RMW, and R64 ≥ RMW, where RMW, R34, R50, and R64 are the radius of maximum wind speed, the radius of 34-knot wind speed, the radius of 50-knot wind speed, and the radius of 64-knot wind speed of the tropical cyclone, respectively. Spatial position screening refers to extracting satellite image data in the satellite image dataset after data cleaning that is located in a set area around the longitude and latitude position of the tropical cyclone path center recorded in the international best track dataset, to obtain the final satellite image dataset. The size of the set area around the longitude and latitude position of the tropical cyclone path center set in this application is 10°×10°, that is, to extract satellite image data within a box area of 5 longitudes and latitudes above, below, to the left, and to the right of the longitude and latitude corresponding to the center point of the tropical cyclone path.

[0049] Step S2, perform spatio-temporal independent field encoding based on the international best track dataset to generate three-channel time-scale field data reflecting the interannual change characteristics, seasonal change laws, and evolution time record characteristics of tropical cyclones, and two-channel space-scale field data characterizing the spatial position information of tropical cyclones, and splice them with the three-channel satellite image data obtained by preprocessing to obtain a tropical cyclone multi-scale comprehensive feature dataset.

[0050] Specifically, the construction process of the tropical cyclone multi-scale comprehensive feature dataset includes:

[0051] Perform spatio-temporal independent field encoding based on the international best track dataset. In the time dimension, generate three-channel time-scale field data, including a long-term time-scale field (year field) reflecting the interannual change characteristics of tropical cyclones, a medium-term time-scale field (month field) reflecting the seasonal change laws of tropical cyclones, and a short-term time-scale field (short-term time series field) capturing the evolution time record characteristics of tropical cyclone synoptic scales; in the spatial dimension, generate two-channel space-scale field data, including a latitude field and a longitude field characterizing the spatial position information of tropical cyclones; among them, all data obtained by spatio-temporal independent field encoding have the same spatial resolution (0.25°×0.25°) as the preprocessed satellite image data, and the value of each grid point obtained by encoding represents the spatio-temporal information corresponding to the grid point position.

[0052] Splice the three-channel time-scale field data, the two-channel space-scale field data, and the infrared channel image data, water vapor channel image data, and CMORPH satellite-inverted precipitation image data in the final satellite image dataset to obtain a tropical cyclone multi-scale comprehensive feature dataset containing 8 channels, such as Figure 2As shown. The multi-scale comprehensive feature dataset of these 8-channel tropical cyclones effectively integrates the historical spatio-temporal evolution information of tropical cyclones recorded in the international best track dataset and the real-time satellite observation features. This multi-dimensional data fusion method enables subsequent models to comprehensively capture the spatio-temporal evolution characteristics of tropical cyclones at multiple scales, providing a solid data foundation for improving model performance. Figure 2 Among the 4 circles from the inside to the outside, they respectively correspond to: RMW (red), R64 (green), R50 (blue), and R34 (purple).

[0053] Step S3, divide the multi-scale comprehensive feature dataset of tropical cyclones into a training set, a validation set, and a test set according to a preset ratio. Input the training set and the validation set into the pre-constructed TC-Resnet model for optimized training to obtain a trained TC-Resnet model; among them, this model consists of a feature extraction layer and a fully connected layer; the feature extraction layer extracts the tropical cyclone scale features of the input image through a convolutional module, a max pooling module, and four groups of pyramid convolutional residual modules connected in sequence; the fully connected layer obtains the feature map output by the feature extraction layer, performs average pooling, flattening, and linear layer classification processing in sequence, and outputs the multi-type scale parameter estimation of the tropical cyclone.

[0054] Among them, the size parameters of the tropical cyclone include the radius of maximum wind speed, the radius of 34-knot wind speed, the radius of 50-knot wind speed, and the radius of 64-knot wind speed. The specific structure of the TC-Resnet model constructed in this application is as Figure 3 shown. The four groups of pyramid convolutional residual modules in the model respectively contain 3 layers, 8 layers, 36 layers, and 3 layers of pyramid convolutional residual blocks. Among them, each pyramid convolutional residual block consists of a first-layer convolutional layer (Conv1), a second-layer pyramid convolutional layer (Pyramidal Convolution, PyConv), a third-layer convolutional layer (Conv2), and a fourth-layer SE attention mechanism module. The data input into the pyramid convolutional residual block is divided into a main path X and a shortcut path Xshortcut. After the data on the main path undergoes three-layer convolutional processing and is input into the SE attention mechanism module for attention weighting, it is concatenated with the data on the shortcut path and input into the next layer for processing. The residual module designed in this way can enhance the feature extraction ability of the model for field variable data.

[0055] Compared with the process of ordinary convolution, which is to input the input feature map into a standard convolution kernel for convolution processing and then obtain the output feature map, this convolution design cannot take into account the defects of features at different scales. In this application, a pyramid convolution layer is added to the residual module. After obtaining the input feature map, the pyramid convolution layer first divides the input feature map into different branches. For the data of each branch, the feature extraction ability branches with different convolution kernel sizes are applied in parallel for convolution, and different numbers of channels are assigned to these convolution kernels. The feature maps obtained by convolving each feature extraction ability branch are concatenated in the data dimension to obtain the final output feature map. Such a design is equivalent to considering separate extraction branch routes for features at different scales. Those larger convolution kernels have a larger receptive field, but also have a larger number of computing parameters and higher requirements for machine performance. Therefore, for the feature extraction ability branch with a larger convolution kernel, fewer channels are allocated, which can well balance the model performance. At the same time, a hierarchical grouping strategy is also introduced in the pyramid convolution layer, that is, for the feature extraction ability branch with a larger convolution kernel, the corresponding number of groups is more. This can construct a hierarchical expression of features at different scales and further reduce the demand for computing resources. Such a multi-scale convolution branch route design of the pyramid convolution layer can show greater advantages in the tropical cyclone multi-scale intelligent monitoring task that needs to pay attention to multi-scale image features.

[0056] An SE (Squeeze-and-Excitation) attention mechanism module is added to the last layer of each pyramid convolution residual block. After obtaining the input feature map, the SE attention mechanism module first performs feature compression conversion, then recalibrates each channel in the feature map after compression conversion to obtain channel attention information, and finally multiplies the channel attention information with the original input feature map channel by channel with a weight coefficient to finally output a feature map with channel attention. This module can assign different weights to each channel of the input feature data, making the model pay more attention to the data of important channels. And there is a design of reduction ratio in the middle of this module, which not only significantly reduces the number of parameters, but also plays the role of an information bottleneck, forcing the network model to learn more discriminative channel attention. Such a design can reduce the influence of feature channels irrelevant to the model fitting result and assign a larger weight to the feature channels with higher utility.

[0057] Furthermore, in the TC-Resnet model, to make the model pay more attention to the key regions on the feature map, the data is processed by a batch normalization layer (BN) and a Mish activation function after each convolution, that is, a batch normalization layer and a Mish activation function are added after each convolutional layer in the model. Among them, compared with the traditional ReLU activation function and its variants shown in Equation (1), the Mish activation function selected in this application exhibits unique mathematical properties, which is a smooth, non-monotonic, and self-normalized activation function. The functional form of Mish is shown in Equation (2):

[0058] (1);

[0059] (2).

[0060] The unboundedness of the Mish activation function can alleviate the vanishing gradient problem. When x is less than 0, y taking a relatively small negative value can alleviate the phenomenon of neuron death to a certain extent and retain some neurons with negative weights. At the same time, the Mish activation function is smoother than the ReLU activation function, so a more stable gradient can be obtained, which has a better effect on the application of the optimizer. Moreover, the Mish activation function has an approximate linear nature. On the one hand, it can maintain its non-linear expression ability while maintaining some characteristics of the linear activation function. For deep networks like Resnet-152, Mish can have particularly excellent gradient flow properties. The TC-Resnet model constructed in this application is particularly deep, so all activation functions are replaced with Mish activation functions to obtain better performance than Relu.

[0061] When training the TC-Resnet model, to facilitate the evaluation of the model performance, the tropical cyclone multi-scale comprehensive feature dataset is divided into a training set, a validation set, and a test set according to the ratio of 80:5:15. During the training process, the training set is used to iteratively adjust the parameters of the model; the validation set is used to guide the tuning of hyperparameters such as the learning rate, facilitate the selection of an appropriate learning rate, and discover problems during model training; the test set is used to evaluate the performance of the model after each iterative training. Such a three-stage training-validation-test process can ensure the reliability of the trained model, facilitate understanding the actual effect of the model, and the data in the test set that has not participated in the training can also be used to evaluate the multi-scale monitoring ability of the model in the actual scenario.

[0062] When optimizing and training the TC-Resnet model, the loss function is calculated based on the estimated multi-type scale parameters of the tropical cyclone predicted by the TC-Resnet model and the true scale parameters of the corresponding type, and the TC-Resnet model is optimized and trained according to the loss function and the Adam optimizer. The Adam optimizer combines the advantages of the momentum method and the adaptive learning rate method and can adaptively adjust the learning rate of the parameters. Among them, the loss function is expressed as

[0063] ;

[0064] Among them, is the number of input feature maps, is the feature map serial number and ; is the estimated scale parameter of the tropical cyclone predicted by the model; is the true value of the scale parameter. Each time the model is trained, the loss function can be calculated from the estimated results of the 4 types of scale parameters monitored by the model and the label true value data, and then the next training iteration of the model can be optimized to obtain a model with continuously optimized performance. This enables the model to continuously learn the input data, optimize the internal parameter weights, and gradually improve the effect of the model on multi-scale monitoring of tropical cyclones. With this end-to-end processing flow, the trained model can achieve direct monitoring from the original satellite image data to the multi-scale parameters of tropical cyclones.

[0065] Step S4, input the test set into the trained TC-Resnet model for tropical cyclone scale monitoring, and predict the multi-scale monitoring results of the tropical cyclone at the target time.

[0066] Furthermore, this method also includes: an additional SE attention mechanism module or attention module is added to the first layer of the TC-Resnet model; among them, the additional SE attention mechanism module is used to assign weights to each type of data input to the model and perform visual analysis of the data weights, and analyze and obtain the importance of each type of data for the multi-scale results of the model monitoring tropical cyclones; the additional attention module is used to perform visual analysis of the model weights after normalizing each type of satellite image data input to the model, and analyze and obtain the key attention areas of the model for each type of satellite image data. By adding the SE attention mechanism module or attention module, it supports visual analysis of input data weights and model weights, can reveal the data types and areas that the model focuses on, and thus can further improve the monitoring performance of the model and optimize model training.

[0067] Table 1 Comparison between the TC-Resnet model and other existing models

[0068]

[0069] To further verify the beneficial effects of the tropical cyclone scale monitoring method based on deep learning and spatio-temporal coding provided by this application, a comparative experiment was conducted between the TC-Resnet model constructed in this application and other existing models. The results of the comparative experiment are shown in Table 1 and Figure 4 as follows. From Table 1 and Figure 4 it can be seen that the TC-Resnet model constructed in this application can simultaneously monitor four tropical cyclone scale parameters, and the accuracy has reached the current best.

[0070] To further verify the monitoring accuracy of the TC-Resnet model constructed in this application in different interval ranges of the four types of scale parameters, for the four types of scales, considering the physical characteristics of the tropical cyclone scale and the quantiles of the four types of scale parameters, for the RMW, since the overall change of the RMW is small, four ranges of values less than 12, 12 - 15, 15 - 20, and greater than 20 were selected. The four value ranges account for 25.79%, 26.47%, 28.12%, and 19.62% of the total samples respectively (considering the integer division effect and only retaining two decimal places). For the R34, since the change range of the 34-knot wind speed is large, four ranges less than or equal to 100, 100 - 130, 130 - 160, and greater than or equal to 160 were selected. Each range accounts for 17.97%, 34.57%, 26.06%, and 21.40% of the total samples respectively. Among them, 100 - 130 is the area with the most samples in the divided interval of this R34. For the R50, considering the value range, it was divided according to less than or equal to 50 (24.42%), 50 - 70 (35.94%), 70 - 90 (21.67%), and greater than or equal to 90 (17.97%). For the R64, it was divided according to less than or equal to 30 (35.67%), 30 - 40 (25.24%), 40 - 50 (18.79%), and greater than or equal to 50 (20.30%). In the evaluation process of the TC-Resnet model, in addition to the mean absolute error (MAE), the mean absolute error weighted by the true value (MAE-weight) was also added as an evaluation index. This index can assign a greater weight to a larger tropical cyclone scale based on the real data, and the obtained results are as Figure 5 shown.

[0071] From Figure 5 it can be seen that for the radius of maximum wind speed RMW, since the value range is generally small, the model monitoring error is low. In the relatively typical scale range of 12 - 15 nautical miles, the model effect reaches the best state. And in the scale range of 15 - 20 nautical miles, the MAE-weight of the model monitoring result is only 3.09 nautical miles. The proportion of these two RMW scale division ranges in the total samples is 54.59%, indicating that the model monitors the typical RMW scale more accurately.

[0072] For the outer tropical cyclone scale R34, the prediction error increases significantly with distance, especially for scales like R34. Starting from the first range, the error gradually increases and reaches its maximum above 160 nautical miles. Compared with several other scales, the R34 has a larger numerical range and greater error, which reflects the inherent challenges in monitoring the scale of the outer wind field. R50 and R64 also follow this pattern, with the error showing a gradual increase from the first division range to the fourth division range. For RMW, for extremely large RMW values (greater than or equal to 20 nautical miles), the model monitoring error is relatively large (MAE-weight is 9.48). For smaller values (less than or equal to 12 nautical miles), the model error is larger than the typical cases in the second and third ranges (MAE-weight is 4.57).

[0073] The monitoring errors of the four types of scale parameters all show the characteristic that the negative deviation increases with the increase of the value, that is, the tropical cyclone scale monitored by the model is smaller than the true label data. This is because the sample size of extreme scale values is small and unbalanced. In the first division range, there are cases where ME and ME-weight are positive. For example, for R34, in the range less than or equal to 100 nautical miles, the ME-weight error of the model monitoring value is 8.16, indicating that in relatively small values of this scale, the monitored value by the model is greater than the true value, which may also be caused by the above reasons. In the typical numerical ranges of the four types of scales, the model monitoring accuracy is good.

[0074] There is a high consistency between the mean absolute error, weighted mean absolute error, mean error weighted based on the true value, and mean error, indicating that the model remains stable in the process of monitoring the tropical cyclone scale after considering the weights given by the true values of the samples.

[0075] Table 2 Experiments with Different Input Data

[0076]

[0077] Furthermore, during verification, multiple satellite single-channel data were separated from the dataset and separately input into the model for training and parameter adjustment. The monitoring effect of the model on tropical cyclones at multiple scales decreased significantly. However, when integrating three satellite data channels into the model training, the monitoring performance of the model was improved to a certain extent. Subsequently, the three satellite data channels were integrated, and the multi-scale spatio-temporal independent field variable information of five channels generated based on IBTrACS was introduced to form an 8-channel dataset, which was put into the model for training, achieving the best effect in this experiment, as shown in Table 2. This indicates that multi-source information fusion plays a significant role in improving the monitoring accuracy of tropical cyclones at multiple scales, and the model performance has also been greatly enhanced.

[0078] In one embodiment, a tropical cyclone scale monitoring device based on deep learning and spatio-temporal coding is provided, including:

[0079] A preprocessing module for obtaining multi-source satellite image data of tropical cyclones and the international best track dataset, and preprocessing the multi-source satellite image data;

[0080] A spatio-temporal coding module for performing spatio-temporal independent field coding based on the international best track dataset to generate three-channel time-scale field data reflecting the interannual variation characteristics, seasonal variation laws, and evolution time record characteristics of tropical cyclones, and two-channel space-scale field data representing the spatial position information of tropical cyclones. By splicing with the three-channel satellite image data obtained through preprocessing, a multi-scale comprehensive feature dataset of tropical cyclones is obtained;

[0081] A model training module for dividing the multi-scale comprehensive feature dataset of tropical cyclones into a training set, a validation set, and a test set according to a preset ratio, and inputting the training set and the validation set into a pre-constructed TC-Resnet model for optimized training to obtain a trained TC-Resnet model; among them, this model consists of a feature extraction layer and a fully connected layer; the feature extraction layer extracts the tropical cyclone scale features of the input image through a convolutional module, a max-pooling module, and four groups of pyramid convolutional residual modules connected in sequence; the fully connected layer obtains the feature map output by the feature extraction layer, performs average pooling, flattening, and linear layer classification processing in sequence, and outputs the multi-type scale parameter estimation of tropical cyclones; among them, the size parameters include the radius of maximum wind speed, the radius of 34-knot wind speed, the radius of 50-knot wind speed, and the radius of 64-knot wind speed;

[0082] A scale monitoring module for inputting the test set into the trained TC-Resnet model for tropical cyclone scale monitoring, and predicting the multi-scale monitoring results of tropical cyclones at the target time.

[0083] For the specific limitations of the tropical cyclone scale monitoring device based on deep learning and spatio-temporal coding, reference can be made to the limitations of the tropical cyclone scale monitoring method based on deep learning and spatio-temporal coding in the above text, which will not be elaborated here. Each module in the above-mentioned tropical cyclone scale monitoring device based on deep learning and spatio-temporal coding can be implemented in whole or in part by software, hardware, and their combination. The above-mentioned modules can be embedded in the processor of the computer device in hardware form or be independent of it, or be stored in the memory of the computer device in software form, so as to facilitate the processor to call and execute the operations corresponding to each of the above modules.

[0084] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope recorded in this specification.

[0085] The above-described embodiments only represent several implementation manners of the present application. The description is relatively specific and detailed, but it should not be construed as a limitation on the scope of the present application. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several modifications and improvements can still be made, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the appended claims.

Claims

1. A tropical cyclone scale monitoring method based on deep learning and spatiotemporal coding, characterized in that: The method comprises: Acquire multi-source satellite image data and international best path data set of tropical cyclones, and pre-process the multi-source satellite image data; Based on the international best path dataset, spatiotemporal independent field encoding is performed to generate three-channel time scale field data in the time dimension, including a long-term time scale field reflecting the interannual variation characteristics of tropical cyclones, a medium-term time scale field reflecting the seasonal variation law of tropical cyclones, and a short-term time scale field capturing the evolution time record characteristics of tropical cyclone weather scales; in the spatial dimension, two-channel spatial scale field data are generated, including a latitude field and a longitude field representing the spatial position information of tropical cyclones; wherein all data obtained by spatiotemporal independent field encoding maintain a consistent spatial resolution with the preprocessed satellite image data, and the numerical value of each grid point obtained by encoding represents the spatiotemporal information corresponding to the grid point position; the three-channel time scale field data, the two-channel spatial scale field data, and the infrared channel image data, the water vapor channel image data, and the CMORPH satellite inversion precipitation image data in the final satellite image dataset obtained by preprocessing are spliced ​​to obtain a tropical cyclone multi-scale comprehensive feature dataset containing 8 channels; The tropical cyclone multi-scale comprehensive feature data set is divided into a training set, a validation set and a test set according to a preset ratio, and the training set and the validation set are input into a pre-built TC-Resnet model for optimization training to obtain a trained TC-Resnet model; wherein the model consists of a feature extraction layer and a fully connected layer; the feature extraction layer extracts the tropical cyclone scale features of the input image through a convolution module, a maximum pooling module and four groups of pyramid convolution residual modules connected in sequence; the fully connected layer obtains the feature map output by the feature extraction layer, performs average pooling, flattening and linear layer classification processing in sequence, and outputs multi-type scale parameter estimation of the tropical cyclone; wherein the last layer of each pyramid convolution residual block contained in each group of pyramid convolution residual modules is an SE attention mechanism module; The test set is input into the trained TC-Resnet model for tropical cyclone scale monitoring, and the multi-scale monitoring results of the tropical cyclone at the target time are predicted.

2. The method according to claim 1, characterized in that Acquire multi-source satellite image data and international best path data set of tropical cyclones, and pre-process the multi-source satellite image data, including: Acquire multi-source satellite image data and international best path data sets of tropical cyclones; wherein the multi-source satellite image data include complementary GridSat-B1 satellite image data, CMORPH satellite inversion precipitation image data and Fengyun series meteorological satellite image data, and the GridSat-B1 satellite image data include infrared channel image data and water vapor channel image data; In the temporal dimension, all satellite image data are aligned to be consistent with the international best path dataset, and in the spatial dimension, all satellite image data are interpolated to a spatial resolution consistent with the CMORPH satellite inverted precipitation image data; After time alignment and spatial interpolation, the multi-source satellite image data are completed, and the completed multi-source satellite image data are dimensionally spliced ​​to construct a satellite image dataset; The satellite image dataset is further cleaned and spatially screened to obtain a final satellite image dataset.

3. The method according to claim 2, characterized in that The satellite image dataset is cleaned and spatially screened to obtain a final satellite image dataset, including: According to the physical characteristics of the tropical cyclone scale parameter, the satellite image data that does not meet the preset scale parameter constraint conditions in the satellite image data set are screened and eliminated; the scale parameter constraint conditions include R34>R50>R64, R34≥RMW, R50≥RMW and R64≥RMW, wherein RMW, R34, R50 and R64 are the maximum wind speed radius, 34-knot wind speed radius, 50-knot wind speed radius and 64-knot wind speed radius of the tropical cyclone respectively; According to the longitude and latitude positions of the tropical cyclone path center recorded in the international best path dataset, satellite image data of a set area around the longitude and latitude positions of the tropical cyclone path center are extracted from the satellite image dataset after data cleaning to obtain a final satellite image dataset.

4. The method according to claim 1, characterized in that: The four groups of pyramid convolution residual modules in the TC-Resnet model include 3-layer, 8-layer, 36-layer and 3-layer pyramid convolution residual blocks respectively; Among them, each pyramid convolution residual block is composed of the first convolution layer, the second pyramid convolution layer, the third convolution layer and the fourth SE attention mechanism module. The data input into the pyramid convolution residual block is divided into a main path and a shortcut path. After the data of the main path is processed by three layers of convolution and input into the SE attention mechanism module for attention weighting, it is spliced ​​with the data of the shortcut path and input into the next layer for processing.

5. The method according to claim 4, characterized in that After obtaining the input feature map, the pyramid convolution layer first divides the input feature map into different branches, applies feature extraction capability branches with different convolution kernel sizes to the data of each branch in parallel for convolution, and concatenates the feature maps obtained by convolution of each feature extraction capability branch in the data dimension to obtain the final output feature map; wherein each feature extraction capability branch is allocated a different number of channels, and for a feature extraction capability branch with a larger convolution kernel, the number of allocated channels is less and the corresponding number of groups is more; After obtaining the input feature map, the SE attention mechanism module first performs feature compression conversion, then recalibrates the features of each channel in the compressed feature map to obtain channel attention information, and finally multiplies the channel attention information and the original input feature map by the weight coefficient channel by channel, and finally outputs a feature map with channel attention.

6. The method according to claim 1 or 4, characterized in that: The data in the TC-Resnet model is processed by a batch normalization layer and a Mish activation function after each convolution.

7. The method according to claim 6, characterized in that The first layer of the TC-Resnet model is also added with an additional SE attention mechanism module or attention module; wherein the additional SE attention mechanism module is used to assign weights to each type of data input to the model and perform data weight visualization analysis, analyzing the importance of each type of data to the model's multi-scale results of monitoring tropical cyclones; the additional attention module is used to perform model weight visualization analysis after normalizing various types of satellite image data input to the model, analyzing and obtaining the model's key focus areas for various types of satellite image data.

8. The method according to claim 1, characterized in that: Optimize the training of the TC-Resnet model, including: The loss function is calculated based on the multi-type scale parameter estimates of the tropical cyclone predicted and output by the TC-Resnet model and the true values ​​of the scale parameters of the corresponding types, and the TC-Resnet model is optimized and trained based on the loss function and the Adam optimizer; wherein the loss function is expressed as: ； in, is the number of input feature maps, is the feature map number and ; Estimates of scale parameters of tropical cyclones for model forecast output; is the true value of the scale parameter.

9. A tropical cyclone scale monitoring device based on deep learning and spatiotemporal coding, characterized in that: The device comprises: A preprocessing module, used to obtain multi-source satellite image data and an international best path data set of tropical cyclones, and preprocess the multi-source satellite image data; A spatiotemporal coding module is used to perform spatiotemporal independent field coding based on the international best path dataset. In the time dimension, three-channel time scale field data are generated, including a long-term time scale field reflecting the interannual variation characteristics of tropical cyclones, a medium-term time scale field reflecting the seasonal variation law of tropical cyclones, and a short-term time scale field capturing the evolution time record characteristics of tropical cyclone weather scales; in the spatial dimension, two-channel spatial scale field data are generated, including a latitude field and a longitude field representing the spatial position information of tropical cyclones; wherein all data obtained by spatiotemporal independent field coding maintain a consistent spatial resolution with the preprocessed satellite image data, and the numerical value of each grid point obtained by coding represents the spatiotemporal information corresponding to the grid point position; the three-channel time scale field data, the two-channel spatial scale field data, and the infrared channel image data, the water vapor channel image data, and the CMORPH satellite inversion precipitation image data in the final satellite image dataset obtained by preprocessing are spliced ​​to obtain a tropical cyclone multi-scale comprehensive feature dataset containing 8 channels; A model training module is used to divide the tropical cyclone multi-scale comprehensive feature data set into a training set, a validation set and a test set according to a preset ratio, and input the training set and the validation set into a pre-built TC-Resnet model for optimization training to obtain a trained TC-Resnet model; wherein the model consists of a feature extraction layer and a fully connected layer; the feature extraction layer extracts the tropical cyclone scale features of the input image through a convolution module, a maximum pooling module and four groups of pyramid convolution residual modules connected in sequence; the fully connected layer obtains the feature map output by the feature extraction layer, performs average pooling, flattening and linear layer classification processing in sequence, and outputs multi-type scale parameter estimation of the tropical cyclone; wherein the last layer of each pyramid convolution residual block contained in each group of pyramid convolution residual modules is an SE attention mechanism module; The scale monitoring module is used to input the test set into the trained TC-Resnet model to perform tropical cyclone scale monitoring and predict the multi-scale monitoring results of the tropical cyclone at the target time.

Images