Intelligent intensity setting method and device based on typhoon hot core structure and storage medium

Abstract

The application discloses an intelligent intensity determination method and device based on a typhoon hot core structure and a storage medium, and relates to the technical field of remote sensing and telemetry and deep learning. According to satellite microwave data, a multi-view typhoon hot core abnormal data set is constructed by the application, and in the model training process, a multi-view multi-scale feature information fusion method is proposed, combined with a single value regression loss function and an ordinal regression loss function to train and optimize the model. The model can finally effectively extract multi-view features, improve the characteristic that the estimated typhoon wind speed is too small, and show good robustness and scalability, so that accurate estimation of the typhoon wind speed is realized.

Description

Technical Field

[0001] This invention relates to the fields of remote sensing and deep learning technologies, and in particular to an intelligent intensity determination method, device and storage medium based on the thermonuclear structure of a typhoon. Background Technology

[0002] Typhoons are among the most severe natural disasters facing humankind. Accurately estimating typhoon wind speed (or typhoon intensity) can help people more effectively predict and manage disasters, reducing the massive losses they cause. Traditional methods of typhoon wind speed estimation first require on-site measurements using instruments equipped with anemometers (weather buoys, aircraft, ships, etc.), which is difficult to conduct long-term observations over vast ocean areas and is too costly. Secondly, techniques such as Dvorak rely heavily on accumulated experience from technicians to judge typhoon intensity, which is highly subjective. Furthermore, the varying structures of reference cloud images in different regions inevitably lead to differences in typhoon intensity estimates. In contrast, satellite remote sensing instruments can continuously observe typhoons unaffected by weather conditions and generate abundant data. Deep learning, with its excellent feature extraction and big data analysis capabilities, has already found some applications in the meteorological field.

[0003] However, existing deep learning-based typhoon intensity determination technologies mostly use satellite cloud image data, which often only focus on the cloud top morphology and structure of the typhoon, ignoring the internal thermal core structure of the typhoon and losing a lot of height information. It is difficult to model and analyze the three-dimensional structure of the typhoon's thermal core and even the entire typhoon. In particular, it is difficult to apply the cloud layer obstruction in the upper part of the typhoon and the vertical morphological characteristics of the typhoon rain belt. Summary of the Invention

[0004] This invention provides an intelligent intensity estimation method, device, and storage medium based on the thermonuclear structure of typhoons. It utilizes a thermonuclear anomaly dataset constructed based on SNPP ATMS (Advanced Microwave Detector) satellite inversion to train a typhoon wind speed regression network model, thereby achieving accurate estimation of typhoon intensity and improving its accuracy.

[0005] On the one hand, this invention provides a method for estimating typhoon intensity based on a thermonuclear anomaly dataset, specifically including the following steps:

[0006] Step 1: Obtain satellite brightness temperature data and typhoon path data;

[0007] Step 2: Preprocess the satellite brightness temperature data and then perform inversion processing to obtain temperature field data; process the temperature field data based on typhoon path data to construct a multi-view typhoon thermal core anomaly dataset and its corresponding typhoon intensity labels;

[0008] Step 3: Estimate wind speed on the thermonuclear anomaly dataset using a deep learning-based regression method to obtain single-value regression and ordinal regression results for wind speed.

[0009] Furthermore, in step 3, when estimating wind speed on the thermonuclear anomaly dataset using a deep learning-based regression method, the typhoon intensity estimation network model used consists of two parts: a multi-view feature fusion network (Fuse(·)) and a wind speed estimation network (Est(·)).

[0010] Among them, the multi-view feature fusion network is used to fuse the input multi-view information at multiple scales to obtain a set of high-dimensional feature encoding vectors r, i.e. x represents the input multiple views. Denotes the real number field, D F Represents the feature dimension of r;

[0011] The wind speed estimation network uses the softmax function to map the input high-dimensional feature vector r into a classification vector c, which can be expressed as: D c The vector dimension (i.e., category dimension) represents the classification vector; and the classification vector c consists of two parts: single-value regression and ordinal regression; wherein, single-value regression is a single numerical value Z that regresses the wind speed to a specified range (preferably 1 to 100 m / s). i,sVR This involves transforming the wind speed estimation task into a classification task, for example, with category dimension D. c =100; Ordinal regression is the process of regressing wind speed into a specified range (preferably 1–100 m / s) in the form of a probability distribution. i，j (j = 1, ..., 100), is transformed into a multi-label classification task, and then the ordinal regression result Z is obtained by calculating the expectation. i，OR ;

[0012] When outputting the final wind speed estimate based on the classification vector c, the final wind speed estimate is obtained by fusing single-value regression and ordinal regression:

[0013]

[0014] Among them, Z i，final This represents the final wind speed estimate, with the weighting coefficient α ranging from [0, 1] and the subscript i indicating the sample number.

[0015] Furthermore, in step 1, the present invention divides the thermonuclear anomaly dataset on atmospheric pressure into profile information in three directions: along the satellite scan line direction and perpendicular to the satellite scan line direction, thereby constituting the multi-view information data required as input to the deep learning network (typhoon intensity estimation network model) in the present invention. Preferably, the data dimensions along the satellite scan line direction and perpendicular to the satellite scan line direction in the pressure direction can be 40×320×320, 21×320×320, and 21×320×320, respectively.

[0016] Furthermore, when training the network parameters based on the multi-view typhoon thermal core anomaly dataset, the loss function used for the typhoon intensity estimation network model is set as follows:

[0017] Loss total =Loss SVR +Loss OR

[0018] Among them, Loss total Loss represents the total loss of the typhoon intensity estimation network model. SVR This represents the loss in a single-valued regression task. OR This represents the loss of the ordinal regression task; and Loss SVR The loss is the weighted mean square error (MSE). weighted Loss OR For adaptive mean residual loss L AMR and κ-weighted loss L WKL The fusion result.

[0019] This invention addresses the asymmetry of risk in typhoon wind speed estimation, specifically the asymmetric penalty for incorrect predictions. It improves the Mean Square Error (MSE) loss to obtain a weighted mean square error (MSE) loss. weighted The loss is calculated as follows:

[0020]

[0021] Among them, y i This is the true value label for wind speed (i.e., the typhoon intensity label). The prediction result (i.e., the final wind speed estimate Z) final ), where i is the sample number and γ is a preset hyperparameter that is set based on the specific application scenario. Preferably, it can be set to γ ​​= 1 / 3.

[0022] This invention addresses the randomness and ambiguity of typhoon growth cycles by introducing wind speed distribution to represent labels with single-valued ambiguities, and incorporating adaptive residual loss to constrain training. The adaptive residual loss consists of two parts: average loss and residual loss. The average loss measures the difference between the average value of the wind speed distribution and the actual wind speed, while the residual loss measures the wind speed probability entropy of the dynamic Top-K distribution. Considering that the accuracy of the Top-K deep learning model has reached a very high level, this invention assumes that y i The loss function is ultimately comprised of two parts: mean loss L, where the sum of wind speed probabilities outside the Top-K classes is considered the residual. m and residual loss L r .

[0023] The adaptive mean residual loss L of the present invention AMR Specifically:

[0024]

[0025]

[0026] L AMR =λ1L m +λ2L r

[0027] Where N represents the number of samples (i.e., batch size), and Top-K represents the top K (preset value) samples with the highest accuracy. y represents the estimated wind speed. i For the true wind speed label, L represents the number of categories, i.e., the number of wind speed intensity levels, for example, corresponding to the value range of 1 to 100 m / s, L∈{1, 2, 3, ..., 100}; λ1 and λ2 represent the mean loss L. m and residual loss L r The weighting coefficients.

[0028] This invention addresses the common problem that cross-entropy loss, used for classification, often neglects information between adjacent classes, making it ineffective for evaluating ordinal regression problems. The κ index, statistically used to measure the consistency among raters in classifying elements into a set of disjoint classes, can be used to address ordinal regression problems by employing κ-weighted loss. By introducing the κ index, the order of classes is considered, and incorrect estimates are penalized based on the distance between the true class and the estimated class. Thus, estimates closer to the true class are considered better than those further away. This invention uses a quadratic weighted κ, where the loss is proportional to the square of the distance.

[0029]

[0030] L WKL =log(1-κ)

[0031] Among them, the parameters κ∈[-1, 1] and L are calculated based on the formula. WKL ∈(-∞, log2], Used to characterize the predictive performance of the typhoon intensity estimation network model. This indicates that the final wind speed estimate Z is based on... i，f inal The typhoon intensity category is obtained by rounding down (either rounding up or down is acceptable, but rounding down is preferred). Here, j represents the true typhoon intensity category of sample i. For example, for an example where typhoon intensity is categorized as 1 to 100, then... like but like but That is, for each sample, the typhoon intensity category predicted by the typhoon intensity estimation network model. This forms a wind speed category with the corresponding true category j. The outer product between the two wind speed classification histogram vectors. Indicates each pair of wind speed categories loss weights, L represents the number of categories.

[0032] Furthermore, step 3 also includes data augmentation processing of the multi-view typhoon thermal core anomaly dataset, followed by wind speed estimation based on a deep learning regression method. Taking advantage of the rotational invariance of typhoons in the pressure direction, this invention randomly rotates and flips the multi-view typhoon thermal core anomaly dataset. Specifically, the rotation angles are set to 0°, 90°, 180°, and 270°, with a probability of 25%; the flip is a left-right flip with a probability of 50%.

[0033] Furthermore, the multi-view feature fusion network of the typhoon intensity estimation network model includes several feature extraction branches. The number of branches is consistent with the number of views included in the multi-view typhoon thermal core anomaly dataset. The multi-view typhoon thermal core anomaly dataset includes one main view and multiple auxiliary views.

[0034] Each feature extraction branch sequentially includes a channel attention module, a convolutional block, a first DarkNet block, and three stacked second DarkNet blocks. Each second DarkNet block in the feature extraction branch containing the main viewpoint is connected to a fusion module. For the feature extraction branches containing the secondary viewpoints, a fusion module is set between the second DarkNet blocks of the same level in every two adjacent branches for feature fusion of each secondary viewpoint feature extraction branch. The fusion result is then input into the fusion module at the corresponding level of the main viewpoint. Furthermore, the output features of the three fusion modules in the main viewpoint feature extraction branch are also input into the multi-scale feature pyramid fusion module. The output features of the multi-scale feature pyramid fusion module and the output features of all feature extraction branches are passed through a concatenation layer to obtain the output features of the multi-view feature fusion network.

[0035] The convolutional block consists of, in sequence: a convolutional layer, a batch normalization layer, and an activation function;

[0036] The first DarkNet block consists of a convolutional block and an ELAN Block, while the second DarkNet block consists of an ELAN Block and an MP Block. The main branch of the ELAN Block is a stacked structure of five CBS Blocks, and the auxiliary branch is a skip connection with one CBS Block. The output features of the auxiliary branch, the output features of the first and third convolutional blocks of the main branch, and the output features of the main branch are concatenated by a splicing layer before being input into another convolutional block to obtain the output of the ELAN Block. The MP Block consists of two branches: one branch is a cascaded max-pooling layer and a convolutional block k, and the other branch is two stacked convolutional blocks. The output of the MP Block is obtained by concatenating the outputs of the two branches through a splicing layer.

[0037] Furthermore, the wind speed estimation network of the typhoon intensity estimation network model consists of two multilayer perceptrons, which are used to obtain the single-value regression results and ordinal regression results of the output features of the multi-view feature fusion network, respectively.

[0038] Furthermore, step 2 specifically includes:

[0039] The atmospheric temperature field distribution is obtained by inverting the preprocessed satellite brightness temperature data;

[0040] The typhoon center is selected based on the best path tracking data, and the temperature field data within the specified pressure range of ±Δ1° from the typhoon center is used as the data source; where Δ1° is the preset first deviation range.

[0041] The clear sky background field temperature of the sample points was selected and calculated using a background field expansion iterative algorithm.

[0042] The background field expansion iterative algorithm can be set as follows: Select the typhoon center, use the center latitude range ±Δ2° (wide strip area) as the background field to search for clear sky points. If there are no clear sky points within the range, expand the latitude range by 1° to search until the number of clear sky points exceeds the specified value. Here, Δ2° is the preset first deviation range. Clear sky is defined as the cloud liquid water path content not exceeding the specified value.

[0043] The background field temperature is obtained by averaging the temperatures at the clear sky points. Among them, T i Let represent the temperature of the i-th clear sky point, and n represent the number of clear sky points;

[0044] The temperature of the clear sky background field is then subtracted from the data source to obtain the temperature anomaly field. Then, profile information is extracted from the pressure direction, the satellite scan line direction, and the direction perpendicular to the satellite scan line to form a multi-view typhoon thermal core anomaly dataset.

[0045] On the other hand, the present invention also provides a computer device, the computer device including a processor and a memory, the memory storing at least one computer program, the at least one computer program being loaded and executed by the processor to implement the intelligent intensity determination method based on the thermonuclear structure of a typhoon provided by the present invention as described above.

[0046] On the other hand, the present invention also provides a computer-readable storage medium storing at least one computer program, which is loaded and executed by a processor to implement the intelligent intensity determination method based on the thermonuclear structure of a typhoon provided by the present invention as described above.

[0047] The technical solution provided by this invention brings at least the following beneficial effects:

[0048] This invention proposes an intelligent typhoon intensity estimation model based on the fusion of multi-view typhoon thermal core anomaly features. The invention constructs a multi-view typhoon thermal core anomaly dataset using satellite microwave data. During model training, a multi-view, multi-scale feature information fusion method is proposed, combined with single-valued regression and ordinal regression loss functions to train and optimize the model. The final model effectively extracts multi-view features, improving the underestimation of typhoon wind speed and exhibiting good robustness and scalability, thus achieving accurate typhoon wind speed estimation. Attached Figure Description

[0049] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0050] Figure 1 A schematic diagram of the processing flow of a typhoon intensity estimation method based on a thermonuclear anomaly dataset provided for implementation of the present invention;

[0051] Figure 2 This is a schematic diagram of the typhoon intensity estimation network model used in an embodiment of the present invention;

[0052] Figure 3 This is a schematic diagram of the attention module used in an embodiment of the present invention;

[0053] Figure 4 This is a schematic diagram of the structure of the multi-view feature pyramid fusion module used in an embodiment of the present invention;

[0054] Figure 5 This is a flowchart illustrating the creation of a multi-view typhoon thermal core anomaly dataset according to an embodiment of the present invention. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be described in detail and completely below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Generally, the components of the embodiments of the present invention described and shown in the accompanying drawings can be arranged and designed using different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of the present invention.

[0056] The present invention provides an intelligent intensity estimation method based on the typhoon thermal core structure. This method is based on multi-view data fusion, two-dimensional convolutional neural network, single-value regression and ordinal regression. Without increasing the number of parameters, this method can make full use of the typhoon structure features of the multi-view thermal core anomaly dataset to improve the accuracy of typhoon intensity estimation.

[0057] like Figure 1 As shown, as one possible implementation, the specific implementation steps of the intelligent intensity determination method based on the typhoon thermonuclear structure provided by this embodiment of the invention include:

[0058] Step S1: Acquire raw satellite data and typhoon path data. In this embodiment, SNPP ATMS satellite microwave data and IBTrACS typhoon optimal path tracking data are acquired;

[0059] Step S2: Preprocess the satellite data.

[0060] This involves preprocessing the satellite brightness temperature data obtained from the microwave detector to remove outliers, such as the dates and regions where they occurred.

[0061] ATMS brightness temperature data is saved in an h5 (a file format used to store scientific data and library files) file format, and needs to be converted to npy (a file format specifically used to store multidimensional array data) format for temperature anomaly data.

[0062] In this embodiment, for brightness temperature data in h5 format, the stored data consists of inaccessible data records and accessible data indexes. The data records are divided into two parts: environmental instrument information and ATMS brightness temperature information, with the unit of brightness temperature being Kelvin (K). A complete day's data consists of data covering two global orbits (ascending and descending). This embodiment first accesses the ATMS brightness temperature data categorized by date using the typhoon date as an index. Then, based on the latitude and longitude of the typhoon center, the ATMS brightness temperature data is filtered by latitude and longitude range, retaining only data within ±10° of the typhoon center.

[0063] Step S3: Create a multi-view typhoon thermal core anomaly dataset and labels.

[0064] The process involves inverting the preprocessed satellite brightness temperature data to obtain the temperature field distribution (temperature field data) on the pressure layer, and then constructing a multi-view thermonuclear anomaly dataset using the temperature field data. Typhoon information is extracted from IBTrACS, mainly including longitude, latitude, maximum sea-level wind speed, and Saffir-Simpson typhoon level, and is divided according to typhoon name and time. The typhoon dataset used in this embodiment spans from 2012 to 2022, covering the North Pacific, South Pacific, and North Atlantic regions. The dataset is divided into a training set (2012-2017), a validation set (2018-2019), and a test set (2019-2022). Based on the satellite data obtained in step S2, a thermonuclear anomaly dataset is obtained by processing the background field iterative expansion algorithm and the anomaly field generation algorithm. View profile data is then obtained from three viewing directions (pressure direction, along the satellite scan line direction, and perpendicular to the satellite scan line direction), with dimensions of 40×320×320, 21×320×320, and 21×320×320, respectively. Of course, in practical applications, other viewing directions can be added. In this embodiment, the pressure direction is the primary viewing direction, and the other viewing directions are auxiliary viewing directions.

[0065] Step S4: Construct and train the typhoon intensity estimation network model.

[0066] A typhoon intensity estimation network model was trained using a dataset of typhoon thermal anomalies from 2012 to 2017. During training, a combination of single-valued regression and ordinal regression losses was used as the loss function for model optimization. When the combined loss converged and decreased to its minimum stationary fluctuation, the network parameters for that round were saved, resulting in the trained typhoon intensity estimation network model. During training, the network's fit was ensured by comparing changes in evaluation metrics between the validation and training sets.

[0067] Step S5, Typhoon intensity estimation test.

[0068] The trained typhoon intensity estimation model was tested on data from 2020 to 2022 to obtain intensity estimation results, which effectively improved the accuracy of typhoon intensity estimation, with a mean absolute error of approximately 3.81 m / s.

[0069] As one possible implementation, the typhoon intensity estimation network model used in this embodiment of the invention is as follows: Figure 2 As shown, it consists of three branches, which extract information from three views (in the pressure direction, along the satellite scan line direction, and perpendicular to the satellite scan line direction). Each branch first uses a channel attention mechanism to emphasize the weight of information from different channels. The backbone is DarkNet for feature extraction, and multi-scale pyramid feature fusion is performed on DarkNet blocks 3, 4, and 5 to obtain multi-view multi-scale fused feature information. This information is then fused with the feature information at each scale to obtain the extracted feature information. Finally, it is input into two multilayer perceptrons to obtain single-value regression results and ordinal regression results, respectively.

[0070] The multi-view feature fusion network of the typhoon intensity estimation network model used in this embodiment of the invention includes several feature extraction branches, the number of which is consistent with the number of views included in the multi-view typhoon thermal core anomaly dataset. Each feature extraction branch includes, in sequence, a channel attention module (Coordinatc Attention Module), a convolutional block (CBSBlock), a first DarkNet block (DarkNet Block 2), and three stacked second DarkNet blocks (DarkNet Blocks 3, 4, and 5).

[0071] The multi-view typhoon thermal core anomaly dataset includes a main view and several secondary views. Each second DarkNet block in the feature extraction branch of the main view is connected to a fusion module. For the feature extraction branches of the secondary views, a fusion module is set between the second DarkNet blocks of the same level of every two adjacent branches for feature fusion of the secondary view feature extraction branches. The fusion result is input into the fusion module of the corresponding level of the main view. The output features of the three fusion modules of the main view feature extraction branch are also input into the multi-scale feature pyramid fusion module. The output features of the multi-scale feature pyramid fusion module and the output features of all feature extraction branches are concatenated to obtain the output features (high-dimensional feature encoding vector r) of the multi-view feature fusion network.

[0072] The typhoon intensity estimation network model used in this embodiment of the invention consists of two multilayer perceptrons, which are used to obtain the single-value regression result and the ordinal regression result of the high-dimensional feature encoding vector r, respectively.

[0073] See Figure 2 In this embodiment of the invention, the CBS Block comprises, in sequence, a convolutional layer (Conv2d), a batch normalization layer (BacthNorm), and an activation function (LeakyReLU). The first DarkNet block comprises a CBS Block and an ELAN Block, and the second DarkNet block comprises an ELAN Block and an MP Block. The main branch of the ELAN Block is a stacked structure of five CBS Blocks, and the auxiliary branch is a skip connection with one CBS Block. The output features of the auxiliary branch, the output features of the first and third CBS Blocks of the main branch, and the output features of the main branch are concatenated by a splicing layer before being input into a CBS Block to obtain the output of the ELAN Block. The MP Block comprises two branches, one of which is a cascaded max pooling layer and a CBS Block, and the other is two stacked CBSBblocks. The output of the MP Block is obtained by concatenating the outputs of the two branches by a splicing layer. Furthermore, the fusion module adopted in this embodiment of the invention specifically involves: stitching the features of two perspectives together through a stitching layer, and then sequentially performing layer normalization and channel convolution operations to obtain the output features of the fusion module.

[0074] In this invention example, the channel attention module adopted is as follows: Figure 3As shown, global pooling is first used to encode spatial information globally. Furthermore, to enable the attention module to obtain more accurate location information and capture long-range spatial relationships, global pooling is decomposed into feature encoding for each channel in the X and Y directions (horizontal and vertical directions), calculated as shown in the following formula:

[0075]

[0076] in, Let x represent the feature encoding of the c-th channel in the Y and X directions, respectively. c (h, i) and x c (j, w) represent the pixel values ​​of the input image of the channel attention module of the c-th channel in the Y and X directions, respectively, and W×H is the feature space size of the input image, i.e., width and height.

[0077] By aggregating features along the two spatial directions through the above transformations, a pair of reverse-perception feature maps can be generated, enabling the CA module (channel attention module) to capture long-distance dependencies along one spatial direction and retain accurate positional relationships in the other spatial direction.

[0078] Based on the two transformations mentioned above, the captured positional relationships are fully utilized to obtain effective inter-channel relationships. First, the results of the two transformations are concatenated, and the concatenated result is sent to a 1×1 convolutional transformation function F1. Then, the result is obtained through a non-linear activation layer, as shown in the following formula.

[0079] f=δ(F1[z h , z w ])

[0080] Where f represents the output feature of the nonlinear activation layer, [·, ·] represents the cascade operation along the spatial dimension, and δ is the nonlinear activation function. This is an intermediate feature map that encodes spatial information in the X and Y directions. r is the scaling factor for controlling the block size, C is the number of channels, and z... h , z w These represent the feature encoding vectors in the Y and X directions, respectively, which are the feature encodings of all channels. This constitutes the feature encoding vector zh.

[0081] Then, the dimensions of space f are split into two independent tensors. F is obtained using two 1×1 convolution transformations. h and F w The feature F obtained after transformation h and F wA tensor with the same number of channels as the input X (i.e., the input image of the channel attention module) is computed as follows, where σ is the activation function.

[0082]

[0083] Where g w g h These are the attention values ​​in the horizontal and vertical directions of the input image, respectively.

[0084] Finally, the output of the CA module is shown below.

[0085]

[0086] Where, x c (i, j) represents the pixel value of the input image at pixel position (i, j). This represents the attention value at horizontal position i. This represents the attention value at vertical position j, and the subscript c is the channel identifier.

[0087] In this invention example, the multi-scale feature pyramid fusion module adopted is as follows: Figure 4 As shown, the input of the multi-scale feature pyramid fusion module contains three levels of features. i∈{1, 2, 3}. Where (W i ×H i ) refers to spatial dimension, C i The channel feature dimensions are preferably set to C1=128, C2=256, and C3=512 in this embodiment.

[0088] The input features of each level are compressed to C1 by a 1×1 channel convolution F1. The features of deeper layers need to be upsampled (Up(·)) to obtain the same dimension as the features of shallower layers. Then, the two are matrix-added to obtain intermediate layer information of different depths, thereby improving the effectiveness of feature information. The formula is shown below.

[0089]

[0090] After embedding, the intermediate layer yields feature vectors for different feature layers, with a dimension of C1×1. The final fused feature information is obtained by matrix addition of the three output feature layer spatial vectors, as shown in the formula below:

[0091] Embedding(Mid i =Linear(Reshape(Mid) i ))

[0092]

[0093] Among them, f y This represents the final fused feature information, i.e., the output features of the multi-scale feature pyramid fusion module. Reshape(·) represents the reshaping function, i.e., reshaping the Mid according to the specified spatial dimension. i Linear(·) is a linear activation function.

[0094] This example uses a large dataset of typhoon thermal core anomalies to train the model, employing a combination of single-valued regression loss and ordinal regression loss: Loss total =Loss SVR +Loss OR Loss SVR =MSE weighted Loss OR =L AMR +L WKL The loss function is calculated as follows:

[0095] 1) Weighted mean square error loss

[0096]

[0097] Among them, y i For truth labels, The model prediction result is given, where i is the i-th sample and γ is a hyperparameter, which is 1 / 3 in this invention.

[0098] 2) Adaptive mean residual loss

[0099] The adaptive residual loss consists of two parts: the mean loss and the residual loss. The mean loss measures the difference between the average wind speed distribution and the actual wind speed, while the residual loss measures the wind speed probability entropy of the dynamic Top-K values ​​in the distribution. Therefore, the loss function ultimately comprises two parts: the mean loss L... m Residual loss L r The calculation process is as follows:

[0100]

[0101]

[0102] L AMR =λ1L m +λ2L r

[0103] in, y represents the estimated wind speed. i Let L be the true value of wind speed, and L∈{1,2,3,..,100} represent the number of categories.

[0104] 3) κ-weighted loss

[0105]

[0106] L WKL =log(1-κ), where L WKL ∈(-∞, log2]

[0107] Among them, the parameters κ∈[-1, 1] and L are calculated based on the formula. WKL ∈(-∞, log2], Used to characterize the predictive performance of the typhoon intensity estimation network model. This indicates that the final wind speed estimate Z is based on... i，f inal The typhoon intensity category is obtained by rounding down (either rounding up or down is acceptable, but rounding down is preferred). Here, j represents the true typhoon intensity category of sample i. For example, for an example where typhoon intensity is categorized as 1 to 100, then... like but like but That is, for each sample, the typhoon intensity category predicted by the typhoon intensity estimation network model. This forms a wind speed category with the corresponding true category j. The outer product between the two wind speed classification histogram vectors. Indicates each pair of wind speed categories loss weights, L represents the number of categories.

[0108] As one possible implementation, the process of creating a multi-view hot core anomaly dataset in step S3 of this embodiment of the invention is as follows: Figure 5 As shown, the specific process is as follows:

[0109] S301, the atmospheric temperature field distribution (1000 hPa to 0 hPa) is obtained by inverting the brightness temperature data (channels 5-15) from the S-NPP ATMS satellite. The inversion calculation formula is shown below.

[0110]

[0111] Where p is pressure, θ is the zenith angle of the sensor; v i It is the ATMS channel index; T b It is the brightness temperature of ATMS; i 1，p , ..., i 2，p It is a temperature-dependent subset of pressure from channels 5-15 of the ATMS satellite; C0, C i It is the regression coefficient.

[0112] S302, the typhoon center is selected based on the best path tracking data, and the temperature field data within the range of ±10° of the typhoon center and the air pressure of 1000hpa to 100hpa is used as the data source.

[0113] S303, the clear sky background field temperature of the sample points is selected and calculated by the background field expansion iterative algorithm. The background field expansion iterative algorithm used in this invention example is to select the typhoon center, use the center latitude range ±5° (wide strip area) as the background field to search for clear sky points. If there are no clear sky points within the range, the latitude range is expanded by 1° to search until the number of clear sky points exceeds 50.

[0114] Clear skies are defined as cloud liquid water path content (CLWP) ≤ 0.0015. The formula for calculating CLWP is shown below.

[0115] CLWP=cosθ[a0+a1 ln(T s -T b23 )+a2 ln(T s -T b31 )

[0116] Where θ is the zenith angle of the sensor, and T s It is the sea surface temperature, usually denoted as 285K, T. b23 and T b31 These represent the observed brightness temperatures of channels 23.8 GHz and 31.4 GHz (i.e., channels 1 and 2 of ATMS), respectively. a0 = 8.240 - (2.622 - 1.846cosθ)cosθ, a1 = 0.754, and a2 = -2.265.

[0117] The background field temperature is obtained by averaging the temperatures at clear-sky points, as shown in the following calculation process:

[0118]

[0119] Where n represents the number of clear sky points.

[0120] S304. Subtract the temperature of the clear sky background field from the data field selected in S302 to obtain the temperature anomaly field. Then, extract the profile information from the pressure direction, the satellite scan line direction, and the direction perpendicular to the satellite scan line to form a multi-view typhoon thermal core anomaly data set.

[0121] S305, Dataset Partitioning.

[0122] All thermonuclear anomaly datasets were divided into three sets according to time: training set (2012–2017), validation set (2018–2019), and test set (2020–2022).

[0123] In an exemplary embodiment, the present invention also provides a computer device including a processor and a memory, the memory storing at least one computer program. The at least one computer program is loaded and executed by one or more processors to implement any of the above-described methods for typhoon intensity estimation based on thermonuclear anomaly datasets.

[0124] In an exemplary embodiment, the present invention also provides a computer-readable storage medium storing at least one computer program, which is loaded and executed by a processor of a computer device to implement any of the above-described methods for typhoon intensity estimation based on thermonuclear anomaly datasets.

[0125] In one possible implementation, the aforementioned computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a compact disc read-only memory (CD-ROM), magnetic tape, and an optical data storage device, etc.

[0126] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

[0127] The above descriptions are merely some embodiments of the present invention. Those skilled in the art can make various modifications and improvements without departing from the inventive concept of the present invention, and these all fall within the scope of protection of the present invention.

Claims

1. A smart intensity determination method based on the thermonuclear structure of a typhoon, characterized in that, Includes the following steps: Step 1: Obtain satellite brightness temperature data and typhoon path data; Step 2: Preprocess the satellite brightness temperature data, and then perform inversion processing to obtain temperature field data; Temperature field data is processed based on typhoon path data to construct a multi-view typhoon thermal core anomaly dataset and its corresponding typhoon intensity labels; Step 3: Estimate wind speed on the thermonuclear anomaly dataset using a deep learning-based regression method to obtain single-value regression and ordinal regression results for wind speed; The typhoon intensity estimation network model used consists of two parts: a multi-view feature fusion network and a wind speed estimation network; Among them, the multi-view feature fusion network is used to fuse the input multi-view information at multiple scales to obtain a set of high-dimensional feature encoding vectors. The wind speed estimation network adopts The function takes the input high-dimensional feature vector Mapping to classification vectors Classification vectors It consists of two parts: single-value regression and ordinal regression; the single-value regression regresses wind speed to a single value within a specified range. , where i represents the sample index; ordinal regression regresses wind speed into a specified range in the form of a probability distribution. , where j represents the discrete value index of the wind speed probability distribution; The final wind speed estimate is obtained by combining single-value regression and ordinal regression. in, This represents the final wind speed estimate, with weighting coefficients. The range of its value is [0,1]; The multi-view feature fusion network includes several feature extraction branches, the number of which is consistent with the number of views included in the multi-view typhoon thermal core anomaly dataset. The views in the multi-view typhoon thermal core anomaly dataset include one main view and multiple auxiliary views. Each feature extraction branch sequentially includes a channel attention module, a convolutional block, a first DarkNet block, and three stacked second DarkNet blocks. Each second DarkNet block in the feature extraction branch containing the main viewpoint is connected to a fusion module. For the feature extraction branches containing the secondary viewpoints, a fusion module is set between the second DarkNet blocks of the same level in every two adjacent branches for feature fusion of each secondary viewpoint feature extraction branch. The fusion result is then input into the fusion module at the corresponding level of the main viewpoint. Furthermore, the output features of the three fusion modules in the main viewpoint feature extraction branch are also input into the multi-scale feature pyramid fusion module. The output features of the multi-scale feature pyramid fusion module and the output features of all feature extraction branches are passed through a concatenation layer to obtain the output features of the multi-view feature fusion network. The convolutional block consists of, in sequence: a convolutional layer, a batch normalization layer, and an activation function; The first DarkNet block consists of a convolutional block and an ELAN Block, and the second DarkNet block consists of an ELAN Block and an MP Block. The main branch of the ELAN Block is a stacked structure of five CBS Blocks, and the auxiliary branch is a skip connection with a CBSB Block. The output features of the auxiliary branch, the output features of the first and third convolutional blocks of the main branch, and the output features of the main branch are concatenated by a splicing layer and then input into a convolutional block to obtain the output of the ELAN Block. The MP Block consists of two branches. One branch is a cascaded max pooling layer and a convolutional block k, and the other branch is two stacked convolutional blocks. The output of the MP Block is obtained by splicing the outputs of the two branches by a splicing layer.

2. The intelligent intensity determination method based on the thermonuclear structure of a typhoon as described in claim 1, characterized in that, Step 2 includes: The atmospheric temperature field distribution is obtained by inverting the preprocessed satellite brightness temperature data; The typhoon center was selected based on the best track tracking data. Temperature field data within a specified pressure range is used as the data source; among which, This is the preset first deviation range; The clear sky background field temperature of the sample points was selected and calculated using a background field expansion iterative algorithm. The background field expansion iterative algorithm is set as follows: select the typhoon center, and calculate the range of center latitude. As a background field, search for clear-sky points. If no clear-sky points are found within the range, expand the latitude range. The search continues until the number of clear-sky points exceeds a specified value; where, This is the preset first deviation range; clear skies are defined as cloud liquid water path content not exceeding a specified value. The background field temperature is obtained by averaging the temperatures at the clear sky points. ,in, Let represent the temperature of the i-th clear sky point, and n represent the number of clear sky points; The temperature of the clear sky background field is then subtracted from the data source to obtain the temperature anomaly field. Then, profile information is extracted from the pressure direction, the satellite scan line direction, and the direction perpendicular to the satellite scan line to form a multi-view typhoon thermal core anomaly dataset.

3. The intelligent intensity determination method based on the thermonuclear structure of a typhoon as described in claim 1, characterized in that, When training the network parameters based on a multi-view typhoon thermal core anomaly dataset, the loss function used for the typhoon intensity estimation network model is set as follows: in, This represents the total loss of the typhoon intensity estimation network model. This represents the loss from a single-value regression task. This represents the loss of the ordinal regression task; and The loss is the weighted mean square error. , For adaptive mean residual loss and -Weighted loss The fusion result.

4. The intelligent intensity determination method based on the typhoon thermonuclear structure as described in claim 3, characterized in that, Adaptive mean residual loss Specifically: Where N is the batch size set during training. This represents the K sample sets with the highest accuracy. This represents the estimated wind speed. Label the true value of wind speed. Indicates the number of categories, used to characterize the number of wind speed intensity levels. , Indicates mean loss and residual loss The weighting coefficients.

5. The intelligent intensity determination method based on the thermonuclear structure of a typhoon as described in claim 3, characterized in that, -Weighted loss Specifically: in, Used to characterize the predictive performance of the typhoon intensity estimation network model. This indicates that the final wind speed estimate is based on the results. The typhoon intensity category obtained after rounding down, where j represents the true typhoon intensity category of sample i, if ,but ,like ,but ; Typhoon Intensity Category The outer product between the histogram vectors of and j, for each pair of typhoon intensity categories loss weights , Indicates the number of categories.

6. The intelligent intensity determination method based on the thermonuclear structure of a typhoon as described in claim 1, characterized in that, Step 3 also includes data augmentation processing on the multi-view typhoon thermal core anomaly dataset, including: randomly rotating and flipping the multi-view typhoon thermal core anomaly dataset; wherein the rotation angle is specifically set as follows: , , and The probability is The flip is a left-right flip with a probability of . .

7. A computer device comprising a processor and a memory, wherein the memory stores at least one computer program, the at least one computer program being loaded and executed by the processor to implement the intelligent intensity determination method based on the thermonuclear structure of a typhoon as described in any one of claims 1 to 6.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores at least one computer program, which is loaded and executed by a processor to implement the intelligent intensity determination method based on the thermonuclear structure of a typhoon as described in any one of claims 1 to 6.

Images