A typhoon multi-parameter monitoring method fusing timing information and physical constraints

Abstract

The present application relates to a kind of typhoon multi-parameter monitoring method of fusing timing information and physical constraint.The method includes: data preprocessing, spatial supervision heat map construction, typhoon multi-parameter prediction model construction, graph output's differentiable parameter inversion, joint loss function construction, model training and reasoning.The present application is aimed at the problem that "center, intensity, maximum wind speed radius are related to each other but are often independently regressed" in infrared image typhoon monitoring, and proposes a training constraint system of "historical prior input + physical constraint wind field heat map + graph output differentiable inversion + point-graph consistency joint loss".Experiments prove that the accuracy of typhoon center positioning, intensity inversion and wind circle radius inversion of the present application is better than that of international advanced methods.

Description

Technical Field

[0001] This invention belongs to the field of marine meteorological satellite remote sensing and artificial intelligence typhoon monitoring technology, and relates to a multi-parameter monitoring method for typhoons that integrates temporal information and physical constraints. Background Technology

[0002] Tropical cyclones or typhoons are significant hazardous weather systems affecting global oceans and coastal areas. Operational typhoon monitoring requires not only obtaining the typhoon's center location but also simultaneously understanding its intensity, characterized by maximum sustained wind speeds (MSGS), and the core wind field scale, characterized by the maximum sustained wind radius. The typhoon center serves as the geometric benchmark for path analysis, early warning location, and storm relative image interpretation; typhoon intensity directly reflects its potential destructive power; and the maximum sustained wind radius is closely related to the eyewall scale, the extent of strong winds near the center, storm surge, and nearshore hazard assessment.

[0003] Current operational monitoring relies on multi-source information, including aircraft reconnaissance, sea surface observation, scatterometers, radar, and optimal path analysis. However, direct wind field observations are not continuous in space and time; aircraft reconnaissance and radar coverage are limited; scatterometer and microwave observations have long revisit cycles and are easily affected by precipitation, field of view, and observation geometry. In contrast, geostationary satellite infrared imagery offers advantages such as high frequency, wide area coverage, and long-term consistency, continuously capturing typhoon eyes, eyewalls, central dense cloud areas, spiral cloud bands, and cloud top brightness and temperature structures, making it a crucial data source for objective typhoon monitoring.

[0004] In recent years, deep learning-based typhoon monitoring methods have been increasingly used to estimate typhoon centers or intensities from infrared satellite imagery. These methods typically input the current infrared imagery into a convolutional neural network or Transformer model and directly regress the center location, intensity, or other scalar parameters. However, a single infrared image can only reflect the current cloud top thermal structure and cannot fully express the typhoon's recent movement direction, intensity changes, and core-scale evolution. For weak typhoons, loosely structured typhoons, and asymmetric typhoons affected by vertical wind shear, there is often a shift between the low-level circulation center and the cold cloud cluster at the cloud top. Relying solely on the current imagery can easily lead to center positioning errors and unstable estimations of intensity and maximum wind speed radius.

[0005] Furthermore, existing methods often treat the typhoon center, intensity, and maximum wind speed radius as independent regression objectives, lacking a mechanism to simultaneously constrain the consistency of spatial structure and physical parameters. Conventional Gaussian heatmaps are primarily used for center location, with their peak located at the typhoon center and decreasing with distance. However, these heatmaps do not represent the radial structure of the typhoon wind field and are difficult to provide information related to the maximum wind speed radius and intensity. Therefore, while Gaussian heatmaps are helpful in learning spatial attention near the center, their ability to constrain the maximum wind speed radius and near-surface wind field structure is limited.

[0006] In summary, existing intelligent typhoon monitoring technologies based on infrared satellite imagery still have the following shortcomings: First, they lack explicit utilization of historical typhoon motion, intensity, and structural evolution information; second, they lack physical consistency constraints between the center, intensity, and maximum wind speed radius; third, conventional center heatmaps cannot represent the radial wind field structure near the maximum wind speed radius; and fourth, they lack generalization stability across ocean basins, typhoons of different intensities, and typhoons at different core scales. Therefore, a method is needed that can integrate historical priors, utilize physically constrained heatmaps, and achieve multi-parameter joint monitoring. Summary of the Invention

[0007] The technical problem this invention aims to solve is: when using only or mainly geostationary satellite infrared imagery for typhoon monitoring, how to make full use of historical typhoon status information and inject the physical structure of the typhoon's radial wind field into the model training process, thereby improving the accuracy, stability, and interpretability of the joint estimation of the current typhoon's center, intensity, and maximum wind speed radius.

[0008] This invention proposes a method, system, and medium for joint monitoring of typhoon multi-parameters based on historical priors and physically constrained heatmaps. The method takes current non-centered infrared satellite imagery and historical typhoon conditions as input. It directly estimates the typhoon center offset, intensity, and maximum wind speed radius using a point output branch, while simultaneously learning a physically constrained wind field heatmap constructed from the typhoon center, intensity, and maximum wind speed radius using a graph output branch. Furthermore, the graph output is made differentiable to the typhoon center, intensity, and maximum wind speed radius, and together with the point output, forms a joint loss function.

[0009] The technical solution provided by this invention is as follows: A method for multi-parameter monitoring of typhoons that integrates temporal information and physical constraints, comprising: Step 1: Collect and process geostationary satellite infrared window channel images and typhoon optimal path data; perform random translation and fixed-size cropping on the infrared images to form non-centered infrared image samples at the current time t, so that the real typhoon center is not fixed in the center of the image; extract the historical typhoon center, intensity and maximum wind speed radius at t-6 h and t-12 h to form historical typhoon state vectors. Step 2: Calculate the radial distance from each pixel to the actual typhoon center within the silhouette image coordinate system; construct a spatial monitoring heatmap based on the actual typhoon center, actual intensity, and actual maximum wind speed radius. Step 3: Construct a multi-parameter typhoon prediction model. This model includes an infrared image coding branch, a historical prior coding branch, a point output branch, and a graph output branch; it outputs at least one point regression vector, representing the directly estimated typhoon center, intensity, and maximum wind circle radius; it also includes an output spatial heat map. Step 4: Convert the output spatial heatmap into a non-negative score map, and perform Softmax normalization on all pixels to obtain the spatial probability distribution; then use Soft-argmax to perform probability-weighted summation on the pixel coordinates to obtain the typhoon center inverted from the map output; use probability-weighted response or high-response region statistics to obtain the typhoon intensity inverted from the map output; use probability-weighted radial distance relative to the inversion center to obtain the maximum wind speed radius inverted from the map output. Step 5: Construct the joint loss function: ; in, For point regression loss; , and These represent the typhoon center, intensity, and maximum wind circle radius directly estimated by the model output, respectively. To determine the consistency loss of graph inversion points, , and This is the typhoon center, intensity, and maximum wind circle radius that can be differentiably inferred from M; the coefficient 0.1 is used to balance the magnitude of heat map loss and point-level loss; Loss due to spatial heat map monitoring; Step 6: During the training phase, input the non-centered infrared image and historical typhoon state vectors into the model, and perform end-to-end optimization of the model parameters according to the joint loss function. During the inference phase, input the current infrared image and historical typhoon state, and output the current typhoon center location, intensity, and maximum wind speed radius. At the same time, a spatial heat map can be output as an interpretive result.

[0010] Preferably, in step two, the spatial heat map is a central heat map or a physically constrained wind field heat map; the central heat map adopts a Gaussian form; the physically constrained wind field heat map is generated using Rankine, Holland, or Willoughby radial wind profiles.

[0011] Preferably, step four involves the following steps: (1) Spatial probability distribution obtained from score map S: ; in, Indicates the first 1 pixel, It is the score value at that pixel. It is the first The score value at each pixel. It is the Softmax temperature; (2) Typhoon center inferred from spatial probability distribution: Let the first The pixel coordinates of each pixel are , , and Let the width and height of the map be respectively. The center offset inferred from the map is defined as: ; ; The typhoon center is estimated as the probability-weighted centroid of the spatial probability distribution map; then, the typhoon intensity ( This is defined as the probability-weighted response of the spatial probability distribution map: ; (3) The radius of the typhoon's maximum wind circle inferred from the spatial probability distribution map: calculated as the average radial distance of the high-response pixel relative to the inferred center: ; in, , These are the coordinates of the typhoon's center, inferred from the map.

[0012] Preferably, in step five, the point regression loss adopts a weighted MAE-MSE hybrid loss: ; in, These are the typhoon centers corresponding to and Axis coordinates, subscript Represents actual ground data, excluding The variables represent the model output; the weights before the four terms are used to balance the numerical scale of the center offset, intensity, and maximum wind circle radius task. , These represent the typhoon intensity and the radius of maximum wind speed, respectively.

[0013] Preferably, in step five, the wind field thermal map monitoring loss adopts the SmoothL1 loss: ; in, and These are the estimated map and the target map, respectively.

[0014] This invention addresses the problem in infrared typhoon imagery monitoring where "center, intensity, and maximum wind speed radius are correlated but often regressed independently." It proposes a method combining "historical prior input + physically constrained wind field heatmap + graphical output with differentiable inversion + point..." Figure 1 The training constraint system of "consistent joint loss" is used. Experiments have shown that the accuracy of this invention in typhoon center location, intensity inversion, and wind circle radius inversion is superior to internationally advanced methods.

[0015] Compared with the prior art, the present invention has the following beneficial effects: (1) Significantly improves the accuracy of multi-parameter joint monitoring of typhoons by infrared satellites. This invention introduces historical typhoon centers, intensities, and maximum wind speed radii at t-6 h and t-12 h, enabling the model to utilize recent motion and structural evolution information, thereby reducing the uncertainty of single infrared images in weak typhoons, asymmetric typhoons, and loosely clouded scenarios. In independent global test samples, the preferred scheme achieves a center positioning error of approximately 30.19 km, an intensity RMSE of approximately 2.35 m / s, and a maximum wind speed radius MAE of approximately 8.72 km.

[0016] (2) Improve the physical consistency of intensity and maximum wind speed radius estimation. Conventional Gaussian heatmaps mainly mark the typhoon center and cannot express the radial wind structure near the RMW. The Rankine, Holland or Willoughby physically constrained wind field heatmaps of this invention are generated by the true center, intensity and RMW, so that the output branches of the graph must encode the typhoon core scale and radial wind field morphology, thereby improving the intensity and RMW estimation.

[0017] (3) The point output and the graph output are mutually constrained, improving interpretability and stability. This invention makes the graph output differentiable inverse into center, intensity, and RMW, and incorporates the inversion result into the loss function, so that the spatial graph is not only a visualization result, but also a structural representation of recoverable physical parameters. This reduces the inconsistency between the point output and the spatial response, and provides a basis for physical interpretation in the inference stage.

[0018] (4) It has good cross-basin generalization ability and convenient engineering implementation. This method does not rely on fixed satellite products, fixed sea basins or fixed network structures, and can be adapted to various models such as convolutional neural networks, Transformers, encoder-decoders, and multi-task networks. Experiments show that historical priors can reduce center positioning error, intensity RMSE and RMW MAE by about 49.4%, 54.5% and 71.2% respectively compared with using only current imagery; after replacing the Gaussian heat map with a physically constrained wind field heat map, RMW estimation and intensity estimation are further improved, among which the Holland wind field heat map has better overall performance. Attached Figure Description

[0019] Figure 1 Heatmaps generated using different heatmap generation methods; Figure 2 This is a schematic diagram of the structure of the typhoon multi-parameter prediction model in an embodiment of the present invention. Detailed Implementation

[0020] The implementation of this patent will be further described below with reference to specific embodiments. It should be understood that the following embodiments are only used to illustrate the preferred implementation of this patent and do not limit the scope of protection of this patent; equivalent substitutions made by those skilled in the art for historical time windows, heatmap profile parameters, network backbone, loss weights, optimizers or data sources without departing from the concept of this patent shall all fall within the scope of protection of this patent.

[0021] This embodiment takes a global typhoon monitoring mission as an example. At the current analysis time t, a non-centered geostationary satellite infrared window image is input, along with typhoon status data from two historical times: t-6 h and t-12 h. The output is the typhoon center offset, typhoon intensity, and maximum wind speed radius at the current time. The overall implementation includes data construction, physical constraint heatmap generation, model output, graph output differentiable inversion, joint loss training, and inference output.

[0022] S110 Data Collection and Sample Construction.

[0023] First, collect geostationary satellite infrared window channel imagery and optimal typhoon track data. Infrared satellite imagery can be from GridSat-B1, Himawari, GOES, FY series, or other similar products, providing spatial structure information such as cloud top brightness temperature, eye region, eyewall, central dense cloud area, and spiral cloud bands. Optimal typhoon track data should include at least the typhoon center longitude, center latitude, maximum sustained wind speed (MST), and maximum wind speed radius. MST should be uniformly converted to m / s, and maximum wind speed radius should be uniformly converted to km.

[0024] For the current moment The infrared images are randomly shifted and cropped at a fixed size to obtain non-centered infrared image samples X1,t, so that the actual typhoon center is not fixed at the center of the image. This processing is used to simulate the situation where the initial center location is uncertain in operational monitoring, and forces the model to learn the offset of the actual typhoon center relative to the cropped image center.

[0025] Further extract the typhoon center, intensity, and maximum wind speed radius at t-6 h and t-12 h as historical prior inputs. The historical state vector is defined as follows: ; in, and These represent the longitude and latitude of the typhoon center, respectively. Indicates typhoon intensity. Indicates the radius of maximum wind speed. Current time. The optimal path record is not used as model input; it is only used to generate supervision labels. The output label for each sample is defined as: ; in, and These represent the horizontal and vertical distances between the actual typhoon center and the center of the silhouette image, respectively. Indicates the current typhoon intensity. This indicates the radius of the current maximum wind speed.

[0026] S120 Physical constraint heatmap target generation.

[0027] In the silhouette image coordinate system, calculate the radial distance from each pixel to the actual typhoon center. Let the pixel coordinates of the actual typhoon center in the silhouette image be... , The image spatial resolution is s The radial distance is calculated using the following formula: .

[0028] Based on radial distance 1. Typhoon intensity and the radius of the current maximum wind speed Construct a spatial monitoring heatmap. This heatmap can be a Gaussian heatmap that only emphasizes the central location, or a Rankine, Holland, or Willoughby wind field heatmap that includes the radial wind field physical structure, such as... Figure 1 As shown.

[0029] (1) Gaussian heatmap ( Gaussian heatmaps are used to provide center localization supervision, and their response decays as the distance from the pixel to the true center increases. ; in, It controls the spatial diffusion range of the Gaussian core. Gaussian heatmaps are mainly used to mark the location of the typhoon center, but do not explicitly express the radial wind field structure near the radius of maximum wind speed.

[0030] (2) Rankine wind field heat map ( Rankine heatmaps enhance with increasing radius within the maximum wind speed radius and decay according to a power law outside the maximum wind speed radius. ; ; in, This represents the outer core decay coefficient. This form allows the model to incorporate the high-response structure near the RMW.

[0031] (3) Holland wind field heat map ( In the preferred embodiment, a Holland heatmap is used as the physical constraint wind field heatmap, and its expression is: ; in, This refers to the shape parameters of Holland. This heatmap exhibits a high response near the radius of maximum wind speed and varies along the radial wind profile of the typhoon on both the inner and outer sides, thus simultaneously carrying information on the center, intensity, and radius of maximum wind speed.

[0032] (4) Willoughby wind field heat map ( Willoughby heatmaps employ a kernel-growing profile and an outer-kernel double-exponential-decreasing profile, with smooth connections in the transition region: ; ; ; ; .

[0033] in, The distance from each grid point in the image to the center of the typhoon. This represents the core heat map within the radius of the maximum wind speed. This represents the outer core heat map beyond the radius of maximum wind speed. and These represent the inner and outer boundaries of the smooth transition region between kernel functions and outer kernel functions, respectively. The kernel region, the heatmap is mainly composed of The decision, in The transition region, according to weight ,right and Perform smooth combination, in The outer core region, the heat map is mainly composed of Decision. Subscript and These represent the kernel region and the outer kernel region, respectively.

[0034] Through the above heatmap construction method, the scalar intensity and maximum wind speed radius in the optimal path are transformed into spatial heatmap supervision, so that the model not only learns the center location, but also learns the spatial structure related to the wind field scale of the typhoon core.

[0035] S130 model output and feature fusion.

[0036] This invention does not limit the specific model backbone structure. For example... Figure 2As shown, in one implementation, the typhoon multi-parameter prediction model includes an infrared image coding branch, a historical prior coding branch, a point output branch, and a graph output branch. The infrared image coding branch can be composed of convolutional and residual bottleneck blocks to extract structural features such as cloud top brightness temperature gradient, eye region, eyewall, central dense cloud region, and spiral cloud band; the historical prior coding branch can be composed of fully connected layers to map the historical typhoon states at t-6 h and t-12 h to the latent feature space.

[0037] Historical prior features are projected onto the same channel dimension as the infrared image features and then broadcast to the spatial dimension of the image features before fusion. Specifically, let the image features extracted from the infrared image branch be... Historical a priori characteristics are .

[0038] ; in, Indicates batch size. The number of channels representing image features. and These represent the spatial height and width of the image features, respectively. Dimensions representing historical a priori characteristics Indicates infrared image information. Representing historical prior information, Represent the real number space. First, pass through a linear projection layer or a fully connected layer. φ(·) Mapping historical prior features to the same channel dimension as infrared image features yields:

[0039] in, This represents the historical prior channel features after projection. φ(·) This indicates a linear projection or a fully connected transformation. (Then...) Transforming the feature into a B'×C×1×1 shape and broadcasting it along the spatial dimension to the same size as the infrared image feature yields: ; in, Reshape(·) This indicates a dimensional transformation operation. Broadcast(·) This indicates a spatial dimension broadcast operation. This represents the historical prior spatial characteristics following the broadcast. Then... Features of infrared images Feature-level fusion is performed. In this embodiment, element-wise addition is preferably used to obtain the fused features: ; in, This represents the fused features. This fusion method is a common implementation of feature-level prior information injection in neural networks, used to introduce prior information such as historical paths, intensity, or movement trends into infrared image features. The point output branch directly outputs the typhoon center offset, typhoon intensity, and maximum wind speed radius based on the fused features; the graph output branch outputs a center heatmap or a physically constrained wind field heatmap based on the fused features.

[0040] Differentiable parameter inversion from the S140 plot output.

[0041] To ensure that the output image not only visually approximates the target heatmap but also reconstructs the typhoon's physical parameters, this embodiment converts the output image M into a non-negative score image S and performs Softmax normalization on all pixels to obtain the spatial probability distribution: ; in, Indicates the first 1 pixel, It is the score value at that pixel. It is the first The score value at each pixel. It is the Softmax temperature; For physically constrained wind field heatmaps 5.0 is a good starting point, but it can also be adjusted based on image resolution and thermal response range.

[0042] After obtaining the spatial probability distribution, the pixel coordinates are summed using a soft-argmax method to obtain the typhoon center shift retrieved from the heatmap: ; .

[0043] The above calculation estimates the typhoon center as a weighted centroid of the spatial probability distribution. Compared to the hard argmax operation, this preserves differentiability and allows the centering error to be backpropagated to the graph output branch. When retrieving the intensity from the graph output, a probabilistically weighted response is used: ; When retrieving the maximum wind speed radius from the graph, the radial distance of each pixel is calculated with the graph inversion center as a reference, and the radial distance of the high-response region is probability-weighted: ; in, , These are the coordinates of the typhoon's center, inferred from the map.

[0044] This inversion process establishes a differentiable relationship between the spatial heat map and three point parameters: center, intensity, and radius of maximum wind speed, providing a basis for subsequent point... Figure 1The loss of consistency provides a computational basis.

[0045] S150 joint loss function training.

[0046] During the training phase, the point output branch and the graph output branch are optimized together. The overall loss consists of point regression loss, heatmap supervision loss, and graph inversion point consistency loss, defined as:

[0047] in, For point regression loss; , and These represent the typhoon center, intensity, and maximum wind circle radius directly estimated by the model output, respectively. To determine the consistency loss of graph inversion points, , and This is the typhoon center, intensity, and maximum wind circle radius that can be differentiably inferred from M; the coefficient 0.1 is used to balance the magnitude of heat map loss and point-level loss; This represents the loss in spatial heat map monitoring.

[0048] The point regression loss employs a weighted MAE-MSE hybrid loss to balance the different dimensions of center offset, intensity, and maximum wind speed radius:

[0049]

[0050]

[0051] in, These are the typhoon centers corresponding to and Axis coordinates, subscript Represents actual ground data (current values), excluding The variables represent the model output (estimates); the weights before the four terms are used to balance the numerical scale of the center offset, intensity, and maximum wind circle radius for the task. , These represent the typhoon intensity and the radius of maximum wind speed, respectively.

[0052] The heatmap-supervised loss employs SmoothL1 loss to improve robustness to local sharp structures such as central peaks and high eyewall responses. ; in, and These represent the model-estimated heatmap and the target heatmap, respectively.

[0053] Through the aforementioned joint loss, the point output branch obtains scalar supervision, the graph output branch obtains spatial structure supervision, and the graph inversion results are also constrained to recover the true typhoon parameters, thereby improving the physical consistency between the center, intensity, and maximum wind speed radius.

[0054] S160 Model Optimization and Online Inference.

[0055] The model is optimized end-to-end using the training samples constructed in step S110. The optimizer can be Adam, SGD, or a variant thereof. In a preferred implementation, the initial learning rate is 1e-4, reduced to 1e-5 using cosine annealing, the batch size is 256, and the training run is 100 epochs; these parameters can be adjusted according to the sample size, model capacity, and business requirements.

[0056] During the inference phase, the model takes into account the current non-centered infrared image and historical typhoon state vectors (t-6 h and t-12 h). The model can output the current typhoon center offset, intensity, and maximum wind speed radius. If the latitude and longitude of the cropped image center are known, the center offset can be converted to the current typhoon center latitude and longitude. The physically constrained wind field heatmap generated by the output branch can serve as an interpretive result, used to demonstrate the central region, core scale, and radial wind field structure of interest to the model.

[0057] S170 Performance Verification and Parameter Optimization.

[0058] Independent test datasets were used to evaluate indicators such as center positioning error, root mean square error of intensity, mean absolute error of maximum wind speed radius, graph inversion intensity error, graph inversion RMW error, consistency error between point output and graph inversion output, and radial wind field profile error.

[0059] In a specific validation instance, global typhoon samples from 2000 to 2020 were used for training, samples from 2021 to 2022 were used for validation, and global samples from 2023 to 2024 were used for independent testing. The model input consisted of current non-centered infrared imagery and historical typhoon states at t-6 h and t-12 h. The output image used Holland physically constrained wind field heatmaps. Test results showed that the model achieved a centering error of approximately 30.19 km, a root mean square error of approximately 2.35 m / s in intensity, and a mean absolute error of approximately 8.72 km in maximum wind speed radius.

[0060] Table 1. Comparison between the method of this invention and existing internationally advanced methods .

[0061] Based on the evaluation results, the historical time window and Softmax temperature can be further adjusted. Holland shape parameters Rankine outer core decay coefficient The system incorporates Willoughby profile parameters, point regression task weights, and heatmap loss weights to adapt to typhoon samples of different ocean basins, intensities, and core scales.

Claims

1. A method for multi-parameter monitoring of typhoons that integrates temporal information and physical constraints, characterized in that, Includes the following steps: Step 1: Collect and process geostationary satellite infrared window channel images and typhoon optimal path data; perform random translation and fixed-size cropping on the infrared images to form non-centered infrared image samples at the current time t, so that the real typhoon center is not fixed in the center of the image; extract the historical typhoon center, intensity and maximum wind speed radius at t-6 h and t-12 h to form historical typhoon state vectors. Step 2: Calculate the radial distance from each pixel to the actual typhoon center within the silhouette image coordinate system; construct a spatial heat map based on the actual typhoon center, actual intensity, and actual maximum wind speed radius. Step 3: Construct a multi-parameter typhoon prediction model. This model includes an infrared image coding branch, a historical prior coding branch, a point output branch, and a graph output branch; it outputs at least one point regression vector, representing the directly estimated typhoon center, intensity, and maximum wind circle radius; it also includes an output spatial heat map. Step 4: Convert the output spatial heatmap into a non-negative score map, and perform Softmax normalization on all pixels to obtain the spatial probability distribution; then use Soft-argmax to perform probability-weighted summation on the pixel coordinates to obtain the typhoon center inverted from the map output. The typhoon intensity inverted from the graph is obtained by using probability-weighted response or high-response region statistics; the maximum wind speed radius inverted from the graph is obtained by using probability-weighted radial distance relative to the inversion center. Step 5: Construct the joint loss function: ； in, For point regression loss; , and These represent the typhoon center, intensity, and maximum wind circle radius directly estimated by the model output, respectively. To determine the consistency loss of graph inversion points, , and This is the typhoon center, intensity, and maximum wind circle radius that can be differentiably inferred from M; the coefficient 0.1 is used to balance the magnitude of heat map loss and point-level loss; Loss due to spatial heat map monitoring; Step 6: During the training phase, input the non-centered infrared image and historical typhoon state vectors into the model, and perform end-to-end optimization of the model parameters according to the joint loss function. During the inference phase, input the current infrared image and historical typhoon state, and output the current typhoon center location, intensity, and maximum wind speed radius. At the same time, output a spatial heat map as an interpretive result.

2. The typhoon multi-parameter monitoring method integrating temporal information and physical constraints according to claim 1, characterized in that, In step two, the spatial heat map is either a central heat map or a physically constrained wind field heat map; the central heat map is in Gaussian form; the physically constrained wind field heat map is generated using Rankine, Holland, or Willoughby radial wind profiles.

3. The typhoon multi-parameter monitoring method integrating temporal information and physical constraints according to claim 1, characterized in that, Step four involves the following steps: (1) Spatial probability distribution obtained from score map S: ； in, Indicates the first 1 pixel, It is the score value at that pixel. It is the first The score value at each pixel. It is the Softmax temperature; (2) Typhoon center inferred from spatial probability distribution: Let the first The pixel coordinates of each pixel are , , and These are the width and height of the map, respectively. ; ； The typhoon center is estimated as the probability-weighted centroid of the spatial probability distribution map; then, the typhoon intensity ( This is defined as the probability-weighted response of the spatial probability distribution map: ； (3) The radius of the maximum wind circle of the typhoon inferred from the spatial probability distribution map ( ): Calculated as the average radial distance of the high-response pixel relative to the inferred center: ； in, , These are the coordinates of the typhoon's center, inferred from the map.

4. The typhoon multi-parameter monitoring method integrating temporal information and physical constraints according to claim 1, characterized in that, In step five, the point regression loss uses a weighted MAE-MSE hybrid loss: ； in, These are the typhoon centers corresponding to and Axis coordinates, subscript Represents actual ground data, excluding The variables represent the model output; the weights before the four terms are used to balance the numerical scale of the center offset, intensity, and maximum wind circle radius task. , These represent the typhoon intensity and the radius of maximum wind speed, respectively.

5. The typhoon multi-parameter monitoring method integrating temporal information and physical constraints according to claim 1, characterized in that, In step five, the wind field heatmap monitoring loss adopts the SmoothL1 loss: ； in, and These are the estimated map and the target map, respectively.

Images