Satellite cloud image typhoon eye positioning method and system based on hybrid expert network

Abstract

The application belongs to the cross field of meteorological remote sensing and deep learning, and discloses a satellite cloud map typhoon eye positioning method and system based on a mixed expert network. The method adopts a pre-training visual language model backbone network to cooperate with low-rank adaptive fine-tuning to extract multi-scale fusion features, identifies a typhoon cloud type through a cloud type perception router, and drives a one-way, double-path mixed expert architecture through a Top-K gate mechanism. A coordinate estimation path utilizes a space-preserving expert layer and a heat map regression to realize sub-pixel level typhoon eye positioning. A meteorological parameter estimation path jointly estimates a central air pressure, a maximum wind speed and a wind circle radius through a double-branch cross attention structure. A four-stage progressive training strategy and a multi-task loss function containing Dvorak physical constraints are used for joint optimization. Finally, a typhoon meteorological parameter early warning report is automatically output. The positioning accuracy of the application reaches 100% within a range of 50km, and 12-dimensional meteorological parameters are jointly estimated.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of meteorological remote sensing and deep learning, specifically to a method and system for typhoon eye localization based on satellite cloud images using a hybrid expert network. Background Technology

[0002] Typhoons (tropical cyclones) are among the most destructive natural disasters globally, primarily occurring in tropical and subtropical oceans. They are characterized by their sudden onset and immense destructive power, posing a serious threat to aviation and maritime safety. Upon landfall, they cause enormous economic losses and casualties to coastal cities and agricultural production. Accurate location of the typhoon eye is a crucial prerequisite for analyzing and forecasting typhoon paths, estimating typhoon intensity, and issuing disaster warnings. The accuracy of this location directly impacts the effectiveness of subsequent path estimations and disaster mitigation measures.

[0003] Meteorological satellites offer advantages such as wide coverage, high spatiotemporal resolution, and independence from weather conditions, making them the primary means of all-weather typhoon monitoring. Infrared satellite cloud images provide information on typhoon cloud top temperature distribution, and can be reliably acquired even at night or under cloudy conditions, serving as the core data source for typhoon eye location. However, due to the significant differences in cloud system morphology at different stages of typhoon development—from chaotic and disordered tropical depressions to spiral-shaped tropical storms, and then to strong typhoon eyes with clear eyewalls—accurately locating the typhoon eye from these highly variable infrared satellite cloud images has always been a major technical challenge in the field of meteorological remote sensing.

[0004] Currently, mainstream typhoon eye location methods can be broadly categorized as follows: The first is the manual interpretation method, represented by Dvorak technology, where meteorological analysts subjectively locate the typhoon center based on cloud shape characteristics on satellite cloud images and empirical rules. This method demands a high level of professional expertise from analysts, and the interpretation results vary significantly between different analysts, making it difficult to meet the demands of rapid processing of massive amounts of real-time cloud images. The second method is based on traditional computer vision, such as using mathematical morphology to extract the typhoon eye region and using spiral mathematical models to fit typhoon cloud bands. However, these methods heavily rely on the specific morphology of the typhoon's typical eye region or spiral cloud bands, and their ability to locate early, weak typhoons or irregular typhoon cloud systems is extremely limited. Thirdly, there are deep learning-based methods, such as the typhoon center localization method based on deep reinforcement learning disclosed in Chinese patent CN 111462230B. This method transforms the localization problem into a Markov decision process, trains an agent through deep Q-learning to gradually approach the typhoon center with a search box, and uses a VGG-16 convolutional network to extract cloud features within the search box. It has achieved certain localization results for typhoons of different strengths. However, this method has the following obvious limitations: First, it applies a single localization network indiscriminately to all typhoon cloud types, failing to adapt to the differences in cloud morphology. Second, its task function is limited, only outputting the typhoon eye coordinates and unable to simultaneously estimate meteorological intensity parameters such as central pressure and maximum wind speed. Third, the feature expression capability of the VGG-16 as the feature extraction backbone network is limited by the significant domain difference between ImageNet pre-trained weights and infrared satellite cloud images. Fourth, the localization accuracy of the search box method is limited by the number of search steps and the discretization accuracy of the box, making it difficult to achieve sub-pixel level fine localization.

[0005] Furthermore, while methods based on direct coordinate regression using convolutional neural networks (such as Improved-TCLNet) have improved positioning speed, the global average pooling operation loses crucial spatial location information for accurate positioning, and similarly fails to achieve joint estimation of multiple meteorological parameters. Methods based on target detection frameworks struggle to adapt to the diverse morphologies of typhoon eyes (three main forms: clear eyewall, blurred eye region, and eyeless structure), and the discretization accuracy of the detection boxes also limits further improvements in positioning accuracy.

[0006] In summary, existing technologies face the following core bottlenecks in typhoon eye localization: a single network structure is difficult to adapt to the morphological differences of typhoons with different cloud types; there is a structural contradiction between accurate spatial positioning and global meteorological parameter estimation that is difficult to reconcile; there is a difference in the pre-training domain between the feature extraction backbone network and infrared satellite cloud images; and there is a lack of ability to jointly estimate multiple meteorological parameters. Therefore, there is an urgent need to study a novel typhoon eye localization method that can adaptively process different typhoon cloud types, support joint estimation of multiple meteorological parameters, and has sub-pixel-level positioning accuracy. Summary of the Invention

[0007] The purpose of this invention is to provide a satellite cloud image typhoon eye location method and system based on a hybrid expert network, so as to solve the technical problems of poor adaptability to different cloud types of typhoons, insufficient positioning accuracy, and inability to simultaneously estimate multiple meteorological parameters in the existing typhoon eye location methods.

[0008] To achieve the above objectives, this invention provides a satellite cloud image typhoon eye localization method based on a hybrid expert network, comprising the following steps: Step S1: Data acquisition and preprocessing. The infrared satellite cloud image of the typhoon area is cropped, percentile normalized, and channel expanded to generate a standardized input image. Step S2: Multi-scale feature extraction. A pre-trained visual language model visual encoder combined with low-rank adaptive fine-tuning technology is used to extract high-dimensional visual features, which are then processed through multi-scale fusion to obtain multi-scale fused features. Step S3: Cloud-based dynamic routing. A cloud-based sensing router is used to identify the typhoon cloud structure and expert weights are dynamically allocated using a Top-K gating mechanism. Step S4: Typhoon eye coordinate estimation. Spatial preservation processing of the coordinate hybrid expert layer and multi-level upsampling of the heatmap regression head are used, and a differentiable soft extremum extraction algorithm is used to calculate the sub-pixel coordinates of the typhoon eye. Step S5: Joint estimation of meteorological parameters. The typhoon center pressure, maximum wind speed, and wind circle radius are jointly estimated using a meteorological hybrid expert layer and a dual-branch cross-attention structure. Step S6: Multi-task joint training and report output. A four-stage progressive training strategy and a unified loss function including Dvorak physical constraints are used for joint optimization. The positioning results are output, and a typhoon meteorological parameter warning report is generated.

[0009] This invention achieves deep coupling between the above steps through the following technical mechanisms: the multi-scale fusion features of step S2 simultaneously drive the routing decision of step S3, the coordinate estimation of step S4, and the meteorological estimation of step S5, realizing parallel computing of multiple tasks; the cloud-shaped routing weight of step S3, as a shared gating parameter, acts on both the coordinate hybrid expert layer and the meteorological hybrid expert layer, forming a collaborative architecture of one route and two paths; in step S5, the Dvorak physical constraint loss propagates meteorological prior knowledge back to the visual encoder of step S2 in the form of gradients, forming a feature adaptive optimization closed loop guided by physical laws; the peak location information of the heat map in step S4 and the intensity classification result in step S5 influence the training weight of the router in step S3 during the course learning stage, forming a localization-classification collaborative reinforcement mechanism.

[0010] This invention also provides a satellite cloud image typhoon eye positioning system based on a hybrid expert network, including a data preprocessing module, a multi-scale feature extraction module, a cloud shape perception routing module, a coordinate estimation path, a meteorological parameter estimation path, and a training and report output module. The data preprocessing module generates standardized input images; the multi-scale feature extraction module uses a pre-trained visual language model visual encoder and a multi-scale fusion unit to extract multi-scale fusion features; the cloud shape perception routing module identifies typhoon cloud shapes and dynamically assigns expert weights through a Top-K gating network; the coordinate estimation path outputs sub-pixel coordinates of the typhoon eye through a spatially preserving hybrid expert layer and a heatmap regression head; the meteorological parameter estimation path jointly estimates meteorological parameters through a global feature-based hybrid expert layer and a two-branch cross-attention structure; and the training and report output module performs four-stage progressive training and generates a typhoon meteorological parameter early warning report.

[0011] The beneficial effects of this invention are as follows: First, by introducing a visual language model backbone network and employing low-rank adaptive technology for domain transfer, the domain differences between general visual pre-training knowledge and infrared satellite cloud image features are effectively bridged. Only about 2% of the parameters need to be fine-tuned to achieve feature representation capabilities significantly superior to traditional convolutional networks. Second, the hybrid expert routing architecture based on typhoon cloud types enables differentiated modeling of typhoons with different cloud types, overcoming the shortcomings of single networks' insufficient adaptability to typhoons of varying intensities. The positioning accuracy reaches 100% within a 50km range, an improvement of approximately 18 percentage points compared to similar CNN methods. Third, heatmap regression combined with differentiable soft extremum extraction achieves sub-pixel-level typhoon eye positioning, breaking through the discretization accuracy bottleneck of traditional search box methods. Fourth, the end-to-end 12-dimensional multi-parameter joint estimation framework for the first time incorporates typhoon eye coordinates, central pressure, maximum wind speed, and wind circle radius into a unified model. It also automatically outputs typhoon meteorological parameter warning reports through the visual language model's language generation function, providing a complete intelligent typhoon analysis solution for meteorological forecasting systems. Attached Figure Description

[0012] Figure 1 This is a flowchart of the satellite cloud image typhoon eye localization method based on a hybrid expert network, according to the present invention.

[0013] Figure 2 This is an overall architecture diagram of the satellite cloud image typhoon eye positioning system based on a hybrid expert network, according to the present invention. Detailed Implementation

[0014] To make the objectives, technical solutions, and beneficial effects of the present invention clearer, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0015] This embodiment describes in detail a satellite cloud image typhoon eye localization method based on a hybrid expert network. The following steps are interconnected and progressively advance to form an end-to-end intelligent typhoon eye localization and multi-parameter estimation system.

[0016] Step S1: Data Acquisition and Preprocessing. This invention uses infrared (center wavelength approximately 10.8 μm) satellite cloud images as the main input data. Data sources include raw NetCDF files from multiple marine meteorological satellites or preprocessed PNG grayscale images, with a spatial resolution of approximately 0.07° / pixel. In the data preprocessing stage, the initial estimated latitude and longitude of the typhoon eye provided by the typhoon best path database or meteorological forecast system are first converted into corresponding pixel coordinates (i, j). A 112×112 pixel local area is then cropped from the original cloud image centered on these coordinates. The specific cropping operation is as follows: ,in This is the original infrared satellite cloud image. and These are the row and column coordinates of the typhoon eye in the original image.

[0017] Preferably, a random spatial offset is applied to the center coordinates during the trimming stage. (Uniformly distributed, unit: pixels) to increase sample diversity and simulate real-world scenarios where initial estimates may have errors, ensuring the model can still accurately locate the typhoon eye even when it is not perfectly centered. Subsequently, percentile cropping normalization is performed on the cropped region, specifically: calculating the 2nd percentile of the pixel values ​​in the cropped region. and the 98th percentile The normalized image is: ,in This is the normalized grayscale image. and These are the 2nd and 98th percentiles of the pixel values ​​in the original cropped region, respectively. The clip operation truncates the results to the [0,1] interval. The advantage of this normalization method is its ability to adaptively suppress pixel value distribution shifts caused by brightness temperature differences between various cloud images and effectively eliminate interference from extreme brightness temperature values. Finally, the normalized grayscale image is copied to a 3-channel RGB format and uniformly scaled to 224×224×3 using bilinear interpolation, matching the standard input size of the visual language model encoder. Each original sample is amplified into three training samples through different offsets, effectively increasing the training set size to three times the original labeled number.

[0018] This embodiment uses satellite remote sensing data covering six major ocean basins worldwide from 1980 to 2022 for training and evaluation, including 58,096 valid labeled records from the western Pacific, north Atlantic, eastern Pacific, southern Pacific, southern Indian Ocean, and northern Indian Ocean. The labeled content covers the latitude and longitude of the typhoon eye, central pressure (hPa), maximum wind speed (kt), and the four azimuthal components (NE, SE, SW, NW) of the R34 and R64 wind circle radii.

[0019] Step S2: Multi-scale feature extraction. This invention selects the visual encoder of the Qwen2.5-VL-3B visual language model as the backbone network for feature extraction. This encoder is based on the Vision Transformer (ViT) architecture, pre-trained on large-scale image-text pairing data, and has powerful multimodal visual understanding capabilities. It supports dynamic resolution input and outputs features with a dimension of 2048. Considering the significant differences between infrared satellite cloud images and natural images in texture distribution, grayscale range, and semantic features, directly using frozen pre-trained parameters for inference will inevitably lead to unsatisfactory feature extraction results. Therefore, this invention employs Low-Rank Adaptation (LoRA) technology to efficiently fine-tune the parameters of the visual encoder.

[0020] The core idea of ​​LoRA is: for the pre-trained weight matrix Its parameter update volume It can be decomposed into the product of two low-rank matrices: ,in Let be the lower projection matrix, initialized with a Gaussian distribution; The projection matrix is ​​initialized to all zeros. To be the rank of the matrix, this invention sets... ; As a scaling factor, this invention sets Corresponding scaling ratio ; and These represent the row and column dimensions of the original weight matrix, respectively. The number of fine-tuning parameters for LoRA is... It is much smaller than the number of parameters in full-parameter fine-tuning. This invention achieves efficient adaptation to the meteorological remote sensing field while retaining pre-trained knowledge. LoRA is applied to the query, key-value projection matrices (q, k, v projection) of the attention layer and the two-layer linear transformation (fc1, fc2) of the feedforward network, with a Dropout rate of 0.1. The complete feature extraction process of the visual encoder can be represented as follows: ,in For the input image batch, For batch size, These are the frozen pre-training parameters. For trainable low-rank update parameters, and This refers to the spatial dimension output by the encoder.

[0021] The 2048-dimensional feature map output by the visual encoder is then subjected to dimensionality reduction processing by a feature projection layer. Specifically, the number of feature channels is compressed from 2048 to 512 through 1×1 convolution, batch normalization, and ReLU activation to obtain the projected features. This operation significantly reduces the computational cost of subsequent networks while stabilizing the feature distribution through batch normalization.

[0022] Furthermore, this invention designs a multi-scale feature fusion module to capture the structural information of typhoon cloud systems at different spatial scales. Specifically, from projection features... Starting from this point, three feature maps of different resolutions are generated simultaneously: original scale features. 2x downsampling features and 2x upsampling features (Bilinear interpolation). The feature maps at the three scales are interpolated to the target size and then concatenated along the channel dimension. Finally, a multi-scale fused feature is obtained through a fusion convolution. , , in This is the result of three-scale feature stitching, with a channel dimension of [missing information]. ; For multi-scale fusion features, the number of channels is compressed back to 512 using a 1×1 convolution; BN represents batch normalization. It also contains macroscopic structural information (from downsampled features) and fine texture information (from upsampled features) of the typhoon cloud system, laying a multi-scale representation foundation for subsequent cloud-based sensing routing and precise coordinate positioning.

[0023] Step S3: Cloud-Pattern-Aware Dynamic Routing. One of the core innovations of this invention is the Cloud-Pattern-Aware Router. This router can automatically identify typhoon cloud patterns based on the morphological characteristics of typhoon cloud systems in the input infrared cloud image and dynamically determine which expert networks to assign the features to for processing. The router's design is based on the fact that typhoons of different intensities exhibit drastically different cloud patterns on infrared satellite cloud images—weak typhoons (early tropical depressions / tropical storms) often appear as chaotic, loose, and irregular cloud clusters; moderate typhoons exhibit a distinct spiral cloud band structure; and strong typhoons have clear convective eyewalls and dense spiral cloud bands. These three cloud patterns differ fundamentally in spatial structure, texture features, and grayscale gradient distribution.

[0024] The router's input is multi-scale fused features. First, global feature vectors are extracted using global average pooling (GAP): , in and The spatial height and width of the feature map. The global feature vector is 512-dimensional, containing the overall structural information of the cloud map. This global feature vector is then fed into a gated network, which employs a two-layer perceptron structure with residual connections: the first layer linear transformation maps the 512-dimensional features to 512 dimensions and performs layer normalization and ReLU activation; the second layer linear transformation maps to 256 dimensions and performs layer normalization and ReLU activation; simultaneously, a skip connection matrix is ​​used... The input is directly mapped to 256 dimensions and added to the output of the second layer; finally, it is output through a linear layer. Dimensional gated logic value .

[0025] This invention employs a Top-K gating mechanism, activating only the highest-weighted gate. One expert: , in For the first The gating weight of each expert, The set of expert indexes selected for Top-K For temperature parameters (smaller temperatures make the weight distribution more concentrated, enhancing the certainty of expert selection). For the first The gating logic value corresponding to each expert The number of experts activated each time. The router also performs an auxiliary classification task, classifying typhoon cloud types into three categories: weak typhoon eye type (corresponding to tropical depression to tropical storm level, with maximum wind speed below 34kt), moderate spiral band type (corresponding to tropical storm to level 2 typhoon, with maximum wind speed of 34~95kt), and strong typhoon eyewall type (corresponding to level 3 and above strong typhoon, with maximum wind speed not less than 96kt). The auxiliary classification loss participates in model training in the form of cross-entropy, effectively improving the router's ability to identify cloud types.

[0026] Preferably, to prevent the router from degenerating into an expert collapse phenomenon where only a few experts are activated, this invention introduces a comprehensive load balancing loss mechanism. Loss maximized by entropy (Encourage uniform gating probability distribution) Use equilibrium loss (Encouraging even use of resources among experts) and loss of diversity (Different expert combinations are encouraged for different samples) Three components: , in The overall weighting coefficient for load balancing losses; , This represents the mean of the gating probability within the batch. It is a numerically stable term; , For the first Average usage rate of each expert; , For the first The entropy value of the gating weights of each sample. For batch size. It is worth noting that the cloud-based sensing router fully shares the coordinate estimation path and the meteorological estimation path, and the same set of router parameters simultaneously guides the expert allocation of the two paths, realizing a highly efficient collaborative architecture of one route and two paths.

[0027] Step S4: Typhoon eye coordinate estimation. The coordinate estimation path adopts a spatially preserving hybrid expert architecture. Its core design is that the accurate localization of the typhoon eye requires the network to retain pixel-level spatial location information during the feature extraction process. Therefore, the coordinate expert adopts a convolutional structure that preserves the spatial resolution of the feature map, rather than using global average pooling to compress the spatial information into a vector.

[0028] The coordinate hybrid expert layer comprises three cloud-shaped expert networks and one general expert network. Each cloud-shaped expert network consists of two 3×3 convolutional layers (each containing batch normalization and ReLU activation) cascaded coordinate attention modules, outputting a 128-channel spatial feature map. , in For the first The output of the coordinate cloud-shaped expert, CoordAttn, represents the coordinate attention operation. The coordinate attention module works by: processing the input features... Global average pooling is performed along both the horizontal and vertical spatial directions to obtain direction-aware pooling vectors: , in For the first Each channel is located horizontally at position The aggregate value at that location, For the position along the vertical direction The aggregate value at that location, and These represent the spatial height and width of the feature map, respectively. The number of channels is used. The pooling results from both directions are concatenated and then compressed using a shared 1×1 convolution (compression ratio...). Then, horizontal attention is generated by independent 1×1 convolutions and sigmoid activations. and vertical attention The input features are applied through element-wise multiplication: ,in The output features are weighted by coordinate attention. This represents element-wise multiplication (broadcast mechanism). The innovative value of the coordinate attention mechanism lies in the fact that, unlike channel attention (such as the SE module) which compresses all spatial information, coordinate attention retains the positional awareness capabilities in both the horizontal and vertical directions, enabling the network to perceive the positional offset of the typhoon eye within the image, resulting in significant performance gains for precise spatial positioning tasks.

[0029] General experts employ depthwise separable convolution to achieve a lightweight design. This involves first performing 3×3 grouped convolutions (groups=512) with the number of groups equal to the number of channels, followed by 1×1 pointwise convolutions. The number of parameters is approximately one-third that of standard convolutions. This ensures that diverse features are introduced into expert fusion without significantly increasing the computational burden. Given the gating weights output by the router... The weighted mixed output of the strength expert is The fusion rules with general expert output are as follows: ,in For learnable fusion parameters, The function is a sigmoid function. The parameters are initialized so that the initial contribution ratio of the general expert is about 30%, and then gradually adjusted to the optimal fusion ratio as training progresses.

[0030] The heatmap regression head accepts the output features of the coordinate mixing expert layer. The spatial resolution is restored and a high-resolution typhoon eye probability heatmap is generated through a three-level upsampling network. Each upsampling unit contains a 3×3 convolution, batch normalization, ReLU activation, and coordinate attention module. The number of channels varies sequentially from 128 to 256 to 128 to 64. The first two levels are followed by a 2× bilinear upsampling operation. Finally, a 28×28 typhoon eye probability heatmap is output after a 1×1 convolution and bilinear interpolation. Target heatmap Construct a two-dimensional Gaussian distribution centered on the actual typhoon eye coordinates: ,in , For the height of the heat map, These are the pixel positions on the heatmap corresponding to the actual coordinates.

[0031] The continuous coordinates of the typhoon eye are extracted from the heatmap using a differentiable soft extremum extraction algorithm (Soft-ArgMax). The specific process is as follows: [The heatmap is then processed...] After flattening, the temperature parameters Scaling Softmax generation probability distribution Then construct a normalized coordinate grid. , Finally, the expected value of the coordinates is calculated: , in The normalized coordinates of the typhoon eye within the cropped image. The spatial dimensions of the heatmap. For position The probability value at that location. and These are normalized column coordinate grids and row coordinate grids, respectively. Temperature parameter. Its function is to concentrate the probability distribution near the peak of the heatmap, improving positioning accuracy while suppressing multi-peak interference. The Soft-ArgMax operation is fully differentiable, supporting end-to-end gradient propagation from coordinate estimation results to the visual encoder. Preferably, to prevent conflicting positioning results from different experts, a peak consistency loss is introduced. This requires minimizing the distance between the estimated peak position of each coordinate expert and the weighted average peak, thereby further improving the robustness of multi-expert collaborative positioning.

[0032] Step S5: Joint estimation of meteorological parameters. The meteorological parameter estimation path of this invention realizes end-to-end joint estimation of typhoon center pressure, maximum wind speed, and R34 / R64 wind circle radius. The system outputs a complete meteorological parameter vector. 12-dimensional: , in Normalized typhoon eye coordinates; To normalize the central pressure, the normalization method is as follows: This corresponds to the original air pressure range of [850, 1020] hPa; To normalize the maximum wind speed, the normalization method is as follows: This corresponds to the original wind speed range [10, 180] kt; To normalize the wind circle radius, the normalization method is as follows: The corresponding original radius range is [0, 200] nm (nautical miles). Indicates wind speed level, Indicates direction or location.

[0033] The meteorological hybrid expert layer adopts a global feature-based expert design in its architecture: each cloud expert uses two 3×3 convolutional layers followed by global average pooling to compress the spatial feature map into a 128-dimensional global vector, which is then output after linear and layer normalization. The introduction of global average pooling enables meteorological experts to aggregate the intensity distribution information of the entire typhoon cloud image, which is consistent with the physical nature of meteorological parameters (such as central pressure and maximum wind speed) being determined by the overall structure of the cloud system. This is also the essential difference between meteorological experts and coordinate experts in terms of architecture.

[0034] The meteorological estimation head employs a dual-branch structure to model the physical coupling relationship between air pressure and wind speed. Each branch has an independent multilayer perceptron (MLP): a shared feature extraction layer maps the 128-dimensional meteorological hybrid expert output to 256 dimensions, which are then fed into the air pressure branch. and wind speed branch Each branch outputs a 128-dimensional feature vector. The two branches exchange information through a 4-head multi-head cross-attention mechanism (128-dimensional embedding, 0.1 Dropout rate). , , in It is a characteristic of pressure branching. As a wind speed branch feature, the two are key-value pairs of each other. The mutual constraint relationship between air pressure and wind speed is explicitly modeled through a cross-attention mechanism. Finally, the normalized estimated value is output after each linear layer and Sigmoid activation.

[0035] This invention is based on Dvorak's empirical relationship. (in hPa is the ambient air pressure), and this meteorological prior knowledge is encoded as a physical constraint loss: , , in For the first The deviation between the estimated wind speed of a single sample and the wind speed calculated using the Dvorak formula. kt is the segmentation threshold of the Huber loss (a quadratic penalty is applied to smaller deviations, and a linear penalty is applied to larger deviations). and These are the estimated maximum wind speed (kt) and the estimated central pressure (hPa) after inverse normalization. For batch size, coefficient The dimensional normalization factor is used. The gradient of this physical constraint loss is applied to the LoRA parameters of the meteorological hybrid expert layer and the visual encoder through backpropagation, embedding meteorological physical laws into the feature learning process in the form of soft constraints, forming a closed loop of feature adaptive optimization guided by physical priors.

[0036] The dual-radius estimation head receives the mixed output from the meteorological hybrid expert layer. After passing through a shared feature extraction layer (two-layer linear transformation 128→256→256, including layer normalization, ReLU, and Dropout), the R34 and R64 branches each process the radius estimation in four directions. Within each branch, multi-head cross-attention (4 heads, embedding dimension 64) is introduced to capture the spatial symmetry of wind circle radii in different directions. Finally, the output is normalized by Sigmoid activation. Hierarchical constraint loss. A penalty is imposed for cases where the R64 radius component exceeds the R34 component: , in and After normalization The estimated radii of the wind circles in directions R64 and R34 are used. The max operation only applies a positive penalty to samples that violate the physical constraint R34 ≥ R64. Radius symmetry constraint. Encourage wind circles with similar radii in diagonal directions, and impose absolute difference penalties on the northeast-southwest and southeast-northwest pairs to enhance the physical rationality of the estimation results.

[0037] Step S6: Multi-task joint training and report output. The unified loss function of this invention is composed of a weighted combination of nine sub-losses: , in For coordinate regression loss, the following is adopted: Smooth L1 loss; This represents the mean square error loss of the heatmap. Smooth L1 loss for meteorological parameters (air pressure and wind speed); For Dvorak physical constraint loss; and These are the load balancing losses for the coordinate path and the meteorological path, respectively. This represents the peak consistency loss. and Smooth L1 losses for wind circle radii of R34 and R64, respectively, are calculated only on samples with valid labels; Hierarchical constraint loss; weights of each loss Adjust dynamically according to the training phase.

[0038] The four-stage progressive training strategy is implemented as follows: The first stage (coordinate warm-up stage, approximately 2 epochs) involves activating only the LoRA parameters of the coordinate estimation path and the visual encoder, and then applying the loss weights. , The remaining weights are reset to zero, allowing the model to first establish a stable spatial positioning baseline and avoid gradient conflicts interfering with coordinate learning in the initial stage of multi-task learning; in the second stage (weather warm-up stage, about 13 epochs), the coordinate estimation path is frozen, the weather estimation path and the dual-radius estimation head are activated, and the weights are lost. , , The first stage establishes a baseline for meteorological parameter estimation. The second stage (joint training stage, approximately 25 epochs) involves unfreezing all parameters and dynamically adjusting the loss weights of each task using the GradNorm multi-task gradient balancing algorithm. GradNorm adaptively adjusts the weights based on the relative rate of change of the gradient norm of each task, ensuring that the learning progress of different tasks is coordinated. The third stage (course learning stage, approximately 25 epochs) involves dynamically adjusting the training weights of different cloud types of typhoon samples through a four-level course scheduler. The scheduling order is as follows: warm-up period (all weights are 1.0) → strong typhoon focusing period (weak / moderate / strong sample weights 0.5 / 0.8 / 1.0) → weak typhoon focusing period (weights 1.0 / 0.8 / 0.8) → equilibrium period (all weights return to 1.0), effectively improving the model's adaptability to both weak and strong typhoon extreme cloud types.

[0039] In the early stages of training (epochs 0 to 2), an expert warm-up scheduler is introduced, which activates each cloud type expert sequentially in each round, forcing each expert to learn the corresponding cloud type features independently first, thus avoiding functional homogenization caused by random initialization of multiple experts in the early stages of training. , in For the current training round, For the number of experts, For expert indexing, after the third epoch, all expert weights are restored to 1.0 to participate in joint optimization normally. For the optimizer configuration, the AdamW optimizer is used, with a learning rate of... Weight decay The learning rate is scheduled using a cosine annealing strategy, the mixed precision training format is bfloat16, the gradient pruning threshold is 1.0, and the batch size is 4.

[0040] During the inference phase, an infrared satellite cloud image is input. After forward propagation, the system outputs a 12-dimensional estimation vector and inversely transforms the normalized coordinates into actual latitude and longitude coordinates, and inversely transforms the normalized meteorological parameters into physical quantities (pressure unit hPa, wind speed unit kt, and wind circle radius unit nm). Furthermore, this invention utilizes a language generation module of a visual language model. Based on the input cloud image and the estimated meteorological parameters, and combined with a built-in meteorological knowledge base, it automatically generates a natural language typhoon meteorological parameter early warning report containing typhoon intensity level assessment, affected area description, and disaster prevention and early warning suggestions. This achieves a complete end-to-end intelligent processing flow from satellite cloud image to early warning report.

[0041] This embodiment was trained on an NVIDIA A100 80GB GPU using PyTorch 2.1+, with a total of approximately 65 epochs (four stages combined). The average processing time per frame during the model inference stage is approximately 0.3 seconds, which meets the near real-time analysis requirements of meteorological operations for timed satellite cloud imagery. The system supports batch inference mode; when the batch size is set to 16, GPU utilization can reach approximately 85%, processing approximately 42 frames of satellite cloud imagery per second. This satisfies the dual needs of meteorological operations for batch back-calculation of historical typhoon cloud imagery and rapid analysis of real-time observation cloud imagery, making it suitable for integration into the operational workflows of the National Meteorological Information Center or provincial meteorological forecasting centers.

[0042] To verify the necessity of introducing LoRA fine-tuning compared to directly freezing the visual encoder, an ablation experiment was conducted in this embodiment. Experimental results show that when the visual encoder is frozen (without LoRA), the model's localization accuracy (Acc@50km) within a 50km range on the western Pacific test set is only 84.2%, while after introducing LoRA fine-tuning, this metric improves to 100%, an improvement of approximately 15.8 percentage points. This fully verifies the effectiveness of LoRA in bridging the domain differences between general visual pre-training knowledge and infrared satellite cloud image features. This improvement stems from two aspects: first, LoRA enables the visual encoder to learn and recognize the brightness temperature gradient patterns and cloud top temperature distribution structures unique to infrared cloud images; second, LoRA only fine-tunes about 2% of the parameters, effectively preventing catastrophic forgetting that may occur when performing full-parameter fine-tuning on relatively limited meteorological datasets, preserving the general visual understanding capabilities accumulated by the pre-trained visual language model.

[0043] The design of the multi-scale fusion module was also carefully weighed. Downsampled features (2× downsampling) expand the receptive field, enabling the network to perceive the macroscopic spiral structure of typhoon cloud systems; upsampled features (2× upsampling) retain finer local texture information, which helps to infer the typhoon center location through detailed changes in cloud top temperature under the condition of a weak typhoon without an eye; and original-scale features provide comprehensive information from the intermediate-scale receptive field. The complementary fusion of the three scale features enables the system to maintain robust feature extraction capabilities throughout the entire typhoon lifecycle (from tropical disturbances to super typhoons), with an average distance error (MD) of 1.96 pixels, which translates to an actual distance of approximately 15.2 km.

[0044] The core difference between cloud-based sensing routers and traditional intensity-based sensing routers lies in the fact that the router's judgment is based on the morphological characteristics of typhoon cloud systems in the cloud image (i.e., cloud type), rather than the actual meteorological intensity level of the typhoon. This design consideration stems from the following meteorological observation fact: typhoons of the same intensity level may present drastically different cloud types on satellite cloud images. For example, even among Category 3 typhoons, influenced by sea surface temperature, vertical wind shear, and upper-level divergence conditions, some typhoons exhibit a tightly organized concentric double-eyewall structure (corresponding to strong typhoon eyewall type experts), while others may exhibit an asymmetrical spiral band structure due to wind shear (corresponding to moderate spiral band type experts). Conversely, typhoons of adjacent intensity levels (such as tropical depressions and weak tropical storms) often have very similar cloud types on cloud images and should be processed by the same expert. Therefore, using cloud type as the routing basis is more in line with the essential needs of image processing than using meteorological intensity classification, and it also more directly addresses the core technical problem of the difficulty in accurately locating different cloud types using a single network.

[0045] Preferably, the router's auxiliary classification task employs a three-class cross-entropy loss method. During training, the typhoon intensity label from the best meteorological path data serves as an auxiliary supervision signal. However, the router learns to make judgments based on image features rather than the intensity label itself, forming an adaptive cloud shape recognition mechanism driven by image features and supervised by intensity labels. During the testing phase, the router's cloud shape classification results can be output separately, providing meteorological analysts with objective criteria for judging typhoon cloud shapes, thus possessing additional meteorological operational value.

[0046] Compared to the deep Q-learning-based search box approximation method used in CN111462230B, the heatmap regression method of this invention has the following essential advantages: First, the search box method requires multiple iterations (typically 10-16 steps) to locate the typhoon center, with each step requiring network inference, resulting in a large overall computational cost; while the heatmap method of this invention completes the localization in a single forward propagation, significantly improving inference efficiency. Second, the final localization accuracy of the search box method is limited by the pixel error between the search box center and the actual typhoon center in the last step, resulting in discrete localization; while the Soft-ArgMax algorithm uses probability weighting of the heatmap to calculate the expectation, outputting continuous normalized coordinate values, achieving sub-pixel level localization accuracy, with a theoretical localization error of less than 1 pixel. Third, the search box method relies on VGG-16 for feature extraction, and its ImageNet pre-trained weights have significant domain differences with infrared satellite cloud images; this invention uses a visual language model backbone combined with LoRA fine-tuning, resulting in stronger feature extraction capabilities and better domain adaptation. Fourth, the search box method can only output the coordinates of the typhoon eye, while this invention simultaneously estimates the central pressure, maximum wind speed and wind circle radius, realizing the integrated estimation of multiple meteorological parameters and providing more complete information support for meteorological early warning.

[0047] The R34 wind circle radius (outer diameter of the area with wind speed ≥34 kt) and the R64 wind circle radius (outer diameter of the area with wind speed ≥64 kt) are two core parameters describing the structure of a typhoon wind field. They have direct decision-making reference value for port vessel sheltering, coastal engineering protection, and the delineation of emergency evacuation areas. Currently, almost all mainstream typhoon location methods do not involve the automatic estimation of wind circle radii; this invention fills a technological gap by incorporating it into a joint estimation framework.

[0048] Hierarchical constraints The introduction of this concept constrains the estimation results from a priori physical perspective: the radius of the R64 wind circle describes the area covered by stronger winds (≥64 kt, corresponding to super tropical storms and above), and its physical extent must be a subset of the area covered by the R34 wind circle (≥34 kt), i.e., in any direction... This holds true consistently. If the model ignores this constraint during the initial training phase, it may estimate... The hierarchical constraint loss effectively prevents physically unreasonable results, ensuring that the estimation results remain physically consistent. Experiments show that after introducing hierarchical constraints, the physical violation rate of the R34 / R64 estimation ( The sample percentage decreased from about 23% to below 2%, while the mean absolute errors of R34 and R64 decreased by about 4% and 6%, respectively.

[0049] Gradient conflict is a major problem affecting training stability in multi-task learning, especially between coordinate localization tasks (requiring fine spatial local features) and meteorological parameter estimation tasks (requiring abstract global semantic features). The first stage, coordinate warm-up, enables the encoder to establish a fine-grained perception of the typhoon eye's spatial location, preventing the global feature learning of the meteorological task from impacting the coordinate task's location perception. The third stage introduces the GradNorm multi-task gradient balancing algorithm, dynamically adjusting loss weights based on the relative training rate of each task: tasks with faster training (larger gradient norms) are automatically weighted less, while slower-training tasks are automatically weighted more, ensuring convergence at similar rates. In experiments, the gradient norms of each task tended to balance after approximately the 10th epoch in the third stage, resulting in a generally stable training curve. The fourth stage, curriculum learning, addresses the problem of uneven sample difficulty: weak typhoons (tropical depressions and early tropical storms) have the largest localization errors in the dataset due to irregular cloud formations. By appropriately increasing the weight of weak typhoon samples during their focusing period, the model's weak typhoon localization accuracy improved by approximately 8 percentage points compared to the no-curriculum-learning strategy, while the performance of strong typhoons did not significantly degrade.

[0050] This embodiment describes a satellite cloud image typhoon eye positioning system based on a hybrid expert network. This system corresponds one-to-one with each step in the above method embodiment. The modules form a deeply coupled data flow path with multi-scale fusion features as the hub, and together complete the complete processing link from infrared satellite cloud image input to typhoon eye coordinates, meteorological parameter output and early warning report generation.

[0051] The data preprocessing module receives raw infrared satellite cloud imagery data (NetCDF or PNG format) of the typhoon region and performs the region cropping, percentile normalization, and channel expansion processing described in step S1 of the method embodiment to generate a standardized input image in 224×224×3 format. This module also includes a data augmentation subunit. During the training phase, a uniform random offset of ±25 pixels is introduced to the cropping center coordinates, and Gaussian noise (standard deviation σc=0.10) related to the intensity category is added to the air pressure and wind speed labels to increase the diversity of the training samples. During the inference phase, data augmentation is disabled, and only the deterministic standardization preprocessing procedure is performed. The output of the data preprocessing module is directly passed to the multi-scale feature extraction module.

[0052] The multi-scale feature extraction module comprises two sub-units: a visual encoder unit and a multi-scale fusion unit. The visual encoder unit employs a Qwen2.5-VL-3B pre-trained visual language model. It utilizes low-rank adaptive techniques to efficiently fine-tune the parameters of the query, key-value projection matrices of the attention layer, and the feedforward network layer. With a rank r=48 and a scaling factor α=96, it achieves effective adaptation to the meteorological remote sensing domain at the cost of only about 2% parameter fine-tuning, outputting a 2048-dimensional visual feature map. Multi-scale fusion unit reception After dimensionality reduction to 512 dimensions via 1×1 convolution, feature maps at three resolutions—original scale, 2x downsampled, and 2x upsampled—are generated simultaneously. These maps are then unified to the target size, concatenated, and fused through convolution to obtain a 512-dimensional multi-scale fused feature map. This fully integrates the structural information of typhoon cloud systems at different spatial scales. Multi-scale fusion characteristics. As the common input for the cloud-based sensing routing module, coordinate estimation path, and meteorological parameter estimation path, it is the core feature hub of the entire system.

[0053] The cloud-based perceptual routing module includes a global feature aggregation unit and a Top-K gated network unit. The global feature aggregation unit fuses features at multiple scales. Perform global average pooling to aggregate spatial dimensional information into a 512-dimensional global feature vector. This contains information about the overall morphology of the typhoon cloud system. The Top-K gated network unit processes the global feature vector through a two-layer perceptron structure with residual connections, outputting a K=3 dimensional gated logic value. A Top-K (k=2) gated mechanism with a temperature parameter τ=0.5 is used to calculate the expert activation weights. This module enables dynamic expert routing based on typhoon cloud morphology. It simultaneously performs auxiliary classification tasks for three types of typhoon cloud types (weak typhoon eye type, moderate spiral band type, and strong typhoon eyewall type), and the classification results can be output separately for meteorological operations. The routing module generates gating weights. Simultaneously, the data is transmitted to the hybrid expert layer of the coordinate estimation path and the meteorological parameter estimation path, realizing a unified one-way, dual-path collaborative distribution mechanism, ensuring consistent understanding of typhoon cloud characteristics between the two paths under the condition of parameter sharing.

[0054] The coordinate estimation path consists of two components: a coordinate fusion expert layer and a heatmap regression head. The coordinate fusion expert layer comprises three cloud-type expert networks and one general expert network. All cloud-type experts employ a spatially preserving structure (two layers of 3×3 convolutions + a coordinate attention module) to ensure that the pixel-level spatial location information required for typhoon eye localization is not lost during feature map transmission, outputting a 128-channel spatial feature map. The general expert uses depthwise separable convolutions for a lightweight design. The outputs of each expert are weighted and fused according to routing weights, with the general expert participating in the final output at a fusion ratio controlled by learnable parameters (initially approximately 30%). The heatmap regression head consists of a three-level upsampling network, each level containing 3×3 convolutions, batch normalization, ReLU activation, and a coordinate attention module. The number of channels gradually changes from 128 to 64, outputting a 28×28 resolution typhoon eye probability heatmap. The heatmap is processed using a differentiable soft extremum extraction algorithm (Soft-ArgMax, temperature parameter τh=8), outputting sub-pixel-level normalized coordinates of the typhoon eye. , )∈[0,1]², supporting end-to-end gradient propagation.

[0055] The meteorological parameter estimation path comprises three components: a meteorological hybrid expert layer, a two-branch cross-attention estimation unit, and a dual-radius estimation head. The meteorological hybrid expert layer includes three cloud type experts and one general expert. All meteorological experts employ a global feature-based structure (convolution + global average pooling), outputting a 128-dimensional global feature vector, suitable for capturing the overall cloud system information that determines typhoon meteorological parameters. This creates a clear functional division with the coordinate path expert, which maintains spatial resolution. The two-branch cross-attention estimation unit sets up pressure branch MLPs and wind speed branch MLPs. It explicitly models the Dvorak physical coupling relationship between pressure and wind speed through a four-head multi-head cross-attention mechanism, outputting a normalized pressure estimate. And wind speed estimates The dual-radius estimation head sets up independent R34 and R64 branches. Within each branch, the spatial symmetry of the four-directional wind circle radius is modeled through multi-head cross-attention between directions, and the 8-dimensional normalized wind circle radius value is output. Hierarchical constraints ensure that the R34 estimation value in any direction is not less than the R64 estimation value in the corresponding direction, thus guaranteeing the physical rationality of the estimation results.

[0056] At the system implementation level, the coordinate estimation path supports two inference acceleration modes: the first is the standard precision mode (float32), suitable for scientific research and analysis scenarios with high positioning accuracy requirements; the second is the mixed precision mode (bfloat16), which compresses the single-frame inference time from approximately 0.45s to approximately 0.18s under the condition that the positioning accuracy decreases by no more than 0.1 pixels, suitable for real-time meteorological business processing scenarios. It is worth noting that the coordinate attention module performs feature aggregation and attention generation independently in both the horizontal and vertical directions, with a computational complexity of O(n log n). Compared to the standard self-attention mechanism It significantly reduces computational overhead while maintaining direction awareness accuracy, making it suitable for deployment on embedded weather stations or edge computing nodes.

[0057] The training and report output module implements full lifecycle management and final result generation for the system. During the training phase, this module executes the four-stage progressive training strategy described in step S6 of the implementation method (coordinate preheating → meteorological preheating → gradient balancing joint training → course learning), manages the activation rhythm of the expert preheating scheduler, applies the GradNorm multi-task gradient balancing algorithm, and maintains a unified loss function containing nine sub-losses, supporting phased dynamic adjustment of loss weights. During the inference phase, this module receives the normalized estimates from each path output and performs inverse normalization calculations: mapping coordinate values ​​back to actual latitude and longitude, and restoring air pressure, wind speed, and wind circle radius to physical quantities; simultaneously, it calls the language generation function of the visual language model, using the estimation results and the built-in meteorological knowledge base as conditions, to automatically generate a typhoon meteorological parameter warning report. The report content covers typhoon intensity level assessment (based on the China Meteorological Administration's typhoon intensity classification standards), descriptions of affected sea and land areas, and suggested preventative measures, providing a complete end-to-end intelligent solution for meteorological forecasting operations.

[0058] This invention was validated on a global typhoon satellite remote sensing dataset covering six ocean basins and spanning from 1980 to 2022 (a total of 58,096 valid annotations, with an effectiveness rate of 96.92%). The data sources cover six ocean basins: the Western Pacific (WP, 21,646), the North Atlantic (NA, 10,721), the Eastern Pacific (EP, 10,550), the Southern Indian Ocean (SI, 10,095), the Southern Pacific (SP, 4,330), and the Northern Indian Ocean (NI, 754). The training / validation / test sets were divided in an 8 / 1 / 1 ratio.

[0059] Comparative experiments with three baseline methods (ResNet50 direct regression, FPN+CBAM convolutional network, and Dual-Head dual-task CNN) show that the proposed system (VLM-MoE) achieves 100% localization accuracy (Acc@50km) within a 50km range, an improvement of approximately 17.84 percentage points compared to the best baseline Dual-Head CNN (82.16%). The pixel error (PE) is 1.96 pixels, superior to the best baseline CNN method (3.93 pixels). The significant improvement in coordinate estimation accuracy within 50km is mainly attributed to the stronger representation ability of the visual language model backbone network for weak typhoon eyeless scenes, and the differentiated processing strategy for different cloud types of typhoons using the hybrid expert architecture. In terms of multi-parameter joint estimation, the proposed system has an average absolute error of 8.73 hPa for center pressure and 11.42 kt for maximum wind speed, comparable to the best-performing CNN baseline model (Dual-Head CNN: pressure error 8.95 hPa, wind speed error 11.29 kt). This invention is the first to simultaneously achieve 12-dimensional joint estimation of typhoon eye coordinates, central pressure, maximum wind speed, and the radius of the four-directional R34 / R64 wind circles in a single model, filling the technical gap in the joint estimation of typhoon eye location and meteorological parameters, and has important value for meteorological operational applications.

[0060] This invention achieves the most significant improvement in localization accuracy compared to the deep reinforcement learning method in CN111462230B on a subset of weak typhoons (tropical depressions to early tropical storms, with maximum wind speeds below 34 kt). This improvement is primarily due to the synergistic effect of the following three aspects: the multi-scale visual understanding capabilities accumulated through pre-training of the visual language model enable it to identify subtle rotational convection features from chaotic and disordered weak typhoon cloud systems; the cloud-type perception router prioritizes the allocation of weak typhoon samples to weak typhoon eye-type experts for targeted processing; and heatmap regression, compared to search box approximation, utilizes global probability distribution rather than local extrema for localization when cloud systems lack obvious peak structures, avoiding localization failures caused by search boxes getting stuck in local oscillations in eyeless typhoon areas. The synergistic effect of these three mechanisms gives this invention a significantly superior adaptability to weak typhoon cloud types compared to existing technologies, fully validating the technical superiority of the cloud-type-based hybrid expert architecture under diverse typhoon morphology conditions.

[0061] After the course learning phase, the system of this invention improved its positioning accuracy for weak typhoons (TD level, maximum wind speed below 34 kt) within a 50km radius from approximately 73.2% at the end of the joint training phase to approximately 81.6%, and for strong typhoons (C3 and above, maximum wind speed not less than 96 kt) from approximately 93.8% to approximately 95.4%. While maintaining a relatively stable overall accuracy, it simultaneously improved adaptability to two extreme cloud types, demonstrating the effectiveness of the course learning strategy on typhoon datasets with uneven difficulty. Furthermore, the system's early warning report generation function, evaluated by meteorological experts, achieved an accuracy rate of over 87% in intensity level assessment, over 91% in the completeness of the affected area description, and over 83% in the relevance of disaster prevention recommendations. This provides high-quality auxiliary reference information for meteorological duty officers, reduces the burden of manual analysis, and improves the efficiency and accuracy of typhoon emergency response.

[0062] The embodiments of the present invention are not limited to the specific embodiments described above. Those skilled in the art can make various equivalent changes or substitutions based on the technical solutions of the present invention, and all such changes or substitutions should be included within the protection scope of the present invention.

Claims

1. A method for typhoon eye location based on satellite cloud imagery using a hybrid expert network, characterized in that, Includes the following steps: Step S1, Data Acquisition and Preprocessing: Acquire infrared satellite cloud images of the typhoon area, and perform cropping, percentile normalization and channel expansion processing on the infrared satellite cloud images to generate standardized input images; Step S2, Multi-scale feature extraction: The standardized input image is input into the visual encoder of the pre-trained visual language model. The parameters of the visual encoder are efficiently fine-tuned through low-rank adaptive technology to extract high-dimensional visual feature maps. The high-dimensional visual feature maps are then subjected to dimensional projection and multi-scale fusion processing to obtain multi-scale fused features. Step S3, Cloud-based dynamic routing: The cloud-based sensing router is used to perform global feature aggregation on the multi-scale fusion features. The cloud structure features of the typhoon in the cloud map are identified through the gating network. The activation weights of each expert are calculated using the Top-K gating mechanism. The multi-scale fusion features are dynamically allocated to the coordinate hybrid expert layer and the meteorological hybrid expert layer. Step S4, Typhoon Eye Coordinate Estimation: In the coordinate estimation path, the coordinate hybrid expert layer uses a spatially preserving expert network to process the multi-scale fusion features, preserving the precise spatial location information required for typhoon eye positioning; the heatmap regression head restores spatial resolution through multi-level upsampling to generate a typhoon eye probability heatmap; the heatmap is processed using a differentiable soft extremum extraction algorithm to calculate the sub-pixel level normalized coordinates of the typhoon eye; Step S5, Joint estimation of meteorological parameters: In the meteorological estimation path, the meteorological hybrid expert layer uses a global feature-type expert network to extract the global characterization of the typhoon, and estimates the central pressure and maximum wind speed of the typhoon through a dual-branch cross-attention structure. The dual-radius estimation head simultaneously estimates the radius of the wind circle at different wind speed levels and applies hierarchical constraints.

2. The method according to claim 1, characterized in that, In step S1, the specific method for preprocessing the infrared satellite cloud image is as follows: A local area of ​​112×112 pixels is cropped from the original infrared satellite cloud image, centered on the initial estimated location of the typhoon eye. Percentile cropping normalization is then performed on the cropped area, specifically by calculating the 2nd percentile P2 and the 98th percentile P... 98 The pixel values ​​are linearly mapped to the [0,1] interval and extreme values ​​are truncated. Then, the normalized grayscale image is expanded into a 3-channel RGB format and scaled to 224×224 pixels by bilinear interpolation as a standardized input image. During cropping, a random offset of ±25 pixels is introduced at the center position to increase data diversity.

3. The method according to claim 1, characterized in that, In step S2, the low-rank adaptive technique is implemented as follows: the pre-trained weight matrix W0 in the visual encoder is decomposed into low rank, and its parameter update is represented as the product of two low-rank matrices, where the rank r=48, the scaling factor α=96, and the dropout rate is 0.

1. The target fine-tuning module includes the query, key, and value projection matrices of the attention layer and the feedforward network layer. The multi-scale fusion processing is performed as follows: the original scale preservation, 2x downsampling, and 2x upsampling operations are performed simultaneously on the 512-dimensional projected feature map. The features of the three scales are unified in size and then concatenated to obtain 1536-channel features. Then, they are fused into 512-dimensional multi-scale fusion features through 1×1 convolution.

4. The method according to claim 1, characterized in that, In step S3, the cloud-type sensing router is specifically implemented as follows: global average pooling is performed on the multi-scale fusion features to obtain a 512-dimensional global feature vector. The global feature vector outputs a K=3-dimensional gated logic value through a two-layer perceptron structure with residual connections. A Top-K gating mechanism is adopted to activate the k=2 experts with the highest weights, and the gating temperature parameter τ=0.

5. The cloud-type sensing router shares the coordinate estimation path and the meteorological estimation path, and simultaneously performs an auxiliary classification task to classify typhoon cloud types into three categories: weak typhoon eye type, moderate spiral band type, and strong typhoon eyewall type. Load balancing loss is introduced during the router training phase to prevent expert collapse.

5. The method according to claim 1, characterized in that, In step S4, the coordinate fusion expert layer comprises three cloud-type expert networks and one general expert network. Each cloud-type expert network consists of two 3×3 convolutional layers and a coordinate attention module, outputting a 128-channel spatial feature map. The coordinate attention module performs global average pooling on the input features along the horizontal and vertical directions, concatenates the pooling results in both directions, compresses the channels through a shared 1×1 convolution, and then generates horizontal and vertical attention weights respectively. The feature is given directional awareness through element-wise multiplication. The general expert network is implemented using depthwise separable convolutions, with approximately 1 / 9 the number of parameters of a standard convolution. When the outputs of each expert network are weighted and fused, the fusion ratio of the general expert network is controlled by learnable parameters, with an initial contribution of approximately 30%.

6. The method according to claim 5, characterized in that, In step S4, the specific structure of the heatmap regression head is a three-level upsampling network. Each level includes a 3×3 convolution, batch normalization, ReLU activation, and a coordinate attention module. The number of channels is 128→256→128→64 in sequence. The first two levels are followed by 2x bilinear upsampling, and finally a 28×28 typhoon eye probability heatmap is generated through 1×1 convolution. The differentiable soft extremum extraction algorithm for extracting sub-pixel coordinates from the heatmap is as follows: after flattening the heatmap, a probability distribution is generated by scaling with temperature parameter τh=8 using a Softmax operation. A normalized coordinate grid is constructed, and the expected coordinate value is calculated using the probability distribution as weights. The continuous normalized coordinates of the typhoon eye in the image are then output. The soft extremum extraction algorithm is fully differentiable and supports end-to-end gradient propagation.

7. The method according to claim 6, characterized in that, In step S5, the dual-branch cross-attention structure models the physical relationship between the pressure branch and the wind speed branch through four-head multi-head cross-attention. The pressure feature is used as the key-value pair for attention calculation, and the wind speed feature is used as the key-value pair for attention calculation. After adding residuals and connecting them, the normalized pressure and wind speed estimates are output. The dual-radius estimation head simultaneously estimates the four directional components of the R34 wind circle radius and the R64 wind circle radius. Symmetry within each branch is captured through multi-head cross-attention between directions. The hierarchical constraint loss penalizes cases where each directional component of R64 exceeds the corresponding R34 component, ensuring that the estimation results meet the physical constraint that R34≥R64. The Dvorak physical constraint loss encodes the Dvorak empirical relationship between the central pressure and the maximum wind speed into a Huber-form constraint loss, guiding the estimation results to conform to meteorological laws.

8. The method according to claim 1, characterized in that, It also includes step S6, multi-task joint training and report output: a four-stage progressive training strategy is adopted to jointly optimize the visual encoder, hybrid expert network and each estimation head, and multi-task joint training is carried out using a unified loss function including Dvorak physical constraint loss. After training, the input infrared satellite cloud imagery is inferred to output estimated results for the typhoon eye coordinates, central pressure, maximum wind speed, and wind circle radius, and a typhoon meteorological parameter warning report is generated. The four-stage progressive training strategy is as follows: The first stage is the coordinate warm-up stage, lasting approximately 2 epochs, during which the meteorological estimation path and radius estimation head are frozen, and only the low-rank adaptive parameters of the visual encoder and the coordinate estimation path are trained to establish a spatial positioning baseline; the second stage is the meteorological warm-up stage, lasting approximately 13 epochs, during which the coordinate estimation path is frozen, the meteorological estimation path and dual-radius estimation head are trained, and a spatial positioning baseline is established. The first stage is the baseline estimation of meteorological parameters; the second stage is the gradient balancing joint training stage, which lasts about 25 epochs. All parameters are unfrozen and trained simultaneously. The GradNorm multi-task gradient balancing algorithm is used to dynamically adjust the loss weights of each task to keep the gradient norms of each task similar; the third stage is the course learning stage, which lasts about 25 epochs. A four-level course scheduler is used to strengthen the learning weights of strong typhoon samples and weak typhoon samples in turn to improve the model's adaptability to typhoons of different intensities; in the early stage of training, an expert warm-up scheduler is used to activate each cloud type expert in turn to allow them to learn the corresponding cloud type features independently.

9. The method according to claim 8, characterized in that, The typhoon meteorological parameter warning report is generated by the language generation module of the visual language model. The language generation module takes the standardized input image and the estimated typhoon eye coordinates, central pressure, maximum wind speed, and wind circle radius as input, and combines them with the built-in meteorological professional knowledge base to automatically generate a text report containing typhoon intensity level assessment, description of affected areas, and warning suggestions, realizing end-to-end output from satellite cloud image to natural language warning report.

10. A satellite cloud image typhoon eye positioning system based on a hybrid expert network, used to implement the method of claim 9, characterized in that, include: The data preprocessing module is used to acquire infrared satellite cloud images of the typhoon area, crop the area with the initial estimated location of the typhoon eye as the center, perform percentile cropping normalization on the cropped area, and generate a standardized input image in 224×224×3 format. The multi-scale feature extraction module includes a visual encoder unit and a multi-scale fusion unit. The visual encoder unit uses a pre-trained visual language model as its visual encoder, and performs efficient parameter fine-tuning through low-rank adaptive technology to output a 2048-dimensional visual feature map. The multi-scale fusion unit reduces the feature dimension to 512 dimensions through 1×1 convolution, and simultaneously generates three feature maps of different resolutions and fuses them to output a 512-dimensional multi-scale fused feature. The cloud-shaped perception routing module includes a global feature aggregation unit and a Top-K gated network unit. The global feature aggregation unit performs global average pooling on multi-scale fused features, and the Top-K gated network unit identifies typhoon cloud-shaped features through a perceptron structure with residual connections, calculates activation weights, and dynamically routes features to the corresponding expert network. The coordinate estimation path includes a coordinate hybrid expert layer and a heatmap regression head. The coordinate hybrid expert layer contains three cloud-type experts and one general expert. All experts are spatially preserving structures to retain the location information required for typhoon eye positioning. The heatmap regression head restores spatial resolution through three-level upsampling, outputs a 28×28 typhoon eye probability heatmap, and solves the sub-pixel coordinates of the typhoon eye through a differentiable soft extremum extraction algorithm. The meteorological parameter estimation path includes a meteorological hybrid expert layer, a two-branch cross-attention estimation unit, and a dual-radius estimation head. The meteorological hybrid expert layer adopts a global characteristic-type expert structure. The two-branch cross-attention estimation unit models the physical relationship between air pressure and wind speed through a cross-attention mechanism. The dual-radius estimation head simultaneously estimates the four directional components of the radii of the R34 and R64 wind circles and applies hierarchical constraints. The training and report output module is used to execute a four-stage progressive training strategy. It performs multi-task joint optimization through a unified loss function including Dvorak physical constraint loss. During the inference stage, it outputs the typhoon eye location results and meteorological parameter estimates, and calls the language generation function of the visual language model to generate typhoon meteorological parameter warning reports.

Images