Typhoon positioning and strength determining method and device based on remote sensing data

Abstract

The invention relates to the technical field of image analysis and processing, in particular to a typhoon positioning and intensity determining method and device based on remote sensing data, and the method comprises the following steps: carrying out the radial gradient enhancement processing of a microwave brightness temperature remote sensing image, and extracting a gradient extreme point distribution diagram of an eye wall region; performing morphological segmentation on the infrared cloud top temperature remote sensing image to generate a binary contour feature map representing an eye region structure; and inputting the gradient extreme point distribution map and the binary contour feature map into an improved convolutional fusion network, and outputting typhoon center coordinates and strength grade parameters. According to the method, the continuous boundary is reconstructed under the complex condition that a double-eye wall or a broken eye wall exists, the adaptive capacity of a positioning algorithm to an asymmetric and fragmented structure is improved, the calculation range is limited through an eye region constraint mechanism, the interference data participation degree is reduced, and higher energy efficiency ratio and engineering deployability are achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of image analysis and processing, and particularly to a typhoon positioning and intensity determination method and device based on remote sensing data. Background Art

[0002] As a tropical cyclone system, a typhoon has an obvious central symmetry structure and significant cloud radiation characteristics. The precise positioning of its central position and the real-time determination of its intensity level are important bases for carrying out disaster warning, path prediction, and marine operation risk assessment. At present, typhoon positioning and intensity determination mainly rely on the distribution characteristics of cloud top temperature or the spatial statistical results of bright temperature extreme points in infrared remote sensing images, and are analyzed in combination with empirical rules or expert recognition means.

[0003] Traditional methods mostly extract the typhoon eye area based on rule templates or global bright temperature thresholds, and it is difficult to effectively deal with non-ideal structure situations such as eyewall fragmentation, double eyewalls, and eccentric annulus, resulting in large positioning deviations or high intensity misjudgment rates.

[0004] Existing algorithms often only rely on a single infrared channel or the aggregation degree of bright temperature extreme points for intensity estimation, lacking the fusion modeling of the geometric structure (such as roundness, continuity) and physical causes (such as gradient mutation) of the typhoon eye area, and both the interpretability and generalization ability are limited.

[0005] Although some typhoon recognition models based on deep networks have good recognition capabilities, they have high computational complexity and are difficult to meet the rapid processing requirements of high-frequency remote sensing images in meteorological services, especially with obvious response lags under high-intensity cloud clusters or large-scale observation conditions. Summary of the Invention

[0006] The present invention provides a typhoon positioning and intensity determination method and device based on remote sensing data, which has the ability of structural self-adaptation, fuses multi-dimensional geometric and physical characteristics, and can meet the requirements of quasi-real-time processing for typhoon positioning and intensity determination, improving the intelligent level of typhoon eye area recognition and intensity determination in complex meteorological scenarios.

[0007] A typhoon positioning and intensity determination method based on remote sensing data includes the following steps: S1, perform radial gradient enhancement processing on a microwave bright temperature remote sensing image, and extract a gradient extreme point distribution map of the eyewall area; S2, perform morphological segmentation on an infrared cloud top temperature remote sensing image to generate a binary contour feature map representing the eye area structure; S3, input the gradient extreme point distribution map and the binary contour feature map into an improved convolutional fusion network, and output the typhoon center coordinates and intensity level parameters; Among them, the improved convolutional fusion network determines the typhoon center position through the spatial density of gradient extreme points, and determines the intensity level through the weighted fusion value of contour roundness and gradient aggregation degree; The structure of the improved convolutional fusion network includes: Deformable convolutional kernel structure: Enhance the weight of the center point of the convolutional kernel and reduce the peripheral weight to form a center-enhanced perception ability; Circular template matching mechanism: Define a rotatable circular template library, and the template diameter dynamically matches the scale range of the typhoon eye area; Typhoon eye area constraint fusion: Embed the prior knowledge of the eye area position in the spatially weighted fusion map to suppress the interference response in the area outside the eyewall.

[0008] Optionally, the S1 specifically includes: S11, construct a polar coordinate system with the initial predicted position of the typhoon as the origin, convert the microwave brightness temperature remote sensing image from the Cartesian coordinate system to the polar coordinate representation, and generate a polar coordinate brightness temperature distribution map; S12, perform a radial gradient convolution operation on the polar coordinate brightness temperature distribution map, use a directional differential operator to calculate the change rate of the brightness temperature along the radius direction, and output a radial gradient intensity map; S13, detect local maximum points in the radial gradient intensity map, screen out spike points, and generate an initial set of gradient extreme points; S14, based on the continuity constraint of the eyewall structure, perform spatial clustering on the initial set of gradient extreme points, remove discrete noise points deviating from the main eyewall ring belt, and finally form a closed gradient extreme point distribution map.

[0009] Optionally, the screening of spike points includes screening points where the gradient intensity exceeds 2 times the standard deviation of the background field.

[0010] Optionally, the spatial clustering of the initial set of gradient extreme points specifically includes: Calculate the number of neighbors around each gradient extreme point, judge the density of the area where it is located. If there are no other neighboring points near a certain gradient extreme point, it means it is noise or an isolated misdetection point and is automatically removed; Through density clustering, group points that are close to each other and continuously distributed in space into one category to form several point clusters. In the point clusters, find the clustering result that best conforms to the typhoon eyewall characteristics, including a closed structure, a round shape, and the largest number of points, to form a candidate eyewall point set; Perform a shape judgment on this candidate eyewall point set. If a ring-shaped or approximately closed structure is formed, output it as the final eyewall boundary map.

[0011] Optionally, the S2 specifically includes: S21. Perform a multi-scale top-hat transform on the infrared image, where the structuring element is a set of concentric rings, extract the cold cloud cover area with a brightness temperature lower than the preset brightness temperature, and generate an enhanced contrast image; S22. Based on the size constraint of the potential eye area of the typhoon, set an adaptive segmentation threshold, and apply the adaptive segmentation threshold to segment on the enhanced contrast image; S23. Perform a topology-aware morphological closing operation on the initial binary mask, use an elliptical structuring element to bridge the contour breaks, and at the same time retain the temperature anomaly holes in the eye area to form a connected eye area mask; S24. Extract the outer boundary chain code of the connected eye area mask, filter the interference area based on the perimeter or roundness, and finally generate a binary contour feature map.

[0012] Optionally, the adaptive segmentation threshold is used to perform adaptive regional threshold segmentation on the enhanced contrast image in combination with the size of the potential eye area of the typhoon. The adaptive segmentation threshold is set as: ; where, is the adaptive segmentation threshold, is the average pixel brightness temperature in the eye area candidate window, is the standard deviation of the brightness temperature in the corresponding area. Binaryize the image according to this adaptive segmentation threshold and output the initial binary mask image, where the bright points correspond to the potential eye area boundary region.

[0013] Optionally, the extracting the outer boundary chain code of the connected eye area mask and filtering the interference area based on the perimeter or roundness in S24 specifically includes: For each boundary contour, calculate its perimeter and geometric roundness , and execute the following filtering rules: Only retain those that satisfy: pixels; of the area; Finally, output the binary contour feature map.

[0014] Optionally, the processing of the improved convolutional fusion network in S3 includes: S31. Dual-channel feature map construction: Use the gradient extreme point distribution map as the first channel and the binary contour feature map as the second channel, and stack them to form a dual-channel input feature map; S32. Spatial density-aware convolution: Use a deformable convolution kernel to process the first channel, where the center weight of the convolution kernel is multiple times the edge weight, strengthen the response of the gradient extreme point aggregation area, and output a gradient aggregation heat map; S33, Geometric Structure Analysis Convolution: Apply a circular template matching convolutional layer to the second channel, calculate the similarity between each position and the standard circle, and output the contour roundness distribution map; S34, Cross-modal Feature Fusion: Concatenate the gradient aggregation heat map and the contour roundness distribution map in the channel dimension, and generate a spatially weighted fusion map through 1×1 convolution.

[0015] Optionally, S3 further includes: Typhoon Center Location: Calculate the centroid coordinates of the gradient aggregation heat map, and its horizontal and vertical coordinate values are the typhoon center coordinates; Intensity Level Determination: Extract the mean value of the spatially weighted fusion map within a pixel region with a predetermined radius centered on the typhoon center coordinates, and map it to an intensity level parameter through the Sigmoid function.

[0016] A typhoon location and intensity determination device based on remote sensing data, used to implement the above-mentioned typhoon location and intensity determination method based on remote sensing data, including the following modules: Image Preprocessing Module: Used to receive and parse microwave brightness temperature remote sensing images and infrared cloud top temperature remote sensing images, perform polar coordinate transformation and radial gradient calculation on the microwave brightness temperature remote sensing images to generate a gradient extreme point distribution map; perform multi-scale morphological processing and structure filtering on the infrared cloud top temperature remote sensing images to generate a binary contour feature map; Feature Fusion Network Module: Includes an improved convolutional fusion network, input two feature maps, and sequentially perform: Extract gradient aggregation features with deformable convolutional kernels; Obtain the geometric structure features of the eye area through circular template matching; Perform cross-modal channel fusion and superimpose the prior information of the eye area to generate a spatially weighted fusion map; Location and Intensity Estimation Module: Calculate the typhoon center coordinates according to the centroid of the gradient aggregation heat map, and combine the average value and roundness distribution of the eye area response region in the fusion map to output the typhoon intensity level parameter.

[0017] Advantages of the present invention: By introducing a combined structure extraction mechanism of "polar coordinate transformation + radial gradient convolution + spatial clustering", the present invention can accurately capture the gradient extreme points in the typhoon eyewall area, and reconstruct its continuous boundary in complex situations with double eyewalls or broken eyewalls, improving the adaptability of the location algorithm to asymmetric and fragmented structures.

[0018] The present invention constructs a cross-modal spatial feature expression system integrating a "gradient aggregation heat map" and a "contour roundness distribution map", and realizes the joint modeling of the structural compactness and geometric roundness of the typhoon eye region by improving the convolutional fusion network. Compared with the traditional method relying on single-channel cloud top brightness temperature or expert rule scoring, it can realize the dual-factor joint quantification of the eyewall intensity and the eye region integrity, and still maintain strong discrimination stability under the interference of high-intensity peripheral cloud systems.

[0019] The present invention significantly optimizes the computational complexity of the image processing process and controls the average single-frame image processing time by means of a one-dimensional radial gradient calculation method in the polar coordinate system and a lightweight convolutional module design (such as deformable convolutional kernels and 1×1 fusion layers), meeting the requirements of the typhoon monitoring system for quasi-real-time response; at the same time, by restricting the calculation range through an eye region constraint mechanism and reducing the participation of interference data, it achieves a higher energy efficiency ratio and engineering deployability. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only those of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0021] Figure 1 It is a schematic flowchart of the method according to an embodiment of the present invention; Figure 2 It is a schematic diagram of each module of the device according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0022] The following will describe the present invention in detail with reference to the drawings and specific embodiments. For some well-known technologies, those skilled in the art can also adopt other alternative methods for implementation; moreover, the drawing part is only for a more specific description of the embodiments and is not intended to specifically limit the present invention.

[0023] As Figure 1 shown, a typhoon positioning and intensity determination method based on remote sensing data includes the following steps: S1, perform radial gradient enhancement processing on the microwave brightness temperature remote sensing image, and extract the distribution map of gradient extreme points in the eyewall area; S2, perform morphological segmentation on the infrared cloud top temperature remote sensing image to generate a binary contour feature map representing the eye region structure; <NULL> S3, input the distribution map of gradient extreme points and the binary contour feature map into the improved convolutional fusion network, and output the typhoon center coordinates and intensity level parameters; Among them, the improved convolutional fusion network determines the typhoon center position through the spatial density of the gradient extreme points, and determines the intensity level through the weighted fusion value of the contour roundness and the gradient aggregation degree; The structure of the improved convolutional fusion network includes: Deformable convolutional kernel structure: Enhance the weight of the central point of the convolutional kernel and reduce the peripheral weight to form the central enhanced perception ability; Circular template matching mechanism: Define a rotatable circular template library, and the template diameter dynamically matches the scale range of the typhoon eye area; Typhoon eye area constraint fusion: Embed the prior knowledge of the eye area position in the spatially weighted fusion map to suppress the interference response in the area outside the eyewall.

[0024] S1 specifically includes: S11, Polar coordinate transformation: Construct a polar coordinate system with the initial predicted position of the typhoon as the origin, and convert the microwave brightness temperature image from the Cartesian coordinate system to polar coordinate representation , generating a polar coordinate brightness temperature distribution map . This transformation enables all subsequent gradient calculations to be performed along the typhoon's radial direction, avoiding the deviation error between the gradient direction in the Cartesian coordinate system and the normal direction of the eyewall.

[0025] The microwave brightness temperature remote sensing image refers to the brightness temperature observation map of the ocean or atmospheric system in the microwave band, which is provided by the microwave radiometer carried by the satellite. The data sources include the Global Precipitation Measurement Microwave Imager (GPM Microwave Imager), the main instrument on the Global Precipitation Observation Mission (GPM) satellite, with high resolution and is particularly suitable for detecting the bright temperature depression characteristics in the eyewall area.

[0026] Processing flow: Obtain Level-1 or Level-2 microwave brightness temperature products; Crop according to the orbit-matched geographical area; Interpolate to unify the spatial resolution; Project and transform to Cartesian coordinates or polar coordinates for analysis.

[0027] The initial predicted position of the typhoon (i.e., the initial estimated point of the typhoon center) can be obtained through various paths and is used to construct the origin of the polar coordinate system, including obtaining it through the numerical prediction center. The sources are such as the forecast products of the National Meteorological Center, which provide hourly or 6-hour center position estimates and are directly used as the origin of the polar coordinate system.

[0028] S12, Radial gradient convolution operation: Perform a one-dimensional radial gradient convolution on the polar coordinate brightness temperature map and calculate the rate of change of the brightness temperature along the radius direction using the directional differential operator, and output the radial gradient intensity map , which is expressed as: ; among which, represents the brightness temperature value in polar coordinates, represents the brightness temperature gradient intensity along the radial direction, represents the first-order differential approximation in the radial direction, and the direction differential operator uses a one-dimensional kernel: [-1, 0, 1], along direction convolution. This design has higher radial specificity and edge response accuracy compared with the conventional two-dimensional Sobel operator.

[0029] S13, local extreme point detection and preliminary screening: In the radial gradient intensity map , detect local maximum points, and screen the points that meet the following conditions to form the initial set of gradient extreme points : ; wherein, is the average gradient intensity of the full-map background field, is the standard deviation of the background field gradient intensity, and the screening condition uses as the threshold. This mechanism can adapt to different typhoon background brightness temperature conditions and effectively improve the recognition ability of weak-gradient eyewall structures.

[0030] S14, spatial clustering and noise removal: Based on the continuity constraint of the eyewall structure, perform spatial density clustering analysis on the initial set , remove outlier discrete points, and retain the points constituting the main eyewall annulus. Finally, output the closed gradient extreme point distribution map , and the specific constraint logic is as follows: Distance threshold: , keep the annulus connected, The value is 10 - 15 km. The diameter of the typhoon eyewall is usually 30 - 80 km, and it is more appropriate to set the point spacing at 1 / 4 - 1 / 8 of its order of magnitude; Point density threshold: Local point density , filter isolated points, The value is 8 - 12. The eyewall structure should have good annular continuity; if there are at least 8 - 12 neighboring points in the neighborhood of each legal point, it can ensure the formation of a "density-connected structure", The smaller it is, the more weak edges are retained but the false detection increases; the larger it is, the smaller the error but there may be missed detection of incomplete eyewalls.

[0031] Finally formed is the "gradient extreme point distribution map" required in step S3.

[0032] Perform spatial density clustering and output the closed The specific process is as follows: (a) Spatial density estimation: Use a fixed radius Calculating the local density based on the number of neighbors within : ; where is the -th point in the initial point set, is the -th point in the initial point set, is 's Euclidean distance or spherical distance, is the neighborhood radius (1.5 km); Set the minimum density threshold , and remove 's isolated points.

[0033] (b) Application of spatial clustering algorithm: Select the DBSCAN algorithm: Input parameters: Neighborhood radius and the minimum number of neighbors MinPts; Operation result: Output multiple density-connected clusters , and select the cluster with the largest area and the highest closure degree (circularity > threshold) as the main eyewall point set.

[0034] (c) Judgment and output of the annulus closure degree: Define the closure degree Roundness of the clustered point set : ; where represents the area of the region enclosed by the point set, represents the perimeter of the boundary formed by the point set; If the closure degree (empirical threshold), then it is recognized as a complete eyewall area and output as the final closure distribution map: ; otherwise, the sub-optimal clusters can be merged to form a combined closed annulus, and " " means "assigned as".

[0035] Typhoons have a central symmetry structure, and the eyewall shows a typical annular distribution. In the traditional Cartesian coordinate system, the gradient calculation directions are fixed as the horizontal direction (x-axis) and the vertical direction (y-axis), which deviate from the true radial diffusion direction of the eyewall, resulting in insensitivity or distortion to the boundary strength response. Therefore, using the polar coordinate system (with the typhoon center as the origin) can ensure that all subsequent analyses are carried out strictly according to the "radial" characteristics of the typhoon structure, significantly improving the ability to restore the structure of the eyewall area. The polar coordinate system conversion of S11 is as follows: S111. First, select a point as the center of the polar coordinate system, generally the initial predicted position of the typhoon, denoted by . This point will be used as the origin of the polar coordinate system, that is .

[0036] S112: For each bright temperature pixel point in the image, its position in the Cartesian coordinate system is , relative to the center point The relative position is: ; ; Then the corresponding polar coordinate position is: , that is, the distance from this point to the center, representing the radius; , that is, the angle from the positive x-axis rotating counterclockwise to the direction of this point, with the unit of radian or degree.

[0037] This transformation process is essentially rearranging the "horizontal-vertical structure" into a "radius-angle structure".

[0038] S113: Construct a polar coordinate grid: In image processing implementation, preset a polar coordinate grid: In the radius direction : From 0 to the maximum observation range (such as 0 - 200 km), sample at equal intervals; In the angle direction : From 0 to 360°, usually sample with an angle accuracy of 1° or higher.

[0039] The corresponding real image position of each polar coordinate grid point is: ; ; S114: Since pixels usually do not exactly fall on integer positions, bilinear interpolation method is needed to sample the bright temperature value on the original image: For each point in the polar coordinate grid , according to its corresponding coordinates, find the four neighboring pixel values in the original image; Obtain the bright temperature value of this point by weighted average according to the distance ; In this way, the resampling from the Cartesian image to the polar coordinate bright temperature distribution map is completed.

[0040] S2 specifically includes: S21, multi-scale top-hat transform: Perform multi-scale top-hat transform on the infrared cloud top temperature remote sensing image , and use a set of concentric rings with different radii as the structural element , which is used to extract the cold cloud cover area with a bright temperature lower than 260K, and output an enhanced contrast image , effectively highlighting the temperature jump structure in the eye area. Specifically, multiple "structural elements" of different sizes are constructed in the image. These structural elements are concentric circles centered around the image center, corresponding to different typhoon eye areas or cloud system scales respectively. Each structural element slides in the image to simulate the trend of "local background brightness temperature". Next, for each structural element scale, a morphological opening operation is performed, that is, erosion first and then dilation, to obtain the background fitting map at this scale. Subtracting the original image from this background image can obtain the "top-hat image" at this scale, that is, the area in the original image where the brightness temperature is significantly lower than the local background. To more comprehensively cover cold cloud cover structures of different sizes, the above operations are repeated for multiple structural elements of different scales, and the top-hat images at all scales are superimposed or the maximum value is taken to generate a final enhanced contrast image. This enhanced image can clearly highlight structural features such as the edge of the cold cloud cover and the brightness temperature mutation in the typhoon eye area, providing a clearer and more stable input for subsequent threshold segmentation and boundary extraction.

[0041] S22, Adaptive threshold segmentation: On the enhanced contrast image , perform adaptive regional threshold segmentation in combination with the potential eye area size of the typhoon, and set the threshold as: ; where is the average brightness temperature of pixels in the eye area candidate window, and is the standard deviation of the brightness temperature in the corresponding area. Binaryize the image according to this threshold, and output the initial binary mask image , where the bright points correspond to the potential eye area boundary region.

[0042] S23, Topology-aware morphological closing operation: Apply a topology-aware morphological closing operation to the initial mask , and the structural element is an ellipse with direction adaptability to repair boundary breaks caused by image occlusion or cloud body fragmentation.

[0043] The major axis direction is perpendicular to the typhoon movement direction (normal alignment), enhancing the repair efficiency; The warm core area with a temperature higher than 280K is set as a protection mask in the closing operation and does not participate in the filling process.

[0044] The processing result is a connected eye area mask that retains the true temperature cavity structure in the eye area.

[0045] Specifically as follows: Generate an ellipse structural element with direction adaptability according to the typhoon movement direction or the main stretching direction of the eye area in the image , where the long axis direction is perpendicular to the typhoon movement direction and is used for patching along the boundary shape direction. Then, perform a morphological closing operation on the initial mask image That is, first dilate and then erode. The dilation operation will expand the mask edge, filling and connecting the originally broken or discontinuous areas; while the erosion operation further refines the edge to ensure that the structure does not expand excessively and restores the true boundary shape. During the execution process, a topology-aware mechanism is applied simultaneously, that is, protecting the areas with higher brightness temperature in the mask (such as the warm core area inside the typhoon eye), and not allowing it to be misfilled by the closing operation. This mechanism ensures that the true cavity structure inside the eye area is retained and will not be wrongly regarded as a "boundary gap" due to morphological operations. Finally, the output connected eye area mask image Has a continuous, closed and complete outer boundary, while retaining the key internal topological features, providing structurally complete basic data for the subsequent accurate extraction of the typhoon eye area contour.

[0046] S24, Contour extraction and region filtering: Extract the outer boundary chain codes of all connected regions from (boundary tracking). For each boundary contour, calculate its perimeter and geometric circularity , and execute the following filtering rules: Only retain those that meet: pixels, . Finally, output the binary contour feature map , as one of the inputs to the convolutional fusion network in step S3.

[0047] The infrared cloud top temperature image comes from the Fengyun series of meteorological satellites (FY-2, FY-4). FY-4A is equipped with an AGRI sensor, which has high spatio-temporal resolution infrared imaging capabilities and provides parameters such as cloud top temperature and cloud height.

[0048] The data processing flow is briefly described as follows: Download the satellite raw observation data; Decode specific infrared bands; Use the radiation conversion formula or the provided brightness temperature product to convert the radiometer output into a brightness temperature image in units of "K"; Resample or project to an equal latitude and longitude grid to meet the analysis input requirements; Output in a standard image format for subsequent use.

[0049] In this embodiment, denote the gradient extreme point distribution map output in S1 as , indicating the gradient extreme response value of the image at the pixel position ; Denote the binary contour feature map output in step S2 as , indicating whether the image belongs to the edge of the eye area contour at the pixel position .

[0050] Specifically, S3 includes: S31, Dual-channel feature map construction: Using the gradient extreme point distribution map as the first channel and the binary contour feature map as the second channel, and superimposing them to form a dual-channel input feature map: .

[0051] S32, Spatial density perception convolution (for the first channel): Applying a 3×3 deformable convolution kernel to the first channel , where the center weight of the convolution kernel is three times the edge weight, used to enhance the response to the eye wall gradient aggregation area, and outputting a gradient aggregation degree heat map : ; where: The convolution kernel weight configuration is: center weight = 0.3, and each of the 8 neighboring points is 0.1; The symbol "*" represents a two-dimensional convolution operation; The output map represents the gradient aggregation degree within each pixel neighborhood.

[0052] S33, Geometric structure analysis convolution (for the second channel): Applying a rotatable circular template matching convolution kernel to the second channel , evaluating the similarity between each position area and the standard circle, and outputting a contour roundness distribution map : ; Among them, represents circular templates constructed with different diameters and rotation angles ; The output map represents the matching degree between the pixel neighborhood and the circular structure.

[0053] S34, After splicing the above two heat maps in the channel dimension, obtaining a spatially weighted fusion map through convolutional fusion: ; Among them, represents the convolutional weight; the output map represents the fusion response intensity combining gradient aggregation and structural roundness.

[0054] S35, Typhoon center positioning: Calculating the centroid coordinates of the gradient aggregation degree heat map [[ID= , this coordinate is the estimated typhoon center point: .

[0055] Fusion map provides structural consistency guarantee for "extracting the eye area with the typhoon center as the center". If there are offsets or non-structural noise points in the original image, the contour roundness signal in the fusion map will guide the structural contour of the sensing area to be closer to the real typhoon eye area.

[0056] S36, intensity level determination: Taking as the center, extract a local fusion area with a radius of 100 pixels, and calculate its average value , and map it to an intensity level parameter through the Sigmoid function : ; wherein, GAI represents the mean value of the gradient aggregation degree heat map in the selected area, CR represents the peak value of the contour roundness distribution map in the selected area, represents the fusion coefficient obtained through training and learning, ; represents the input variable of the Sigmoid function, that is, .

[0057] GAI and CR are values extracted from two original images, but the semantic basis of these values is strengthened by the fusion map . That is to say: the fusion map has mixed "structural" (roundness) and "intensity" (aggregation degree) at the spatial level; making the two index values more stable and the distribution more consistent; avoiding problems such as inconsistent signals and inaccurate spatial alignment.

[0058] As Figure 2 shown, a typhoon positioning and intensity determination device for implementing the above typhoon positioning and intensity determination method includes the following modules: Image preprocessing module: used to receive and parse microwave brightness temperature remote sensing images and infrared cloud top temperature remote sensing images, perform polar coordinate transformation and radial gradient calculation on the microwave brightness temperature remote sensing images to generate a gradient extreme point distribution map; perform multi-scale morphological processing and structure filtering on the infrared cloud top temperature remote sensing images to generate a binary contour feature map; Feature fusion network module: including an improved convolutional fusion network, inputting two feature maps, and sequentially performing: Using deformable convolutional kernels to extract gradient aggregation features; Performing circular template matching to obtain eye area geometric structure features; Performing cross-modal channel fusion and superimposing eye area prior information to generate a spatially weighted fusion map; Positioning and intensity estimation module: Calculate the typhoon center coordinates according to the centroid of the gradient aggregation heat map, and combine the average value and circularity distribution of the eye response area in the fusion map to output the typhoon intensity level parameters.

[0059] The present invention covers any alternatives, modifications, equivalent methods and solutions made on the essence and scope of the present invention. In order to enable the public to have a thorough understanding of the present invention, specific details are described in detail in the following preferred embodiments of the present invention, and those skilled in the art can fully understand the present invention without the description of these details. In addition, well-known methods, processes, procedures, components and circuits are not described in detail in order to avoid unnecessary confusion to the essence of the present invention.

[0060] The above are only the preferred embodiments of the present invention. It should be pointed out that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present invention.

Claims

1. A typhoon positioning and intensity determination method based on remote sensing data, characterized in that, It includes the following steps: S1. Perform radial gradient enhancement processing on the microwave brightness temperature remote sensing image, and extract the distribution map of gradient extreme points in the eyewall area; S2. Perform morphological segmentation on the infrared cloud top temperature remote sensing image to generate a binary contour feature map representing the eye area structure; S3. Input the distribution map of gradient extreme points and the binary contour feature map into an improved convolutional fusion network, and output the typhoon center coordinates and intensity level parameters; Among them, the improved convolutional fusion network determines the typhoon center position through the spatial density of gradient extreme points, and determines the intensity level through the weighted fusion value of contour roundness and gradient aggregation degree; The structure of the improved convolutional fusion network includes: Deformable convolutional kernel structure: Enhance the weight of the center point of the convolutional kernel and reduce the peripheral weight to form a center-enhanced perception ability; Circular template matching mechanism: Define a rotatable circular template library, and the template diameter dynamically matches the scale range of the typhoon eye area; Typhoon eye area constraint fusion: Embed the prior knowledge of the eye area position in the spatially weighted fusion map to suppress the interference response in the area outside the eyewall.

2. The typhoon positioning and intensity determination method based on remote sensing data according to claim 1, characterized in that The specific content of S1 includes: S11. Construct a polar coordinate system with the initial predicted position of the typhoon as the origin, convert the microwave brightness temperature remote sensing image from the Cartesian coordinate system to the polar coordinate representation, and generate a polar coordinate brightness temperature distribution map; S12. Perform a radial gradient convolution operation on the polar coordinate brightness temperature distribution map, use a directional differential operator to calculate the brightness temperature change rate along the radius direction, and output a radial gradient intensity map; S13. Detect local maximum points in the radial gradient intensity map, screen out spike points, and generate an initial set of gradient extreme points; S14. Based on the continuity constraint of the eyewall structure, perform spatial clustering on the initial set of gradient extreme points, remove discrete noise points deviating from the main eyewall annulus, and finally form a closed distribution map of gradient extreme points.

3. A typhoon positioning and intensity determination method based on remote sensing data according to claim 2, characterized in that, The screening of spike points includes screening points where the gradient intensity exceeds 2 times the standard deviation of the background field.

4. A typhoon positioning and intensity determination method based on remote sensing data according to claim 2, characterized in that, The specific content of performing spatial clustering on the initial set of gradient extreme points includes: Calculate the number of neighbors around each gradient extreme point, judge the density of the area where it is located. If there are no other neighboring points near a certain gradient extreme point, it means it is noise or an isolated misdetection point and is automatically removed; Through density clustering, group points that are close to each other and continuously distributed in space into one category to form several point clusters, and find the clustering result that best conforms to the typhoon eyewall characteristics in the point clusters, including closed structure, round shape, and the largest number of points, to form a candidate eyewall point set; Perform a shape judgment on this candidate eyewall point set. If a ring-shaped or approximately closed structure is formed, output it as the final eyewall boundary map.

5. A typhoon positioning and intensity determination method based on remote sensing data according to claim 1, characterized in that, The specific content of S2 includes: S21. Perform a multi-scale top-hat transform on the infrared image, where the structural element is a set of concentric rings, extract the cold cloud cover area where the brightness temperature is lower than the preset brightness temperature, and generate an enhanced contrast image; S22. Based on the size constraint of the potential typhoon eye area, set an adaptive segmentation threshold, and apply the adaptive segmentation threshold to perform segmentation on the enhanced contrast image; S23. Perform a topological perception-based morphological closing operation on the initial binary mask, use an elliptical structural element to bridge contour breaks, and at the same time retain the temperature anomaly holes in the eye area to form a connected eye area mask. S24. Extract the outer boundary chain code of the connected eye region mask, filter out the interference regions based on the perimeter or circularity, and finally generate a binary contour feature map.

6. The typhoon positioning and intensity determination method based on remote sensing data according to claim 5, characterized in that, The adaptive segmentation threshold is used to perform adaptive regional threshold segmentation on the enhanced contrast image in combination with the size of the potential eye region of the typhoon. The adaptive segmentation threshold is set as: ; wherein, is the adaptive segmentation threshold, is the average brightness temperature of pixels within the candidate window of the eye region, is the standard deviation of the brightness temperature in the corresponding region. The image is binarized according to this adaptive segmentation threshold, and an initial binary mask image is output, where the bright points correspond to the potential eye region boundary area.

7. A typhoon positioning and intensity determination method based on remote sensing data according to claim 5, characterized in that, In S24, the extraction of the outer boundary chain code of the connected eye region mask and the filtering of the interference regions based on the perimeter or circularity specifically include: For each boundary contour, calculate its perimeter and geometric roundness , and perform the following filtering rules: Only retain those that satisfy: Pixel; area; Finally, output the binary contour feature map.

8. A typhoon positioning and intensity determination method based on remote sensing data according to claim 1, characterized in that, The processing of the improved convolutional fusion network in S3 includes: S31. Dual-channel feature map construction: Use the gradient extreme point distribution map as the first channel and the binary contour feature map as the second channel, and stack them to form a dual-channel input feature map; S32. Spatial density-aware convolution: Use a deformable convolution kernel to process the first channel, where the weight at the center of the convolution kernel is multiple times that of the edge weights, strengthen the response in the gradient extreme point aggregation region, and output a gradient aggregation heat map; S33. Geometric structure analysis convolution: Apply a circular template matching convolutional layer to the second channel, calculate the similarity between each position and the standard circle, and output a contour circularity distribution map; S34. Cross-modal feature fusion: Concatenate the gradient aggregation heat map and the contour circularity distribution map in the channel dimension, and generate a spatially weighted fusion map through 1×1 convolution.

9. A typhoon positioning and intensity determination method based on remote sensing data according to claim 8, characterized in that, S3 also includes: Typhoon center positioning: Calculate the centroid coordinates of the gradient aggregation heat map, and its horizontal and vertical coordinate values are the typhoon center coordinates; Intensity level determination: Extract the mean value of the spatially weighted fusion map within a pixel region with a predetermined radius centered on the typhoon center coordinates, and map it to an intensity level parameter through the Sigmoid function.

10. A typhoon positioning and intensity determination device based on remote sensing data, which is used to implement a typhoon positioning and intensity determination method based on remote sensing data as described in any one of claims 1-9, characterized in that, It includes the following modules: Image preprocessing module: Used to receive and parse the microwave brightness temperature remote sensing image and the infrared cloud top temperature remote sensing image, perform polar coordinate transformation and radial gradient calculation on the microwave brightness temperature remote sensing image to generate a gradient extreme point distribution map; perform multi-scale morphological processing and structure filtering on the infrared cloud top temperature remote sensing image to generate a binary contour feature map; Feature fusion network module: Includes an improved convolutional fusion network, which inputs two feature maps and sequentially performs: Extract gradient aggregation features using a deformable convolution kernel; Obtain the eye region geometric structure features through circular template matching; Perform cross-modal channel fusion and superimpose the eye region prior information to generate a spatially weighted fusion map; Positioning and intensity estimation module: Calculate the typhoon center coordinates based on the centroid of the gradient aggregation heat map, and combine the average value and circularity distribution of the eye region response area in the fusion map to output the typhoon intensity level parameter.

Images