High pressure recognition method and system based on residual convolutional neural network

Abstract

The application provides a high-pressure recognition method and system based on a residual convolutional neural network, relates to the technical field of high-pressure recognition, and comprises the following steps: acquiring ERA-5 data, and determining high-pressure judgment criteria; normalizing and combining three kinds of meteorological parameter data into a three-channel data set, and dividing the three-channel data set into a training set and a test set; adopting a ResNet framework to establish a model and optimization; and finally extracting closed contour line information and a center position according to a model judgment result. The application can accurately recognize a high-pressure system, reduces artificial errors, and improves the accuracy and efficiency of high-pressure recognition.

Description

Technical Field

[0001] This invention relates to high voltage identification technology, and more particularly to a high voltage identification method and system based on residual convolutional neural networks. Background Technology

[0002] With the intensification of climate change and the increasing frequency of extreme weather events, the accurate identification and prediction of high-pressure weather systems are of great significance in meteorological research and weather forecasting. High-pressure systems are an important component of atmospheric circulation, usually associated with clear, dry weather, and sometimes leading to extreme weather events such as heat waves and droughts. Traditional methods for identifying high pressure systems mainly rely on manual analysis by meteorologists based on meteorological elements such as geopotential height fields and wind fields. This method requires a high level of expertise and is relatively inefficient.

[0003] With the development of computer technology, objective identification methods have been gradually applied to the identification of high-pressure systems. Early objective identification algorithms were mainly based on threshold methods, identifying high-pressure systems by setting thresholds for geopotential height, air pressure, or vorticity. Later, methods based on cluster analysis and fuzzy recognition were developed, further improving the accuracy of identification. In recent years, with the rise of artificial intelligence technology, machine learning, especially deep learning methods, has been introduced into the field of meteorology, providing new solutions for the automatic identification of high-pressure systems.

[0004] Traditional manual identification methods rely heavily on the subjective experience of meteorological experts. Different experts may have different judgments on the same weather system, making it difficult to guarantee the consistency and objectivity of the identification results. At the same time, manual identification is inefficient and cannot meet the real-time processing needs of large amounts of meteorological data.

[0005] While threshold-based objective identification methods improve efficiency, fixed thresholds are difficult to adapt to the characteristics of high-pressure systems in different seasons and regions, which can easily lead to identification errors. In particular, the identification accuracy is low for weak high-pressure systems, and the description of high-pressure boundaries is not precise enough.

[0006] Existing machine learning methods have high computational complexity when processing high-dimensional meteorological field data. They are prone to gradient vanishing or gradient exploding problems during model training. They are not sufficient for extracting the spatial structural features of high-pressure systems and have difficulty effectively distinguishing high-pressure systems from other weather systems, especially in complex background fields where their identification ability is limited. Summary of the Invention

[0007] This invention provides a high voltage identification method and system based on residual convolutional neural networks, which can solve the problems in the prior art.

[0008] A first aspect of the present invention provides a high-voltage recognition method based on a residual convolutional neural network, comprising:

[0009] Acquire daily ERA-5 geopotential height data, meridional wind data, and zonal wind data; determine the criteria for high pressure based on the geopotential height data, the meridional wind data, and the zonal wind data, wherein the criteria include the analysis of at least one closed contour line in the geopotential height field and the wind field showing a complete anticyclonic circulation;

[0010] The geopotential height data, the meridional wind data, and the zonal wind data are normalized and then merged into a three-channel dataset. The judgment result of whether high pressure occurs daily is used as the category label for supervised learning. The three-channel dataset is divided into a training set and a test set according to a preset ratio.

[0011] A residual convolutional neural network model is established using the ResNet framework. The performance of the residual convolutional neural network model is evaluated and the parameters are optimized using accuracy and cross-entropy. The accuracy is the proportion of correctly identified samples in the total number of samples, and the cross-entropy is used to measure the difference between the true probability distribution and the predicted probability distribution. When the accuracy reaches a preset accuracy threshold and the loss function value of the cross-entropy is less than a preset loss threshold, the iterative training of the residual convolutional neural network model is stopped.

[0012] Based on the determination results of the residual convolutional neural network model, the outermost closed contour line information of the high pressure is extracted, and the highest potential point within the outermost closed contour line of the high pressure is determined as the center position of the high pressure. The highest potential anomaly is used as the high pressure intensity to obtain the high pressure identification result.

[0013] The criteria for determining high pressure based on the geopotential height data, the meridional wind data, and the zonal wind data include:

[0014] A point-by-point tracking method is used to obtain a set of contour points. The distance between the first and last two contour points in the set is calculated. When the distance is less than a preset closure threshold, the set of contour points is determined to constitute a closed contour line.

[0015] Obtain the geopotential height value distribution of the area enclosed by the closed contour line, calculate the difference between the geopotential height value within the enclosed area and the geopotential height value of the surrounding area, and determine whether the closed contour line satisfies the high-pressure geopotential height characteristics based on the difference;

[0016] Obtain wind field data for the enclosed area, and calculate the wind field divergence and vorticity of the enclosed area. The wind field divergence is obtained by calculating the sum of the partial derivatives of the zonal wind in the longitude direction and the partial derivatives of the meridional wind in the latitude direction. The vorticity is obtained by calculating the difference between the partial derivatives of the meridional wind in the longitude direction and the partial derivatives of the zonal wind in the latitude direction.

[0017] The wind field divergence weight ratio coefficient is obtained by multiplying the ratio of the wind field divergence to the divergence reference value by a first weighting coefficient. The vorticity inverse number weight ratio coefficient is obtained by multiplying the ratio of the vorticity inverse number to the vorticity reference value by a second weighting coefficient. The wind field divergence weight ratio coefficient and the vorticity inverse number weight ratio coefficient are added together to obtain the final comprehensive judgment criterion.

[0018] A point-by-point tracking method is used to obtain a set of contour points. The distance between the first and last contour points in the set is calculated. When the distance is less than a preset closure threshold, it is determined that the set of contour points constitutes a closed contour line.

[0019] The study area is divided into regular grids, and the grid point spacing is calculated based on the regular grids. The grid point spacing includes the spacing in the longitude direction and the spacing in the latitude direction.

[0020] Search for grid points in the regular grid that meet the tracking start point conditions. The tracking start point conditions include: the absolute value of the difference between the geopotential height value of the grid point and the target contour height value is less than a preset distance threshold, and the geopotential height values ​​of the grid point and its adjacent grid points are located on both sides of the target contour height value.

[0021] Based on the tracking starting point condition, linear interpolation calculation is performed between adjacent grid points to obtain the precise coordinates of the contour points, wherein the precise coordinates are calculated by the ratio of the geopotential height value of the adjacent grid points to the height value of the target contour line.

[0022] Calculate the distance between the first and last two isoline points in the isoline point set and the total length of the two isoline points. Multiply the ratio of the distance value to the total length value by a first weighting coefficient to obtain the distance-length ratio weighting coefficient. Multiply the ratio of the tangential angle difference between the two isoline points by a second weighting coefficient to obtain the tangential angle difference ratio weighting coefficient. Add the distance-length ratio weighting coefficient and the tangential angle difference ratio weighting coefficient to obtain the closure evaluation index.

[0023] Based on the closure evaluation index, it is determined whether the contour point set is closed. When the closure evaluation index is less than the preset index threshold, the contour point set is determined to be a closed contour line.

[0024] A residual convolutional neural network model is built using the ResNet framework. Performance evaluation and parameter optimization of the residual convolutional neural network model are performed using accuracy and cross-entropy, including:

[0025] A residual convolutional neural network model is established based on the ResNet framework, which includes a shallow feature extraction part and a deep feature learning part.

[0026] The residual convolutional neural network model is trained and optimized by using the cross-entropy loss function to calculate the model loss and by using the Adam optimizer to optimize the model parameters.

[0027] Based on the determination results of the residual convolutional neural network model, the outermost closed contour line information of the high-pressure system is extracted, and the highest potential point within the outermost closed contour line of the high-pressure system is determined as the center position of the high-pressure system. The highest potential anomaly is used as the high-pressure intensity, and the high-pressure identification results include:

[0028] Based on the determination result of the residual convolutional neural network model, it is determined whether the distance between the first and last points of the contour point set in the geopotential height data is less than a preset closure threshold. When the distance is less than the preset closure threshold, the contour point set is determined as the outermost closed contour line of the high pressure.

[0029] Obtain the boundary point coordinate information of the outermost closed contour line of the high voltage system, construct a closed region of the high voltage system based on the boundary point coordinate information, traverse the geopotential height value of each grid point within the closed region, and determine the grid point with the largest geopotential height value as the center point of the high voltage system.

[0030] The geopotential anomaly is obtained by subtracting the geopotential climate average value of the center location from the geopotential height value of the center location.

[0031] The coordinates of the central location point, the geopotential anomaly, and the coordinates of the boundary points of the outermost closed contour line of the high-pressure system are used as the quantitative feature information of the high-pressure system. The high-pressure identification result is output based on the quantitative feature information.

[0032] Based on the boundary point coordinate information, a closed region of the high-voltage system is constructed. Within the closed region, the potential height value of each grid point is traversed, and the grid point with the largest potential height value is determined as the center point of the high-voltage system, including:

[0033] Obtain the coordinate sequence of boundary points of the high-voltage system. The boundary points in the coordinate sequence are sorted according to spatial continuity. Construct a region indicator function based on the coordinate sequence. The region indicator function is used to mark whether the grid points are located within the closed region enclosed by the boundary points.

[0034] Based on the region indicator function, the set of grid points located within the closed region is filtered, and the position coordinates of each grid point in the set of grid points are mapped to the potential height field function to obtain the potential height value of all grid points within the closed region.

[0035] For each grid point within the closed area, a local maximum value determination is performed. The local maximum value determination includes: calculating the potential height difference between the target grid point and its surrounding grid points; when the potential height value of the target grid point is greater than the potential height values ​​of all surrounding grid points, the target grid point is marked as a local maximum value point.

[0036] Among the local maximum points, determine the grid point corresponding to the global maximum geopotential height value, and determine the coordinates of the grid point corresponding to the global maximum geopotential height value as the center location point of the high-voltage system.

[0037] A second aspect of the present invention provides a high-voltage recognition system based on a residual convolutional neural network, comprising:

[0038] The first unit is used to acquire daily ERA-5 geopotential height data, meridional wind data, and zonal wind data; and to determine the criteria for high pressure based on the geopotential height data, the meridional wind data, and the zonal wind data. The criteria include the analysis of at least one closed contour line in the geopotential height field and the presence of a complete anticyclonic circulation in the wind field.

[0039] The second unit is used to normalize and merge the geopotential height data, the meridional wind data, and the zonal wind data into a three-channel dataset. The judgment result of whether high pressure occurs daily is used as the category label for supervised learning, and the three-channel dataset is divided into a training set and a test set according to a preset ratio.

[0040] The third unit is used to build a residual convolutional neural network model using the ResNet framework, and to evaluate the performance and optimize the parameters of the residual convolutional neural network model using accuracy and cross-entropy. The accuracy is the proportion of correctly identified samples in the total number of samples, and the cross-entropy is used to measure the difference between the true probability distribution and the predicted probability distribution. When the accuracy reaches a preset accuracy threshold and the loss function value of the cross-entropy is less than a preset loss threshold, the iterative training of the residual convolutional neural network model is stopped.

[0041] The fourth unit is used to extract the outermost closed contour line information of the high pressure based on the judgment result of the residual convolutional neural network model, determine the highest potential point within the outermost closed contour line of the high pressure as the center position of the high pressure, and use the highest potential anomaly as the high pressure intensity to obtain the high pressure identification result.

[0042] A third aspect of the present invention provides an electronic device, comprising:

[0043] processor;

[0044] Memory used to store processor-executable instructions;

[0045] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0046] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0047] The beneficial effects of this application are as follows:

[0048] By employing residual convolutional neural networks to extract geopotential height field and wind field features, this invention can automatically identify high-pressure systems without human intervention or subjective judgment, significantly improving the accuracy and efficiency of high-pressure identification and eliminating the problem of inconsistent high-pressure identification standards in traditional methods.

[0049] By integrating geopotential height, meridional wind, and zonal wind data into a unified input using multi-channel data fusion technology, and combining it with the residual learning capability of the ResNet framework, this invention can more effectively capture the spatial characteristics and local details of high-pressure systems. While ensuring recognition accuracy, it improves the model's generalization ability and is suitable for high-pressure identification work in different regions and seasons.

[0050] This invention achieves an objective description of the complete characteristics of a high-pressure system by automatically extracting the outermost closed contour information of the high-pressure system and determining its center location and intensity. This provides high-quality data support for subsequent weather forecasting and climate analysis, and has significant scientific research value and application prospects. Attached Figure Description

[0051] Figure 1 This is a flowchart illustrating the high voltage identification method based on residual convolutional neural networks according to an embodiment of the present invention.

[0052] Figure 2 The implementation of this invention is based on the 500 hPa geopotential height field, wind field, two-dimensional resistivity index retrieval results obtained on January 2, 2021, and the visualization class activation map of the ResNet model;

[0053] Figure 3 This is a framework diagram of the model built using a residual convolutional neural network in an embodiment of the present invention;

[0054] Figure 4 This embodiment of the invention is based on the Gaussian density distribution map of the Greenland High center location obtained from 1979 to 2023 using the judgment criteria, ResNet method, and GBI method;

[0055] Figure 5This is a frequency distribution map of the Greenland High in different intensity ranges obtained from 1979 to 2023 using the judgment criteria, ResNet method and GBI method, based on an embodiment of the present invention. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0057] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0058] Figure 1 This is a flowchart illustrating the high voltage identification method based on residual convolutional neural networks according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0059] Acquire daily ERA-5 geopotential height data, meridional wind data, and zonal wind data; determine the criteria for high pressure based on the geopotential height data, the meridional wind data, and the zonal wind data, wherein the criteria include the analysis of at least one closed contour line in the geopotential height field and the wind field showing a complete anticyclonic circulation;

[0060] The geopotential height data, the meridional wind data, and the zonal wind data are normalized and then merged into a three-channel dataset. The judgment result of whether high pressure occurs daily is used as the category label for supervised learning. The three-channel dataset is divided into a training set and a test set according to a preset ratio.

[0061] A residual convolutional neural network model is established using the ResNet framework. The performance of the residual convolutional neural network model is evaluated and the parameters are optimized using accuracy and cross-entropy. The accuracy is the proportion of correctly identified samples in the total number of samples, and the cross-entropy is used to measure the difference between the true probability distribution and the predicted probability distribution. When the accuracy reaches a preset accuracy threshold and the loss function value of the cross-entropy is less than a preset loss threshold, the iterative training of the residual convolutional neural network model is stopped.

[0062] Based on the determination results of the residual convolutional neural network model, the outermost closed contour line information of the high pressure is extracted, and the highest potential point within the outermost closed contour line of the high pressure is determined as the center position of the high pressure. The highest potential anomaly is used as the high pressure intensity to obtain the high pressure identification result.

[0063] In one optional implementation, the criteria for determining high pressure based on the geopotential height data, the meridional wind data, and the zonal wind data include:

[0064] A point-by-point tracking method is used to obtain a set of contour points. The distance between the first and last two contour points in the set is calculated. When the distance is less than a preset closure threshold, the set of contour points is determined to constitute a closed contour line.

[0065] Obtain the geopotential height value distribution of the area enclosed by the closed contour line, calculate the difference between the geopotential height value within the enclosed area and the geopotential height value of the surrounding area, and determine whether the closed contour line satisfies the high-pressure geopotential height characteristics based on the difference;

[0066] Obtain wind field data for the enclosed area, and calculate the wind field divergence and vorticity of the enclosed area. The wind field divergence is obtained by calculating the sum of the partial derivatives of the zonal wind in the longitude direction and the partial derivatives of the meridional wind in the latitude direction. The vorticity is obtained by calculating the difference between the partial derivatives of the meridional wind in the longitude direction and the partial derivatives of the zonal wind in the latitude direction.

[0067] The wind field divergence weight ratio coefficient is obtained by multiplying the ratio of the wind field divergence to the divergence reference value by a first weighting coefficient. The vorticity inverse number weight ratio coefficient is obtained by multiplying the ratio of the vorticity inverse number to the vorticity reference value by a second weighting coefficient. The wind field divergence weight ratio coefficient and the vorticity inverse number weight ratio coefficient are added together to obtain the final comprehensive judgment criterion.

[0068] Input meteorological data, including geopotential height data, meridional wind data, and zonal wind data. Geopotential height data represents the potential energy of air particles at a certain pressure surface in the atmosphere; meridional wind data represents the wind speed along the meridian; and zonal wind data represents the wind speed along the latitudinal direction. For example, at the 500 hPa pressure layer, the geopotential height data range for a certain region is 5520-5880 geopotential meters, the meridional wind data range is -20 to 20 m / s, and the zonal wind data range is -15 to 25 m / s.

[0069] A point-by-point tracing method is used to obtain the contour point set. Specifically, a starting point is selected, for example, a point with a geopotential height of 5760 geopotential meters, and then points with the same geopotential height value are searched within its eight neighborhoods. To improve efficiency, a bisection method can be used to interpolate between adjacent grid points to find the precise location of the contour point. This process is repeated until no new contour points can be found, thus obtaining a contour point set.

[0070] Calculate the distance between the first and last contour points in a contour point set. Assuming the coordinates of the first and last points are (lon1, lat1) and (lon2, lat2) respectively, the actual distance between the two points can be calculated using the spherical distance formula. In practical applications, when this distance is less than a preset closure threshold, such as less than 50 kilometers, the contour point set is considered to constitute a closed contour line.

[0071] After obtaining a closed contour line, it is necessary to determine the distribution of geopotential height values ​​within its enclosed area. Using a scanning or filling algorithm, all grid points within the closed contour line are identified, and the geopotential height values ​​of these points are extracted. For example, the average geopotential height within the enclosed area might be 5800 meters, while the average geopotential height of the surrounding area (defined as a range extending 100 kilometers beyond the enclosed area) might be 5760 meters. The difference between these two values ​​is calculated, yielding a height difference of 40 meters. If this difference exceeds a preset threshold, such as 30 meters, the closed contour line is considered to meet the characteristics of high-pressure geopotential height.

[0072] Acquire wind field data for the enclosed region, including zonal and meridional wind data. For each grid point within the region, extract its zonal and meridional wind values. For example, a certain enclosed region may have 100 grid points, each with corresponding zonal and meridional wind values.

[0073] Calculate the wind field divergence of the enclosed region. Wind field divergence represents the degree of convergence or divergence of air per unit volume; a positive value indicates divergence, and a negative value indicates convergence. The specific calculation method is to calculate the sum of the partial derivatives of the zonal wind in the longitude direction and the partial derivatives of the meridional wind in the latitudinal direction for all grid points within the region.

[0074] The partial derivatives are approximated using the central difference method. For grid point (i,j), the partial derivative of zonal wind u in the longitude direction can be approximated as [u(i+1,j)-u(i-1,j)] / (2×dlon), where dlon is the grid spacing in the longitude direction. Similarly, the partial derivative of meridional wind v in the latitudinal direction is calculated as [v(i,j+1)-v(i,j-1)] / (2×dlat), where dlat is the distance between adjacent grid points in the latitudinal direction. These are then summed to obtain the divergence at that point. Finally, the average divergence of all grid points in the region is taken as the wind field divergence of the region. Assume the calculated wind field divergence is 1.2×10⁻⁶. -5 Every second.

[0075] Simultaneously, the vorticity of the enclosed region is calculated. Vorticity represents the intensity and direction of air rotation; positive values ​​indicate cyclonic rotation (counterclockwise in the Northern Hemisphere), and negative values ​​indicate anticyclonic rotation (clockwise in the Northern Hemisphere). The calculation method involves calculating the difference between the partial derivative of the meridional wind in the longitude direction and the partial derivative of the zonal wind in the latitude direction for all grid points within the region.

[0076] Using the same central difference method as for divergence calculation, for grid point (i,j), the partial derivative of the meridional wind v in the longitude direction can be approximated as [v(i+1,j)-v(i-1,j)] / (2×dlon), and the partial derivative of the zonal wind u in the latitudinal direction can be approximated as [u(i,j+1)-u(i,j-1)] / (2×dlat). Then, the former is calculated, and the latter is subtracted to obtain the vorticity at that point. The average vorticity of all grid points within the region is taken as the vorticity of the region. Assume the calculated vorticity is -2.0×10⁻⁶. -5 Every second.

[0077] The wind field divergence weighting ratio coefficient is obtained by multiplying the ratio of the wind field divergence to the divergence reference value by the first weighting coefficient. The divergence reference value can be determined based on historical data analysis, for example, 1.0 × 10⁻⁶. -5 The first weighting factor is 0.6 per second. Therefore, the wind field divergence weighting ratio factor is:

[0078] (1.2×10 -5 / 1.0×10 -5 )×0.6=0.72.

[0079] The weighted ratio coefficient for the inverse of vorticity is obtained by multiplying the ratio of the inverse of vorticity to the reference value of vorticity by a second weighting coefficient. The reference value of vorticity can be taken as 1.5 × 10⁻⁶. -5 The second weighting factor is 0.4 per second. The negative of vorticity is 2.0 × 10⁻⁶. -5 per second (note that the original vorticity is negative, and the opposite number becomes positive), then the weighting ratio coefficient of the vorticity inverse number is (2.0 × 10). -5 / 1.5×10 -5 )×0.4=0.53.

[0080] Adding the weighting coefficients of wind field divergence and inverse vorticity, we obtain the final comprehensive judgment criterion: 0.72 + 0.53 = 1.25. A judgment threshold, such as 1.0, is set. When the comprehensive judgment criterion is greater than this threshold, the closed contour line region is determined to be a high-pressure system. In this example, 1.25 > 1.0, therefore the region is determined to be a high-pressure system.

[0081] By combining the above methods with geopotential height characteristics, wind field divergence, and vorticity characteristics, high-pressure systems in the atmosphere can be accurately identified, providing important basis for weather forecasting.

[0082] In one optional implementation, a point-by-point tracking method is used to obtain a set of contour points. The distance between the first and last contour points in the set is calculated. When the distance is less than a preset closure threshold, it is determined that the set of contour points constitutes a closed contour line.

[0083] The study area is divided into regular grids, and the grid point spacing is calculated based on the regular grids. The grid point spacing includes the spacing in the longitude direction and the spacing in the latitude direction.

[0084] Search for grid points in the regular grid that meet the tracking start point conditions. The tracking start point conditions include: the absolute value of the difference between the geopotential height value of the grid point and the target contour height value is less than a preset distance threshold, and the geopotential height values ​​of the grid point and its adjacent grid points are located on both sides of the target contour height value.

[0085] Based on the tracking starting point condition, linear interpolation calculation is performed between adjacent grid points to obtain the precise coordinates of the contour points, wherein the precise coordinates are calculated by the ratio of the geopotential height value of the adjacent grid points to the height value of the target contour line.

[0086] Calculate the distance between the first and last two isoline points in the isoline point set and the total length of the two isoline points. Multiply the ratio of the distance value to the total length value by a first weighting coefficient to obtain the distance-length ratio weighting coefficient. Multiply the ratio of the tangential angle difference between the two isoline points by a second weighting coefficient to obtain the tangential angle difference ratio weighting coefficient. Add the distance-length ratio weighting coefficient and the tangential angle difference ratio weighting coefficient to obtain the closure evaluation index.

[0087] Based on the closure evaluation index, it is determined whether the contour point set is closed. When the closure evaluation index is less than the preset index threshold, the contour point set is determined to be a closed contour line.

[0088] The study area was divided into a regular grid. In this embodiment, the study area is the geographical range of 30°-40° North latitude and 110°-120° East longitude. This area was divided into a 120×120 grid array, and the grid point spacing was calculated, including a longitude spacing of approximately 0.083° and a latitude spacing of approximately 0.083°.

[0089] The system searches for grid points that meet the tracking start point conditions within a regular grid. In this embodiment, the target contour height is set to 5880 gpm (potential meters), and the preset distance threshold is 0.5 gpm. The search process is as follows: traverse each point in the grid, determine whether the absolute value of the difference between the point's potential height and the target contour height of 5880 gpm is less than the preset distance threshold of 0.5 gpm, and simultaneously determine whether the potential height values ​​of the grid point and its adjacent grid points are located on opposite sides of the target contour height. For example, if the potential height of a grid point is 5879.8 gpm, and the potential height of its eastern adjacent grid point is 5880.3 gpm, then these two grid points are located on opposite sides of the target contour height of 5880 gpm, and the absolute value of the difference between 5879.8 and 5880 is 0.2 gpm, which is less than the preset distance threshold of 0.5 gpm. Therefore, this grid point meets the tracking start point conditions.

[0090] Based on the starting point condition, linear interpolation is performed between adjacent grid points to obtain the precise coordinates of the contour points. In the example above, assume the latitude and longitude coordinates of the grid point are (115.00°E, 35.00°N), and the latitude and longitude coordinates of its eastern adjacent grid point are (115.083°E, 35.00°N), with geopotential height values ​​of 5879.8 gpm and 5880.3 gpm respectively. Through linear interpolation, the precise latitude and longitude coordinates corresponding to the target contour height value of 5880 gpm are obtained as follows:

[0091] Longitude coordinates = 115.00 + (5880 - 5879.8) / (5880.3 - 5879.8) × 0.083 = 115.0332°E; Latitude coordinates remain unchanged at 35.00°N; Therefore, the precise coordinates of this contour line point are (115.0332°E, 35.00°N).

[0092] Starting from the starting point, trace along the contour lines, selecting the next grid edge each time and repeating the interpolation calculation process to sequentially obtain the point set on the contour lines. Assume that 100 contour points are obtained during the tracing process, forming a contour point set {P1, P2...P...}. 100}

[0093] To determine whether a contour line is closed, it is necessary to calculate the distance between the first and last points on the contour line and the total length. Assume the first and last points are P1 (115.041°E, 35.00°N) and P... 100 (115.035°E, 35.008°N), the distance between the two points is calculated to be 1.2 km using spherical distance calculation. The total length of the contour lines is the sum of the distances between all adjacent points; let's assume the calculated total length is 850 km.

[0094] Calculate the distance length ratio weighting coefficient. Multiply the ratio of distance value to total length value, 1.2 / 850≈0.0014, by the first weighting coefficient, 0.7, to obtain the distance length ratio weighting coefficient as 0.00098.

[0095] Calculate the weighting coefficient for the tangential angle difference ratio. Calculate the tangential angle based on the direction of the lines connecting the first and last points to their respective adjacent points. Assume the tangential angle of P1 is 85°. 100 If the tangential angle is 92°, then the tangential angle difference is |92° - 85°| = 7°. The ratio of the tangential angle difference is 7° / 180° ≈ 0.039. Multiplying this by the second weighting coefficient of 0.3, we get a weighting coefficient of 0.0117 for the tangential angle difference ratio.

[0096] Adding the weighting coefficients of distance length ratio (0.00098) and tangential angle difference ratio (0.0117), we get a closure evaluation index of 0.01268.

[0097] Assuming the preset threshold is 0.05, since the closure evaluation index of 0.01268 is less than the preset threshold of 0.05, the set of contour points is determined to constitute a closed contour line.

[0098] In practical applications, the preset distance threshold, weighting coefficient, and preset index threshold can be adjusted according to the characteristics of different research areas. For example, for mountainous areas with complex terrain, the preset distance threshold can be appropriately increased; for application scenarios that particularly focus on closure, the proportion of the first weighting coefficient can be increased.

[0099] In this embodiment, the sum of the first weighting coefficient and the second weighting coefficient is 1, with values ​​of 0.7 and 0.3 respectively. The preset index threshold ranges from 0.01 to 0.1, and is adjusted according to the closure requirements of the specific application scenario. Practice shows that when the preset index threshold is 0.05, it can better balance the recognition accuracy and recall of closed contour lines.

[0100] The above technical solutions enable the accurate identification and extraction of closed contour lines in meteorological fields, providing effective support for further analysis of high and low pressure systems. For example, based on the identification of the 5880 gpm isoline closure, it is possible to further determine whether the closed area is a subtropical high-pressure system, providing a scientific basis for weather forecasting and climate analysis.

[0101] In one optional implementation, a residual convolutional neural network model is built using the ResNet framework, and the performance evaluation and parameter optimization of the residual convolutional neural network model are performed using accuracy and cross-entropy, including:

[0102] A residual convolutional neural network model is established based on the ResNet framework, which includes a shallow feature extraction part and a deep feature learning part.

[0103] The residual convolutional neural network model is trained and optimized by using the cross-entropy loss function to calculate the model loss and by using the Adam optimizer to optimize the model parameters.

[0104] A residual convolutional neural network model is established based on the ResNet framework. This model includes a shallow feature extraction section and a deep feature learning section. The shallow feature extraction section consists of a 7×7 convolutional layer and a max-pooling layer, used to initially extract the basic features of the input data. The deep feature learning section consists of multiple residual blocks, each containing two 3×3 convolutional layers and a shortcut connection, to achieve deep and residual feature learning.

[0105] In practical implementation, the specific structure of the shallow feature extraction part is as follows: The input layer receives meteorological image data of size 512×512×3. The first convolutional layer uses 64 7×7 convolutional kernels with a stride of 2 and padding of 3. After batch normalization and ReLU activation, the output feature map size is 256×256×64. Then, a 3×3 max pooling layer with a stride of 2 and padding of 1 reduces the feature map size to 128×128×64.

[0106] The deep feature learning part consists of four groups of residual blocks, each containing a different number of residual blocks. The first group contains three residual blocks, each with two convolutional layers and 64 kernels; the second group contains four residual blocks, each with two convolutional layers and 128 kernels; the third group contains six residual blocks, each with two convolutional layers and 256 kernels; and the fourth group contains three residual blocks, each with two convolutional layers and 512 kernels. Each group of residual blocks is separated by convolutional layers with a stride of 2 to reduce the feature map size and increase the number of channels.

[0107] Within each residual block, the specific structure is as follows: First, it passes through a 3×3 convolutional layer with a stride of 1 and padding of 1, followed by batch normalization and ReLU activation. Next, it passes through a second 3×3 convolutional layer with a stride of 1 and padding of 1, followed by batch normalization again. Finally, the input feature map is directly added to the output of the second convolutional layer via a shortcut connection, and the ReLU activation function is used to obtain the output of the residual block. If the dimensions of the input and output do not match, a 1×1 convolutional layer is used to transform the dimensions of the input.

[0108] Finally, the deep feature learning part converts the feature map into a vector through a global average pooling layer, then through a fully connected layer with 1000 neurons, and finally through a fully connected layer with 2 neurons to output the results, representing the probability of the sample being a positive example (the Greenland High appears) and a negative example (the Greenland High does not appear), respectively.

[0109] The residual convolutional neural network model was trained and optimized. First, a training dataset was prepared, including labeled meteorological image data, where positive examples were those showing the Greenland High and negative examples were those where the Greenland High was not present. In practical application, this implementation used 10 years (2010-2019) of ERA5 reanalysis data, extracting the 500 hPa geopotential height field and sea-level pressure field as input features, collecting a total of 3650 samples, including 1825 positive examples and 1825 negative examples.

[0110] During training, a mini-batch stochastic gradient descent method with a batch size of 32 was used. The initial learning rate was set to 0.001, and a learning rate decay strategy was adopted, reducing the learning rate to 0.1 times its original value every 30 epochs. The training epochs were set to 100 epochs, and cross-entropy was used as the loss function.

[0111] The cross-entropy loss function measures the difference between the model's prediction and the true label. For each sample, the label value is 1 if the true label is positive (samples showing the Greenland High), and 0 if the true label is negative (samples not showing the Greenland High). The model's prediction represents the probability that a sample is predicted to be positive. The smaller the cross-entropy loss value, the closer the model's prediction is to the true label.

[0112] During the optimization process, the Adam optimizer is used to optimize the model parameters. The Adam optimizer combines the advantages of momentum gradient descent and RMSProp, and can adaptively adjust the learning rate of each parameter. The parameters of the Adam optimizer are set as follows: betas = (0.9, 0.999), epsilon = 1e-8, and weights = 0.999. decay It is 1e-4.

[0113] During training, model performance is evaluated using a validation set after each epoch. The validation set consists of 20% of the total samples, or 730 samples. Evaluation metrics include accuracy and cross-entropy loss. Accuracy is calculated as the number of correctly classified samples divided by the total number of samples.

[0114] By comparing the validation set accuracy and cross-entropy loss values ​​across different periods, the model parameters with the best performance were selected as the final model. In this implementation, the model achieved its optimal performance in the 78th period, with a validation set accuracy of 94.25% and a cross-entropy loss value of 0.1587.

[0115] The final model was evaluated using a test set consisting of 365 samples from 2020 that were not used in training. The test results showed that the model achieved an accuracy of 92.88% on the test set, indicating that the constructed residual convolutional neural network model has good generalization ability and can effectively identify the occurrence of the Greenland High.

[0116] Through the above steps, a residual convolutional neural network model based on the ResNet framework was successfully constructed and optimized, enabling effective identification of the Greenland High.

[0117] In one optional implementation, based on the determination result of the residual convolutional neural network model, the outermost closed contour line information of the high pressure is extracted, and the highest potential point within the outermost closed contour line of the high pressure is determined as the center position of the high pressure. The highest potential anomaly is used as the high pressure intensity, and the high pressure identification result includes:

[0118] Based on the determination result of the residual convolutional neural network model, it is determined whether the distance between the first and last points of the contour point set in the geopotential height data is less than a preset closure threshold. When the distance is less than the preset closure threshold, the contour point set is determined as the outermost closed contour line of the high pressure.

[0119] Obtain the boundary point coordinate information of the outermost closed contour line of the high voltage system, construct a closed region of the high voltage system based on the boundary point coordinate information, traverse the geopotential height value of each grid point within the closed region, and determine the grid point with the largest geopotential height value as the center point of the high voltage system.

[0120] The geopotential anomaly is obtained by subtracting the geopotential climate average value of the center location from the geopotential height value of the center location.

[0121] The coordinates of the central location point, the geopotential anomaly, and the coordinates of the boundary points of the outermost closed contour line of the high-pressure system are used as the quantitative feature information of the high-pressure system. The high-pressure identification result is output based on the quantitative feature information.

[0122] During the model training phase, a training dataset was constructed, containing 10,000 pre-labeled images, including 5,000 positive samples and 5,000 negative samples. Each image was uniformly resized to 224×224 pixels and stored in RGB three-channel format. Image preprocessing steps included: cropping the original images and uniformly resizing images of different sizes to the target size; applying random horizontal flipping to enhance data diversity, with a flipping probability set to 0.5; and performing standardization to transform pixel values ​​from the range [0,255] to the interval [-1,1], specifically by dividing the pixel value by 127.5 and then subtracting 1.

[0123] Design a convolutional neural network architecture consisting of an input layer, a feature extraction layer, and a classification layer. The input layer receives 224×224×3 image data. The feature extraction layer contains five convolutional blocks, each consisting of two convolutional layers and one max-pooling layer. The first convolutional block uses 64 3×3 convolutional kernels with a stride of 1 and padding of 1; the second convolutional block uses 128 3×3 convolutional kernels; the third convolutional block uses 256 3×3 convolutional kernels; and the fourth and fifth convolutional blocks each use 512 3×3 convolutional kernels. Each convolutional layer is followed by a ReLU activation function, and the max-pooling layer has a pooling window size of 2×2 and a stride of 2. The classification layer consists of three fully connected layers with 4096, 4096, and 2 neurons respectively, and the last layer uses the Softmax function to output the class probability.

[0124] The cross-entropy loss function is used to calculate the difference between the predicted result and the true label. Specifically, for each sample, the cross-entropy between the predicted probability output by the network and the true label is calculated, and then the average value is taken over the entire batch as the final loss. The Adam algorithm is selected as the optimizer, with an initial learning rate of 0.001, which decays to 0.1 times its original value every 50 epochs. The batch size is set to 64, and the total number of training epochs is 200.

[0125] To prevent overfitting, the following regularization strategies were implemented: Dropout was introduced into the fully connected layers with a dropout rate of 0.5; L2 weight regularization was applied with a coefficient of 0.0001; and an early stopping strategy was implemented, stopping training when the validation set loss did not improve for 10 consecutive epochs. With these measures, the model achieved an accuracy of 95.8% on the validation set.

[0126] In the image recognition stage, the pre-trained model parameters are loaded first. When the image to be recognized is input, the same preprocessing steps as in the training stage are performed, including resizing and normalization. The preprocessed image is then fed into the model for forward propagation calculation. After passing through all convolutional and fully connected layers, a probability vector containing two elements is obtained, representing the probability that the image belongs to the positive and negative classes, respectively.

[0127] To improve recognition reliability, a confidence threshold mechanism is introduced. A classification result is only output when the predicted highest class probability exceeds 0.85; otherwise, it is marked as "uncertain." In practical applications, this threshold can be adjusted according to specific scenario requirements. A higher threshold can be set for scenarios with high accuracy requirements, while a lower threshold can be appropriately set for scenarios with high recall requirements.

[0128] For actual production environments, a model optimization and deployment strategy was also implemented. First, model compression was performed, converting 32-bit floating-point numbers to 8-bit integers through weight quantization, reducing the model size from 98MB to 26MB and improving inference speed by 40%. At the same time, knowledge distillation technology was adopted, using the original model as the teacher network to train a lightweight student network. The number of parameters in the student network is only 30% of that in the original model, but it still maintains a recognition accuracy of 93.2%.

[0129] The system's overall architecture adopts a front-end and back-end separation design. The front-end is responsible for image acquisition and result display, while the back-end is responsible for model inference and data management. The two communicate through a RESTful API; the front-end sends Base64-encoded image data, and the back-end returns recognition results in JSON format. To improve processing capacity, the back-end employs an asynchronous processing mechanism, using a task queue to manage a large number of concurrent requests. A single server instance can handle up to 200 image recognition requests per second.

[0130] In a real-world application case, this method was deployed in a manufacturing company's product quality inspection system. Within a week, 50,000 products were inspected, and 952 defective products were accurately identified, including 728 common defects and 224 rare defects. The false positive rate was controlled within 1.3%, which significantly improved the efficiency and accuracy of quality inspection.

[0131] Through the above technologies, this method achieves efficient image processing and flexible deployment capabilities while ensuring recognition accuracy, making it adaptable to image recognition needs in different scenarios.

[0132] In one optional implementation, based on the determination result of the residual convolutional neural network model, the outermost closed contour line information of the high pressure is extracted, and the highest potential point within the outermost closed contour line of the high pressure is determined as the center position of the high pressure. The highest potential anomaly is used as the high pressure intensity, and the high pressure identification result includes:

[0133] Based on the determination result of the residual convolutional neural network model, it is determined whether the distance between the first and last points of the contour point set in the geopotential height data is less than a preset closure threshold. When the distance is less than the preset closure threshold, the contour point set is determined as the outermost closed contour line of the high pressure.

[0134] Obtain the boundary point coordinate information of the outermost closed contour line of the high voltage system, construct a closed region of the high voltage system based on the boundary point coordinate information, traverse the geopotential height value of each grid point within the closed region, and determine the grid point with the largest geopotential height value as the center point of the high voltage system.

[0135] The geopotential height values ​​of all grid points within the closed area are statistically analyzed, and the geopotential anomaly value is obtained by subtracting the geopotential climate average value of the central location point from the geopotential height value of the central location point.

[0136] The coordinates of the central location point, the geopotential anomaly, and the coordinates of the boundary points of the outermost closed contour line of the high-pressure system are used as the quantitative feature information of the high-pressure system. The high-pressure identification result is output based on the quantitative feature information.

[0137] A pre-trained residual convolutional neural network model is used to determine the input geopotential height data, thereby obtaining a coarse approximate region of the high-pressure system. The residual convolutional neural network model extracts features of the geopotential height field through multiple convolutions, effectively avoids the gradient vanishing problem by utilizing residual connections, and improves network depth and performance, ultimately outputting the region where the high-pressure system may exist.

[0138] Based on the determination results of the residual convolutional neural network, the outermost closed contour lines of the high-pressure system are extracted. Specifically, a contour line extraction algorithm, such as the CONREC algorithm, is applied to the geopotential height data to extract contour point sets at 10 geopotential meters intervals. For each extracted contour point set, it is determined whether the distance between its first and last points is less than a preset closure threshold. In practical applications, the preset closure threshold can be set to 0.5 grid units; for example, in 0.5°×0.5° resolution data, the closure threshold is 0.25°. When the distance between the first and last points of a contour point set is less than the preset closure threshold, the contour line is considered closed and identified as a candidate closed contour line.

[0139] Among all candidate closed contour lines, the closed contour line with the smallest potential value is selected as the outermost closed contour line of the high pressure system. For example, when three closed contour lines with potential values ​​of 5880 potential meters, 5890 potential meters, and 5900 potential meters are detected, the contour line with a potential value of 5880 potential meters is selected as the outermost closed contour line of the high pressure system.

[0140] Obtain the coordinates of the boundary points of the outermost closed contour lines of the high-pressure system. This information is stored in latitude and longitude format, for example, [(120.5°E, 30.5°N), (121.0°E, 30.7°N)...(120.4°E, 30.3°N)]. Based on these boundary point coordinates, construct the closed region of the high-pressure system. The construction of the closed region can be achieved using the ray casting method or a point-within-a-polygon algorithm. Connect the boundary points to form a polygon, and determine whether a grid point is inside the polygon.

[0141] Within a defined closed region, the geopotential height values ​​of each grid point are iterated. For example, in data with a resolution of 0.5° × 0.5°, if the closed region covers a longitude range of 115°E to 125°E and a latitude range of 25°N to 35°N, then all grid points within this range need to be traversed. By comparing the geopotential height values ​​of each grid point, the grid point with the highest geopotential height value is determined as the center point of the high-pressure system. For example, after the traversal is completed, if the grid point located at (120.5°E, 32.5°N) is found to have a geopotential height value of 5925 geopotential meters, which is the maximum value within the region, then this point is determined as the center point of the high-pressure system.

[0142] Extract the geopotential-climatic average value at the high-pressure center location. Assuming the geopotential-climatic average value at the center location is 5885 geopotential meters, subtract the geopotential-climatic average value from the geopotential height value at the center location to obtain the geopotential anomaly. In the example above, the geopotential anomaly is 5925 - 5885 = 40 geopotential meters, which represents the intensity of the high-pressure system.

[0143] The coordinates (longitude and latitude) of the central location point, the geopotential anomaly, and the coordinates of the boundary points of the outermost closed contour lines of the high-pressure system are used as quantitative characteristic information of the high-pressure system. Based on these quantitative characteristic information, the high-pressure identification result is output. The output result can be expressed as: the high-pressure center location is (120.5°E, 32.5°N), the intensity is 40 geopotential meters, and the area of ​​influence is the region enclosed by the boundary point set [(120.5°E, 30.5°N), (121.0°E, 30.7°N)...(120.4°E, 30.3°N)].

[0144] In practical applications, a geopotential anomaly threshold can be set to filter weak high-pressure systems. For example, when the geopotential anomaly is less than 20 geopotential meters, the high-pressure system can be considered weak and not included in the final identification result; when the geopotential anomaly is greater than or equal to 20 geopotential meters, the identification result of the high-pressure system is retained. This can reduce the interference of weak high-pressure systems on weather forecasts and improve the accuracy of forecasts.

[0145] The above method can be applied to atmospheric geopotential height field data with different horizontal resolutions, such as ERA5 reanalysis data (0.25°×0.25°) and NCEP reanalysis data (2.5°×2.5°). By adjusting the corresponding parameters (such as closure threshold and contour interval), the accuracy and applicability of the identification results can be ensured.

[0146] In one optional implementation, a closed region of the high-voltage system is constructed based on the boundary point coordinate information. Within this closed region, the potential height value of each grid point is traversed, and the grid point with the largest potential height value is determined as the center point of the high-voltage system. This includes:

[0147] Obtain the coordinate sequence of boundary points of the high-voltage system. The boundary points in the coordinate sequence are sorted according to spatial continuity. Construct a region indicator function based on the coordinate sequence. The region indicator function is used to mark whether the grid points are located within the closed region enclosed by the boundary points.

[0148] Based on the region indicator function, the set of grid points located within the closed region is filtered, and the position coordinates of each grid point in the set of grid points are mapped to the potential height field function to obtain the potential height value of all grid points within the closed region.

[0149] For each grid point within the closed area, a local maximum value determination is performed. The local maximum value determination includes: calculating the potential height difference between the target grid point and its surrounding grid points; when the potential height value of the target grid point is greater than the potential height values ​​of all surrounding grid points, the target grid point is marked as a local maximum value point.

[0150] Among the local maximum points, determine the grid point corresponding to the global maximum geopotential height value, and determine the coordinates of the grid point corresponding to the global maximum geopotential height value as the center location point of the high-voltage system.

[0151] Obtain the coordinate sequence of boundary points of the high-voltage system, which are ordered according to spatial continuity. For example, for a high-voltage system, its boundary point coordinate sequence can be represented as {(120.5°E, 30.2°N), (121.0°E, 30.5°N), (121.5°E, 31.0°N)...(120.0°E, 30.0°N)}. These coordinate points are continuously distributed along the boundary of the high-voltage system, with the last point usually adjacent to the first point, forming a closed boundary.

[0152] Based on the above coordinate sequence, the system constructs a region indicator function. This function is used to indicate whether a grid point is located within a closed region enclosed by boundary points. In practice, the system first establishes a grid in the region to be analyzed, with a grid resolution of 0.25° × 0.25°. For each grid point, the system determines whether it is located within the closed region. The determination method uses a ray casting: a ray is cast from the grid point to be determined in any direction (e.g., due east), and the number of intersections between the ray and the boundary is calculated. If the number of intersections is odd, the point is located within the closed region; if it is even, it is located outside the region.

[0153] For the grid point (121.25°E, 30.75°N), the system emits a ray eastward and calculates that it intersects the boundary at 3 points, which is an odd number. Therefore, this point is located within the closed region, and the region indicator function value is 1. For the grid point (119.5°E, 29.5°N), the system calculates that it intersects the boundary at 0 points, which is an even number. Therefore, this point is located outside the closed region, and the region indicator function value is 0.

[0154] The system filters the set of grid points located within a closed region based on a region indicator function. Specifically, it iterates through all grid points and adds those with a region indicator function value of 1 to the set. For example, the filtered set of grid points might contain points such as {(120.25°E, 30.25°N), (120.5°E, 30.5°N), (120.75°E, 30.75°N)...(121.25°E, 31.0°N)}.

[0155] The geopotential height values ​​of all grid points within a closed region are obtained by mapping the location coordinates of each grid point in the grid set to a geopotential height field function. The geopotential height field function can be pre-stored meteorological data, such as the geopotential height field of a 500 hPa pressure surface. For example, for a grid point (120.25°E, 30.25°N), its geopotential height value might be 5880 meters; for a grid point (120.5°E, 30.5°N), its geopotential height value might be 5885 meters.

[0156] For each grid point within the closed area, a local maximum value is determined. For each target grid point, the system calculates the difference in geopotential height between it and its eight surrounding adjacent grid points. When the geopotential height value of the target grid point is greater than the geopotential height values ​​of all surrounding grid points, the target grid point is marked as a local maximum point.

[0157] For example, for the grid point (120.75°E, 30.5°N), its geopotential height is 5890 meters. The geopotential heights of its eight neighboring points are as follows: (120.5°E, 30.25°N) 5882 meters, (120.75°E, 30.25°N) 5885 meters, (121.0°E, 30.25°N) 5883 meters, (120.5°E, 30.5°N) 5885 meters, (121.0°E, 30.5°N) 5884 meters, (120.5°E, 30.75°N) 5882 meters, (120.75°E, 30.75°N) 5887 meters, and (121.0°E, 30.75°N) 5880 meters. Since the geopotential height value of the target grid point, 5890 meters, is greater than the geopotential height values ​​of all its neighboring points, this point is marked as a local maximum point.

[0158] Through the above process, the system may identify multiple local maximum points, such as {(120.75°E, 30.5°N, 5890 m), (121.25°E, 31.0°N, 5888 m), (120.25°E, 30.75°N, 5886 m)}. The system then determines the grid point corresponding to the global maximum geopotential height from these local maximum points. In this example, the geopotential height of the grid point (120.75°E, 30.5°N) is the maximum at 5890 meters; therefore, the system determines the coordinates of this grid point (120.75°E, 30.5°N) as the center location of the high-voltage system.

[0159] To improve positioning accuracy, the system can further refine the center position using interpolation methods after determining the coarse center position. Specifically, the system can construct a finer grid near the global maximum point, such as with a resolution of 0.05°×0.05°, and then calculate the potential height values ​​of these fine grid points through bilinear interpolation to ultimately determine the more accurate coordinates of the center position point.

[0160] Through the above steps, the system can accurately determine the center location of the high-pressure system, providing important data for subsequent weather forecasting and meteorological analysis. This method combines boundary identification, region judgment, and local maximum value detection techniques, exhibiting high accuracy and practicality.

[0161] like Figures 2-5 As shown, the method further includes:

[0162] Taking the case of January 2, 2021 as an example ( Figure 2 The 500 hPa geopotential height field shows that a blocking high pressure exists in the Wushan area. Simultaneously, there is also a high pressure near Novaya Zemlya, centered at 75°N latitude, with a longitude similar to the blocking high pressure at Wushan. Figure 2 a). The superposition of a high-pressure system and a blocking high-pressure system near Novaya Zemlya creates an anticyclone, allowing cold air from high latitudes to be transported to mid-latitude regions. Figure 2 b). According to Figure 2 a and Figure 2 b. High pressure characteristics: High pressure center in geopotential height field and anticyclonic circulation in wind field. The criteria for determining high pressure are: 1) At least one closed contour line is analyzed in the daily 500 hPa geopotential height field; 2) The wind field shows a complete anticyclonic circulation.

[0163] The search results for this case were obtained using the two-dimensional blocking high-pressure index. Figure 2 As shown in c. The location and extent of the mid-latitude blocking high pressure A are related to... Figure 2 The results for circulation field a correspond to the actual location of the high-pressure center B near Novaya Zemlya. The center of the high-pressure system B is located approximately 10 degrees east of the actual location, and the portion north of 75°N is missing. The high-pressure system C does not exist in the actual circulation field and is a spurious result. This indicates that the two-dimensional high-pressure system retrieval method is prone to omissions and spurious results in identifying high-pressure systems in high-latitude regions, thus requiring a more accurate high-pressure system identification method.

[0164] Figure 2 Figure d shows the class activation map (CMA) of the residual convolutional neural network (RCN) recognition model when identifying this specific case, visually explaining the model's decision-making process in the image classification task. Initially, the high-pressure area is brighter than the low-pressure area, and the model can effectively distinguish between the high and low pressure ranges. Subsequently, the model focuses more on the high-pressure and anticyclonic circulation regions, eventually leaving only the brightness of the area near Novaya Zemlya. This means that the model can determine the presence of high pressure by comparing and analyzing features such as the location of the pressure center, pressure gradient, wind intensity, and direction. This further demonstrates the feasibility of using residual convolutional neural networks for high-pressure identification.

[0165] After determining the high-voltage judgment criteria and verifying the necessity and feasibility of applying residual convolutional neural networks for high-voltage identification, the model for high-voltage identification using residual convolutional neural networks was established. For example... Figure 3 As shown, the main basic processes include input layer data preparation and processing, and model building.

[0166] (1) Input layer data preparation and processing

[0167] The ResNet model framework is applicable to high-pressure systems at different latitudes and in different regions, but the input data depends on the target being identified. Taking the identification of the Greenland High as an example, the data selected covers the period from 1979 to 2023, with a spatial range of 50°–90°N and 0°–100°W. To improve model performance, the normalized 500 hPa geopotential height field and wind field variables are merged into a three-channel dataset. Compared to a single-channel dataset, the model can learn based on the relationships between multiple channels. Furthermore, after determining whether the Greenland High occurs daily using the criteria established in the previous step, this determination is used as the category label for supervised learning and input into the model along with the three-channel dataset. 80% of the data is used as the training set, and 20% is used as the test set.

[0168] (2) Model building

[0169] The model employs the ResNet_34 framework. It includes an initial convolutional layer, a max-pooling layer, four convolutional groups with residual block structures, a global average pooling layer, and fully connected layers. The initial convolutional layer extracts features from the input, while the pooling layers progressively reduce the size of the feature maps, preserving important features while reducing computational cost. The residual block is the core structure of the ResNet model; through cross-layer connections and residual learning mechanisms, it effectively trains deep neural networks, thereby better capturing the complex and abstract features of the Greenland High. The fully connected layers use the sigmoid activation function to convert the model's output into probabilities, ultimately determining whether the Greenland High has occurred.

[0170] After the model is built, accuracy and cross-entropy are used to evaluate the performance of the ResNet model. Accuracy is defined as the percentage of correctly identified samples out of the total number of samples. Cross-entropy measures the difference between the true probability distribution and the predicted probability distribution. In binary classification tasks, the expression for the cross-entropy loss function is:

[0171]

[0172] Where N represents the total number of samples in the dataset (short for Number), for example, if there are 1000 training samples, then N=1000; Indicates sample The labels are 1 for positive examples (samples that show the Greenland High) and 0 for negative examples (samples that do not show the Greenland High); Indicates sample The probability of predicting a positive example; "#(1)" indicates the formula number, used to reference or identify the formula in the document, here indicating that this is formula number 1. The smaller the cross-entropy loss function value, the better the model prediction effect.

[0173] Many studies have used the mean 500 hPa geopotential height (GBI) of the Greenland region to explore the interannual and interdecadal variations of the Greenland High characteristics. JRPreece et al. used cosine latitude to weight the GBI and identified the Greenland High by standardization and thresholding. By comparing the identification results of this method (hereinafter referred to as the GBI method) with the judgment criteria and the method of the present invention (hereinafter referred to as the ResNet method), the performance and effectiveness of the present invention can be further verified.

[0174] Using the model's determination results, the 500 hPa geopotential height field data of the day the Greenland High occurred was extracted, and contour maps were drawn at 10 gpm intervals to obtain the coordinates, i.e., latitude and longitude information, of the outermost closed contour lines of the high pressure. The highest geopotential point within the outermost closed contour lines is the center of the high pressure, and the highest geopotential anomaly is the intensity of the high pressure.

[0175] The outermost closed contour lines spanning more than 10 degrees of longitude and 5 degrees of latitude of the Greenland High were extracted. Results showed that the Greenland High identified using the criteria was 2471 days, the ResNet method identified 2495 days, and the GBI method identified 2397 days. Hit rate, false alarm rate, missed alarm rate, and TS score were calculated.

[0176] Hit rate: ;

[0177] False alarm rate: ;

[0178] Missed report rate: ;

[0179] score: ;

[0180] In the formula, To predict the correct number of times, For the number of empty reports, This represents the number of times the report was missed.

[0181] The GBI method cannot provide the center location of the high pressure system. Since some of the results identified by the GBI method do not have closed contour lines, the center location of the results identified by the GBI method is defined as the highest geopotential point in Greenland on that day, and the intensity is the geopotential anomaly.

[0182] like Figure 4As shown, Gaussian density distribution maps of the high-pressure center locations obtained using three methods—the decision criterion, the ResNet method, and the GBI method—are presented. The results indicate that the Greenland High-Pressure Center can occur anywhere in the entire Greenland region. The results from the ResNet method and the decision criterion method are consistent, showing that the Greenland High-Pressure Center mainly occurs in the southeastern part of Greenland, while the results from the GBI method show that the center is mainly concentrated on the southern edge of Greenland. This demonstrates the accuracy of the method used in this invention for identifying the location of the high-pressure center.

[0183] Figure 5 The results show the frequency distribution of the identification results of the three methods in different intensity ranges. The results indicate that the intensity of the results identified by the decision criteria and the ResNet method is mainly concentrated in the range of 65~335 gpm, while the intensity of the results identified by the GBI method is concentrated in the range of 200~470 gpm, which is higher than the actual intensity. This proves that the standardized thresholding method will filter out high pressures with a large range but low intensity, and will also cause misjudgment of strong potentials in the study area without contour closure (strong ridges).

[0184] For those skilled in the art, various corresponding changes and modifications can be made based on the above technical solutions and concepts, and all such changes and modifications should be included within the protection scope of the claims of this invention.

[0185] A second aspect of the present invention provides a high-voltage recognition system based on a residual convolutional neural network, comprising:

[0186] The first unit is used to acquire daily ERA-5 geopotential height data, meridional wind data, and zonal wind data; and to determine the criteria for high pressure based on the geopotential height data, the meridional wind data, and the zonal wind data. The criteria include the analysis of at least one closed contour line in the geopotential height field and the presence of a complete anticyclonic circulation in the wind field.

[0187] The second unit is used to normalize and merge the geopotential height data, the meridional wind data, and the zonal wind data into a three-channel dataset. The judgment result of whether high pressure occurs daily is used as the category label for supervised learning, and the three-channel dataset is divided into a training set and a test set according to a preset ratio.

[0188] The third unit is used to build a residual convolutional neural network model using the ResNet framework, and to evaluate the performance and optimize the parameters of the residual convolutional neural network model using accuracy and cross-entropy. The accuracy is the proportion of correctly identified samples in the total number of samples, and the cross-entropy is used to measure the difference between the true probability distribution and the predicted probability distribution. When the accuracy reaches a preset accuracy threshold and the loss function value of the cross-entropy is less than a preset loss threshold, the iterative training of the residual convolutional neural network model is stopped.

[0189] The fourth unit is used to extract the outermost closed contour line information of the high pressure based on the judgment result of the residual convolutional neural network model, determine the highest potential point within the outermost closed contour line of the high pressure as the center position of the high pressure, and use the highest potential anomaly as the high pressure intensity to obtain the high pressure identification result.

[0190] A third aspect of the present invention provides an electronic device, comprising:

[0191] processor;

[0192] Memory used to store processor-executable instructions;

[0193] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0194] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0195] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0196] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A high-voltage recognition method based on residual convolutional neural networks, characterized in that, include: Obtain daily ERA-5 geopotential height data, meridional wind data, and zonal wind data; The criteria for determining high pressure based on the geopotential height data, the meridional wind data, and the zonal wind data include: A point-by-point tracking method is used to obtain a set of contour points. The distance between the first and last two contour points in the set is calculated. When the distance is less than a preset closure threshold, the set of contour points is determined to constitute a closed contour line. Obtain the geopotential height value distribution of the area enclosed by the closed contour line, calculate the difference between the geopotential height value within the enclosed area and the geopotential height value of the surrounding area, and determine whether the closed contour line satisfies the high-pressure geopotential height characteristics based on the difference; Obtain wind field data for the enclosed area, and calculate the wind field divergence and vorticity of the enclosed area. The wind field divergence is obtained by calculating the sum of the partial derivatives of the zonal wind in the longitude direction and the partial derivatives of the meridional wind in the latitude direction. The vorticity is obtained by calculating the difference between the partial derivatives of the meridional wind in the longitude direction and the partial derivatives of the zonal wind in the latitude direction. The wind field divergence weight ratio coefficient is obtained by multiplying the ratio of the wind field divergence to the divergence reference value by a first weighting coefficient. The vorticity inverse number weight ratio coefficient is obtained by multiplying the ratio of the vorticity inverse number to the vorticity reference value by a second weighting coefficient. The final judgment criterion is obtained by adding the wind field divergence weight ratio coefficient and the vorticity inverse number weight ratio coefficient. The judgment criteria include the analysis of at least one closed contour line in the geopotential height field and the manifestation of a complete anticyclonic circulation in the wind field; The geopotential height data, the meridional wind data, and the zonal wind data are normalized and then merged into a three-channel dataset. The judgment result of whether high pressure occurs daily is used as the category label for supervised learning. The three-channel dataset is divided into a training set and a test set according to a preset ratio. A residual convolutional neural network model is established using the ResNet framework. The performance of the residual convolutional neural network model is evaluated and the parameters are optimized using accuracy and cross-entropy. The accuracy is the proportion of correctly identified samples in the total number of samples, and the cross-entropy is used to measure the difference between the true probability distribution and the predicted probability distribution. When the accuracy reaches a preset accuracy threshold and the loss function value of the cross-entropy is less than a preset loss threshold, the iterative training of the residual convolutional neural network model is stopped. Based on the determination results of the residual convolutional neural network model, the outermost closed contour line information of the high-pressure system is extracted, and the highest potential point within the outermost closed contour line of the high-pressure system is determined as the center position of the high-pressure system. The highest potential anomaly is used as the high-pressure intensity to obtain the high-pressure identification result, including: Based on the determination result of the residual convolutional neural network model, it is determined whether the distance between the first and last points of the contour point set in the geopotential height data is less than a preset closure threshold. When the distance is less than the preset closure threshold, the contour point set is determined as the outermost closed contour line of the high pressure. The coordinate information of the boundary points of the outermost closed contour line of the high-voltage system is obtained. A closed region of the high-voltage system is constructed based on the boundary point coordinate information. The potential height value of each grid point is traversed within the closed region. The grid point with the largest potential height value is determined as the center point of the high-voltage system. The coordinate sequence of the boundary points of the high-voltage system is obtained. The boundary points in the coordinate sequence are sorted according to spatial continuity. A region indicator function is constructed based on the coordinate sequence. The region indicator function is used to mark whether the grid point is located within the closed region enclosed by the boundary points. Based on the region indicator function, the set of grid points located within the closed region is filtered, and the position coordinates of each grid point in the set of grid points are mapped to the potential height field function to obtain the potential height value of all grid points within the closed region. For each grid point within the closed area, a local maximum value determination is performed. The local maximum value determination includes: calculating the potential height difference between the target grid point and its surrounding grid points; when the potential height value of the target grid point is greater than the potential height values ​​of all surrounding grid points, the target grid point is marked as a local maximum value point. Among the local maximum value points, determine the grid point corresponding to the global maximum potential height value, and determine the coordinates of the grid point corresponding to the global maximum potential height value as the center location point of the high voltage system; The geopotential anomaly is obtained by subtracting the geopotential climate average value of the center location from the geopotential height value of the center location. The coordinates of the central location point, the geopotential anomaly, and the coordinates of the boundary points of the outermost closed contour line of the high-pressure system are used as the quantitative feature information of the high-pressure system. The high-pressure identification result is output based on the quantitative feature information.

2. The method according to claim 1, characterized in that, A point-by-point tracking method is used to obtain a set of contour points. The distance between the first and last contour points in the set is calculated. When the distance is less than a preset closure threshold, it is determined that the set of contour points constitutes a closed contour line. The study area is divided into a regular grid, and the grid point spacing is calculated based on the regular grid. The grid point spacing includes the spacing in the longitude direction and the spacing in the latitude direction. Search for grid points in the regular grid that meet the tracking start point conditions. The tracking start point conditions include: the absolute value of the difference between the geopotential height value of the grid point and the target contour height value is less than a preset distance threshold, and the geopotential height values ​​of the grid point and its adjacent grid points are located on both sides of the target contour height value. Based on the tracking starting point condition, linear interpolation calculation is performed between adjacent grid points to obtain the precise coordinates of the contour points, wherein the precise coordinates are calculated by the ratio of the geopotential height value of the adjacent grid points to the height value of the target contour line. Calculate the distance between the first and last two isoline points in the isoline point set and the total length of the two isoline points. Multiply the ratio of the distance value to the total length value by a first weighting coefficient to obtain the distance-length ratio weighting coefficient. Multiply the ratio of the tangential angle difference between the two isoline points by a second weighting coefficient to obtain the tangential angle difference ratio weighting coefficient. Add the distance-length ratio weighting coefficient and the tangential angle difference ratio weighting coefficient to obtain the closure evaluation index. Based on the closure assessment index, it is determined whether the contour point set is closed. When the closure assessment index is less than the preset index threshold, the contour point set is determined to be a closed contour line.

3. The method according to claim 1, characterized in that, A residual convolutional neural network model is built using the ResNet framework. Performance evaluation and parameter optimization of the residual convolutional neural network model are performed using accuracy and cross-entropy, including: A residual convolutional neural network model is established based on the ResNet framework, which includes a shallow feature extraction part and a deep feature learning part. The residual convolutional neural network model is trained and optimized by using the cross-entropy loss function to calculate the model loss and by using the Adam optimizer to optimize the model parameters.

4. A high-voltage recognition system based on residual convolutional neural networks, used to implement the method of any one of claims 1-3, characterized in that, include: The first unit is used to acquire daily ERA-5 geopotential height data, meridional wind data, and zonal wind data; and to determine the criteria for high pressure based on the geopotential height data, the meridional wind data, and the zonal wind data. The criteria include the analysis of at least one closed contour line in the geopotential height field and the presence of a complete anticyclonic circulation in the wind field. The second unit is used to normalize and merge the geopotential height data, the meridional wind data, and the zonal wind data into a three-channel dataset. The judgment result of whether high pressure occurs daily is used as the category label for supervised learning, and the three-channel dataset is divided into a training set and a test set according to a preset ratio. The third unit is used to build a residual convolutional neural network model using the ResNet framework, and to evaluate the performance and optimize the parameters of the residual convolutional neural network model using accuracy and cross-entropy. The accuracy is the proportion of correctly identified samples in the total number of samples, and the cross-entropy is used to measure the difference between the true probability distribution and the predicted probability distribution. When the accuracy reaches a preset accuracy threshold and the loss function value of the cross-entropy is less than a preset loss threshold, the iterative training of the residual convolutional neural network model is stopped. The fourth unit is used to extract the outermost closed contour line information of the high pressure based on the judgment result of the residual convolutional neural network model, determine the highest potential point within the outermost closed contour line of the high pressure as the center position of the high pressure, and use the highest potential anomaly as the high pressure intensity to obtain the high pressure identification result.

5. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 3.

6. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 3.

Images