A Machine Learning-Based Method for Sea Wind Front Recognition

Abstract

This invention relates to the field of meteorological monitoring technology, specifically to a sea front identification method based on machine learning, comprising the following steps: S1, acquiring meteorological station, radiosonde station, and radar echo data of the target area, extracting meteorological element gradients, abrupt change intensity, and land-sea difference, and extracting radar echo features to form multi-source feature data; S2, collecting sea front and non-sea front data to construct samples and dividing them into training, validation, and test sets; S3, jointly analyzing meteorological features to identify the land-sea interaction boundary and constructing a sea front probability field; S4, constructing a machine learning model to fuse multi-modal features and training and evaluating it; S5, inputting multi-source data in real time for sea front identification and outputting early warnings. This invention, by fusing multi-source meteorological data and combining it with a machine learning model, achieves accurate identification and real-time early warning of sea fronts, thereby improving the accuracy and automation level of sea front monitoring.

Description

Technical Field

[0001] This invention relates to the field of meteorological monitoring technology, and in particular to a method for identifying sea fronts based on machine learning. Background Technology

[0002] With the continuous advancement of urbanization in coastal areas, the impact of land-sea thermal differences and local circulation activities on regional weather changes is becoming increasingly significant. Among these, sea breeze fronts are a typical example of land-sea interaction weather systems. Sea breeze fronts typically manifest as frontal structures formed during the advance of sea breeze air masses towards land, and their formation is closely related to various meteorological factors such as land-sea temperature differences, wind field changes, and local convection activity. The occurrence of sea breeze fronts is often accompanied by abrupt temperature changes, wind direction shifts, and localized precipitation or enhanced convection, significantly impacting weather forecasting, aviation and maritime safety, and urban meteorological services in coastal areas. Therefore, accurately identifying and monitoring the occurrence and evolution of sea breeze fronts has become an important research topic in the field of meteorological monitoring and forecasting.

[0003] Currently, the identification of sea wind fronts largely relies on manual experience or threshold judgment methods based on single meteorological elements, such as temperature changes, sudden wind direction changes, or radar echo characteristics. While these methods can reflect the local characteristics of sea wind fronts to some extent, they typically utilize only single or limited meteorological information, making it difficult to fully characterize the complex thermodynamic coupling features during the formation of sea wind fronts. Furthermore, due to the strong spatiotemporal variability in the spatial structure and evolution of sea wind fronts, traditional methods are prone to problems such as low identification accuracy, insufficient stability, and low automation under complex meteorological backgrounds, making it difficult to meet the needs of refined meteorological monitoring and real-time early warning. Summary of the Invention

[0004] This invention provides a machine learning-based method for identifying sea wind fronts. By fusing multi-source meteorological information such as ground meteorological station observation data, radiosonde station observation data, and radar echo images, it extracts structured meteorological features and radar image features, and constructs a probability field for sea wind front occurrence by combining the differences between land and sea meteorological elements. Based on this, a machine learning identification model that integrates multimodal features is used to identify and warn of sea wind fronts, thereby improving the accuracy and stability of sea wind front identification and realizing automated monitoring and real-time early warning of sea wind fronts.

[0005] A machine learning-based method for identifying sea wind fronts includes the following steps:

[0006] S1. Acquire multi-source data of the target area, including observation data from ground meteorological stations, observation data from radiosonde stations, and radar echo images. Perform missing value imputation, outlier removal, and standardization on the ground meteorological station observation data, and extract structured meteorological features, including meteorological element gradients, meteorological element abrupt change intensity, and differences between land and sea meteorological elements. At the same time, enhance and denoise the radar echo images and extract radar image features, including echo morphology features and echo intensity features, to form a multi-source meteorological feature dataset.

[0007] S2. Collect multi-source data corresponding to historical sea breeze front events as positive samples, select multi-source data corresponding to non-sea breeze front events as negative samples, and construct a sample dataset. Expand the samples by data augmentation of structured meteorological features and radar image features, and divide the sample dataset into training set, validation set and test set.

[0008] S3. Based on the multi-source meteorological feature dataset, perform spatial joint analysis on the difference between land and sea meteorological elements, temperature gradient and wind speed gradient in the target area, identify the abrupt change zone of land and sea meteorological elements and determine the location of the land-sea interaction boundary, and construct the probability field of sea wind front occurrence by combining the spatial distribution characteristics of radar echoes, extract the probability field features of sea wind front to characterize the potential distribution area of ​​sea wind front in space and its evolution trend.

[0009] S4. Construct a machine learning recognition model, which integrates structured meteorological features, radar image features and sea front probability field features through an attention mechanism, and uses a deep neural network to extract high-level semantic features for classification and recognition. The machine learning recognition model is trained using a training set, and the recognition performance of the machine learning recognition model is evaluated and optimized by combining a validation set and a test set.

[0010] S5 collects multi-source data of the target area in real time, and obtains the sea wind front identification result and its confidence level through the trained machine learning identification model. When the identification confidence level is higher than the identification confidence level threshold, it outputs a sea wind front identification warning and combines radar echo images and meteorological station spatial distribution information to determine the approximate location, occurrence time and related meteorological element characteristics of the sea wind front. Otherwise, it outputs no sea wind front identification result and continues to monitor.

[0011] Optionally, S1 includes:

[0012] S11: Acquire multi-source data of the target area, including observation data from ground meteorological stations, observation data from radiosonde stations, and radar echo images, and perform time synchronization processing on multi-source data from different sources, while completing spatial alignment according to a unified spatial coordinate system;

[0013] S12, fill in the missing values ​​in the ground meteorological station observation data, and estimate the ground meteorological station observation data at the missing position based on the ground meteorological station observation data at adjacent time nodes using the Lagrange interpolation method;

[0014] S13, After completing the missing value filling, outlier removal and standardization processing are performed on the ground meteorological station observation data;

[0015] S14, extracting structured meteorological features based on standardized ground meteorological station observation data, including meteorological element gradients, meteorological element abrupt change intensity, and differences between land and sea meteorological elements;

[0016] S15 performs image enhancement and denoising on the radar echo image and extracts radar image features, including echo intensity features and echo morphology features.

[0017] S16 integrates structured meteorological features with radar image features to form a multi-source meteorological feature dataset.

[0018] Optionally, S15 includes:

[0019] S151, regarding time radar echo image Contrast enhancement is performed to highlight the echo structure, and Gaussian filtering is used to suppress speckle noise and local high-frequency interference, resulting in an enhanced and denoised radar echo image. ;

[0020] S152, Based on the enhanced and denoised radar echo image, calculate the echo intensity characteristics, including the mean echo intensity and the variance of echo intensity;

[0021] S153 obtains a binary mask of the echo region based on threshold segmentation on the enhanced and denoised radar echo image, extracts the echo morphology features of the maximum connected echo region, and uses them to characterize the spatial structure of the echo, including area, perimeter and compactness.

[0022] Optionally, S2 includes:

[0023] S21. Collect a list of historical sea breeze front events (including start and end times and affected areas). Within the time window of each sea breeze front event, extract multi-source data corresponding to the target area as positive samples. Simultaneously, select time windows from non-sea breeze front periods that are comparable to the positive samples in terms of season, weather conditions, and topographic background, and extract corresponding multi-source data as negative samples. Control the ratio of positive to negative samples to form a labeled initial sample set. ;

[0024] S22, the initial sample set is enhanced with structured features and radar images respectively. The structured meteorological features are expanded by adding small random perturbations and constructing combined features, while the radar echo images are expanded by rotation, flipping, cropping, and scaling, thus obtaining the enhanced sample set. ;

[0025] S23, the enhanced sample set The samples are randomly shuffled, and the augmented sample set is then processed while maintaining the ratio of positive to negative samples. The dataset is divided into training, validation, and test sets, with the following ratios: .

[0026] Optionally, S3 includes:

[0027] S31. Based on the multi-source meteorological feature dataset, a spatial joint analysis is performed on the difference between land and sea meteorological elements, temperature gradient, and wind speed gradient within the target area to construct a joint intensity field that reflects the coupled changes of land-sea differences, thermal gradient, and dynamic gradient. By screening the spatial clustering areas of the joint intensity field, candidate areas of abrupt transition zones where land and sea meteorological elements rapidly transition from weak to strong are identified.

[0028] S32, within the candidate region of the abrupt change zone, the joint intensity field is profiled and scanned along the principal normal direction of the coast and land to determine the spatial location where the joint intensity reaches its peak or the gradient change rate is the largest, and this spatial location is connected in two-dimensional space to form a continuous boundary line or boundary zone, thereby obtaining the location of the sea-land interaction boundary of the target region. At the same time, the continuity, curvature abrupt change and fracture segment of the boundary are smoothed and connected to form a stable boundary constraint for the construction of the probability field.

[0029] S33. Combining the location of the land-sea interaction boundary with the spatial distribution characteristics of radar echoes, a probability field for the occurrence of sea wind fronts is constructed within the boundary neighborhood. By fusing the joint intensity field, the boundary neighborhood distance attenuation factor, and the linear structure and intensity distribution characteristics of the radar echo image, the probability value of the occurrence of sea wind fronts for each spatial grid is obtained. The probability field features of sea wind fronts, including probability peak value, probability bandwidth, probability gradient, and probability centerline orientation, are extracted from the probability field to characterize the potential spatial distribution area of ​​sea wind fronts and their evolution trend over time.

[0030] Optionally, S31 includes:

[0031] S311, Based on the multi-source meteorological feature dataset, the difference between land and sea meteorological elements, temperature gradient and wind speed gradient in the target area are mapped to a unified spatial grid to form three types of feature grids, and the three types of grids are normalized.

[0032] S312 weighted and fused the normalized land-sea meteorological element difference grid, temperature gradient grid, and wind speed gradient grid to construct a joint intensity field. ;

[0033] S313 extracts high-value regions from the joint intensity field and performs spatial connectivity domain screening on the high-value regions, retaining high-intensity connected regions that are clustered in a band and meet the continuity requirements, as candidate areas for abrupt transition zones where marine and land meteorological elements rapidly transition from weak to strong.

[0034] Optionally, S32 includes:

[0035] S321, within the candidate region of the mutation zone, the local principal normal direction is determined based on the spatial gradient direction of the joint intensity field, and a one-dimensional profile line is constructed along the corresponding principal normal direction, starting from each sampling center point in the candidate region of the mutation zone, thereby converting the two-dimensional joint intensity field into a joint intensity sequence that varies along the normal direction.

[0036] S322, along the constructed normal profile line, scan the joint intensity sequence to determine the profile position where the joint intensity reaches the peak or the gradient change rate is the largest, and take this position as the local land-sea interaction boundary feature point to obtain the initial boundary point sequence reflecting the spatial distribution of the land-sea interaction boundary;

[0037] S323: The initial boundary point sequence is sorted and connected in two-dimensional space according to the proximity relationship to form the initial boundary line. For the local jaggedness, curvature abrupt change and broken segment in the initial boundary line, the smoothing and connectivity correction method is used to optimize it to obtain the location of the land-sea interaction boundary.

[0038] Optionally, S33 includes:

[0039] S331, based on the location of the sea-land interaction boundary, constructs a boundary neighborhood within a preset range on both sides of the boundary, calculates the joint intensity field, boundary neighborhood distance attenuation factor, and linear structure and intensity distribution of radar echo image for each spatial grid in the boundary neighborhood, and performs normalization processing.

[0040] S332, within the boundary neighborhood, the normalized joint intensity field, the linear structure features and intensity distribution features of the radar echo image, and the boundary neighborhood distance attenuation factor are fused to obtain the probability value of the occurrence of the sea wind front corresponding to each spatial grid, thereby forming the probability field of the occurrence of the sea wind front.

[0041] S333, based on the constructed probability field of sea wind front occurrence, performs feature extraction on the probability field to obtain probability field features describing the spatial distribution and evolution trend of sea wind front, including probability peak value, probability bandwidth, probability gradient and probability centerline orientation.

[0042] Optionally, S4 includes:

[0043] S41, input the sample dataset into the machine learning recognition model, extract the structured meteorological features, radar image features and sea wind front probability field features of each sample in the sample dataset, and perform weight allocation and fusion through the attention mechanism to generate a fused feature vector.

[0044] S42 inputs the fused feature vector into the deep neural network, extracts high-level semantic features layer by layer using multi-layer nonlinear mapping, and completes the classification and recognition of sea wind fronts and non-sea wind fronts through the output layer, outputting the predicted probability of the corresponding category.

[0045] S43 uses the training set to supervise the training of the machine learning recognition model, constrains the deviation between the predicted results and the true labels through the cross-entropy loss function, and updates the parameters of the machine learning recognition model using the gradient descent method. During the training process, the convergence and generalization ability of the machine learning recognition model are monitored in conjunction with the validation set, and overfitting is avoided through an early stopping strategy. After training is completed, the recognition performance of the machine learning recognition model is evaluated using the test set, and the parameters of the machine learning recognition model are optimized and adjusted based on the evaluation results.

[0046] Optionally, S5 includes:

[0047] S51 collects multi-source data of the target area in real time, extracts the structured meteorological features, radar image features and sea wind front probability field features at the current moment, and generates the fused feature vector at the current moment.

[0048] S52, input the fused feature vector of the current moment into the trained machine learning recognition model, output the predicted probability of the sea wind front category, and use the predicted probability as the recognition confidence;

[0049] S53, when the identification confidence level is higher than the identification confidence level threshold, the approximate location of the sea front is determined by combining the distribution of high-value echo areas in the current radar echo image and the spatial distribution information of ground meteorological stations, and the current time is recorded as the occurrence time of the sea front. At the same time, the corresponding relevant meteorological element characteristics are output, including the difference between land and sea meteorological elements, temperature gradient, wind speed gradient, probability field peak value and probability bandwidth. When the identification confidence level is lower than the identification confidence level threshold, the result of no sea front identification is output, and the multi-source data at the current time is continued to be included in the time-series monitoring process to achieve continuous dynamic tracking of the target area.

[0050] The beneficial effects of this invention are:

[0051] This invention achieves effective fusion of multi-source meteorological information and high-quality feature construction by uniformly acquiring, synchronizing, and spatially aligning multi-source meteorological data, including ground meteorological station observation data, radiosonde station observation data, and radar echo images. It also incorporates missing value imputation, outlier removal, and standardization processing, while extracting structured meteorological features such as meteorological element gradients, meteorological element abrupt change intensity, and differences between land and sea meteorological elements, as well as radar echo intensity and morphological features. This enables a more comprehensive reflection of the thermal, dynamic, and echo structure features during the formation of sea wind fronts, thereby improving the reliability of basic data and the ability to express features in sea wind front identification.

[0052] This invention constructs a joint intensity field reflecting the coupled changes of land-sea differences, thermal gradients, and dynamic gradients by spatially analyzing the differences between land and sea meteorological elements, temperature gradients, and wind speed gradients. Based on this, it identifies the location of the land-sea interaction boundary and constructs a probability field for the occurrence of sea wind fronts by combining the spatial distribution characteristics of radar echoes. It also extracts probability field features such as probability peak value, probability bandwidth, probability gradient, and probability centerline orientation, thereby achieving a quantitative characterization of the potential distribution area and evolution trend of sea wind fronts. This allows sea wind front identification to not only rely on machine learning models but also to combine meteorological physical mechanisms for auxiliary judgment, thus improving the accuracy and stability of the identification results.

[0053] This invention constructs a machine learning recognition model that integrates structured meteorological features, radar image features, and sea front probability field features. It utilizes an attention mechanism to adaptively weight and fuse multimodal features, and combines deep neural networks for advanced semantic feature extraction and classification. The model performance is evaluated and optimized on training, validation, and test sets. In practical applications, it identifies and analyzes real-time multi-source meteorological data. When the recognition confidence reaches a preset threshold, it automatically outputs a sea front identification warning, its spatial location, and related meteorological features. This achieves real-time automatic identification and warning of sea fronts, improving the automation level and timeliness of sea front monitoring. Attached Figure Description

[0054] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only for this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0055] Figure 1 This is a schematic diagram of the identification method according to an embodiment of the present invention;

[0056] Figure 2 This is a schematic diagram of the probability field construction for the occurrence of sea wind fronts in an embodiment of the present invention. Detailed Implementation

[0057] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. Those skilled in the art may employ other alternative methods to implement some well-known technologies; moreover, the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.

[0058] like Figures 1-2 As shown, a machine learning-based method for identifying sea fronts includes the following steps:

[0059] S1. Acquire multi-source data of the target area, including observation data from ground meteorological stations, observation data from radiosonde stations, and radar echo images. Perform missing value imputation, outlier removal, and standardization on the ground meteorological station observation data, and extract structured meteorological features, including meteorological element gradients, meteorological element abrupt change intensity, and differences between land and sea meteorological elements. At the same time, enhance and denoise the radar echo images and extract radar image features, including echo morphology features and echo intensity features, to form a multi-source meteorological feature dataset.

[0060] S2. Collect multi-source data corresponding to historical sea breeze front events as positive samples, select multi-source data corresponding to non-sea breeze front events as negative samples, and construct a sample dataset. Expand the samples by data augmentation of structured meteorological features and radar image features, and divide the sample dataset into training set, validation set and test set.

[0061] S3, based on a multi-source meteorological feature dataset, performs spatial joint analysis on the difference between land and sea meteorological elements, temperature gradient, and wind speed gradient in the target area, identifies abrupt change zones of land and sea meteorological elements and determines the location of the land-sea interaction boundary, and constructs a probability field for the occurrence of sea wind fronts by combining the spatial distribution characteristics of radar echoes, extracts the characteristics of the probability field of sea wind fronts, and describes the potential distribution area of ​​sea wind fronts in space and their evolution trend.

[0062] S4. Construct a machine learning recognition model, which integrates structured meteorological features, radar image features and sea front probability field features through an attention mechanism, and uses a deep neural network to extract high-level semantic features for classification and recognition. The machine learning recognition model is trained using a training set, and the recognition performance of the machine learning recognition model is evaluated and optimized by combining a validation set and a test set.

[0063] S5 collects multi-source data of the target area in real time, and obtains the sea wind front identification result and its confidence level through the trained machine learning identification model. When the identification confidence level is higher than the identification confidence level threshold, it outputs a sea wind front identification warning and combines radar echo images and meteorological station spatial distribution information to determine the approximate location, occurrence time and related meteorological element characteristics of the sea wind front. Otherwise, it outputs no sea wind front identification result and continues to monitor.

[0064] S1 includes:

[0065] S11: Acquire multi-source data of the target area, including observation data from ground meteorological stations, observation data from radiosonde stations, and radar echo images, and perform time synchronization processing on multi-source data from different sources, while completing spatial alignment according to a unified spatial coordinate system;

[0066] S12, missing values ​​in the ground weather station observation data are filled in. The missing data is estimated using Lagrange interpolation based on observations from adjacent time points, and is represented as follows:

[0067] ;

[0068] ;

[0069] in, In time Observational data from ground meteorological stations, The ground weather station observation data obtained through interpolation calculations, Given the known observation time points, Let Lagrange basis functions be used. The interpolation order;

[0070] S13, after completing the missing value imputation, outlier removal and standardization are performed on the ground meteorological station observation data, as shown below:

[0071] ;

[0072] ;

[0073] in, Data from ground-based meteorological stations. For standardized ground weather station observation data, , These represent the mean and standard deviation of observation data from ground weather stations. , These represent the minimum and maximum values ​​of observation data from ground weather stations, respectively.

[0074] S14, based on standardized ground meteorological station observation data, extracts structured meteorological features, including meteorological element gradients, meteorological element abrupt change intensity, and land-sea meteorological element differences, expressed as:

[0075] ;

[0076] ;

[0077] ;

[0078] ;

[0079] ;

[0080] in, For the gradient of meteorological elements, , This consists of standardized ground meteorological station observation data from adjacent observation points. , These represent the distances between adjacent observation points. The intensity of sudden changes in meteorological elements, For time intervals, The difference between land and sea meteorological elements. This represents the average value of standardized ground-based meteorological station observations for the marine region. This represents the average value of standardized ground weather station observations over a land area. Standardized ground-based meteorological station observation data for marine areas. Standardized ground meteorological station observation data for land areas. This refers to the number of meteorological observation stations located on the ocean side of the target area. This refers to the number of meteorological observation stations located on the land side of the target area;

[0081] S15 performs image enhancement and denoising on the radar echo image and extracts radar image features, including echo intensity features and echo morphology features.

[0082] S16 integrates structured meteorological features with radar image features to form a multi-source meteorological feature dataset.

[0083] S15 includes:

[0084] S151, regarding time radar echo image Contrast enhancement is performed to highlight the echo structure, and Gaussian filtering is used to suppress speckle noise and local high-frequency interference, resulting in an enhanced and denoised radar echo image. , is represented as:

[0085] ;

[0086] ;

[0087] ;

[0088] in, For reflectivity channel in pixel The value, The enhanced reflectivity pixel value. This is the denoised reflectance image. , These are the minimum and maximum reference values ​​for the reflectivity channel, respectively. To crop the values ​​to , For Gaussian kernel, For Gaussian kernel scaling parameters, , These are the relative displacement coordinates within the core;

[0089] S152, based on the enhanced and denoised radar echo image, calculate the echo intensity characteristics, including the mean echo intensity and the variance of the echo intensity, expressed as:

[0090] ;

[0091] ;

[0092] in, For a moment The average echo intensity For a moment The variance of echo intensity , These represent the height and width of the radar echo image, respectively.

[0093] S153, based on threshold segmentation, a binary mask of the echo region is obtained on the enhanced and denoised radar echo image. The echo morphology features of the most connected echo region are extracted to characterize the spatial structure of the echo, including area, perimeter, and compactness, as follows:

[0094] ;

[0095] ;

[0096] ;

[0097] ;

[0098] in, This is a binary echo mask (1 represents the echo pixel, and 0 represents the background). This is an indicator function; it returns 1 if the condition is true, and 0 otherwise. Echo segmentation threshold, For mask The set of pixels in the largest connected echo region. To connect the echo region The set of boundary pixels, The area of ​​the echo region. The perimeter of the echo region. For firmness;

[0099] The echo segmentation threshold is expressed as:

[0100] (1) First define the boundary neighborhood zone (background statistical area): Let the set of pixels of the land-sea interaction boundary in the radar image coordinate system be defined. Take the width as The neighborhood band is represented as:

[0101] ;

[0102] in, For the background statistical area pixel set, The minimum Euclidean distance from a pixel to the boundary set;

[0103] (2) Calculate the reflectance statistics after denoising and enhancement on the background statistical area, expressed as:

[0104] ;

[0105] ;

[0106] in, This represents the number of pixels in the background statistics area. , These represent the mean and standard deviation of the background statistical area reflectance, respectively.

[0107] (3) Adaptively set the echo segmentation threshold, expressed as:

[0108] ;

[0109] in, This is the threshold coefficient.

[0110] S2 includes:

[0111] S21. Collect a list of historical sea breeze front events (including start and end times and affected areas). Within the time window of each sea breeze front event, extract multi-source data corresponding to the target area as positive samples. Simultaneously, select time windows from non-sea breeze front periods that are comparable to the positive samples in terms of season, weather conditions, and topographic background, and extract corresponding multi-source data as negative samples. Control the ratio of positive to negative samples to form a labeled initial sample set. , is represented as:

[0112] ;

[0113] ;

[0114] ;

[0115] in, For the first Multi-source meteorological characteristics of each sample The labels are: 1 for samples from a sea breeze front and 0 for samples from non-sea breeze fronts. The initial total number of samples, The number of positive samples. The number of negative samples;

[0116] S22, the initial sample set is enhanced with structured features and radar images respectively. The structured meteorological features are expanded by adding small random perturbations and constructing combined features, while the radar echo images are expanded by rotation, flipping, cropping, and scaling, thus obtaining the enhanced sample set. Specifically, it includes:

[0117] (1) For the initial sample set Structured meteorological feature subvectors for each sample By superimposing controlled-amplitude micro-Gaussian random perturbations onto it, perturbed structured features are generated. To simulate the feature fluctuations caused by observation noise and environmental perturbations, thereby expanding the distribution range of structured feature samples and improving model robustness, it is expressed as:

[0118] ;

[0119] in, This is the perturbation noise vector. With a mean of 0 and a covariance of The multivariate Gaussian distribution, The intensity of the structured feature perturbation. It is the identity matrix;

[0120] (2) Based on the perturbed structured features, the structured feature components are further combined to generate ratio features and product features that can characterize the coupling relationship of elements. The generated combined features are then spliced ​​into the structured feature vector to achieve the expansion of structured feature dimensions and information enhancement.

[0121] set up For the selected feature components The construction is represented as:

[0122] ;

[0123] And then concatenate the combined features with the original features:

[0124] ;

[0125] in, The number of structured feature dimensions. , These are the th elements in the structured feature vector. , One portion, As a ratio combination feature, For product combination features, It is the stability constant. For multiple The ratio eigenvectors formed For multiple The product eigenvectors formed This is the combined and enhanced structured meteorological feature vector;

[0126] (3) For the initial sample set The radar echo image corresponding to each sample. Perform geometric transformations such as rotation, flipping, cropping, and scaling to obtain enhanced radar images. The sample labels are kept unchanged; then, the corresponding radar image features are extracted from the enhanced radar images. and with structured enhancement features Concatenation forms an enhanced sample feature vector Thus, the enhanced sample set is obtained. ;

[0127] For the The first sample generates the second... An enhanced version: ;

[0128] The set of enhancement operators includes:

[0129] ;

[0130] Enhanced sample feature vectors: ;

[0131] Enhanced sample set: ;

[0132] in, For the first An image enhancement operator, For rotation operators, , , These are the horizontal flip and vertical flip operators, respectively. For the trimming operator, For cropping ratio or cropping window parameters, For scaling operators, This is the scaling factor. To enhance the multi-source meteorological characteristics, For the first The number of enhanced versions generated per sample;

[0133] S23, the enhanced sample set The samples are randomly shuffled, and the augmented sample set is then processed while maintaining the ratio of positive to negative samples. The dataset is divided into training, validation, and test sets, with the following ratios: , is represented as:

[0134] ;

[0135] ;

[0136] ;

[0137] ;

[0138] ;

[0139] ;

[0140] ;

[0141] in, , , These are the training set, validation set, and test set, respectively. To enhance the total number of samples, For the union, For intersection, It is an empty set.

[0142] S3 includes:

[0143] S31, based on a multi-source meteorological feature dataset, performs spatial joint analysis on the difference between land and sea meteorological elements, temperature gradient, and wind speed gradient within the target area, constructs a joint intensity field that reflects the coupled changes of land-sea differences, thermal gradient, and dynamic gradient, and identifies candidate areas of abrupt transition zones where land and sea meteorological elements rapidly transition from weak to strong by screening the spatial clustering areas of the joint intensity field.

[0144] S32, within the candidate region of the abrupt change zone, the joint intensity field is profiled and scanned along the principal normal direction of the coast and land to determine the spatial location where the joint intensity reaches its peak or the gradient change rate is the largest. This spatial location is then connected in two-dimensional space to form a continuous boundary line or boundary zone, thereby obtaining the location of the sea-land interaction boundary of the target region. At the same time, the continuity, curvature abrupt change and fracture segment of the boundary are smoothed and connected to form a stable boundary constraint for the construction of the probability field.

[0145] S33 combines the location of the land-sea interaction boundary with the spatial distribution characteristics of radar echoes to construct a probability field for the occurrence of sea wind fronts within the boundary neighborhood. By fusing the joint intensity field, the boundary neighborhood distance attenuation factor, and the linear structure and intensity distribution characteristics of the radar echo image, the probability value of the occurrence of sea wind fronts for each spatial grid is obtained. The probability field features of sea wind fronts, such as probability peak, probability bandwidth, probability gradient, and probability centerline orientation, are extracted from the probability field to characterize the potential spatial distribution area of ​​sea wind fronts and their evolution trend over time.

[0146] S31 includes:

[0147] S311, based on a multi-source meteorological feature dataset, maps the differences between land and sea meteorological elements, temperature gradients, and wind speed gradients within the target area to a unified spatial grid, forming three types of feature grids. These three types of grids are then normalized and represented as follows:

[0148] ;

[0149] in, For spatial grid positions, For any feature grid, , The difference between land and sea meteorological elements. For temperature gradient, For wind speed gradient, , These represent the minimum and maximum values ​​of the feature raster within the target region, respectively. These are the normalized feature raster values;

[0150] S312 weighted and fused the normalized land-sea meteorological element difference grid, temperature gradient grid, and wind speed gradient grid to construct a joint intensity field. This allows the joint intensity field to exhibit a higher spatial response to regions with significant land-sea differences and simultaneous enhancement of both thermal and dynamic gradients, thus reflecting the coupled variation characteristics of sea wind front formation, as expressed in:

[0151] ;

[0152] in, , , These are, respectively, normalized land-sea meteorological element difference raster, temperature gradient raster, and wind speed gradient raster. , , These are the corresponding weighting coefficients;

[0153] S313 extracts high-value regions from the joint intensity field and performs spatial connectivity domain screening on these high-value regions. High-intensity connected regions exhibiting zonal clustering and meeting continuity requirements are retained as candidate areas for abrupt transition zones where marine and land meteorological elements rapidly transition from weak to strong. This is represented as:

[0154] Binarization of high-value regions: ;

[0155] Set of connected components: ;

[0156] Clustering and filtering: ;

[0157] in, It is a high-value binary mask for the joint intensity field (1 represents the high-intensity region, and 0 represents the non-high-intensity region). This is an indicator function; it returns 1 if the condition is true, and 0 otherwise. For the joint intensity field threshold, Extracting operators for connected components For the first A set of connected pixels / grids The number of connected components. The area of ​​the connected region. The minimum area threshold for connected components. Candidate regions for mutation bands;

[0158] Minimum area threshold of connected components Represented as:

[0159] ;

[0160] in, The length of the land-sea interaction boundary within the target area. The average grid spacing of the spatial grid in the target region. This is the area retention factor. This is the width adjustment coefficient.

[0161] S32 includes:

[0162] S321, within the candidate region of the mutation zone, the local principal normal direction is determined based on the spatial gradient direction of the joint intensity field. Taking each sampling center point within the candidate region of the mutation zone as a starting point, a one-dimensional profile line is constructed along the corresponding principal normal direction, thereby converting the two-dimensional joint intensity field into a joint intensity sequence varying along the normal direction, as shown below:

[0163] ;

[0164] ;

[0165] ;

[0166] ;

[0167] in, For the joint intensity field at position The gradient vector at that point, , The joint strength field is respectively in direction, Partial derivatives in direction, Let be the magnitude of the gradient vector. For position The unit vector located along the coastal continental normal direction. The parameters on the normal profile line are spatial points, Here is the arc length parameter along the principal normal direction. Half the length of the cross section For position The joint intensity sequence on the corresponding normal profile;

[0168] S322, along the constructed normal profile line, scan the joint intensity sequence to determine the profile location where the joint intensity reaches its peak or the gradient change rate is the largest, and use this location as the local land-sea interaction boundary feature point to obtain the initial boundary point sequence reflecting the spatial distribution of the land-sea interaction boundary, represented as:

[0169] Peak positioning: ;

[0170] Location of maximum gradient rate of change: ;

[0171] Local land-sea interaction boundary feature points: ;

[0172] Initial boundary point sequence: ;

[0173] in, These are the optimal parameter values ​​for the corresponding boundary positions in the normal profile. The rate of change of the combined strength along the normal profile, These are local land-sea interaction boundary feature points. For the initial boundary point sequence, For the first Each sampling center point This represents the number of sampling center points within the candidate region that participate in the profile scan.

[0174] S323, the initial boundary point sequence is sorted and connected in two-dimensional space according to proximity to form the initial boundary line. Then, for local jaggedness, abrupt curvature changes, and broken segments in the initial boundary line, smoothing and connectivity correction methods are used for optimization to obtain the location of the land-sea interaction boundary, specifically including:

[0175] (1) Initial boundary curve representation: Let the sequence of boundary points arranged in spatial order be:

[0176] ;

[0177] in, ;

[0178] (2) Boundary smoothing: ;

[0179] (3) Calculation of discrete curvature: ;

[0180] (4) Criterion for connectivity of broken segments: for adjacent boundary points and If satisfied If it is greater than 1, it is determined that it can be directly connected. Then, interpolation connectivity is performed on the broken segments:

[0181] ;

[0182] (5) Final stability boundary constraints:

[0183] ;

[0184] in, For the initial set of boundary line points, For the first An initial boundary point, , The initial number of boundary points. These are the smoothed boundary points. The moving average window radius is... For the first Discrete curvature at each boundary point To prevent the stability constant from having a denominator of zero, The maximum distance threshold between points that are allowed to be directly connected. Transition boundary points generated by interpolation of the fracture segment. These are the interpolation coefficients. This is the corrected location of the land-sea interaction boundary. This is the total number of boundary points after correction;

[0185] ;

[0186] in, The distance between the corresponding scanning base points of two adjacent normal profiles. This is the average of the local offsets of adjacent boundary points. The base spacing magnification factor, This is the local fluctuation correction coefficient.

[0187] S33 includes:

[0188] S331, based on the location of the land-sea interaction boundary, constructs a boundary neighborhood within a preset range on both sides of the boundary. For each spatial grid within the boundary neighborhood, calculates the joint intensity field, the boundary neighborhood range attenuation factor, and the linear structure and intensity distribution of the radar echo image, and performs normalization processing, expressed as follows:

[0189] (1) The boundary line between land and sea is Then the sea-land interaction boundary neighborhood area Defined as:

[0190] ;

[0191] (2) Boundary neighborhood distance: ;

[0192] (3) Distance attenuation factor: ;

[0193] (4) Radar echo linear structure factor normalization: Let the radar linear structure characteristics be... The radar intensity distribution characteristics are Then, after normalization, it can be expressed as:

[0194] ;

[0195] ;

[0196] (5) Normalization of joint intensity field:

[0197] ;

[0198] in, For spatial grid points within the target area, Spatial grid points To the boundary line The minimum distance, The width of the boundary neighborhood. Boundary line Any boundary point on, For point Time The Euclidean distance between them The boundary neighborhood distance decay factor, For distance attenuation control parameters, , These are the normalized radar linear structure characteristics and intensity distribution characteristics, respectively. , These represent the minimum and maximum values ​​of the radar linear structure characteristics, respectively. , These represent the minimum and maximum values ​​of the radar intensity distribution characteristics, respectively. The normalized joint intensity field, , These are the minimum and maximum values ​​of the joint intensity field, respectively;

[0199] S332, within the boundary neighborhood, the normalized joint intensity field, the linear structure features and intensity distribution features of the radar echo image, and the boundary neighborhood range attenuation factor are fused to obtain the sea wind front occurrence probability value corresponding to each spatial grid, thus forming the sea wind front occurrence probability field, expressed as:

[0200] ;

[0201] ;

[0202] in, Spatial grid points The fusion response value, Spatial grid points The probability value of a sea front occurring at that location. This is the parameter for adjusting the slope of the probability mapping. For probability mapping offset parameters, , , , These are the corresponding weighting coefficients;

[0203] ;

[0204] in, For the probability field of sea breeze front occurrence. For the boundary neighborhood spatial grid set;

[0205] S333, based on the constructed probability field of sea wind front occurrence, feature extraction is performed on the probability field to obtain probability field features describing the spatial distribution and evolution trend of the sea wind front, including probability peak value, probability bandwidth, probability gradient, and probability centerline orientation, expressed as:

[0206] ;

[0207] ;

[0208] ;

[0209] ;

[0210] ;

[0211] ;

[0212] ;

[0213] ;

[0214] ;

[0215] in, This represents the peak probability. For probability gradient, Let be the gradient vector of the probability field. , The probability fields are respectively in , Partial derivatives in the direction, For probability bandwidth, For a probability field that satisfies The area of ​​the high-probability region. As a probability threshold, Let be the direction angle of the center line of the probability field. For high probability value regions in and Covariance in direction The number of grid points in the high-probability region participating in the calculation. , The high probability regions are respectively in and Variance in direction, , For the first Coordinates of grid points in a high-probability region. , These are all grid points in the high-probability region. and The average value of the coordinates;

[0216] ;

[0217] in, , Let be the average probability value and standard deviation of the probability field in the boundary neighborhood, respectively. This is the threshold adjustment coefficient.

[0218] S4 includes:

[0219] S41, input the sample dataset into the machine learning recognition model, extract the structured meteorological features, radar image features, and sea wind front probability field features for each sample in the sample dataset, and perform weight allocation and fusion through an attention mechanism to generate a fused feature vector, specifically including:

[0220] (1) Let the first The structured atmospheric image features, radar image features, and sea wind front probability field features of each sample are as follows: , and First, the three types of features are mapped to the same dimensional empty space:

[0221] ;

[0222] ;

[0223] ;

[0224] (2) Calculate the attention score for each modality feature:

[0225] ;

[0226] ;

[0227] ;

[0228] (3) Normalize the attention scores to obtain the attention weights:

[0229] ;

[0230] ;

[0231] ;

[0232] (4) The final fused feature vector is obtained:

[0233] ;

[0234] in, , , These are the hidden layer representations of structured atmospheric image features, radar image features, and sea wind front probability field features after linear mapping. , , These are the corresponding mapping weight matrices. , , These are the corresponding bias vectors. For attention scoring vectors, , , The attention scores are for structured atmospheric features, radar image features, and sea wind front probability field features, respectively. , , These are the attention weights for structured atmospheric features, radar image features, and sea wind front probability field features, respectively. To fuse feature vectors;

[0235] S42 inputs the fused feature vector into a deep neural network, extracts high-level semantic features layer by layer using multi-layer nonlinear mapping, and completes the classification and recognition of sea wind fronts and non-sea wind fronts through the output layer, outputting the predicted probability of the corresponding category, as shown below:

[0236] ;

[0237] ;

[0238] ;

[0239] in, For the first The sample at the th Layer feature representation, For input layer features, For the first Layer network weight matrix, For the first Layer network bias vector, It is the ReLU activation function. For the first The predicted class probability vector of each sample. , These are the output layer weight matrix and bias vector, respectively. For the first The sample at the th Layer feature representation, The number of hidden layers in a deep neural network;

[0240] S43, supervised training of the machine learning recognition model is performed using the training set. The deviation between the predicted results and the true labels is constrained by the cross-entropy loss function, and the parameters of the machine learning recognition model are updated using gradient descent. During training, the convergence and generalization ability of the machine learning recognition model are monitored using the validation set, and overfitting is avoided through an early stopping strategy. After training, the recognition performance of the machine learning recognition model is evaluated using the test set, and the parameters of the machine learning recognition model are optimized and adjusted based on the evaluation results, as shown below:

[0241] ;

[0242] ;

[0243] ;

[0244] ;

[0245] ;

[0246] in, The value of the cross-entropy loss function. The number of samples in the training set. For the first The true labels for each training sample are: 1 for sea breeze fronts and 0 for non-sea breeze fronts. For the first The probability of a training sample being predicted as a sea breeze front. For the first The set of model parameters at the next iteration For the first The set of model parameters after the next iteration. For learning rate, The gradient of the loss function with respect to the model parameters. The model's recognition accuracy on the test set. To predict the number of correctly predicted samples in the test set, The total number of samples in the test set. For the actual number of cases, The number of false positives The number of false negatives For accuracy, This refers to the recall rate.

[0247] S5 includes:

[0248] S51 collects multi-source data of the target area in real time, extracts the structured meteorological features, radar image features and sea wind front probability field features at the current moment, and generates the fused feature vector at the current moment.

[0249] S52, input the fused feature vector of the current moment into the trained machine learning recognition model, output the predicted probability of the sea wind front category, and use the predicted probability as the recognition confidence;

[0250] S53, when the identification confidence level is higher than the identification confidence level threshold, the approximate location of the sea front is determined by combining the distribution of high-value echo areas in the current radar echo image and the spatial distribution information of ground meteorological stations, and the current time is recorded as the occurrence time of the sea front. At the same time, the corresponding relevant meteorological element characteristics are output, including the difference between land and sea meteorological elements, temperature gradient, wind speed gradient, probability field peak value and probability bandwidth. When the identification confidence level is lower than the identification confidence level threshold, the result of no sea front identification is output, and the multi-source data at the current time is continued to be included in the time series monitoring process to achieve continuous dynamic tracking of the target area.

[0251] The approximate location of the sea breeze front is determined: Let the grid set of radar high-value regions satisfying the echo threshold condition be . The approximate location of the sea wind front can be represented by the centroid of this region as follows:

[0252] ;

[0253] in, , These are the approximate coordinates of the current location of the sea breeze front;

[0254] ;

[0255] in, To identify the confidence threshold, , These represent the mean and standard deviation of the predicted probability of the sea breeze front in the validation set. This is the threshold adjustment coefficient.

[0256] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0257] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for identifying sea wind fronts based on machine learning, characterized in that, Includes the following steps: S1. Acquire multi-source data of the target area, including observation data from ground meteorological stations, observation data from radiosonde stations, and radar echo images. Perform missing value imputation, outlier removal, and standardization on the ground meteorological station observation data, and extract structured meteorological features, including meteorological element gradients, meteorological element abrupt change intensity, and differences between land and sea meteorological elements. At the same time, enhance and denoise the radar echo images and extract radar image features, including echo morphology features and echo intensity features, to form a multi-source meteorological feature dataset. S2. Collect multi-source data corresponding to historical sea breeze front events as positive samples, select multi-source data corresponding to non-sea breeze front events as negative samples, and construct a sample dataset. Expand the samples by data augmentation of structured meteorological features and radar image features, and divide the sample dataset into training set, validation set and test set. S3. Based on the multi-source meteorological feature dataset, perform spatial joint analysis on the difference between land and sea meteorological elements, temperature gradient and wind speed gradient in the target area, identify the abrupt change zone of land and sea meteorological elements and determine the location of the land-sea interaction boundary, and construct the probability field of sea wind front occurrence by combining the spatial distribution characteristics of radar echoes, extract the probability field features of sea wind front to characterize the potential distribution area of ​​sea wind front in space and its evolution trend. S4. Construct a machine learning recognition model, which integrates structured meteorological features, radar image features and sea front probability field features through an attention mechanism, and uses a deep neural network to extract high-level semantic features for classification and recognition. The machine learning recognition model is trained using a training set, and the recognition performance of the machine learning recognition model is evaluated and optimized by combining a validation set and a test set. S5 collects multi-source data of the target area in real time, and obtains the sea wind front identification result and its confidence level through the trained machine learning identification model. When the identification confidence level is higher than the identification confidence level threshold, it outputs a sea wind front identification warning and combines radar echo images and meteorological station spatial distribution information to determine the approximate location, occurrence time and related meteorological element characteristics of the sea wind front. Otherwise, it outputs no sea wind front identification result and continues to monitor.

2. The sea wind front identification method based on machine learning according to claim 1, characterized in that, S1 includes: S11: Acquire multi-source data of the target area, including observation data from ground meteorological stations, observation data from radiosonde stations, and radar echo images, and perform time synchronization processing on multi-source data from different sources, while completing spatial alignment according to a unified spatial coordinate system; S12, fill in the missing values ​​in the ground meteorological station observation data, and estimate the ground meteorological station observation data at the missing position based on the ground meteorological station observation data at adjacent time nodes using the Lagrange interpolation method; S13, After completing the missing value filling, outlier removal and standardization processing are performed on the ground meteorological station observation data; S14, extracting structured meteorological features based on standardized ground meteorological station observation data, including meteorological element gradients, meteorological element abrupt change intensity, and differences between land and sea meteorological elements; S15 performs image enhancement and denoising on the radar echo image and extracts radar image features, including echo intensity features and echo morphology features. S16 integrates structured meteorological features with radar image features to form a multi-source meteorological feature dataset.

3. The sea wind front identification method based on machine learning according to claim 2, characterized in that, S15 includes: S151, regarding time radar echo image Contrast enhancement is performed to highlight the echo structure, and Gaussian filtering is used to suppress speckle noise and local high-frequency interference, resulting in an enhanced and denoised radar echo image. ; S152, Based on the enhanced and denoised radar echo image, calculate the echo intensity characteristics, including the mean echo intensity and the variance of echo intensity; S153 obtains a binary mask of the echo region based on threshold segmentation on the enhanced and denoised radar echo image, extracts the echo morphology features of the maximum connected echo region, and uses them to characterize the spatial structure of the echo, including area, perimeter and compactness.

4. The sea wind front identification method based on machine learning according to claim 1, characterized in that, S2 includes: S21. Collect a list of historical sea breeze front events. Within the time window of each sea breeze front event, extract multi-source data corresponding to the target area as positive samples. Simultaneously, select time windows from non-sea breeze front periods that are comparable to the positive samples in terms of season, weather conditions, and topographic background, and extract corresponding multi-source data as negative samples. Control the ratio of positive to negative samples to form a labeled initial sample set. ; S22, the initial sample set is enhanced with structured features and radar images respectively. The structured meteorological features are expanded by adding small random perturbations and constructing combined features. The radar echo image is expanded by rotation, flipping, cropping and scaling, thus obtaining the enhanced sample set. ; S23, the enhanced sample set The samples are randomly shuffled, and the augmented sample set is then processed while maintaining the ratio of positive to negative samples. The dataset is divided into training, validation, and test sets, with the following ratios: .

5. The sea wind front identification method based on machine learning according to claim 1, characterized in that, S3 includes: S31. Based on the multi-source meteorological feature dataset, a spatial joint analysis is performed on the difference between land and sea meteorological elements, temperature gradient, and wind speed gradient within the target area to construct a joint intensity field that reflects the coupled changes of land-sea differences, thermal gradient, and dynamic gradient. By screening the spatial clustering areas of the joint intensity field, candidate areas of abrupt transition zones where land and sea meteorological elements rapidly transition from weak to strong are identified. S32, within the candidate region of the abrupt change zone, the joint intensity field is profiled and scanned along the principal normal direction of the coast and land to determine the spatial location where the joint intensity reaches its peak or the gradient change rate is the largest, and this spatial location is connected in two-dimensional space to form a continuous boundary line or boundary zone, thereby obtaining the location of the sea-land interaction boundary of the target region. At the same time, the continuity, curvature abrupt change and fracture segment of the boundary are smoothed and connected to form a stable boundary constraint for the construction of the probability field. S33. Combining the location of the land-sea interaction boundary with the spatial distribution characteristics of radar echoes, a probability field for the occurrence of sea wind fronts is constructed within the boundary neighborhood. By fusing the joint intensity field, the boundary neighborhood distance attenuation factor, and the linear structure and intensity distribution characteristics of the radar echo image, the probability value of the occurrence of sea wind fronts for each spatial grid is obtained. The probability field features of sea wind fronts, including probability peak value, probability bandwidth, probability gradient, and probability centerline orientation, are extracted from the probability field to characterize the potential spatial distribution area of ​​sea wind fronts and their evolution trend over time.

6. The sea wind front identification method based on machine learning according to claim 5, characterized in that, S31 includes: S311, Based on the multi-source meteorological feature dataset, the difference between land and sea meteorological elements, temperature gradient and wind speed gradient in the target area are mapped to a unified spatial grid to form three types of feature grids, and the three types of grids are normalized. S312 weighted and fused the normalized land-sea meteorological element difference grid, temperature gradient grid, and wind speed gradient grid to construct a joint intensity field. ; S313 extracts high-value regions from the joint intensity field and performs spatial connectivity domain screening on the high-value regions, retaining high-intensity connected regions that are clustered in a band and meet the continuity requirements, as candidate areas for abrupt transition zones where marine and land meteorological elements rapidly transition from weak to strong.

7. The sea wind front identification method based on machine learning according to claim 6, characterized in that, S32 includes: S321, within the candidate region of the mutation zone, the local principal normal direction is determined based on the spatial gradient direction of the joint intensity field, and a one-dimensional profile line is constructed along the corresponding principal normal direction, starting from each sampling center point in the candidate region of the mutation zone, thereby converting the two-dimensional joint intensity field into a joint intensity sequence that varies along the normal direction. S322, along the constructed normal profile line, scan the joint intensity sequence to determine the profile position where the joint intensity reaches the peak or the gradient change rate is the largest, and take this position as the local land-sea interaction boundary feature point to obtain the initial boundary point sequence reflecting the spatial distribution of the land-sea interaction boundary; S323: The initial boundary point sequence is sorted and connected in two-dimensional space according to the proximity relationship to form the initial boundary line. For the local jaggedness, curvature abrupt change and broken segment in the initial boundary line, the smoothing and connectivity correction method is used to optimize it to obtain the location of the land-sea interaction boundary.

8. The sea wind front identification method based on machine learning according to claim 7, characterized in that, S33 includes: S331, based on the location of the sea-land interaction boundary, constructs a boundary neighborhood within a preset range on both sides of the boundary, calculates the joint intensity field, boundary neighborhood distance attenuation factor, and linear structure and intensity distribution of radar echo image for each spatial grid in the boundary neighborhood, and performs normalization processing. S332, within the boundary neighborhood, the normalized joint intensity field, the linear structure features and intensity distribution features of the radar echo image, and the boundary neighborhood distance attenuation factor are fused to obtain the probability value of the occurrence of the sea wind front corresponding to each spatial grid, thereby forming the probability field of the occurrence of the sea wind front. S333, based on the constructed probability field of sea wind front occurrence, performs feature extraction on the probability field to obtain probability field features describing the spatial distribution and evolution trend of sea wind front, including probability peak value, probability bandwidth, probability gradient and probability centerline orientation.

9. The sea wind front identification method based on machine learning according to claim 1, characterized in that, S4 includes: S41, input the sample dataset into the machine learning recognition model, extract the structured meteorological features, radar image features and sea wind front probability field features of each sample in the sample dataset, and perform weight allocation and fusion through the attention mechanism to generate a fused feature vector. S42 inputs the fused feature vector into the deep neural network, extracts high-level semantic features layer by layer using multi-layer nonlinear mapping, and completes the classification and recognition of sea wind fronts and non-sea wind fronts through the output layer, outputting the predicted probability of the corresponding category. S43 uses the training set to supervise the training of the machine learning recognition model, constrains the deviation between the predicted results and the true labels through the cross-entropy loss function, and updates the parameters of the machine learning recognition model using the gradient descent method. During the training process, the convergence and generalization ability of the machine learning recognition model are monitored in conjunction with the validation set, and overfitting is avoided through an early stopping strategy. After training is completed, the recognition performance of the machine learning recognition model is evaluated using the test set, and the parameters of the machine learning recognition model are optimized and adjusted based on the evaluation results.

10. A method for identifying sea wind fronts based on machine learning according to claim 9, characterized in that, S5 includes: S51 collects multi-source data of the target area in real time, extracts the structured meteorological features, radar image features and sea wind front probability field features at the current moment, and generates the fused feature vector at the current moment. S52, input the fused feature vector of the current moment into the trained machine learning recognition model, output the predicted probability of the sea wind front category, and use the predicted probability as the recognition confidence; S53, when the identification confidence level is higher than the identification confidence level threshold, the approximate location of the sea front is determined by combining the distribution of high-value echo areas in the current radar echo image and the spatial distribution information of ground meteorological stations, and the current time is recorded as the occurrence time of the sea front. At the same time, the corresponding relevant meteorological element characteristics are output, including the difference between land and sea meteorological elements, temperature gradient, wind speed gradient, probability field peak value and probability bandwidth. When the identification confidence level is lower than the identification confidence level threshold, the result of no sea front identification is output, and the multi-source data at the current time is continued to be included in the time-series monitoring process to achieve continuous dynamic tracking of the target area.

Images