A weather feature image clustering and fuzzy fusion non-weather target recognition method

Abstract

The application discloses a meteorological feature image clustering and fuzzy fusion non-meteorological target identification method. Firstly, the millimeter wave cloud radar meteorological data is preprocessed. Then, meteorological data parameter characteristic images are generated, and two-dimensional meteorological data arrays are converted into two-dimensional pseudo-color images. By analyzing the distribution characteristics of meteorological targets and non-meteorological targets on the data images, the meteorological data parameter characteristic images are divided into different potential target regions by using an image segmentation large model, parameter characteristics in the potential target regions are extracted, and average radial texture and inter-bank change degree of a reflectivity factor, regional average value and variance of a radial velocity, and regional average value of a velocity spectrum width are calculated. A fuzzy logic fusion algorithm is used to determine fuzzy membership functions of statistical information, fuzzy membership degrees of each data point are calculated by setting fuzzy rules, and a weighted summation method is used to de-fuzzify to obtain a judgment value compared with a threshold value. The application improves the accuracy of identifying non-meteorological targets.

Description

Technical Field

[0001] The present invention belongs to the field of computer vision and artificial intelligence, and relates to a method for identifying non-meteorological targets using millimeter-wave cloud radar, and more particularly to a method for identifying non-meteorological targets using meteorological feature image clustering and fuzzy fusion. Background Art

[0002] Meteorological services play an important role in disaster prevention and mitigation, agricultural production, transportation, energy supply and allocation, urban construction and planning, public services and life convenience, sustainable research and technological innovation.

[0003] With the steady development of the economy and science and technology, meteorological radars are becoming increasingly advanced, and meteorological observation stations have been built in various locations. In actual applications, radar observations not only include meteorological targets (such as clouds and precipitation), but also detect a large number of non-meteorological targets (such as flocks of birds, insects, ground clutter, etc.). These non-meteorological targets can interfere with the accuracy of radar data. Intelligent meteorological target recognition methods have become a major trend and are in strong demand in the development of meteorological services. Classifying and labeling large-scale meteorological data can help improve the accuracy of meteorological observations, support refined weather forecasts and disaster warnings, assist in atmospheric and environmental science research, expand the application scenarios of millimeter-wave radar, and promote the integrated development of intelligent recognition technology and radar technology. How to identify non-meteorological targets from massive amounts of complex meteorological data has become an important research direction.

[0004] Millimeter-wave cloud radar, a radar system that utilizes electromagnetic waves in the millimeter-wave band (30-300 GHz), is widely used for meteorological observations, particularly for detecting clouds and precipitation. Compared to traditional S-band or C-band radars, millimeter-wave cloud radars offer higher spatial resolution and sensitivity, enabling the detection of meteorological elements such as tiny water droplets and ice crystals. Millimeter-wave cloud radars are not only suitable for diverse weather conditions, including clear skies, clouds, and precipitation, but can also penetrate thin clouds and light precipitation, providing effective data in both clear skies and light precipitation, and are particularly stable in light rain or foggy conditions. Their narrow beam and low detection altitude allow them to simultaneously acquire multiple parameters, including reflectivity, velocity, and spectral width. Combined with polarization technology, they can further derive information on target shape, size, and velocity. These characteristics make millimeter-wave cloud radars a valuable tool for refined meteorological observations and the study of cloud physics properties. They also exhibit potential applications in low-altitude aircraft monitoring and ecological research.

[0005] Research on intelligent recognition of non-meteorological targets using millimeter-wave cloud radar focuses on the challenge of accurately identifying and processing non-meteorological targets in complex environments. Millimeter-wave cloud radar generates a large amount of echo data. The echo signals of non-meteorological targets (such as insects, birds, and buildings) may overlap with those of meteorological targets in terms of shape and intensity, making it difficult for the radar system to effectively distinguish them during detection. This interferes with the detection of meteorological targets, reducing data quality and analysis accuracy. Therefore, real-time processing of this echo data and enabling rapid identification and classification are key to ensuring accurate meteorological observations.

[0006] Machine learning and deep learning methods require large amounts of labeled data for training, but in practice, it is difficult to obtain and label large-scale, comprehensive non-meteorological target data, especially for rare target types. Without sufficient data support, it is difficult for the model to learn stable features, resulting in poor recognition accuracy and generalization ability. Traditional algorithms are often inefficient when processing these large-scale data and cannot meet real-time requirements, resulting in an inability to respond in a timely manner, affecting real-time monitoring and analysis in practical applications. When faced with different geographical regions, climatic conditions and detection environments, the characteristics of radar echoes will change, which means that after the model is trained in a certain scenario, it may not be able to adapt to data from different scenarios. In order to improve recognition accuracy, it is often necessary to combine with other sensors (such as lidar, cameras, etc.), but due to differences in data types, resolutions and sampling rates between different sensors, data fusion faces technical challenges.

[0007] Using millimeter-wave cloud radar meteorological data to identify non-meteorological targets combines the distribution characteristics of both meteorological and non-meteorological targets. While these distribution characteristics can be used to identify non-meteorological targets, when the echo characteristics of non-meteorological and meteorological targets are similar, the segmentation results can be inaccurate, increasing the false positive rate. This is particularly true for insects and small cloud droplets. Furthermore, noisy data can be easily misclassified as a separate non-meteorological target, leading to increased misclassification and interference, thus compromising overall recognition accuracy.

[0008] Therefore, improving the accuracy of non-meteorological target recognition in meteorological data remains a major challenge. Summary of the Invention

[0009] To address the shortcomings of existing technologies, this paper proposes a non-meteorological target recognition method that combines meteorological feature image clustering with fuzzy fusion. This method uses a large image segmentation model to cluster the multi-parameter feature images of millimeter-wave cloud radar meteorological data. A fuzzy logic fusion algorithm is then used to fusion-judge the multi-feature clustering results, achieving high-precision recognition of non-meteorological targets. This method fully utilizes the global feature aggregation capabilities of the large image segmentation model and the local rule judgment of fuzzy logic, thereby optimizing the recognition accuracy of non-meteorological targets.

[0010] A non-meteorological target recognition method based on meteorological feature image clustering and fuzzy fusion, the specific steps are as follows:

[0011] Step 1: Preprocessing of millimeter-wave cloud radar meteorological data

[0012] Collect the raw meteorological echo data of a certain site's millimeter-wave cloud radar in THI mode throughout the day.

[0013] Since the scattered point clutter and radial thin line clutter in the raw weather echo data are mostly Gaussian noise caused by birds and electromagnetic interference, they can adversely affect the subsequent identification of non-meteorological targets. This clutter is removed from the raw weather echo data to obtain filtered weather echo data. Meteorological data parameter features are extracted from the filtered weather echo data, including reflectivity factor Z, velocity V, spectral width W, signal-to-noise ratio (SNR), and other parameter features. These features are then normalized to allow clustering analysis of different parameter features using a unified scale.

[0014] Step 2: Generation of millimeter-wave cloud radar meteorological data parameter feature images

[0015] Using pseudo-color processing, the normalized meteorological data parameter characteristics (reflectivity factor Z, velocity V, spectral width W, signal-to-noise ratio SNR, etc.) are mapped to different pseudo-colors according to their numerical values ​​to produce characteristic images of each meteorological data parameter. Image segmentation and clustering methods are used to analyze the characteristics of meteorological targets and non-meteorological targets on each meteorological data parameter characteristic image.

[0016] Step 3: Clustering of potential targets in millimeter-wave cloud radar meteorological data parameter feature images

[0017] The Segment Anything (SAM) image segmentation model is fine-tuned in a one-shot manner, and unsupervised clustering is performed on the parameter feature images of each meteorological data. According to the continuity and consistency of the parameter features in the temporal dimension, each meteorological data parameter feature image is divided into different regions. Each region represents different echo features, such as background region, meteorological echo region and non-meteorological echo region. The meteorological echo region and non-meteorological echo region are regarded as potential target areas.

[0018] Step 4: Extract parameter feature statistics of potential target areas

[0019] Based on the potential target areas divided in step 3, statistical information of the meteorological data parameter characteristics in each area is extracted, including the average radial texture of the reflectivity factor, the radial variation of the echo intensity between libraries, the regional average of the radial velocity, the regional variance of the radial velocity, and the regional average of the velocity spectrum width, and serves as the input of the subsequent fuzzy logic fusion algorithm.

[0020] Step 5: Non-meteorological target recognition using fuzzy logic fusion algorithm

[0021] Based on the parameter characteristic statistics of each potential target area obtained in step 4, the fuzzy membership function of each meteorological data parameter characteristic statistical information is determined. Based on the typical parameter characteristics of non-meteorological targets, fuzzy logic rules are set to calculate the non-meteorological target fuzzy membership of each data point in the potential target area based on the different parameter characteristic statistical information. These fuzzy memberships are weighted and summed to obtain the non-meteorological target likelihood score for each data point. Based on a pre-set threshold, each data point in the potential target area is determined to be either a non-meteorological target or a meteorological target. The threshold value is adjusted to make the fuzzy boundaries of the potential target area and the points within the area that are difficult to clearly classify clearer and more specific, thereby improving the accuracy of non-meteorological target identification.

[0022] The present invention has the following beneficial effects:

[0023] This method combines the global data parameter feature clustering capabilities of a large-scale image segmentation model with the local rule-based judgment of a fuzzy logic fusion algorithm, resulting in more accurate non-meteorological target identification. By utilizing multiple parameter features of millimeter-wave cloud radar meteorological data, such as reflectivity factor, velocity, spectral width, and signal-to-noise ratio, followed by noise filtering, this method reduces the impact of noise on the overall non-meteorological target recognition system, effectively accommodating complex echo scenarios under varying meteorological conditions and enhancing the model's robustness in changing environments. The large-scale image segmentation model can delineate potential target regions based on the parameter features of meteorological data. The fuzzy logic fusion algorithm allows for flexible adjustment of the weights of statistical information from different parameter features and the definition of fuzzy logic rules based on the specific characteristics of non-meteorological echoes. This flexibility allows for adapting to varying observation conditions or radar equipment characteristics, generating a comprehensive non-meteorological target likelihood score for each data point within each potential target region, making subsequent processing and analysis more convincing. The fusion of the large-scale image segmentation model and the fuzzy logic algorithm can better handle regions with blurred cluster boundaries or transitional regions. Particularly in the transition region between weather and non-meteorological echoes, the model enables more refined classification through fuzzy membership judgment, significantly improving the accuracy of identifying non-meteorological targets in raw millimeter-wave cloud radar weather data. By integrating statistical information from multiple parameter features, it can more comprehensively capture the differences between non-meteorological and weather target echoes, reducing the rate of misjudgment. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 This is a flow chart of the non-meteorological target recognition method based on meteorological feature image clustering and fuzzy fusion;

[0025] Figure 2 Flowchart of the method for fine-tuning the Segment Anything model;

[0026] Figure 3 It is the algorithm flow chart of fuzzy logic algorithm;

[0027] Figure 4 The meteorological map measured for a certain site 1 is the result obtained directly by image segmentation;

[0028] Figure 5 The meteorological map measured for a certain site 2 is the result obtained directly by image segmentation;

[0029] Figure 6 The meteorological map measured for a certain site 3 is the result obtained directly by image segmentation;

[0030] Figure 7 The result map obtained by directly using image segmentation for the weather map measured at a certain site 4;

[0031] Figure 8 This is the result obtained by clustering and fuzzy fusion algorithm of meteorological characteristic images of a certain site 1;

[0032] Figure 9 This is the result obtained by clustering and fuzzy fusion algorithm of meteorological characteristic images of a certain site 2;

[0033] Figure 10 This is the result obtained by clustering and fuzzy fusion algorithm of meteorological characteristic images of a certain site 3;

[0034] Figure 11 This is the result obtained by clustering and fuzzy fusion algorithm of meteorological characteristic images of a certain site 4. DETAILED DESCRIPTION

[0035] The method of the present invention is further explained below with reference to the accompanying drawings. Figure 1 As shown in FIG, a non-meteorological target recognition method based on meteorological feature image clustering and fuzzy fusion is proposed. The specific steps are as follows:

[0036] Step 1: Preprocessing of millimeter-wave cloud radar meteorological data

[0037] Import raw meteorological echo data from a millimeter-wave cloud radar operating in THI mode at a specific site and store the four parameter features—reflectivity factor Z, radial velocity V, spectral width W, and signal-to-noise ratio (SNR)—in array matrix format. THI mode is a time-altitude indication mode, commonly used for vertical profile scanning. In this mode, the millimeter-wave cloud radar vertically scans the atmosphere at a fixed azimuth angle to capture meteorological target information at different altitudes. Each array dimension in the two-dimensional matrix of the raw meteorological echo data parameters is 1440 × 500. The vertical dimension represents the number of bins, with a maximum observation range of 500 bins, each 30 meters long, meaning a maximum observation altitude of 15 kilometers. The horizontal dimension represents time, with a temporal resolution of 1 minute. The 1440 × 500 dimension represents meteorological echo data from 0 to 15 kilometers observed over a 24-hour period, ensuring that each parameter feature matrix has the same dimensions.

[0038] Noise filtering: Weather radar data often contains discrete, isolated noise signals near the ground. It is necessary to filter the noise to reduce the adverse effects of noise clutter in subsequent image segmentation model division and fuzzy logic fusion algorithm analysis. The formula is:

[0039]

[0040] Where k represents the distance bin of the reflectivity factor selected from the weather radar data; N represents the number of distance bins with valid values ​​in the sliding window (N is 3×3); N t is the total number of all values ​​contained in the selected sliding window; P k It indicates the percentage of the effective echo values ​​in the selected sliding window to the total library number. k If the value is less than a certain threshold (initial setting is 75%), it can be determined that the numerical point is isolated and discrete, identified as noise clutter and removed (set to NaN value). After filtering, the reflectivity factor Z, velocity V, spectral width W, signal-to-noise ratio SNR and other parameter features of the meteorological echo data are normalized and mapped to the range of 0 to 1 so that different parameter features can be clustered under the same dimension.

[0041] Step 2: Generation of millimeter-wave cloud radar meteorological data parameter feature images

[0042] The parametric features of the filtered meteorological echo data (reflectivity factor Z, velocity V, spectral width W, signal-to-noise ratio SNR) are mapped to different pseudo colors according to the numerical values ​​to generate parameter feature images of each meteorological data. This ensures that the parameter feature image format of the millimeter-wave cloud radar meteorological data is suitable for the input of the large image segmentation model, facilitates the analysis of the distribution characteristics of meteorological targets and non-meteorological targets on the pseudo-color image, and enables the model to better utilize image features to divide regions.

[0043] When performing pseudo-color mapping, a Cartesian coordinate system is used, with the horizontal axis set as the time axis, with each scale being 1 minute and containing 1440 points, corresponding to 1440 minutes, or 24 hours, of observation data; the vertical axis is set as the altitude axis, with each scale being 30 meters and containing 500 points, corresponding to 15 kilometers.

[0044] For color mapping, a color mapping scheme is selected for each parameter feature. For reflectivity factor Z, a "viridis" color band is commonly used to easily identify different echo intensities and clearly present strong and weak contrasts; for velocity V, a "coolwarm" color band is used to distinguish between positive and negative velocity values; for spectral width W, a "plasma" color band is used to highlight regions of varying velocity spectral widths; and for signal-to-noise ratio (SNR), a "YlGnBu" color band is used to indicate high and low signal-to-noise ratios. Two-dimensional color maps of each parameter feature are generated using these color mapping schemes. These color mapping schemes create visually distinct differences between different parameters, enabling more effective interpretation and comparison of millimeter-wave cloud radar weather echo data.

[0045] Step 3: Clustering of potential targets in millimeter-wave cloud radar meteorological data parameter feature images

[0046] For each parametric feature image generated in step 2, analyze the distribution characteristics of non-meteorological targets one by one on each parametric feature image. Non-meteorological targets, such as clear-air echoes, have a low reflectivity factor, a wide distribution range, and a strip-like shape. They are generally low in altitude (1-4 km). Their radial velocities are relatively smooth or close to zero and evenly distributed, indicating low momentum. Their spectral width is small, and their signal-to-noise ratio is also low. For example, ground clutter has a high reflectivity factor; ground clutter is generally static, with velocities close to zero; its spectral width is small because its velocity changes little; and its signal-to-noise ratio is high.

[0047] The Segment Anything (SAM) model is used as an image segmentation model to process the millimeter wave cloud radar meteorological data parameter feature image. It is a segmentation model based on deep learning. Its main task is to classify each pixel in the parameter feature image into different regions, such as background region, meteorological echo region, and non-meteorological echo region, so as to achieve the segmentation of the parameter feature image content. The SAM image segmentation model is fine-tuned in a one-shot manner to make it suitable for the regional clustering task of meteorological parameter feature images. The fine-tuning method process is as follows: Figure 2 As shown, specifically:

[0048] First, a reference meteorological feature image I is given R (such as reflectivity factor map) and non-meteorological target area mask image Then, the image encoder ImEncoder in the trained SAM image segmentation model is used to extract the reference meteorological feature image I R Extract pixel value and position distribution of each pixel in the meteorological feature image I as semantic feature vector. The process is expressed as follows:

[0049] F I =ImEncoder(I)

[0050] F R =ImEncoder(I R )

[0051] in Represent the reference meteorological characteristic image I R and the semantic feature vector of the meteorological feature image I to be tested, h represents the height of the image, w represents the width of the image, and c represents the number of channels of the image.

[0052] Then, the non-meteorological target area mask image M is obtained based on the reference meteorological feature image. R , calculate the non-meteorological target characteristics T p , the formula is:

[0053]

[0054] in, Represents the matrix dot product.

[0055] Next, calculate T P With F I The cosine similarity between them is used to obtain the confidence probability map S of the meteorological feature image I to be tested. p , using the statistical average confidence probability map S p Aggregate into an overall confidence map

[0056] Select the two pixels with the highest and lowest confidence values ​​in the overall confidence map S, denoted as P h and P l .P h Indicates the most likely center pixel point of the non-meteorological target area, P l P is the pixel farthest from the non-meteorological target. h and P l These are treated as positive and negative cues and fed into the cue encoder in the SAM image segmentation model, guiding it to cluster non-meteorological targets. This approach allows the SAM model to cluster contiguous regions around positive cues, allowing it to segment potential target regions such as data-free background, meteorological targets, and non-meteorological targets.

[0057] Step 4: Extract parameter feature statistics of potential target areas

[0058] Through step 3, each potential target area is segmented. Different clusters in the area may represent different meteorological targets. The corresponding values ​​in the area are obtained according to the pixel points of the regional image. One pixel corresponds to one value. The corresponding statistical information of these regional values ​​is extracted, including the reflectivity factor and the average radial texture (T DBZ ), the radial variation of echo intensity between reservoirs (S PIN ), the regional average of radial velocity (M DVE ), regional variance of radial velocity (S DVE ) and the regional average of the velocity spectrum width (M DSW ), the formula is as follows:

[0059]

[0060] Among them, N A 、N R Indicates the defined range in the direction of azimuth and distance, N A ×N R is the size of the window; Z i,j is the reflection factor of any point, T DBZ Mainly reflects the local change of echo intensity; Z thresh The threshold for echo intensity change between reservoirs is usually set at 2-5dB. PIN It reflects the consistency of the echo intensity change along the radial direction; Indicates intermediate parameters; Indicates the radial velocity value of a point in the velocity graph within the sliding window range. represents the average value within the sliding window size; S DVE represents the regional variance of radial velocity; M DSW represents the average value of the velocity spectrum width within the sliding window; W i,j Represents the spectral width at any point, V i,j Indicates the velocity of any point; for the characteristic quantity related to the echo intensity, set N A =3, N R =3; for the characteristic quantities related to radial velocity and velocity spectrum width, set N A =3, N R = 9. The above five statistical information serve as input for the subsequent fuzzy logic fusion algorithm.

[0061] Step 5: Non-meteorological target recognition using fuzzy logic fusion algorithm

[0062] After the image segmentation model divides the potential target area, there will be fuzzy boundaries and points in the area that are difficult to clearly classify. The fuzzy logic fusion algorithm is used to further subdivide the non-meteorological echo class and the meteorological echo class, which can help identify subtle differences and distinguish the echo type more accurately. The process is as follows: Figure 3 shown.

[0063] Based on the five parameter feature statistics extracted in step 4, a fuzzy membership function is defined to represent the different value ranges of the parameter feature statistics. The membership function of the fuzzy logic algorithm is defined: the output value range of the membership function is 0 to 1, and the larger the value, the higher the possibility that the point is a non-meteorological target. Based on the probability distribution of the five parameter feature statistics, a trapezoidal broken line is used to represent the membership function of each parameter feature statistical information to achieve the fuzzification of the parameter feature statistical information. The range of the trapezoidal membership function points and the trapezoidal inflection points of each parameter feature statistical information are roughly determined to ensure that the value range is between 0 and 1.

[0064] Fuzzy rules are designed to describe non-meteorological targets. These targets are identified based on the following criteria: clear-air echoes are located at low altitudes, have low radial velocities and spectral widths, and exhibit a low reflectivity factor. Ground object echoes also primarily appear in the lower layers, have radial velocities and spectral widths close to zero, exhibit no apparent motion, but exhibit a high reflectivity factor. Meteorological targets, such as precipitation echoes, are identified based on their shape, vertical structure, evolution, and motion.

[0065] Rule 1: If the average radial texture of the reflectivity factor is lower than the set minimum threshold, the radial variation of the echo intensity is within the set range, the regional average of the radial velocity is lower than the set minimum threshold, the regional variance of the radial velocity is lower than the set minimum variance threshold, and the regional average of the velocity spectrum width is lower than the set minimum spectrum width setting threshold, then it is judged to be a clear-air echo from a non-meteorological target.

[0066] Rule 2: If the reflectivity factor has a high average radial texture, the echo intensity has a small radial variation, the radial velocity area average is close to 0, the velocity variance is low, and the spectral width is low, then it is determined to be a ground object echo among non-meteorological targets.

[0067] The parametric characteristic statistics of each data point within the potential target area are substituted into the defined membership function to calculate the membership degree of each data point. The obtained membership degrees are defuzzified, that is, converted into specific numerical outputs. The defuzzified results are calculated using a weighted summation method. Each parametric characteristic statistic is assigned a different weight coefficient based on its influence, and all weight coefficients sum to 1 to obtain the final judgment value for each data point. A judgment threshold is set (0.6 for clear-air echoes and 0.4 for ground object echoes). When the defuzzified judgment value exceeds the judgment threshold, the data point is classified as a non-meteorological target echo; when it is below the judgment threshold, it is classified as a meteorological target echo. By adjusting the threshold value and the weight coefficients of the characteristic parametric statistics, the fuzzy boundaries of the potential target area and points within the area that are difficult to clearly classify are further clarified and defined, thereby improving the accuracy of non-meteorological target recognition.

[0068] Taking the meteorological data collected at sites 1, 2, 3, and 4 as an example, the results obtained by directly fine-tuning the SAM image segmentation model are as follows: Figure 4 、 Figure 5 、 Figure 6 、 Figure 7 As shown in the figure, the boundary area of ​​the non-meteorological target marked by the dotted rectangle is fuzzy, that is, the low-altitude non-meteorological target is not accurately identified and is mistakenly judged as a meteorological target. However, the meteorological data collected by a certain station 1, 2, 3, and 4 are clustered with meteorological feature images and fuzzy fusion to obtain the results of non-meteorological target recognition. Figure 8 、 9 As shown in Figures 10 and 11, the fuzzy boundary of the non-meteorological target area is optimized, and the non-meteorological targets can be accurately identified.

Claims

1. A non-meteorological target recognition method based on meteorological feature image clustering and fuzzy fusion, characterized in that: For the multi-parameter feature data of millimeter-wave cloud radar meteorological data, a large image segmentation model is used to achieve feature clustering of multi-parameter feature images. Then, a fuzzy logic fusion algorithm is used to achieve fusion judgment of the multi-feature clustering results and complete high-precision recognition of non-meteorological targets. The method specifically includes the following steps: Step 1: Preprocessing of millimeter wave cloud radar meteorological data; 1-1. Collect raw meteorological echo data from a certain site in millimeter-wave cloud radar THI mode throughout the day; 1-2. Filter the original meteorological echo data and extract meteorological data parameter features of the filtered meteorological echo data, including four parameter features: reflectivity factor Z, radial velocity V, spectral width W, and signal-to-noise ratio SNR. After normalization, save the data in array matrix form. Step 2: Generating parameter feature images of millimeter-wave cloud radar meteorological data; 2-1. Using pseudo-color processing, normalized meteorological data parameter features are mapped to different pseudo-colors according to their numerical values ​​to produce characteristic images of each meteorological data parameter; 2-2. Use image segmentation and clustering methods to analyze the characteristics of meteorological targets and non-meteorological targets on the characteristic images of various meteorological data parameters; Step 3: Clustering of potential targets in millimeter-wave cloud radar meteorological data parameter feature images 3-1. One-shot fine-tuning of the SAM image segmentation model; 3-2. Perform unsupervised clustering on each meteorological data parameter feature image. Based on the continuity and consistency of the parameter features in the temporal dimension, each meteorological data parameter feature image is divided into different regions. Each region represents a different echo feature, including background region, meteorological echo region, and non-meteorological echo region. The meteorological echo region and non-meteorological echo region are identified as potential target regions. Step 4: Extract parameter feature statistics of potential target areas Based on the potential target areas divided, statistical information on the characteristics of meteorological data parameters in each target area is extracted, including the average radial texture of the reflectivity factor, the radial variation of the echo intensity between reservoirs, the regional average of the radial velocity, the regional variance of the radial velocity, and the regional average of the velocity spectrum width. These information is used as input for the subsequent fuzzy logic fusion algorithm. Step 5: Non-meteorological target recognition using fuzzy logic fusion algorithm 5-1. Determine the fuzzy membership function of the statistical information of each meteorological data parameter characteristic based on the statistical information of each potential target area obtained; 5-2. Based on the typical parameter characteristics of non-meteorological targets, fuzzy logic rules are set to calculate the non-meteorological target fuzzy membership of each data point within the potential target area based on the statistical information of different parameter characteristics. The weighted sum of these fuzzy memberships is used to obtain the non-meteorological target possibility score for each data point. 5-3. Based on the pre-set threshold, each data point in the potential target area is determined to be a non-meteorological target or a meteorological target. The threshold value is adjusted to make the fuzzy boundaries of the potential target area and the points in the area that are difficult to clearly classify clearer and more specific, thereby improving the accuracy of non-meteorological target identification.

2. The non-meteorological target recognition method based on meteorological feature image clustering and fuzzy fusion according to claim 1 is characterized in that: The specific process of step 3-1 fine-tuning is as follows: First, a reference meteorological feature image I is given R and non-meteorological target area mask images Then, the image encoder ImEncoder in the trained SAM image segmentation model is used to extract the reference meteorological feature image I R Extract pixel value and position distribution of each pixel in the meteorological feature image I as semantic feature vector. The process is expressed as follows: F I =ImEncoder(I) F R =ImEncoder(I R ) in Represent the reference meteorological characteristic image I R and the semantic feature vector of the meteorological feature image I to be tested, h represents the height of the image, w represents the width of the image, and c represents the number of channels of the image; Then, according to the non-meteorological target area mask image M of the reference meteorological feature image R , calculate the non-meteorological target characteristics T p , the formula is: in, represents the matrix dot product; Next, calculate T P With F I The cosine similarity between them is used to obtain the confidence probability map S of the meteorological feature image I to be tested. p , using the statistical average confidence probability map S p Aggregate into an overall confidence map 3. The non-meteorological target recognition method based on meteorological feature image clustering and fuzzy fusion according to claim 2 is characterized in that: Step 3-2 is implemented as follows: Select the two pixels with the highest and lowest confidence values ​​in the overall confidence map S, denoted as P h and P l ;P h Indicates the most likely center pixel point of the non-meteorological target area, P l P is the pixel farthest from the non-meteorological target. h and P l They are regarded as positive and negative prompt points and input into the prompt encoder in the SAM image segmentation model to guide the SAM image segmentation model to complete the regional clustering of non-meteorological targets. In this way, the SAM model will complete the continuous region clustering around the positive prompt points and segment out different potential target areas.

4. The non-meteorological target recognition method based on meteorological feature image clustering and fuzzy fusion according to claim 3 is characterized in that: Step 4 is implemented as follows: Through step 3, we can get each potential target area. Different clusters in the area may represent different meteorological targets. We can get the corresponding values ​​in the area according to the pixel points of the regional image. One pixel corresponds to one value. We can extract the corresponding statistical information of these regional values, including the reflectivity factor and the average radial texture T. DBZ , the radial variation of echo intensity between reservoirs S PIN , the regional average of the radial velocity M DVE , regional variance of radial velocity S DVE and the regional average value of the velocity spectrum width M DSW , the formula is as follows: Among them, N A 、N R Indicates the defined range in the direction of azimuth and distance, N A ×N R is the size of the window; Z i,j is the reflection factor of any point, T DBZ Mainly reflects the local change of echo intensity; Z thresh is the threshold of echo intensity change between reservoirs, S PIN It reflects the consistency of the echo intensity change along the radial direction; Indicates intermediate parameters; Indicates the radial velocity value of a point in the velocity graph within the sliding window range. represents the average value within the sliding window size; S DVE represents the regional variance of radial velocity; M DSW represents the average value of the velocity spectrum width within the sliding window; W i,j Represents the spectral width at any point, V i,j Indicates the velocity at any point.

Images