Fine monitoring edge computing system and method based on weather radar IQ data

Abstract

This invention relates to the fields of radar remote sensing and edge computing technology, specifically to an edge computing system and method for fine-grained monitoring of insects and birds based on weather radar IQ data. The system is deployed at a weather radar station and directly acquires raw complex IQ data. The system includes: an edge IQ data preprocessing module for clutter suppression and time-frequency domain transformation; a multi-dimensional biofeature extraction module for extracting micro-Doppler features, dual-polarization parameter features, and radar cross-section features; a cascaded classification and identification module using a two-stage classifier to filter non-biogenic echoes and distinguish between insects and birds; a spatiotemporal trajectory tracking module for obtaining target trajectories; a biomass inversion and product generation module for inverting biomass density and generating structured monitoring products; and an edge cloud collaboration module for uploading products to the cloud. This invention directly processes raw complex IQ data through an edge computing architecture, preserving fine features such as micro-Doppler, to achieve fine-grained identification, real-time monitoring, and quantitative analysis of insect and bird targets.

Description

Technical Field

[0001] This invention relates to the fields of radar remote sensing and edge computing technology, specifically to an edge computing system and method for fine monitoring of insects and birds based on weather radar IQ data. Background Technology

[0002] Aerial ecological monitoring is of great significance for early warning of agricultural pests and diseases, aviation safety, and ecological research. Weather radar networks, as a mature atmospheric remote sensing infrastructure, can transmit echoes that can be scattered by aerial biological targets, making large-scale aerial ecological monitoring possible. Related research and operational trials have been conducted both domestically and internationally, but many technical bottlenecks remain.

[0003] Existing technologies suffer from the following drawbacks: They rely on meteorological product data for analysis, and the phase information and micro-Doppler characteristics of the original signals are lost on average during product generation, making it impossible to distinguish between insects and birds, let alone perform species subdivision; threshold-based fuzzy logic recognition methods have low accuracy and poor adaptability, with a significant drop in accuracy when biological targets are mixed with light precipitation; centralized processing architectures require uploading nationwide radar data to a central processing center, making massive IQ data transmission difficult and resulting in long processing delays, failing to meet real-time early warning requirements; and they lack the ability to invert fine information such as biomass density and migration trajectories from the raw signal level. These problems limit the application of weather radar in operational ecological monitoring. Summary of the Invention

[0004] This invention provides an edge computing system and method for fine monitoring of insects and birds based on weather radar IQ data, aiming to solve the problems of severe information loss, insufficient identification accuracy, and poor real-time performance in existing weather radar ecological monitoring.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] This invention relates to an edge computing system for fine-grained monitoring of insects and birds based on weather radar IQ data, comprising:

[0007] The edge IQ data preprocessing module is deployed on the weather radar station side. It accesses the weather radar digital intermediate frequency receiver through the edge computing node to obtain the raw complex IQ data, and performs ground clutter suppression and time-frequency domain transformation on the raw complex IQ data to obtain preprocessed IQ data.

[0008] The multidimensional biometric feature extraction module extracts multidimensional features from the preprocessed IQ data, including micro-Doppler features, dual polarization parametric features, and radar cross section features; wherein, the micro-Doppler features include periodic frequency modulation components generated by insect wing flapping and bird wing flapping.

[0009] The cascaded classification and recognition module uses a cascaded classifier to fuse and distinguish the micro-Doppler features, dual polarization parameter features, and radar cross section features. The cascaded classifier includes a first-level classifier and a second-level classifier. The first-level classifier filters out non-biological echoes, and the second-level classifier fuses the micro-Doppler features and dual polarization parameter features to distinguish between insect targets and bird targets.

[0010] The spatiotemporal trajectory tracking module tracks the trajectory of the insect and bird targets to obtain their spatiotemporal trajectories.

[0011] The biomass inversion and product generation module inverts biomass density and generates structured monitoring products based on the classification of insect and bird targets, the radar cross section characteristics, and the spatiotemporal trajectory.

[0012] The edge cloud collaboration module uploads the structured monitoring product to the cloud data center after the edge computing node completes data processing.

[0013] As a preferred embodiment of the present invention, the ground clutter suppression in the edge IQ data preprocessing module includes:

[0014] Acquire sounding data and numerical model wind field data provided by meteorological departments;

[0015] A dynamic clutter map is constructed based on the aforementioned radiosonde data and numerical model wind field data;

[0016] Based on the dynamic clutter map, an adaptive digital filter is used to filter the original complex IQ data to suppress fixed ground clutter and radio frequency interference.

[0017] As a preferred embodiment of the present invention, the time-frequency domain transformation in the edge IQ data preprocessing module includes:

[0018] The raw complex IQ data after ground clutter suppression is processed by pulse pair processing or fast Fourier transform to generate a power spectral density containing amplitude information, phase information and spectral characteristic information.

[0019] As a preferred embodiment of the present invention, the multidimensional biometric feature extraction module extracts micro-Doppler features in the following manner:

[0020] Perform a short-time Fourier transform on the preprocessed IQ data to generate a time-frequency spectrum.

[0021] The time-spectrum diagram is decomposed using a subspace decomposition method to separate the main Doppler velocity component and the micro-Doppler sideband component;

[0022] The frequency, amplitude, and modulation period are extracted from the microDoppler sideband components;

[0023] Among them, the micro-Doppler sideband component generated by insect wing flapping exhibits periodic high-frequency modulation characteristics, while the micro-Doppler sideband component generated by bird wing flapping exhibits low-frequency non-stationary modulation characteristics.

[0024] As a preferred embodiment of the present invention, the multidimensional biometric feature extraction module extracts dual polarization parameter features in the following ways:

[0025] The echo signals of the horizontal polarization channel and the vertical polarization channel are extracted from the preprocessed IQ data, respectively.

[0026] The differential reflectivity is obtained by calculating the ratio of the echo power of the horizontal polarization channel to the echo power of the vertical polarization channel.

[0027] Calculate the phase difference between the echo signal of the horizontal polarization channel and the echo signal of the vertical polarization channel to obtain the differential phase;

[0028] Calculate the correlation coefficient between the echo signal of the horizontal polarization channel and the echo signal of the vertical polarization channel;

[0029] Spatial texture analysis is performed on the differential reflectivity, differential phase, and correlation coefficient to extract texture statistical features.

[0030] As a preferred embodiment of the present invention, the multidimensional biometric feature extraction module extracts radar cross-section features in the following ways:

[0031] The echo power is calculated from the preprocessed IQ data;

[0032] The radar cross section is calculated using radar equations based on the echo power and radar system parameters.

[0033] Extract the temporal variation characteristics of the radar cross section.

[0034] As a preferred embodiment of the present invention, the cascaded classification and recognition module performs fusion discrimination in the following manner:

[0035] The dual polarization parameter features and the radar cross section features are input into the first-level classifier;

[0036] The first-level classifier filters out turbulent echoes based on the high variability of the texture statistical features of the dual polarization parametric features, where the correlation coefficient is lower than a preset threshold and the differential reflectivity is high.

[0037] The first-level classifier filters out abnormal propagation echoes and retains candidate biological targets based on the non-physical jump in the differential phase.

[0038] The micro-Doppler features, dual polarization parametric features, and radar cross section features of the candidate biological targets are input into the second-level classifier;

[0039] The second-level classifier distinguishes between insect targets and bird targets based on the frequency range and spectral characteristics of the periodic modulation components in the micro-Doppler features, combined with the texture statistics of the dual polarization parametric features and the temporal variation characteristics of the radar cross section.

[0040] Based on the frequency characteristics of the micro-Doppler features and the amplitude range of the radar cross section, the insect targets are subdivided into moths or locusts, and the bird targets are subdivided into large birds or small birds.

[0041] As a preferred embodiment of the present invention, the spatiotemporal trajectory tracking module operates as follows:

[0042] A sequence of target points is established in a three-dimensional spherical coordinate system, which includes distance, azimuth, and elevation.

[0043] Calculate the correlation probability between the detected target point trace at the current moment and the historical trajectory;

[0044] Data association is performed based on the association probability. Trajectory initiation is performed for newly emerging target points, trajectory maintenance is performed for already associated target points, and trajectory termination is performed for continuous unassociated trajectories.

[0045] Output the target's continuous spatiotemporal trajectory, flight speed, flight direction, and flight altitude.

[0046] As a preferred embodiment of the present invention, the operation mode of the biomass inversion and product generation module includes:

[0047] Establish a correlation model between radar cross section and biological body size;

[0048] Based on the aforementioned correlation model, the radar cross section of the detected target is converted into the size of a biological individual;

[0049] The number of targets and the size of the biological individuals in different spatial grids within the monitoring airspace are statistically analyzed, and the biomass density of the spatial grids is calculated.

[0050] Cluster analysis is performed on the spatiotemporal trajectories to identify migration paths, clustering areas, and diffusion directions;

[0051] The classification labels, spatiotemporal trajectories, biomass density, migration paths, aggregation areas, and diffusion directions of the insect and bird targets are encapsulated into structured monitoring products in JSON or GIS layer format.

[0052] This invention also proposes a method for fine-grained edge computing of insect and bird monitoring based on weather radar IQ data, including:

[0053] Edge computing nodes are deployed at the weather radar station. The edge computing nodes are connected to the digital intermediate frequency receiver of the weather radar to obtain raw complex IQ data. Ground clutter suppression and time-frequency domain transformation are performed on the raw complex IQ data to obtain preprocessed IQ data.

[0054] Multi-dimensional features are extracted from the preprocessed IQ data, including micro-Doppler features, dual polarization parametric features, and radar cross section features; wherein, the micro-Doppler features include periodic frequency modulation components generated by insect wing flapping and bird flapping.

[0055] A cascaded classifier is used to fuse and distinguish the micro-Doppler features, dual polarization parameter features, and radar cross section features. The cascaded classifier includes a first-level classifier and a second-level classifier. The first-level classifier filters out non-biological echoes, and the second-level classifier fuses the micro-Doppler features and dual polarization parameter features to distinguish between insect targets and bird targets.

[0056] Trajectory tracking is performed on the insect and bird targets to obtain their spatiotemporal trajectories;

[0057] Based on the classification of insect and bird targets, the radar cross section characteristics, and the spatiotemporal trajectory, biomass density is retrieved and structured monitoring products are generated.

[0058] After the edge computing node completes data processing, the structured monitoring product is uploaded to the cloud data center.

[0059] The beneficial effects of this invention are:

[0060] 1. This invention directly processes raw complex IQ data from radar, fully preserving the amplitude, phase, and micro-Doppler information of biological targets, and extracting wing flapping and flapping features that are completely lost in traditional product data. Through the fusion of multi-dimensional features from micro-Doppler, dual polarization, and radar cross-section, it achieves fine differentiation and species subdivision between insects and birds, significantly improving identification accuracy and robustness.

[0061] 2. By adopting an edge computing architecture to offload data processing to the weather radar station side, the bandwidth bottleneck of remote transmission of massive amounts of IQ data is eliminated, reducing end-to-end latency from hours to minutes or even seconds. Edge nodes only upload highly compressed structured monitoring products, significantly reducing the amount of data and making large-scale ecological monitoring operations across the national radar network economically feasible.

[0062] 3. By using a correlation model between radar cross section and organism size, the system enables quantitative inversion of biomass density, individual size, and migration trajectory, providing refined monitoring information. Combined with a cascaded classifier and spatiotemporal trajectory tracking, the system can identify migration paths, aggregation areas, and diffusion directions, meeting the application needs of real-time pest and disease control and immediate aviation safety alerts. Attached Figure Description

[0063] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0064] Figure 1 This is a schematic diagram of the edge computing system for fine monitoring of insects and birds based on weather radar IQ data according to the present invention;

[0065] Figure 2 This is a flowchart illustrating the edge computing method for fine-grained monitoring of insects and birds based on weather radar IQ data, as per the present invention. Detailed Implementation

[0066] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0067] Example 1: As Figure 1 As shown, the present invention provides an edge computing system for fine-grained monitoring of insects and birds based on weather radar IQ data, comprising:

[0068] The edge IQ data preprocessing module is deployed on the weather radar station side. It accesses the weather radar digital intermediate frequency receiver through the edge computing node to obtain the raw complex IQ data, and performs ground clutter suppression and time-frequency domain transformation on the raw complex IQ data to obtain preprocessed IQ data.

[0069] Specifically, the edge IQ data preprocessing module is deployed at the weather radar station and directly connects to the weather radar's digital intermediate frequency receiver via an edge computing node to acquire raw complex IQ data in real time. The raw complex IQ data is the in-phase component output by the radar receiver. and orthogonal components The complex sequence formed is represented as:

[0070] ;

[0071] in, for The original complex IQ signal at time t, for In-phase components at time, for Orthogonal components at time points, Sampling time, The imaginary unit is used. This complex sequence fully preserves the amplitude and phase information of the echo signal, serving as the original data basis for subsequent extraction of micro-features of biological targets.

[0072] Furthermore, the ground clutter suppression in the edge IQ data preprocessing module includes:

[0073] Acquire sounding data and numerical model wind field data provided by meteorological departments;

[0074] A dynamic clutter map is constructed based on the aforementioned radiosonde data and numerical model wind field data;

[0075] Based on the dynamic clutter map, an adaptive digital filter is used to filter the original complex IQ data to suppress fixed ground clutter and radio frequency interference.

[0076] Specifically, edge computing nodes acquire radiosonde data and numerical model wind field data provided by meteorological departments through standard meteorological data interfaces. The radiosonde data includes temperature, humidity, wind speed, and wind direction information at different altitudes, while the numerical model wind field data is a three-dimensional wind field distribution output by a numerical weather prediction model.

[0077] Radial velocities of meteorological targets at different distances, azimuths, and elevations are estimated using radiosonde wind field information, and a velocity-spatial distribution map is established as a dynamic clutter map. This dynamic clutter map identifies areas at different spatial locations that may be occupied by meteorological echoes and their corresponding radial velocity ranges. Specifically, based on the altitude in the radiosonde data... wind speed at the location and wind direction Combined with the azimuth angle of the radar beam direction and elevation angle Calculate the radial velocity of the meteorological target at that location:

[0078] ;

[0079] The calculated radial velocity Spatial location Correspondingly, among which For the distance library, a dynamic clutter map is generated. Compared to traditional static clutter maps, which can only identify fixed ground features, dynamic clutter maps can reflect the spatial distribution and movement of meteorological echoes in real time.

[0080] Based on the dynamic clutter map, an adaptive digital filter is used to filter the original complex IQ data. The adaptive digital filter can employ either the Least Mean Square (LMS) algorithm or the Recursive Least Squares (RLS) algorithm. The key is to adjust the suppression strategy according to the dynamic clutter map: for the meteorological echo areas marked by the dynamic clutter map, the filter suppresses the corresponding radial velocity range. The spectral components within, where The velocity tolerance is determined based on the wind speed measurement uncertainty of the radiosonde data, typically taken as 1-2 m / s; for ground clutter, the filter suppresses the area near zero frequency. Fixed echo components within the m / s range; for radio frequency interference, interference signals at specific frequencies are suppressed by frequency domain notch filtering.

[0081] Traditional weather radar clutter suppression can over-suppress low-speed moving targets. However, guided by dynamic clutter maps, filters can remove weather clutter while preserving the micro-motion signal components unique to biological targets, particularly the micro-Doppler sidebands in the spectrum. These micro-Doppler sidebands are typically distributed in the range of several to tens of hertz on either side of the main Doppler peak. If they are mistakenly filtered out by traditional filters, subsequent insect and bird identification becomes impossible. The filtered IQ data sequence is obtained after clutter suppression. .

[0082] Furthermore, the time-frequency domain transformation in the edge IQ data preprocessing module includes:

[0083] The raw complex IQ data after ground clutter suppression is processed by pulse pair processing or fast Fourier transform to generate a power spectral density containing amplitude information, phase information and spectral characteristic information.

[0084] Specifically, for the IQ data sequence after clutter suppression Perform time-frequency domain transformation. For continuously received signals... Perform a Fast Fourier Transform (FFT) on the IQ data of each pulse:

[0085] ;

[0086] in, The total number of pulses participating in the transformation. The sampling time corresponding to each pulse. ( ) is the first Each pulse corresponds to a complex IQ signal value at the sampling time. For Doppler frequency, This is the power spectrum in the frequency domain. Calculate the power spectral density:

[0087] ;

[0088] in, For power spectral density, The amplitude squared of the power spectrum in the frequency domain. This represents the total number of pulses participating in the transformation. The power spectral density contains three types of information: amplitude information is... Characterization, reflecting the target echo intensity; phase information is provided by The phase angle is represented, preserving the phase change characteristics of the signal; the spectral feature information is... The frequency distribution is characterized, where the main Doppler peak corresponds to the radial velocity of the target, and the micro-Doppler sidebands correspond to the micro-motion characteristics of the target.

[0089] To further extract time-varying micro-Doppler features, a short-time Fourier transform (STFT) was performed on the IQ data to generate a time-spectrum graph:

[0090] ;

[0091] in, for Time, frequency The time-frequency spectrum at that location. For complex IQ data sequences after ground clutter suppression The value at time, For The sliding time window function centered on the center. For integration variables, The center moment of the time window, For Doppler frequency, The window function and window length directly affect the time-frequency resolution: This implementation uses the Hanning window, and the window length is determined based on the micro-Doppler frequency range to be detected. For S-band weather radar, the pulse repetition frequency is typically around 1000Hz. To distinguish between insect wingbeat frequencies (20-80Hz) and bird wingbeat frequencies (2-10Hz), a frequency resolution better than 10Hz is required. Therefore, the window length is set to 10-30 pulse repetition cycles. This implementation preferably uses 20 pulse repetition cycles. A window length that is too short (less than 10 pulse cycles) will result in insufficient frequency resolution, making it impossible to distinguish the micro-Doppler sidebands; a window length that is too long (more than 50 pulse cycles) will result in a loss of time resolution, making it impossible to capture the instantaneous changes in wingbeats or flapping. (Time-frequency spectrum diagram) The vertical axis represents frequency, the horizontal axis represents time, and the color intensity represents energy, which can clearly show the periodic frequency modulation patterns produced by the flapping of insect wings or the flapping of birds' wings.

[0092] After the above processing, the obtained preprocessed IQ data includes IQ data sequences after clutter suppression. Power spectral density and time spectrum These data fully preserve the time-domain, frequency-domain, and time-frequency-domain characteristics of biological targets, providing a comprehensive data foundation for subsequent multidimensional feature extraction. Compared with traditional meteorological product data (which only contains single parameters such as average reflectivity and average velocity), preprocessed IQ data retains the complete phase information and instantaneous frequency change characteristics of the signal, making it possible to extract fine features such as micro-Doppler.

[0093] The multidimensional biometric feature extraction module extracts multidimensional features from the preprocessed IQ data, including micro-Doppler features, dual polarization parametric features, and radar cross section features; wherein, the micro-Doppler features include periodic frequency modulation components generated by insect wing flapping and bird wing flapping.

[0094] Furthermore, the multidimensional biometric extraction module extracts micro-Doppler features in the following ways:

[0095] Perform a short-time Fourier transform on the preprocessed IQ data to generate a time-frequency spectrum.

[0096] The time-spectrum diagram is decomposed using a subspace decomposition method to separate the main Doppler velocity component and the micro-Doppler sideband component;

[0097] The frequency, amplitude, and modulation period are extracted from the microDoppler sideband components;

[0098] Among them, the micro-Doppler sideband component generated by insect wing flapping exhibits periodic high-frequency modulation characteristics, while the micro-Doppler sideband component generated by bird wing flapping exhibits low-frequency non-stationary modulation characteristics.

[0099] Specifically, a short-time Fourier transform is performed on the preprocessed IQ data to generate a time-frequency spectrum. The time-spectrum diagram is decomposed using a subspace decomposition method to separate the main Doppler velocity component and the micro-Doppler sideband component.

[0100] Construct the time-spectrum graph into a matrix Perform singular value decomposition on it:

[0101] ;

[0102] in, For the time spectrum diagram The discretized two-dimensional matrix has rows corresponding to time frames and columns corresponding to frequency points. and It is a singular vector matrix. It is a singular value diagonal matrix. This represents the conjugate transpose. After performing singular value decomposition on the time-spectrum matrix, a separation threshold is set based on the energy distribution of the singular values. (Setting the singular value threshold) ( (This is the maximum singular value). This threshold corresponds to the energy difference between the main Doppler component and the micro-Doppler sideband components. Typically, the main Doppler component accounts for 70%-90% of the total energy. Components greater than the threshold are classified as main Doppler components, while those less than the threshold but greater than the noise level (set as...) are classified as... The component with a signal-to-noise ratio of approximately 20 dB is classified as a micro-Doppler sideband component, thus achieving signal subspace separation.

[0103] Frequency extraction from microDoppler sideband components Amplitude and modulation period The standard deviation is calculated from the modulation period sequence at consecutive time points. Quantify the stability of micro-Doppler.

[0104] The micro-Doppler characteristics of insect wingbeats are: wingbeat frequency High periodic stability in the 20-80Hz range ( The time-spectral graph shows regular periodic stripes. The micro-Doppler characteristics of bird wing flapping are: wingbeat frequency... Low periodic stability in the 2-10Hz range ( The time-spectrum diagram shows an irregular modulation pattern.

[0105] The micro-Doppler eigenvectors are: .

[0106] Furthermore, the multidimensional biometric extraction module extracts dual polarization parametric features in the following ways:

[0107] The echo signals of the horizontal polarization channel and the vertical polarization channel are extracted from the preprocessed IQ data, respectively.

[0108] The differential reflectivity is obtained by calculating the ratio of the echo power of the horizontal polarization channel to the echo power of the vertical polarization channel.

[0109] Calculate the phase difference between the echo signal of the horizontal polarization channel and the echo signal of the vertical polarization channel to obtain the differential phase;

[0110] Calculate the correlation coefficient between the echo signal of the horizontal polarization channel and the echo signal of the vertical polarization channel;

[0111] Spatial texture analysis is performed on the differential reflectivity, differential phase, and correlation coefficient to extract texture statistical features.

[0112] Specifically, horizontal polarization channel echo signals are extracted from the preprocessed IQ data. and vertical polarization channel echo signal .

[0113] Calculate differential reflectance:

[0114] ;

[0115] in, This is the echo signal from the horizontally polarized channel. This is the echo signal from the vertical polarization channel. Differential reflectivity reflects the difference in a target's scattering ability in the horizontal and vertical directions, and is measured in dB.

[0116] Calculate the differential phase:

[0117] ;

[0118] in, This represents the phase angle when taking a complex number. Differential phase represents the phase difference between the echo signals from the horizontal and vertical polarization channels, expressed in degrees (°).

[0119] Calculate the correlation coefficient:

[0120] ;

[0121] in, The correlation coefficient, To indicate complex conjugate, This represents the time average over 10-20 consecutive pulses. (Biological target) The value is usually no higher than 0.90, while the precipitation target is usually higher than 0.95.

[0122] Spatial texture analysis was performed on differential reflectance, differential phase, and correlation coefficient. The current pixel was selected as the center. Neighborhood window, calculate the mean and standard deviation of each polarization parameter: and Differential reflectance within the neighborhood Spatial mean and standard deviation; and Differential phases within the neighborhood Spatial mean and standard deviation; and Correlation coefficients within the neighborhood The spatial mean and standard deviation. Due to their uneven spatial distribution, biological targets exhibit high variability with a large standard deviation.

[0123] The eigenvectors of the dual polarization parameters are:

[0124] .

[0125] Furthermore, the multidimensional biometric extraction module extracts radar cross-section features in the following ways:

[0126] The echo power is calculated from the preprocessed IQ data;

[0127] The radar cross section is calculated using radar equations based on the echo power and radar system parameters.

[0128] Extract the temporal variation characteristics of the radar cross section.

[0129] Specifically, the echo power is calculated from the preprocessed IQ data. ,in To suppress clutter in the IQ data sequence, time averaging of multiple consecutive pulses is performed during actual calculations to obtain a stable power estimate. The radar cross section is then calculated based on the radar equations. :

[0130] ;

[0131] in, Radar cross section, For transmission power, For antenna gain, For radar wavelength, The target distance is denoted as , and all parameters are known to the radar system.

[0132] Extracting radar cross section The temporal variation characteristics of the radar cross section at continuous time points. Calculate the time-series mean Time series standard deviation The dominant frequency was extracted using Fourier transform. .

[0133] The temporal fluctuations of biological targets are significant. Insects Typical value to m², At 20-80Hz; birds Typical value to m², Between 2-10Hz.

[0134] Radar cross section eigenvector The calculation method is as follows: .

[0135] in, for Time series mean for Time series standard deviation for The dominant oscillation frequency obtained by Fourier transforming a time series.

[0136] Combining the above three types of features, a fused feature vector is output. It has a total of 16 feature components, which provide multi-dimensional discrimination basis for subsequent cascaded classification.

[0137] The cascaded classification and recognition module uses a cascaded classifier to fuse and distinguish the micro-Doppler features, dual polarization parameter features, and radar cross section features. The cascaded classifier includes a first-level classifier and a second-level classifier. The first-level classifier filters out non-biological echoes, and the second-level classifier fuses the micro-Doppler features and dual polarization parameter features to distinguish between insect targets and bird targets.

[0138] Furthermore, the cascaded classification and recognition module performs fusion discrimination in the following ways:

[0139] The dual polarization parameter features and the radar cross section features are input into the first-level classifier;

[0140] The first-level classifier filters out turbulent echoes based on the high variability of the texture statistical features of the dual polarization parametric features, where the correlation coefficient is lower than a preset threshold and the differential reflectivity is high.

[0141] The first-level classifier filters out abnormal propagation echoes and retains candidate biological targets based on the non-physical jump in the differential phase.

[0142] The micro-Doppler features, dual polarization parametric features, and radar cross section features of the candidate biological targets are input into the second-level classifier;

[0143] The second-level classifier distinguishes between insect targets and bird targets based on the frequency range and spectral characteristics of the periodic modulation components in the micro-Doppler features, combined with the texture statistics of the dual polarization parametric features and the temporal variation characteristics of the radar cross section.

[0144] Based on the frequency characteristics of the micro-Doppler features and the amplitude range of the radar cross section, the insect targets are subdivided into moths or locusts, and the bird targets are subdivided into large birds or small birds.

[0145] Specifically, the dual polarization parameter features and radar cross section features are input into the first-level classifier. The first-level classifier is based on the support vector machine (SVM) or random forest algorithm, which uses polarization parameters and physical features to quickly filter out non-biological interference.

[0146] The first-level classifier is based on the correlation coefficient. Below the preset threshold, differential reflectance texture standard deviation It exhibits high variability and filters out turbulent echoes. The specific criterion is: when... and When the value reaches dB, it is identified as a turbulent echo and filtered out. Turbulent echoes are generated due to the inhomogeneity of the refractive index caused by atmospheric turbulence, and their polarization characteristics exhibit low correlation and high spatial variability.

[0147] The first-level classifier uses the differential phase... If a non-physical jump occurs, abnormal propagation echoes are filtered out. The specific criterion is: calculate the differential phase gradient between adjacent distance bins. ,in The differential phase difference between adjacent storage units. The distance between adjacent libraries. When the absolute value of the gradient exceeds a threshold (usually set to...). When the distance is less than 1 km, it is identified as an anomalous propagation echo and filtered out. Anomalous propagation echoes are usually caused by anomalous propagation conditions such as atmospheric waveguides and superrefractive patterns, and the differential phase exhibits a drastic non-physical jump.

[0148] After screening by the first-level classifier, candidate biological targets are retained, with a filtering rate typically reaching 60%-80%, effectively reducing the computational burden of subsequent processing.

[0149] The micro-Doppler features, dual-polarization parameter features, and radar cross-section features of candidate biological targets are input into the second-level classifier. The second-level classifier employs a deep learning model to fuse multi-dimensional features for refined identification. The deep learning model is an attention-enhanced neural network, including a feature fusion layer, an attention layer, and a classification layer.

[0150] The second-level classifier distinguishes insect targets from bird targets based on the frequency range and spectral characteristics of the periodic modulation components in the micro-Doppler features, combined with the texture statistics of the dual polarization parametric features and the temporal variation characteristics of the radar cross section.

[0151] The criterion for identifying insect targets is: wingbeat frequency. High modulation period stability in the 20-80Hz range ( ), with a smaller radar cross section ( m²) and dominant frequency With wingbeat frequency Close, correlation coefficient Typically in the range of 0.7-0.85. The time spectrum shows regular periodic stripes with uniform stripe spacing.

[0152] The criterion for identifying bird targets is: wingbeat frequency. In the 2-10Hz range, the modulation period stability is low. ), with a large radar cross section ( m²) and dominant frequency It exhibits low-frequency characteristics, and the correlation coefficient is low. Typically in the range of 0.75-0.90. The time-spectrum graph shows an irregular modulation pattern, reflecting the non-periodicity of the flapping wings.

[0153] Based on the frequency characteristics of micro-Doppler features and the amplitude range of the radar cross section, insect targets are further subdivided into moths or locusts, and bird targets are further subdivided into large birds or small birds.

[0154] Moths are characterized by wingbeat frequencies typically between 40-80 Hz and radar cross-sections. exist to m² range, microDoppler intensity Relatively weak. Locusts are characterized by wingbeat frequencies typically between 20-50 Hz and a low radar cross-section. exist to m² range, microDoppler intensity Relatively strong.

[0155] Large birds are characterized by: wingbeat frequencies typically between 2-5 Hz and radar cross-sections. Greater than m², differential reflectivity High (typically greater than 2 dB). Small birds are characterized by: wingbeat frequencies typically between 5-10 Hz, and radar cross-sections. exist to Within a m² area, the spatial distribution exhibits clustering, and the texture standard deviation... Relatively large.

[0156] The second-level classifier outputs a classification label (insect / bird and its subclasses) and classification confidence score for each target. The confidence score is calculated using the softmax function of the neural network output layer, with a value ranging from 0 to 1. When the confidence score is below a threshold (usually set to 0.7), it is marked as an "undetermined" category for manual review.

[0157] Through a two-stage filtering process using a cascaded classifier, progressive identification from raw echoes to fine-grained classification is achieved, ensuring both computational efficiency and classification accuracy. The first-stage classifier quickly filters out a large amount of non-biological interference, while the second-stage classifier focuses on the fine-grained identification of biological targets. The two stages work together, and the identification accuracy typically reaches over 85%.

[0158] The spatiotemporal trajectory tracking module tracks the trajectory of the insect and bird targets to obtain their spatiotemporal trajectories.

[0159] Furthermore, the spatiotemporal trajectory tracking module operates as follows:

[0160] A sequence of target points is established in a three-dimensional spherical coordinate system, which includes distance, azimuth, and elevation.

[0161] Calculate the correlation probability between the detected target point trace at the current moment and the historical trajectory;

[0162] Data association is performed based on the association probability. Trajectory initiation is performed for newly emerging target points, trajectory maintenance is performed for already associated target points, and trajectory termination is performed for continuous unassociated trajectories.

[0163] Output the target's continuous spatiotemporal trajectory, flight speed, flight direction, and flight altitude.

[0164] Specifically, the spatiotemporal trajectory tracking module tracks the insect and bird targets to obtain their spatiotemporal trajectories. This module employs an improved Joint Probabilistic Data Association (JPDA) filtering and Multiple Hypothesis Tracking (MHT) algorithm for trajectory management in a three-dimensional spherical coordinate system.

[0165] A sequence of target points is established in a three-dimensional spherical coordinate system. This three-dimensional spherical coordinate system has the radar as its origin and includes the range... Azimuth and elevation angle Three coordinate components. Distance Radial distance from the target to the radar, azimuth angle The angle of elevation of the target on the horizontal plane. The pitch angle of the target relative to the horizontal plane.

[0166] For the The first scan time detected the first The location of each target point is represented as: , The first The first scan time detected the first The distance, azimuth, and elevation angles of each target point are given, with the superscript T indicating matrix transpose. Converting the 3D spherical coordinates to Cartesian coordinates facilitates subsequent calculations.

[0167] ;

[0168] ;

[0169] ;

[0170] in, , , These represent the three-dimensional position components of the target in Cartesian coordinates. The radial distance to the target. It is the azimuth angle. The angle of elevation.

[0171] Simultaneously record the radial velocity of each point. Classification labels and feature vectors are used to form complete point information.

[0172] Then, the correlation probability between the detected target point and the historical trajectory is calculated. For the _th Based on historical trajectories, predict the location of each trajectory according to its motion state (position, velocity). Kalman filtering is used for motion state prediction, assuming that the target is moving at a constant velocity or with uniform acceleration.

[0173] Calculate the predicted location Compared with the actual detection points Distance between:

[0174] ;

[0175] in, For the first Time of the first The first point and the second The Euclidean distance between the predicted locations of the historical trajectories. Denotes the Euclidean norm. The actual location of the detected spot. For the first The historical trajectory in the first Predicted location at any given time.

[0176] Set association threshold This is typically determined based on the target's speed and radar scanning interval. For insect targets, The distance is typically set at 200-500 meters; for bird targets, It is usually set at 500-1000 meters. When When calculating the association probability:

[0177] ;

[0178] in, For the first The first point and the second The correlation probability of historical trajectories Let be the Euclidean distance between the two. This represents the standard deviation of the association threshold. The consistency of the classification labels is also considered; if the classification labels of the points and the trajectory are inconsistent, the association probability decreases by 50%.

[0179] Data association is performed based on the aforementioned association probabilities. The JPDA algorithm is used to handle association ambiguity in multi-object scenarios. When the association probabilities of a point with multiple trajectories are all high, JPDA calculates the joint probability of various association hypotheses and selects the association scheme with the highest probability. When ambiguity still exists in the association, the MHT algorithm is used to maintain multiple hypotheses and delay the decision until subsequent observations eliminate the ambiguity.

[0180] A new trajectory is initiated for newly emerging target points. When a point cannot be associated with an existing trajectory for 2-3 consecutive scan times and meets the minimum signal-to-noise ratio requirement, a new trajectory is initialized. The initial state of the new trajectory is the position and velocity of the point, and the initial covariance is set according to the measurement accuracy.

[0181] Trajectory maintenance is performed on the associated target points. The trajectory state is updated using a Kalman filter, fusing predicted and measured values.

[0182] ;

[0183] in, The updated trajectory state includes the target's position, velocity, and acceleration components. For trajectory indexing, For the scan time index, For the first The trajectory in the first State estimation at time; The Kalman gain is determined by the prediction covariance and the observation noise covariance, reflecting the correction weight of the observation information on the state estimation. The observation residual is the difference between the actual detected position and the predicted position.

[0184] Terminate the trajectory for consecutive, unrelated trajectories. When a trajectory is continuous... Each scan time (usually) If no associated points are obtained (set to 5-10), the target is determined to have disappeared or left the monitoring range, the trajectory is terminated and archived.

[0185] Output the target's continuous spatiotemporal trajectory, flight speed, flight direction, and flight altitude.

[0186] The spatiotemporal trajectory is a sequence of the position points of the trajectory at each time step. ,in Record the target's movement path for the total number of scan times contained in the trajectory.

[0187] Flight speed includes radial velocity (obtained directly from Doppler measurements) and tangential velocity (Calculated based on position changes at adjacent time points). The target's total flight speed is:

[0188] ;

[0189] in, For the target total flight speed, Radial velocity, This represents the tangential velocity.

[0190] The flight direction is calculated using the vector difference between positions at adjacent moments and is expressed as the azimuth angle. and pitch angle .

[0191] Flight altitude through the angle of elevation of the target position and distance calculate:

[0192] ;

[0193] in, For flight altitude, This represents the altitude of the radar antenna.

[0194] The output of the spatiotemporal trajectory tracking module is a set of trajectory information for each target, including trajectory number, classification label, spatiotemporal trajectory sequence, flight speed, flight direction, and flight altitude, providing basic data for subsequent biomass inversion and situational analysis. The continuity and accuracy of trajectory tracking directly affect the reliability of migration path identification and biomass estimation.

[0195] The biomass inversion and product generation module inverts biomass density and generates structured monitoring products based on the classification of insect and bird targets, the radar cross section characteristics, and the spatiotemporal trajectory.

[0196] Furthermore, the operation of the biomass inversion and product generation module includes:

[0197] Establish a correlation model between radar cross section and biological body size;

[0198] Based on the aforementioned correlation model, the radar cross section of the detected target is converted into the size of a biological individual;

[0199] The number of targets and the size of the biological individuals in different spatial grids within the monitoring airspace are statistically analyzed, and the biomass density of the spatial grids is calculated.

[0200] Cluster analysis is performed on the spatiotemporal trajectories to identify migration paths, clustering areas, and diffusion directions;

[0201] The classification labels, spatiotemporal trajectories, biomass density, migration paths, aggregation areas, and diffusion directions of the insect and bird targets are encapsulated into structured monitoring products in JSON or GIS layer format.

[0202] Specifically, a correlation model between radar cross section and biological body size was established. This correlation model, based on electromagnetic scattering theory and measured data, reflects the correspondence between the size of an individual organism and its radar cross section.

[0203] For insect targets, an empirical formula is used:

[0204] ;

[0205] in, The length of the insect's body (in cm). Radar cross section (unit: m²). is the empirical fitting coefficient for the content dimension. Both are dimensionless indices, obtained by fitting measured insect body length and radar cross-section data. For moths, Usually, 100 is used. The value is usually 0.33; for locusts, Usually, 120 is used. Typically, 0.33 is used. Individual mass is estimated through body length, using... ,in The density coefficient is usually taken as... to g / cm³.

[0206] For bird targets, different association models are used:

[0207] ;

[0208] in, Individual bird mass (in grams). and This is an empirical coefficient. For large birds, Usually, 500 is used. Usually, 0.5 is used; for small birds, Usually, 300 is used. The value is usually 0.5.

[0209] Based on the aforementioned correlation model, the detected target radar cross section This can be converted into biological individual size or mass, enabling the inversion from electromagnetic scattering parameters to biophysical parameters.

[0210] The number of targets and the size of the biological individuals in different spatial grids within the monitoring airspace are statistically analyzed, and the biomass density of the spatial grids is calculated.

[0211] The monitored airspace is divided into a three-dimensional grid. Horizontally, it is divided by distance and azimuth, with a typical grid spacing of 1-5 km for distance and 1°-5° for azimuth. Vertically, it is divided by altitude, with a typical grid spacing of 100-500 m. The volume of each grid is approximately:

[0212] ;

[0213] in, The distance to the grid interval, The distance from the grid center to the radar. The azimuth grid spacing is in radians. The height grid interval is used, and this approximation formula is applicable to cases with small height grid intervals.

[0214] Count the number of targets detected in each grid cell Calculate the target number density:

[0215] ;

[0216] in, For the target number density, This represents the number of targets detected within the grid. This represents the volume of the grid, expressed in units of cells / m³.

[0217] Combine individual quality of each goal Calculate biomass density:

[0218] ;

[0219] in, Biomass density, expressed in g / m³ or kg / km³. For the first in the grid Individual quality of each goal This represents the number of targets detected within the grid. This represents the volume of the grid. Biomass density reflects the total amount of organisms per unit space and is an important indicator for assessing the size of insect or bird populations.

[0220] Considering the volume effect and beam filling factor of radar detection, a correction is made for biomass density. The radar beam increases in volume at long distances, resulting in different detection efficiency for targets of the same density at different distances. The correction factor is:

[0221] ;

[0222] in, For beam filling correction factors, The actual volume of the target group is estimated from the three-dimensional spatial distribution range of the target group output by the spatiotemporal trajectory tracking module. Let be the volume of the radar beam illuminated at the target location, and be a known parameter of the radar system. The corrected biomass density is: .

[0223] Cluster analysis is performed on the spatiotemporal trajectories to identify migration paths, clustering areas, and diffusion directions.

[0224] Density clustering algorithms (such as DBSCAN) are used to spatially cluster trajectory points to identify target clusters. Clustering parameters include neighborhood radius. (Usually set to 1-3km) and minimum number of points (Typically set to 5-10). Clustering results identify high-density clusters of biological targets, which may be gathering points or habitats during migration.

[0225] Statistical analysis of trajectory directions is performed to identify the dominant migration direction. The flight direction angles of all trajectories are calculated, and direction histograms or wind rose diagrams are plotted; the peak direction is the dominant migration direction. By analyzing migration directions at different altitude levels, vertically stratified migration behavior is revealed.

[0226] Migration paths are identified through spatiotemporal evolution analysis of trajectories. Trajectory segments that are close in time and spatially connected are linked to form complete migration corridors. The width of the migration path is determined by the lateral distribution range of the trajectory, typically ranging from several kilometers to tens of kilometers.

[0227] The direction of diffusion is determined by analyzing the departure direction of targets within the aggregation area. The distribution of trajectories originating from the aggregation area is statistically analyzed; the dominant direction is the diffusion direction, reflecting the diffusion pattern of insect or bird swarms from the gathering point to the surrounding areas.

[0228] The classification tags, spatiotemporal trajectories, biomass density, migration paths, aggregation areas, and diffusion directions of the insect and bird targets are packaged into a structured monitoring product.

[0229] Structured monitoring products use either JSON or GIS layer format. JSON format facilitates machine parsing and data exchange between systems, while GIS layer format facilitates visualization within a geographic information system.

[0230] The GIS layer format uses Shapefile or GeoJSON, and includes point layers (target locations), line layers (trajectories and migration paths), and polygon layers (aggregates and biomass density rasters).

[0231] The generated structured monitoring product data volume is reduced by 3-4 orders of magnitude compared to the original complex IQ data, greatly reducing the pressure on data transmission and storage. The high-level semantic information contained in the product (such as "a swarm of locusts with a density of 150 g / km³ has appeared in a certain area and is moving in a southeast direction") directly serves business applications without secondary processing, meeting the needs of real-time early warning and decision support.

[0232] The edge cloud collaboration module uploads the structured monitoring product to the cloud data center after the edge computing node completes data processing.

[0233] Specifically, after the edge cloud collaboration module completes data processing at the edge computing node, it uploads the structured monitoring product to the cloud data center.

[0234] Edge nodes asynchronously upload the generated structured monitoring products to provincial or national cloud data centers via the network. These structured monitoring products are in JSON or GIS layer format, with a single scan typically producing data ranging from several KB to several MB. The upload bandwidth requirement is hundreds of Kbps to several Mbps, far lower than the bandwidth required to transmit raw complex IQ data. The upload frequency is synchronized with the radar scan cycle, typically every 5-10 minutes.

[0235] The edge cloud collaboration module manages task scheduling and algorithm model updates for edge nodes. Deep learning models trained and optimized based on historical data from multiple sites are distributed incremental model parameter files to edge nodes, which automatically load the new model parameters for online updates, typically weekly or monthly.

[0236] The edge cloud collaboration module supports trajectory relay with neighboring radar edge nodes. When a target moves out of the detection range of this station, the edge node sends trajectory information to neighboring nodes, and the neighboring nodes establish tracking associations in the expected area, realizing continuous trajectory tracking across stations.

[0237] After receiving products uploaded by each edge node, the cloud data center performs large-scale situational fusion, historical data storage and long-term series analysis, and sends the optimized classification parameters and early warning instructions to the edge nodes, forming a closed-loop working mode of edge-cloud collaboration.

[0238] This embodiment directly processes raw complex IQ data through an edge computing architecture, fully preserving fine features such as micro-Doppler, and achieving accurate differentiation between insects and birds. The system completes real-time processing at the edge node and only uploads structured monitoring products to the cloud, significantly reducing data transmission pressure and processing latency, and providing a feasible technical solution for operational ecological monitoring of large-scale weather radar networks.

[0239] Example 2: In response to the need for monitoring locust migration in summer, the agricultural department of a certain province adopted the following methods: Figure 2 The edge computing method for fine-grained monitoring of insects and birds based on weather radar IQ data, as shown, deployed edge computing nodes at three S-band weather radar stations.

[0240] Large-scale migrations of the Oriental migratory locust frequently occur in this region during the summer. Traditional monitoring methods struggle to capture the swarm dynamics in real time, leading to delayed control responses. After deploying the system of this invention, edge computing nodes directly connect to the digital intermediate frequency receivers of each radar station, acquiring raw complex IQ data in real time and completing the entire processing locally.

[0241] In a typical locust migration event, the system detected periodic modulation signals with wingbeat frequencies in the range of 30-45Hz from IQ data using a micro-Doppler feature extraction module. Combined with radar cross-section characteristics, a cascaded classifier accurately identified the locust targets and eliminated interference from weak precipitation echoes occurring simultaneously. The spatiotemporal trajectory tracking module continuously tracked the swarm movement in three-dimensional space, identifying the main migration direction as southwest to northeast, with flight altitudes concentrated between 500-1200 meters. The biomass inversion module calculated the spatial distribution of swarm density based on the radar cross-section, identifying high-density aggregation areas.

[0242] The entire monitoring process, from data acquisition to the generation of early warning products, has a latency of less than 5 minutes. Edge nodes only upload structured JSON-formatted products to the provincial data center, with a single upload data volume of less than 1MB. In contrast, if the traditional solution were used to upload raw complex IQ data to the center for processing, the data volume per station per scan would exceed 500MB, and the transmission and processing latency would exceed 1 hour, making it difficult to meet real-time prevention and control requirements.

[0243] Application results show that this invention, by directly processing raw complex IQ data at the weather radar station and extracting fine features such as micro-Doppler, effectively distinguishes locusts from targets such as birds and precipitation, with significantly higher identification accuracy than traditional product data-based methods. The edge computing architecture significantly reduces data transmission pressure and processing latency, enabling real-time early warning. Quantitative inversion of biomass density and migration trajectories provides a scientific basis for prevention and control decisions. The system operates stably, providing strong technical support for agricultural pest and disease monitoring in the region.

[0244] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An edge computing system for fine-grained monitoring of insects and birds based on weather radar IQ data, characterized in that, include: The edge IQ data preprocessing module is deployed on the weather radar station side. It accesses the weather radar digital intermediate frequency receiver through the edge computing node to obtain the raw complex IQ data, and performs ground clutter suppression and time-frequency domain transformation on the raw complex IQ data to obtain preprocessed IQ data. The multidimensional biometric feature extraction module extracts multidimensional features from the preprocessed IQ data, including micro-Doppler features, dual polarization parametric features, and radar cross section features; wherein, the micro-Doppler features include periodic frequency modulation components generated by insect wing flapping and bird wing flapping. The multidimensional biometric feature extraction module extracts micro-Doppler features in the following ways: Perform a short-time Fourier transform on the preprocessed IQ data to generate a time-frequency spectrum. The time-spectrum diagram is decomposed using a subspace decomposition method to separate the main Doppler velocity component and the micro-Doppler sideband component; Extract the frequency, amplitude, and modulation period from the microDoppler sideband components; Among them, the micro-Doppler sideband component generated by insect wing flapping exhibits periodic high-frequency modulation characteristics, while the micro-Doppler sideband component generated by bird wing flapping exhibits low-frequency non-stationary modulation characteristics. The cascaded classification and recognition module uses a cascaded classifier to fuse and distinguish the micro-Doppler features, dual polarization parameter features, and radar cross section features. The cascaded classifier includes a first-level classifier and a second-level classifier. The first-level classifier filters out non-biological echoes, and the second-level classifier fuses the micro-Doppler features and dual polarization parameter features to distinguish between insect targets and bird targets. The spatiotemporal trajectory tracking module tracks the trajectory of the insect and bird targets to obtain their spatiotemporal trajectories. The biomass inversion and product generation module inverts biomass density and generates structured monitoring products based on the classification of insect and bird targets, the radar cross section characteristics, and the spatiotemporal trajectory. The edge cloud collaboration module uploads the structured monitoring product to the cloud data center after the edge computing node completes data processing.

2. The edge computing system for fine-grained monitoring of insects and birds based on weather radar IQ data according to claim 1, characterized in that, Ground clutter suppression in the edge IQ data preprocessing module includes: Acquire sounding data and numerical model wind field data provided by meteorological departments; A dynamic clutter map is constructed based on the aforementioned radiosonde data and numerical model wind field data; Based on the dynamic clutter map, an adaptive digital filter is used to filter the original complex IQ data to suppress fixed ground clutter and radio frequency interference.

3. The edge computing system for fine-grained monitoring of insects and birds based on weather radar IQ data according to claim 1, characterized in that, The time-frequency domain transformation in the edge IQ data preprocessing module includes: The raw complex IQ data after ground clutter suppression is processed by pulse pair processing or fast Fourier transform to generate a power spectral density containing amplitude information, phase information and spectral characteristic information.

4. The edge computing system for fine-grained monitoring of insects and birds based on weather radar IQ data according to claim 1, characterized in that, The multidimensional biometric feature extraction module extracts dual-polarization parametric features in the following ways: The echo signals of the horizontal polarization channel and the vertical polarization channel are extracted from the preprocessed IQ data, respectively. The differential reflectivity is obtained by calculating the ratio of the echo power of the horizontal polarization channel to the echo power of the vertical polarization channel. Calculate the phase difference between the echo signal of the horizontal polarization channel and the echo signal of the vertical polarization channel to obtain the differential phase; Calculate the correlation coefficient between the echo signal of the horizontal polarization channel and the echo signal of the vertical polarization channel; Spatial texture analysis is performed on the differential reflectivity, differential phase, and correlation coefficient to extract texture statistical features.

5. The edge computing system for fine-grained monitoring of insects and birds based on weather radar IQ data according to claim 1, characterized in that, The multidimensional biometric feature extraction module extracts radar cross-section features in the following ways: The echo power is calculated from the preprocessed IQ data; The radar cross section is calculated using radar equations based on the echo power and radar system parameters. Extract the temporal variation characteristics of the radar cross section.

6. The edge computing system for fine-grained monitoring of insects and birds based on weather radar IQ data according to claim 4, characterized in that, The cascaded classification and recognition module performs fusion discrimination in the following ways: The dual polarization parameter features and the radar cross section features are input into the first-level classifier; The first-level classifier filters out turbulent echoes based on the high variability of the texture statistical features of the dual polarization parametric features, where the correlation coefficient is lower than a preset threshold and the differential reflectivity is high. The first-level classifier filters out abnormal propagation echoes and retains candidate biological targets based on the non-physical jump in the differential phase. The micro-Doppler features, dual polarization parametric features, and radar cross section features of the candidate biological targets are input into the second-level classifier; The second-level classifier distinguishes between insect targets and bird targets based on the frequency range and spectral characteristics of the periodic modulation components in the micro-Doppler features, combined with the texture statistics of the dual polarization parametric features and the temporal variation characteristics of the radar cross section. Based on the frequency characteristics of the micro-Doppler features and the amplitude range of the radar cross section, the insect targets are further subdivided into moths or locusts, and the bird targets are further subdivided into large birds or small birds.

7. The edge computing system for fine-grained monitoring of insects and birds based on weather radar IQ data according to claim 1, characterized in that, The spatiotemporal trajectory tracking module operates as follows: A sequence of target points is established in a three-dimensional spherical coordinate system, which includes distance, azimuth, and elevation. Calculate the correlation probability between the detected target point trace at the current moment and the historical trajectory; Data association is performed based on the association probability. Trajectory initiation is performed for newly emerging target points, trajectory maintenance is performed for already associated target points, and trajectory termination is performed for continuous unassociated trajectories. Output the target's continuous spatiotemporal trajectory, flight speed, flight direction, and flight altitude.

8. The edge computing system for fine-grained monitoring of insects and birds based on weather radar IQ data according to claim 1, characterized in that, The biomass inversion and product generation module operates as follows: Establish a correlation model between radar cross section and biological body size; Based on the aforementioned correlation model, the radar cross section of the detected target is converted into the size of a biological individual; The number of targets and the size of the biological individuals in different spatial grids within the monitoring airspace are statistically analyzed, and the biomass density of the spatial grids is calculated. Cluster analysis is performed on the spatiotemporal trajectories to identify migration paths, clustering areas, and diffusion directions; The classification labels, spatiotemporal trajectories, biomass density, migration paths, aggregation areas, and diffusion directions of the insect and bird targets are encapsulated into structured monitoring products in JSON or GIS layer format.

9. A method for fine-grained edge computing of insects and birds based on weather radar IQ data, characterized in that, include: Edge computing nodes are deployed at the weather radar station. The edge computing nodes are connected to the digital intermediate frequency receiver of the weather radar to obtain raw complex IQ data. Ground clutter suppression and time-frequency domain transformation are performed on the raw complex IQ data to obtain preprocessed IQ data. Multi-dimensional features are extracted from the preprocessed IQ data, including micro-Doppler features, dual polarization parametric features, and radar cross section features; wherein, the micro-Doppler features include periodic frequency modulation components generated by insect wing flapping and bird flapping. The extraction methods for the micro-Doppler features include: Perform a short-time Fourier transform on the preprocessed IQ data to generate a time-frequency spectrum. The time-spectrum diagram is decomposed using a subspace decomposition method to separate the main Doppler velocity component and the micro-Doppler sideband component; Extract the frequency, amplitude, and modulation period from the microDoppler sideband components; Among them, the micro-Doppler sideband component generated by insect wing flapping exhibits periodic high-frequency modulation characteristics, while the micro-Doppler sideband component generated by bird wing flapping exhibits low-frequency non-stationary modulation characteristics. A cascaded classifier is used to fuse and distinguish the micro-Doppler features, dual polarization parameter features, and radar cross section features. The cascaded classifier includes a first-level classifier and a second-level classifier. The first-level classifier filters out non-biological echoes, and the second-level classifier fuses the micro-Doppler features and dual polarization parameter features to distinguish between insect targets and bird targets. Trajectory tracking is performed on the insect and bird targets to obtain their spatiotemporal trajectories; Based on the classification of insect and bird targets, the radar cross section characteristics, and the spatiotemporal trajectory, biomass density is retrieved and structured monitoring products are generated. After the edge computing node completes data processing, the structured monitoring product is uploaded to the cloud data center.

Images