A ground clutter adaptive suppression method and system for fixed-platform weather radar

Abstract

This invention discloses an adaptive ground clutter suppression method and system for fixed-platform weather radar, relating to the field of weather radar technology. The method includes the following steps: processing the raw I / Q signals received by the radar to obtain a multi-dimensional hybrid feature vector; constructing and maintaining a dynamic feature library based on a Gaussian mixture model-Hidden Markov model (GMM-HMM); intelligently classifying the echo sources of each resolution unit, outputting category labels and confidence levels; applying differentiated signal suppression strategies to different types of units based on the category labels and confidence levels; performing a physical constraint-based quality assessment on the suppressed data, and using the assessment results to optimize the classifier threshold, filter parameters, and model parameters of the dynamic feature library. This method enables accurate identification and separation of clutter and meteorological targets.

Description

Technical Field

[0001] This invention relates to the field of weather radar technology, and in particular to an adaptive ground clutter suppression method and system for fixed-platform weather radar. Background Technology

[0002] With the increasing demand for more refined meteorological observation networks, the fixed deployment of high-performance radars (including radars transplanted from mobile platforms) has become an important means to fill monitoring gaps and improve spatiotemporal resolution. However, this shift in application scenarios has brought about a fundamental technical challenge: how to effectively extract weak meteorological target signals from a strong ground clutter background when the radar platform is completely stationary.

[0003] Traditional clutter suppression techniques mainly fall into two paradigms, but neither is suitable for this new scenario: (1) Platform motion-based method: typically applied to airborne and vehicle-mounted radar. The principle is to use the relative motion between the platform and the ground to separate stationary clutter from moving meteorological targets in the Doppler frequency domain, and then filter them out through techniques such as space-time adaptive processing (STAP). When the platform is stationary, the relative velocity is zero, and this method is completely ineffective.

[0004] (2) Static map-based method: Commonly used for some fixed radar stations. A static "clutter map" is generated through long-term statistics or prior measurements, and then subtracted during real-time processing. This method has inherent limitations: Poor environmental adaptability: It cannot adapt to dynamic changes such as new buildings, vegetation growth and withering, and wind-induced ground movement, resulting in the suppression effect decaying over time or producing false alarms.

[0005] Severe loss of meteorological information: In order to suppress strong clutter near zero velocity, frequency domain "apodization" or fixed threshold filters are usually used. This will indiscriminately filter out real low-speed precipitation (such as stratus clouds and drizzle), resulting in the loss of key meteorological information.

[0006] Reliance on ideal prior information: Its performance is heavily dependent on high-precision digital elevation model (DEM) and geographic information system (GIS) data, and its performance drops sharply in the absence of data or in complex urban environments.

[0007] On a fixed platform, ground clutter and meteorological targets highly overlap in the Doppler domain (both concentrated near zero / low velocity) and spatial domain (interleaving each other), and the statistical characteristics of the clutter dynamically change with environment, season, and weather. Existing algorithms lack the ability to intelligently separate such complex, time-varying coupled signals. Therefore, there is an urgent need for a novel clutter suppression solution that can dynamically sense the environment, intelligently learn features, adaptively and accurately suppress clutter, while preserving meteorological information to the maximum extent. Summary of the Invention

[0008] The technical problem to be solved by the present invention is how to provide an adaptive ground clutter suppression method for fixed platform weather radar that can achieve accurate identification and separation of clutter and meteorological targets.

[0009] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: an adaptive ground clutter suppression method for fixed platform weather radar, comprising the following steps: S1: Process the raw I / Q signal received by the radar, and extract time-domain stability features, frequency-domain spectrum features, spatial context features and polarization features in parallel for each resolution unit to form a multi-dimensional hybrid feature vector; S2: Construct and maintain a dynamic feature library based on Gaussian Mixture Model-Hidden Markov Model (GMM-HMM) to describe the spatiotemporal statistical distribution and evolution of clutter characteristics of each resolving unit; S3: Based on the improved random forest classifier, the multidimensional hybrid feature vector and the prior probability of the dynamic feature library are fused to intelligently classify the echo source of each resolution unit and output the category label and confidence score. S4: Based on the category label and confidence level, apply differentiated signal suppression strategies to different types of units, including feature cancellation for high-confidence clutter, adaptive IIR filtering for suspected clutter, and conformal FIR filtering for meteorological targets. S5: Perform a physical constraint-based quality assessment on the suppressed data, and optimize the classifier threshold, filter parameters, and model parameters of the dynamic feature library based on the assessment results.

[0010] The present invention also discloses a ground clutter adaptive suppression system for a fixed platform weather radar, comprising: Multidimensional hybrid feature vector construction module: For the raw I / Q signal received by the radar, time-domain stability features, frequency-domain spectral features, spatial context features and polarization features are extracted in parallel for each resolution unit to form a multidimensional hybrid feature vector; Dynamic Feature Library Construction Module: Used to build and maintain a dynamic feature library based on Gaussian Mixture Model-Hidden Markov Model (GMM-HMM), which describes the spatiotemporal statistical distribution and evolution of clutter characteristics of each resolving unit; Intelligent classification module: Based on an improved random forest classifier, it integrates the multidimensional hybrid feature vector and the prior probability of the dynamic feature library to intelligently classify the echo source of each resolution unit and output the category label and confidence score. Signal suppression module: used to apply differentiated signal suppression strategies to different types of units according to the category label and confidence level, including feature cancellation for high-confidence clutter, adaptive IIR filtering for suspected clutter, and conformal FIR filtering for meteorological targets; Quality assessment and optimization module: Used to perform physical constraint-based quality assessment on the suppressed data, and to optimize the classifier threshold, filter parameters, and model parameters of the dynamic feature library based on the assessment results.

[0011] The beneficial effects of adopting the above technical solution are as follows: 1) High suppression accuracy and low information loss: Through intelligent classification, precise targeted suppression is achieved. While achieving an average clutter suppression ratio of more than 38dB, the retention rate of low-speed precipitation is increased to more than 96%, which is far superior to traditional methods.

[0012] 2) Strong environmental adaptability: Utilizing the GMM-HMM model and online learning mechanism, the system can automatically track long-term environmental changes (such as buildings and vegetation) and short-term fluctuations (such as wind power) without the need for manual recalibration.

[0013] 3) The algorithm has good robustness and real-time performance: It adopts parallel feature extraction and incremental learning strategies, and the single-frame processing latency is less than 10ms on a typical computing platform, which meets the real-time requirements of business.

[0014] 4) Excellent system interpretability and scalability: The design based on feature engineering and probabilistic models makes the decision-making process transparent and debuggable. The framework is easy to integrate new features or sensor data and adaptable to different radar models. Attached Figure Description

[0015] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0016] Figure 1 This is an overall flowchart of the method described in Embodiment 1 of the present invention; Figure 2 This is a schematic diagram of the spatiotemporal evolution model of the GMM-HMM dynamic feature library in the method described in Embodiment 1 of the present invention; Figure 3 This is a schematic diagram of the system described in Embodiment 2. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0018] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0019] Example 1 Overall, such as Figure 1 As shown in the figure, an embodiment of the present invention discloses an adaptive ground clutter suppression method for a fixed platform weather radar, the method comprising the following steps: S1: Process the raw I / Q signal received by the radar, and extract time-domain stability features, frequency-domain spectrum features, spatial context features and polarization features in parallel for each resolution unit to form a multi-dimensional hybrid feature vector; S2: Construct and maintain a dynamic feature library based on Gaussian Mixture Model-Hidden Markov Model (GMM-HMM) to describe the spatiotemporal statistical distribution and evolution of clutter characteristics of each resolving unit; S3: Based on the improved random forest classifier, the multidimensional hybrid feature vector and the prior probability of the dynamic feature library are fused to intelligently classify the echo source of each resolution unit and output the category label and confidence score. S4: Based on the category label and confidence level, apply differentiated signal suppression strategies to different types of units, including feature cancellation for high-confidence clutter, adaptive IIR filtering for suspected clutter, and conformal FIR filtering for meteorological targets. S5: Perform a physical constraint-based quality assessment on the suppressed data, and optimize the classifier threshold, filter parameters, and model parameters of the dynamic feature library based on the assessment results.

[0020] The above steps will be explained in detail below with specific methods. Furthermore, in step S1, during the multi-dimensional hybrid feature extraction, four types of feature vectors are extracted in parallel from the original I / Q data for each distance-orientation resolution unit: Among them, the time-domain stability characteristics include the long-term mean (μ) and variance (σ). 2 ), kurtosis (K) and autocorrelation decay time (τ).

[0021] Frequency domain spectral characteristics include Doppler spectral entropy (H(s)), spectral symmetry (S_sym), and spectral kurtosis (K_s), which are calculated using the following formulas: ; Normalized power spectrum; The spatial context features include local spatial consistency (C_local) and terrain height matching degree (M_terrain).

[0022] Polarization characteristics (if supported by radar) include differential reflectivity (Z_DR) and co-polarization correlation coefficient (ρ_hv).

[0023] In step S2, the construction of the GMM-HMM model includes: using data from the dry season, fitting a Gaussian mixture model to each cell using the expectation-maximization algorithm, and training a hidden Markov model describing the diurnal or seasonal variations of the features using the forward-backward algorithm.

[0024] Furthermore, the GMM model is used to cluster the feature distribution of each unit, and the HMM model is used to describe its evolution over time (such as diurnal and seasonal variations): GMM model: K=3 (usually corresponding to three types: strong static clutter, time-varying clutter, and noise).

[0025] Hidden MM model: Define hidden states (e.g., 24 hours in a day), and learn the state transition probabilities using the Baum-Welch algorithm. The formula is as follows: The set of hidden states S = {s1, s2, ..., s24} corresponds to the 24 time periods of a day; The state transition probability matrix A = [a_ij]{24×24}, where a_ij represents the probability of transitioning from state i to state j, satisfying ∑{j=1}^{24} a_ij = 1; The observation probability function b_j(o) = P(o_t = o | s_t = j) is given by the Gaussian mixture model; The model parameters λ = (π, A, B) are iteratively optimized using the Baum-Welch algorithm, where the parameter update formula is: ; Let be the joint probability of being in state i at time t and in state j at time t+1. Let be the marginal probability of being in state i at time t.

[0026] This hybrid model can more accurately characterize the dynamic statistical properties of clutter over time and the environment, such as... Figure 2 The diagram shown is a schematic of the spatiotemporal evolution model of the GMM-HMM dynamic feature library.

[0027] In step S3, the improvements of the improved random forest classifier include: introducing a cost-sensitive learning mechanism, adding spatiotemporal continuity constraints when splitting decision tree nodes, and calibrating the output classification probability.

[0028] Furthermore, a random forest classifier containing T decision trees was constructed, and three key improvements were made: Cost-sensitive learning: Set a higher misclassification cost for "misclassifying weather as clutter" and prioritize the protection of weather signals.

[0029] Spatiotemporal constraint splitting: When selecting splitting features for decision tree nodes, spatial adjacency and temporal continuity constraints are introduced.

[0030] Probability calibration: The Platt Scaling method is used to calibrate the raw probabilities output by the classifier, thereby improving the reliability of the confidence level.

[0031] In step S4, the cutoff frequency of the adaptive IIR filter is dynamically adjusted based on the confidence level that the current cell is identified as clutter. The adjustment formula is: ω_c = ω_base + Δω P_clutter, where P_clutter is the clutter probability estimate.

[0032] Furthermore, based on the "label" and "confidence" output by the classifier, three levels of differential suppression are performed: High-confidence deterministic clutter: Employing a "feature cancellation" algorithm. Based on the clutter complex signal estimated by the GMM model, precise cancellation is performed in the I / Q domain; Medium-confidence suspected / time-varying clutter: An adaptive IIR filter with adjustable parameters is used. The filter cutoff frequency is dynamically adjusted according to the clutter probability; Low confidence level / classified as a meteorological target: Use a minimum phase distortion FIR conformal filter with only slight smoothing to preserve signal characteristics to the maximum extent.

[0033] In step S5, the quality assessment is based on a comprehensive index function J, which is a weighted combination of clutter suppression ratio, precipitation retention rate, velocity field continuity score, false alarm rate, and missed detection rate. The system achieves closed-loop adaptation by optimizing the value of this function.

[0034] Design a multi-objective evaluation function to quantify the processing effect and drive system self-optimization: The comprehensive evaluation index J = w1·CR + w2·PR + w3·VCS - w4·FAR - w5·MDR; Where: CR - clutter suppression ratio, PR - precipitation retention rate, VCS - velocity field continuity score, FAR - false alarm rate, MDR - missed detection rate, and w_i is the weight.

[0035] Based on the changes in the J value, the decision threshold of the classifier, the parameters of the filter, and the update rate of the GMM-HMM model are dynamically adjusted through gradient descent or heuristic rules to form a closed-loop optimization.

[0036] Example 2 like Figure 3 As shown, corresponding to the method described in Embodiment 1, this embodiment of the invention also discloses a ground clutter adaptive suppression system for a fixed platform weather radar, comprising: Multidimensional hybrid feature vector construction module 101: For the original I / Q signal received by the radar, time-domain stability features, frequency-domain spectrum features, spatial context features and polarization features are extracted in parallel for each resolution unit to form a multidimensional hybrid feature vector; Dynamic Feature Library Construction Module 102: Used to construct and maintain a dynamic feature library based on Gaussian Mixture Model-Hidden Markov Model (GMM-HMM) to describe the spatiotemporal statistical distribution and evolution of clutter characteristics of each resolving unit; Intelligent classification module 103: Based on an improved random forest classifier, it integrates the multidimensional hybrid feature vector and the prior probability of the dynamic feature library to intelligently classify the echo source of each resolution unit and output the category label and confidence score. Signal suppression module 104: is used to apply differentiated signal suppression strategies to different types of units according to the category label and confidence level, including feature cancellation for high-confidence clutter, adaptive IIR filtering for suspected clutter, and conformal FIR filtering for meteorological targets; Quality assessment and optimization module 105: Used to perform physical constraint-based quality assessment on the suppressed data, and to optimize the threshold of the classifier, the parameters of the filter, and the model parameters of the dynamic feature library based on the assessment results.

[0037] It should be noted that the specific implementation methods of each module in the system described in this embodiment can refer to the corresponding steps of the method described in Embodiment 1, and will not be repeated here.

[0038] Application Example 1 This application example discloses an adaptive ground clutter suppression method for a fixed-platform weather radar, the method comprising the following steps: 1) System initialization and offline training: Input: After the radar station is installed, collect radar baseband data for M consecutive days (M≥7 recommended) of clear, rainless days.

[0039] Processing flow (Algorithm 1): For each resolution unit, feature sequences from different time periods of the day are extracted.

[0040] The Expectation-Maximization (EM) algorithm is used to fit a K=3 GMM model (parameter θ) to each unit to characterize its feature clustering.

[0041] Using the Baum-Welch algorithm, a Hidden Markov Model (HMM) with 24 hidden states (corresponding to 24 hours in a day) is trained for each unit to describe the daily variation pattern of features.

[0042] An improved random forest classifier was trained using offline data, and its hyperparameters were optimized.

[0043] The trained GMM parameters θ, HMM parameters λ, random forest model, and calculated spatial correlation matrix are all stored in the initial dynamic feature library.

[0044] 2) Real-time online processing flow: Input: The radar I / Q data frame at the current moment.

[0045] Processing flow (Algorithm 2): Step 1: Parallel Feature Extraction The multidimensional hybrid feature vector Ft is computed in parallel for each resolution unit.

[0046] Step 2: Joint Probability Classification For unit (i,j): a) Calculate the GMM likelihood: P_gmm=GMM_Probability(Ft,θ_ij) b) Calculate the HMM likelihood: P_hmm = HMM_Probability(Ft, λ_ij, current_hour) c) Calculate the fusion probability: P_fusion = α·P_gmm + β·P_hmm d) Obtain the random forest classification probability: P_rf = RandomForest_Predict(Ft_subset) e) Decision-making logic: If P_fusion > Th_high and P_rf > 0.8 → label = "deterministic clutter", confidence level = 0.95 If P_fusion > Th_low or P_rf > 0.5 → label = "suspected clutter", confidence level = P_rf Otherwise → label = "weather target", confidence level = 1 - P_rf Step 3: Layered Suppression Based on labels and confidence levels: - "Deterministic Clutter": Performs feature cancellation.

[0047] - "Suspected noise": Use strong suppression IIR filtering if the confidence level is high, and use weak suppression IIR filtering if the confidence level is low.

[0048] - "Meteorological Target": Perform FIR conformal filtering.

[0049] Step 4: Quality Assessment and Feedback Calculate physical parameters such as the continuity of the velocity field in the output data. If the continuity does not meet the threshold, trigger feedback: Fine-tune the classification thresholds Th_high and Th_low; Update the GMM parameters θ_ij online with a small learning rate η; If the number of misclassified samples accumulates to a certain level, it will trigger a local retraining of the random forest.

[0050] 3) Algorithm complexity and performance guarantees: This invention features optimized design for real-time performance: Feature extraction: O(N) complexity, highly parallelizable.

[0051] GMM-HMM calculation: Pre-calculation and probability table lookup are used to reduce the amount of real-time calculation.

[0052] Random forest: Fixed-depth decision-making with controllable complexity.

[0053] Filtering operations: Recursion or short convolutions are used, which are highly efficient.

[0054] Under typical configuration (1000x1000 beam units, modern multi-core CPU), a single-frame processing latency of <10ms can be achieved.

[0055] 4) Examples and Test Results A continuous 90-day test at an X-band fixed weather radar station (urban environment) shows that the algorithm of this invention has a significant performance improvement compared to the traditional static map method and fixed filtering method: Clutter rejection ratio (CR): 38.2 dB (conventional method: 21.5 dB) Low-velocity precipitation retention (PR): 96.7% (conventional method: 78.3%) Classification accuracy: 95.4% (Traditional feature method: 82.1%) False Alarm Rate (FAR): 0.8% (Traditional Method: 3.2%) System adaptive cycle: In the event of significant environmental changes (such as new buildings), the system automatically adjusts and restores its optimal performance within 24 hours.

[0056] It should be noted that although the above embodiments have been described, those skilled in the art can make other changes and modifications to these embodiments once they understand the basic inventive concept. Therefore, the above description is only an embodiment of the present invention and does not limit the scope of patent protection of the present invention. Any equivalent structural or procedural transformations made using the content of the present invention specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. An adaptive ground clutter suppression method for fixed-platform weather radar, characterized in that... Includes the following steps: S1: Process the raw I / Q signal received by the radar, and extract time-domain stability features, frequency-domain spectrum features, spatial context features and polarization features in parallel for each resolution unit to form a multi-dimensional hybrid feature vector; S2: Construct and maintain a dynamic feature library based on Gaussian Mixture Model-Hidden Markov Model (GMM-HMM) to describe the spatiotemporal statistical distribution and evolution of clutter characteristics of each resolving unit; S3: Based on the improved random forest classifier, the multidimensional hybrid feature vector and the prior probability of the dynamic feature library are fused to intelligently classify the echo source of each resolution unit and output the category label and confidence score. S4: Based on the category label and confidence level, apply differentiated signal suppression strategies to different types of units, including feature cancellation for high-confidence clutter, adaptive IIR filtering for suspected clutter, and conformal FIR filtering for meteorological targets. S5: Perform a physical constraint-based quality assessment on the suppressed data, and optimize the classifier threshold, filter parameters, and model parameters of the dynamic feature library based on the assessment results.

2. The adaptive ground clutter suppression method for fixed-platform weather radar as described in claim 1, characterized in that, In S1: Temporal stability characteristics include long-term mean μ and variance σ 2 , kurtosis K and autocorrelation decay time τ; Frequency domain spectral features include Doppler spectral entropy H(s), spectral symmetry S_sym, and spectral kurtosis K_s; Spatial context features include local spatial consistency (C_local) and terrain height matching (M_terrain); Polarization characteristics include differential reflectivity Z_DR and co-polarization correlation coefficient ρ_hv.

3. The adaptive ground clutter suppression method for fixed-platform weather radar as described in claim 1, characterized in that, The method for constructing the GMM-HMM model in S2 includes: Using data from periods without precipitation, a Gaussian mixture model is fitted to each cell using the expectation-maximization algorithm, and a hidden Markov model describing the diurnal or seasonal variations of the features is trained using the forward-backward algorithm.

4. The adaptive ground clutter suppression method for fixed-platform weather radar as described in claim 1, characterized in that, The method for constructing the GMM-HMM model in S2 includes: The GMM model is used to cluster the feature distribution of each unit, and the HMM model is used to describe its evolution over time. GMM model: Where K=3; Hidden MM model: Define hidden states and learn the state transition probabilities using the Baum-Welch algorithm. The formula is as follows: The set of hidden states S = {s1, s2, ..., s24} corresponds to the 24 time periods of a day; The state transition probability matrix A = [a_ij]{24×24}, Where a_ij represents the probability of transitioning from state i to state j, satisfying ∑{j=1}^{24} a_ij = 1; The observation probability function b_j(o) = P(o_t = o | s_t = j) is given by the Gaussian mixture model; The model parameters λ = (π, A, B) are iteratively optimized using the Baum-Welch algorithm, where the parameter update formula is: ， Let be the joint probability of being in state i at time t and in state j at time t+1. Let be the marginal probability of being in state i at time t.

5. The adaptive ground clutter suppression method for fixed-platform weather radar as described in claim 1, characterized in that, In step S3, the improved random forest classifier implementation method includes: introducing a cost-sensitive learning mechanism, adding spatiotemporal continuity constraints when splitting decision tree nodes, and calibrating the output classification probability.

6. The adaptive ground clutter suppression method for fixed-platform weather radar as described in claim 1, characterized in that, The implementation method of the improved random forest classifier in step S3 includes: Cost-sensitive learning: Sets a higher misclassification cost for misclassifying weather as clutter, prioritizing the protection of weather signals; Spatiotemporal constraint splitting: When selecting splitting features for decision tree nodes, spatial adjacency and temporal continuity constraints are introduced; Probability calibration: The raw probabilities output by the classifier are calibrated using the Platt Scaling method.

7. The adaptive ground clutter suppression method for fixed-platform weather radar as described in claim 1, characterized in that, In step S4, the cutoff frequency of the adaptive IIR filter is dynamically adjusted based on the confidence level that the current cell is identified as clutter. The adjustment formula is as follows: ω_c = ω_base + Do P_clutter; Where P_clutter is the clutter probability estimate.

8. The adaptive ground clutter suppression method for fixed-platform weather radar as described in claim 1, characterized in that, The differentiated signal suppression strategies applied to different types of units in step S4 include: Based on the label and confidence level output by the classifier, perform three levels of differential suppression: High-confidence deterministic clutter: A feature cancellation algorithm is used to accurately cancel the clutter complex signal estimated by the GMM model in the I / Q domain; Medium-confidence suspected / time-varying clutter: An adaptive IIR filter with adjustable parameters is used, and the filter cutoff frequency is dynamically adjusted according to the clutter probability; Low confidence / classified as a meteorological target: Use a minimum phase distortion FIR conformal filter with only slight smoothing to preserve signal characteristics to the maximum extent.

9. The adaptive ground clutter suppression method for fixed-platform weather radar as described in claim 1, characterized in that, The physical constraint-based quality assessment and feedback in S5 includes the following steps: Design a multi-objective evaluation function to quantify the processing effect and drive system self-optimization: The comprehensive evaluation index J = w1·CR + w2·PR + w3·VCS - w4·FAR - w5·MDR; Where: CR - clutter suppression ratio, PR - precipitation retention rate, VCS - velocity field continuity score, FAR - false alarm rate, MDR - missed detection rate, w_i is the weight; Based on the changes in the J value, the decision threshold of the classifier, the parameters of the filter, and the update rate of the GMM-HMM model are dynamically adjusted through gradient descent or heuristic rules to form a closed-loop optimization.

10. A ground clutter adaptive suppression system for a fixed-platform weather radar, characterized in that... include: Multidimensional hybrid feature vector construction module: For the raw I / Q signal received by the radar, time-domain stability features, frequency-domain spectral features, spatial context features and polarization features are extracted in parallel for each resolution unit to form a multidimensional hybrid feature vector; Dynamic Feature Library Construction Module: Used to build and maintain a dynamic feature library based on Gaussian Mixture Model-Hidden Markov Model (GMM-HMM), which describes the spatiotemporal statistical distribution and evolution of clutter characteristics of each resolving unit; Intelligent classification module: Based on an improved random forest classifier, it integrates the multidimensional hybrid feature vector and the prior probability of the dynamic feature library to intelligently classify the echo source of each resolution unit and output the category label and confidence score. Signal suppression module: used to apply differentiated signal suppression strategies to different types of units according to the category label and confidence level, including feature cancellation for high-confidence clutter, adaptive IIR filtering for suspected clutter, and conformal FIR filtering for meteorological targets; Quality assessment and optimization module: Used to perform physical constraint-based quality assessment on the suppressed data, and to optimize the classifier threshold, filter parameters, and model parameters of the dynamic feature library based on the assessment results.

Images