An adaptive noise suppression method and system for radar echo signal

Abstract

The application relates to the technical field of radar echo denoising, in particular to an adaptive suppression method and system for filtering noise of a radar echo signal; radar station position and beam pointing data are acquired, forest vegetation data and environmental interference data of a target area pointed by the beam are synchronously collected, and radar environment data are formed; the radar environment data are identified through a multi-band decision model to obtain multi-modal waveform parameters; based on the multi-modal waveform parameters, a radar device is controlled to periodically detect the target area to obtain radar echo data; the radar echo data and the radar environment data are identified, dynamic spectrum features in the radar echo data are extracted, the dynamic spectrum features are subjected to clutter suppression processing by using the radar environment data, a confidence probability that early fire exists in the target area is calculated, and spatiotemporal consistency joint verification is performed; the application carries out fire early warning through the confidence probability of the fire.

Description

Technical Field

[0001] This invention relates to the field of radar echo denoising technology, specifically to an adaptive suppression method and system for filtering noise in radar echo signals. Background Technology

[0002] Early warning of forest fires using radar systems has become an important research direction. Existing radar fire detection technologies primarily rely on emitting electromagnetic waves and receiving backscattered echoes from the target area to identify smoke plumes generated by early fires or drastic changes in local atmospheric refractive index caused by high temperatures. However, in practical applications, especially in complex terrain environments such as forests, existing technologies face significant technical bottlenecks.

[0003] Early-stage fires generate extremely weak radar echo signals, which are easily drowned out by strong background noise and clutter. Dynamic clutter generated by wind-driven vegetation (such as leaves and branches) is a major source of interference. Its Doppler spectrum characteristics are highly similar to the turbulent flow of fire plumes, making it difficult for traditional Doppler filtering methods to effectively distinguish between them, resulting in high false alarm and missed alarm rates. Existing radar detection systems mostly employ fixed operating parameters and signal processing algorithms, lacking the ability to adapt to changes in the detection environment. Forest vegetation type, canopy density, and real-time environmental factors such as wind speed and humidity all significantly affect the propagation and scattering characteristics of radar waves. Fixed systems cannot maintain optimal detection performance in variable environments.

[0004] Traditional identification methods typically rely solely on the intensity of the echo or a single spectral feature, failing to fully explore and utilize the multi-dimensional and deep-level feature information contained in the echo signal that can characterize the unique physical mechanism of the fire, resulting in insufficient robustness and accuracy of identification.

[0005] To address this, an adaptive method and system for suppressing noise in radar echo signal filtering are proposed. Summary of the Invention

[0006] The purpose of this invention is to provide an adaptive suppression method and system for filtering noise in radar echo signals. By constructing an adaptive suppression model, radar echo data and radar environment data are identified, dynamic spectral features in the radar echo data are extracted, and clutter suppression processing is performed on the dynamic spectral features using the radar environment data to calculate the confidence probability of early fire in the target area.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] An adaptive method for suppressing filtering noise in radar echo signals, comprising:

[0009] The radar station location and beam pointing data are acquired, and forest vegetation data and environmental interference data of the target area pointed to by the beam are collected simultaneously to form radar environmental data; the radar environmental data are identified by a multi-band decision model to obtain multi-modal waveform parameters, including X-band parameters and W-band parameters.

[0010] Based on multimodal waveform parameters, the radar equipment is controlled to periodically detect the target area and obtain radar echo data.

[0011] An adaptive signal suppression model is constructed to identify radar echo data and radar environment data, extract dynamic spectral features from the radar echo data, and use the radar environment data to perform clutter suppression processing on the dynamic spectral features to calculate the confidence probability of early fire in the target area.

[0012] A spatiotemporal consistency joint verification is performed on the confidence probability, and the confirmed fire source is vertically located and the fire type is determined by combining the forest vegetation data, and an early warning information is issued.

[0013] The radar environmental data includes forest vegetation data and environmental interference data;

[0014] The forest vegetation data includes: digital elevation models of the target area, vegetation cover data, and vegetation distribution data including tree species, average tree height, leaf type, and canopy density, obtained by calling and parsing data from a geographic information system; the environmental disturbance data includes: wind speed data, wind direction data, atmospheric temperature, relative humidity, and rainfall data.

[0015] The multi-band decision model includes a first decision module, a second decision module, and a third decision module;

[0016] The first decision module, based on the digital elevation model, vegetation coverage and canopy density in the forest vegetation data, and the calculated X-band path attenuation coefficient, adaptively adjusts the pulse width and peak power of the X-band transmitted waveform with the criterion of maximizing the residual signal-to-noise ratio after the signal energy penetrates the vegetation canopy.

[0017] The second decision module, based on the Mie scattering theory model of standard smoke particles and combined with real-time atmospheric humidity and W-band path attenuation coefficient, adaptively adjusts the center frequency and bandwidth of the W-band transmitted waveform with the criterion of maximizing smoke particle detection.

[0018] The third decision module dynamically adjusts the pulse repetition frequency of the radar equipment based on the wind speed and wind direction data in the environmental interference data.

[0019] The multimode waveform parameters are obtained based on the pulse width and peak power of the X-band transmitted waveform, the center frequency and bandwidth of the W-band transmitted waveform, and the pulse repetition frequency.

[0020] The adaptive signal suppression model includes an echo feature extraction module, which extracts features from radar echo data to obtain dynamic spectrum features; the dynamic spectrum features include cross-band structural features and turbulent chaos features.

[0021] The cross-band structural features utilize X-band echo signals to process and generate a three-dimensional structural background mask representing the tree trunk, main branches, and ground surface; leveraging the high sensitivity of W-band signals to microparticles, the spatial location and motion intensity of dynamic scatterers are identified, resulting in a dynamic scatterer map; the dynamic scatterers include vegetation leaves and dust particles; and the cross-band structural features are generated using the W-band dynamic scatterer map and the X-band structural background mask.

[0022] The turbulent chaotic characteristics are obtained by performing high-resolution time-frequency analysis on the W-band echo, calculating and extracting parameters to characterize the thermal turbulent chaotic characteristics above the fire source, including the energy entropy of the time-frequency distribution, the spectral fractal dimension, and the Lyapunov exponent characterizing the nonlinear evolution of the spectrum.

[0023] The signal adaptive suppression model includes an echo data processing module, which performs clutter suppression processing on dynamic spectral characteristics using radar environmental data, including:

[0024] Based on the wind speed and wind direction data in the environmental interference data and the vegetation distribution data in the forest vegetation data, a predicted vegetation clutter spectrum is generated.

[0025] Based on the environmental interference data, a predicted atmospheric precipitation particle interference spectrum is generated; based on the vegetation clutter spectrum and the atmospheric precipitation particle interference spectrum, environmental background noise is generated; and an adaptive filtering algorithm is used to process the radar echo data according to the environmental background noise.

[0026] The prediction process for the vegetation clutter spectrum includes:

[0027] An interactive model between radar beams and three-dimensional vegetation canopy is constructed. The digital elevation model, average tree height, and canopy density from the forest vegetation data are input, and the predicted vegetation clutter spectrum is calibrated and spatially corrected.

[0028] A mapping model between leaf type and microDoppler spectrum is constructed. The tree species and leaf type in the forest vegetation data are input, and the spectral morphology of the predicted vegetation clutter spectrum is adjusted.

[0029] The signal adaptive suppression model includes a fire identification and early warning module, which includes a probabilistic inference model based on machine learning.

[0030] The dynamic spectrum characteristics after noise suppression processing are used as the main input, and environmental factors related to fire risk level in the radar environmental data are fused as auxiliary input.

[0031] Based on the decision function and network weights learned through data training within the probabilistic reasoning model, the input multidimensional feature data is nonlinearly mapped into a quantitative confidence probability value representing the existence of an early fire in the target area.

[0032] The process of performing spatiotemporal consistency joint verification on the confidence probability includes: performing a time domain continuity verification to determine whether the confidence probability is continuously higher than the main threshold within a preset continuous detection period; performing a spatial domain clustering verification by using a spatial clustering algorithm to determine whether the monitoring units higher than the main threshold form a connected cluster of a preset size in space; and performing a physical-logical consistency verification to verify whether the spatiotemporal evolution trend of the connected cluster is consistent with the wind field data in the environmental interference data.

[0033] An adaptive noise suppression system for radar echo signals includes:

[0034] The parameter decision module acquires radar station location and beam pointing data, and simultaneously collects forest vegetation data and environmental interference data of the target area pointed to by the beam, forming radar environmental data; the radar environmental data is identified through a multi-band decision model to obtain multi-modal waveform parameters, including X-band parameters and W-band parameters.

[0035] The radar identification module, based on multi-modal waveform parameters, controls the radar equipment to periodically detect the target area and obtain radar echo data.

[0036] The fire detection module constructs a signal adaptive suppression model, identifies radar echo data and radar environment data, extracts dynamic spectrum features from the radar echo data, and uses the radar environment data to perform clutter suppression processing on the dynamic spectrum features, thereby calculating the confidence probability of an early fire in the target area.

[0037] The fire early warning module performs a joint spatiotemporal consistency check on the confidence probability, combines the forest vegetation data to vertically locate the confirmed fire source and determine the fire type, and issues early warning information.

[0038] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0039] 1. This invention achieves active environmental adaptation for radar detection through a multi-band decision model, ensuring that the radar waveform is already optimized for the specific environment and detection task at the source of signal transmission, thereby maximizing the signal-to-noise ratio of effective fire signals in the original echo data; compared with radar systems using fixed parameters, it can significantly improve the ability to capture early weak fire signals in complex and dynamic forest environments.

[0040] 2. This invention effectively separates fire signals from vegetation backgrounds through cross-band structural features, solving the problem of spatial overlap; furthermore, through turbulent chaotic features, it distinguishes fire turbulence, meteorological turbulence, and vegetation movement from the intrinsic physical laws of motion, solving the problem of similar spectral features; it greatly improves the discriminative power and signal-to-noise ratio of features, providing high-quality input for subsequent machine learning models, and is the core technical support for achieving low false alarms and high-precision identification.

[0041] 3. This invention constructs a clutter suppression method based on physical model prediction and adaptive cancellation, which can track environmental changes in real time and dynamically filter out specific clutter spectra that vary with the environment, significantly improving the signal-to-noise ratio of the signal. Compared with traditional fixed bandpass / bandstop filters, it is more intelligent and precise. It can suppress environmental noise to the greatest extent while retaining weak fire signals, greatly reducing the difficulty of subsequent processing and the false alarm rate. Attached Figure Description

[0042] Figure 1 This is a flowchart illustrating an adaptive suppression method for filtering noise in radar echo signals according to the present invention.

[0043] Figure 2 This is a logical schematic diagram of an adaptive suppression method for radar echo signal filtering noise according to the present invention;

[0044] Figure 3 This is a schematic diagram of the structure of an adaptive suppression system for filtering noise in radar echo signals according to the present invention. Detailed Implementation

[0045] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0046] Example 1:

[0047] This invention proposes an adaptive method for suppressing noise in radar echo signal filtering. The process of the method is as follows: Figure 1As shown, the logic of the method is as follows: Figure 2 As shown, it specifically includes:

[0048] The radar station location and beam pointing data are acquired, and forest vegetation data and environmental interference data of the target area pointed to by the beam are collected simultaneously to form radar environmental data. The radar environmental data is identified by a multi-band decision model to obtain multi-modal waveform parameters, including X-band parameters and W-band parameters.

[0049] Preferably, the radar environmental data includes forest vegetation data and environmental interference data;

[0050] The forest vegetation data includes: digital elevation models of the target area, vegetation cover data, and vegetation distribution data including tree species, average tree height, leaf type, and canopy density, obtained by calling and parsing data from a geographic information system; the environmental disturbance data includes: wind speed data, wind direction data, atmospheric temperature, relative humidity, and rainfall data.

[0051] Specifically, a query request containing the geographic coordinates and extent of the target area is sent to the geographic information system server. The server returns a digital elevation model of the area, a vegetation cover layer, and vegetation attribute data including parameters such as tree species (e.g., coniferous forest, broadleaf forest), average tree height, leaf morphology (needle-like, broadleaf), and canopy density. Environmental disturbance data is acquired in real time through local sensors or meteorological data services, including wind speed and direction data that drive vegetation swaying, atmospheric temperature and relative humidity that affect electromagnetic wave attenuation, and rainfall data that may cause strong scattering interference.

[0052] This invention clearly defines the specific content and source of the data, ensuring the accuracy and reliability of the input parameters and providing the necessary prerequisites for the effective operation of the subsequent multi-band decision model and signal adaptive suppression model. It enhances the feasibility of the method and the accuracy of the final fire identification, avoids the risk of model failure due to ambiguous environmental parameter definitions, and provides a solid and operable data foundation for the entire technical solution.

[0053] Preferably, the multi-band decision model includes a first decision module, a second decision module, and a third decision module;

[0054] The first decision module, based on the digital elevation model, vegetation coverage and canopy density in the forest vegetation data, and the calculated X-band path attenuation coefficient, adaptively adjusts the pulse width and peak power of the X-band transmitted waveform with the criterion of maximizing the residual signal-to-noise ratio after the signal energy penetrates the vegetation canopy.

[0055] The second decision module, based on the Mie scattering theory model of standard smoke particles and combined with real-time atmospheric humidity and W-band path attenuation coefficient, adaptively adjusts the center frequency and bandwidth of the W-band transmitted waveform with the criterion of maximizing smoke particle detection.

[0056] The third decision module dynamically adjusts the pulse repetition frequency of the radar equipment based on the wind speed and wind direction data in the environmental interference data.

[0057] The multimode waveform parameters are obtained based on the pulse width and peak power of the X-band transmitted waveform, the center frequency and bandwidth of the W-band transmitted waveform, and the pulse repetition frequency.

[0058] Specifically, the first decision module optimizes the X-band, which is an optimization process with penetration as the target. It receives vegetation coverage, canopy density and digital elevation model data, combines the attenuation model of the X-band in the atmosphere and vegetation, and performs iterative calculations through optimization algorithms (such as particle swarm optimization) to find a pulse width and peak power combination that can make the signal penetrate the canopy and have the highest remaining signal-to-noise ratio.

[0059] The second decision module optimizes the W-band, which is an optimization process aimed at sensitivity. Based on Mie scattering theory, it describes the interaction between electromagnetic waves and particles of similar size and wavelength (such as smoke and dust). It adjusts the atmospheric attenuation model according to real-time humidity data, and then performs scanning calculations within a preset frequency range to find the center frequency and bandwidth with the largest scattering cross section for standard smoke particles and the highest detection sensitivity.

[0060] The third decision module optimizes the pulse repetition frequency, which is an optimization process aimed at eliminating ambiguity. It directly calculates the maximum Doppler frequency shift that may be caused by vegetation swaying based on real-time wind speed data and the Doppler effect principle. In order to avoid speed ambiguity (high-speed moving leaves are misjudged as low-speed targets), a sufficiently high pulse repetition frequency is dynamically selected to ensure that the unambiguous speed range can completely cover the estimated maximum vegetation movement speed.

[0061] This invention achieves active environmental adaptation for radar detection through a multi-band decision model, ensuring that the radar waveform is already optimized for the specific environment and detection task at the source of signal transmission, thereby maximizing the signal-to-noise ratio of effective fire signals in the original echo data. Compared with radar systems using fixed parameters, it can significantly improve the ability to capture early weak fire signals in complex and dynamic forest environments.

[0062] Furthermore, the multimodal waveform parameters also include a dynamic waveform synthesis strategy;

[0063] The dynamic waveform synthesis strategy includes: dynamically synthesizing and interleaving X-band and W-band transmission pulse sequences within a coherent processing interval based on the evaluation results of real-time signal-to-noise ratio and clutter distribution; when vegetation clutter is strong, a high-repetition-rate X-band-W-band alternating transmission mode is adopted to obtain near-synchronous structural background and dynamic particle information; when the plume signal is weak, a W-band pulse cluster mode is adopted to concentrate energy on smoke and dust detection in a short time; the composite waveform sequence is synthesized in real time by an arbitrary waveform generator and transmitted to achieve optimal dynamic allocation of detection resources in the time and frequency domains.

[0064] Specifically, before each coherent processing interval begins, the clutter intensity and target signal-to-noise ratio in the current beam direction are quickly assessed using echo data from the previous cycle. The assessment results are then input into the waveform synthesis strategy module, which makes decisions based on preset rules or a reinforcement learning agent. For example, if high-dynamic vegetation clutter caused by strong winds is detected, the module outputs an alternating pulse sequence command of "XWXW...". If a suspected weak plume is detected, a pulse cluster command of "WWWX..." is output. This command is sent to the radar's arbitrary waveform generator, which generates and outputs a complex, non-uniform pulse sequence containing different frequency bands and pulse widths in real time according to the command. The receiver then synchronously separates and processes the echoes of different frequency bands according to the transmission timing.

[0065] This invention advances adaptive capabilities from macroscopic parameter settings to microscopic pulse-level real-time waveform synthesis and scheduling, enabling refined management of detection resources. By dynamically interweaving waveforms of different frequency bands within an extremely short timescale, it trades time for performance, maximizing the suppression of time mismatch effects. In environments with strong clutter, it achieves near-synchronous acquisition of structural background (X-band) and dynamic particle information (W-band), thus enabling near-perfect background cancellation. It significantly improves the ability to extract weak signals in highly dynamic environments, further lowering the system's detectable signal-to-noise ratio lower limit.

[0066] Based on multimodal waveform parameters, the radar equipment is controlled to periodically detect the target area and obtain radar echo data.

[0067] An adaptive signal suppression model is constructed to identify radar echo data and radar environment data, extract dynamic spectral features from the radar echo data, and use the radar environment data to perform clutter suppression processing on the dynamic spectral features to calculate the confidence probability of early fire in the target area.

[0068] The adaptive signal suppression model includes an echo feature extraction module, which extracts features from radar echo data to obtain dynamic spectrum features; the dynamic spectrum features include cross-band structural features and turbulent chaos features.

[0069] The cross-band structural features utilize X-band echo signals to process and generate a three-dimensional structural background mask representing the tree trunk, main branches, and ground surface; leveraging the high sensitivity of W-band signals to microparticles, the spatial location and motion intensity of dynamic scatterers are identified, resulting in a dynamic scatterer map; the dynamic scatterers include vegetation leaves and dust particles; and the cross-band structural features are generated using the W-band dynamic scatterer map and the X-band structural background mask.

[0070] The turbulent chaotic characteristics are obtained by performing high-resolution time-frequency analysis on the W-band echo, calculating and extracting parameters to characterize the thermal turbulent chaotic characteristics above the fire source, including the energy entropy of the time-frequency distribution, the spectral fractal dimension, and the Lyapunov exponent characterizing the nonlinear evolution of the spectrum.

[0071] Furthermore, the cross-band structural feature utilizes the difference in response of the X-band and W-band to different targets. In practice, the X-band echo, which has strong penetrability and is sensitive to coarse structures, is processed first. A static three-dimensional structural background mask containing only tree trunks, main branches, and the ground surface is generated using a three-dimensional imaging algorithm (such as a back projection algorithm). Subsequently, the W-band echo, which is sensitive to small particles, is processed to generate a dynamic scatterer map containing all dynamic scatterers (swaying leaves, airborne dust particles). Based on the spatial registration of the W-band dynamic scatterer map and the X-band structural background mask, differential or logical operations can be performed to effectively remove the fixed tree trunk background and some leaf echoes attached to the branches, thereby highlighting the dynamic targets of non-vegetation structures suspended in space, i.e., potential dust plumes.

[0072] Turbulent chaos features aim to capture the unique physical properties of turbulence generated by rising hot air above a fire source. For suspicious regions identified in the W-band echo, high-resolution time-frequency analysis tools (such as the Wigner-Ville distribution or continuous wavelet transform) are used to generate their dynamic spectra. Then, a series of nonlinear dynamic parameters are identified: energy entropy is used to quantify the disorder and randomness of the spectrum; spectral fractal dimension is used to describe the self-similarity and complexity of the spectrum at different scales; and the Lyapunov exponent is used to measure the predictability of the system's evolution over time, with high positive values ​​representing chaotic states.

[0073] Specifically, based on the dynamic spectrogram, the energy entropy quantifying spectral disorder is first calculated by analyzing the uniformity and dispersion of energy distribution. Secondly, the spectral fractal dimension characterizing its morphological features is calculated by analyzing the geometric complexity and irregularity of the dynamic spectrogram. Furthermore, the phase space of the Doppler time-series signal of the monitoring unit within the suspicious region is reconstructed. First, the optimal time delay and embedding dimension parameters are determined using a standardization method. Then, based on these two key parameters, the original one-dimensional time-series signal is expanded into a high-dimensional state trajectory that fully reveals its inherent dynamic characteristics. On this high-dimensional state trajectory, a specific algorithm suitable for small data samples is used to track the evolution of trajectory points with similar initial positions over time, and the average rate of separation and divergence of these trajectory paths is calculated to obtain the Lyapunov exponent.

[0074] Fire thermal turbulence exhibits highly random, complex, and chaotic characteristics in these metrics, which are quite different from wind-induced vegetation swaying (which is more periodic).

[0075] This invention effectively isolates fire signals from vegetation backgrounds through cross-band structural features, solving the problem of spatial overlap. Furthermore, by utilizing turbulent chaotic features, it distinguishes between fire turbulence, meteorological turbulence, and vegetation movement based on the inherent physical laws of motion, thus addressing the issue of similar spectral features. This significantly improves the discriminative power and signal-to-noise ratio of the features, providing high-quality input for subsequent machine learning models and serving as the core technological support for achieving low false alarms and high-precision identification.

[0076] Furthermore, the dynamic spectral characteristics may also include microphysical morphological characteristics based on full polarization information; the process of acquiring the microphysical morphological characteristics includes:

[0077] The radar device is controlled to transmit electromagnetic waves in two orthogonal polarization states, horizontal (H) and vertical (V), in a fully polarized mode, and simultaneously receive echo signal components of the same polarization (HH,VV) and cross polarization (HV,VH).

[0078] Based on the fully polarized echo data, differential reflectivity, linear depolarization ratio, and differential propagation phase are calculated and extracted. The microphysical morphological features are then fused with the cross-band structural features and turbulent chaotic features to form an enhanced dynamic spectral feature vector for subsequent clutter suppression and probabilistic solution, in order to distinguish scatterers with similar dynamic characteristics but different microphysical morphologies, such as dust particles, raindrops, and insect swarms.

[0079] Specifically, at the radar hardware level, the antenna system needs to be equipped with orthogonal mode couplers and dual-channel receivers to support dual-polarization synchronous transmission and reception. At the signal processing level, the system first calculates the covariance matrix from the echoes of the four polarization channels (HH, VV, HV, VH). Subsequently, polarization parameters are calculated through different combinations of matrix elements: differential reflectivity is obtained from the power ratio of the HH and VV channels, reflecting the average shape of the scatterer; linear depolarization ratio is obtained from the power ratio of the HV and HH channels, reflecting the irregularity and tumbling motion of the scatterer; differential propagation phase is obtained by analyzing the rate of change of the phase difference between the HH and VV channels along the distance, which is sensitive to the liquid water content along the path. Finally, this set of polarization parameters is concatenated with the extracted features to form a higher-dimensional feature vector, which is then input into the subsequent machine learning model.

[0080] Traditional spectral features sometimes fail to distinguish between fire smoke and raindrops or insect swarms in severe convective weather. This invention introduces polarization features to enable refined microphysical identification, allowing machine learning models to easily distinguish features from different situations. This fundamentally solves the false alarm problem caused by rainfall or biological clusters, enhances the robustness of the method under complex weather conditions and its ability to work in all weather conditions, and significantly improves the accuracy of identification.

[0081] The signal adaptive suppression model includes an echo data processing module, which performs clutter suppression processing on dynamic spectral characteristics using radar environmental data, including:

[0082] Based on the wind speed and wind direction data in the environmental interference data and the vegetation distribution data in the forest vegetation data, a predicted vegetation clutter spectrum is generated.

[0083] Based on the environmental interference data, a predicted atmospheric precipitation particle interference spectrum is generated; based on the vegetation clutter spectrum and the atmospheric precipitation particle interference spectrum, environmental background noise is generated; and an adaptive filtering algorithm is used to process the radar echo data according to the environmental background noise.

[0084] Furthermore, vegetation clutter spectrum prediction is a model-based prediction process; it receives wind speed and direction from environmental data, as well as parameters such as tree species, canopy density, and leaf type from vegetation data, calls a pre-established vegetation micro-Doppler model library, and predicts the Doppler spectrum (i.e., vegetation clutter spectrum) generated by vegetation swaying under this environment based on the combination of input parameters.

[0085] The atmospheric precipitation particle interference spectrum prediction process uses real-time meteorological data such as rainfall rate and humidity, and employs scattering models (such as Rayleigh scattering or Mie scattering models) to predict the interference echo spectrum generated by particles such as raindrops and fog.

[0086] The two predicted interference spectra are combined to form a complete environmental background noise baseline. Then, an adaptive filtering algorithm, such as the least mean square error algorithm or the recursive least squares algorithm, is used as a reference input to process the actual received radar echo data. The filter automatically adjusts its internal parameters to suppress the components of the output signal related to the reference noise to the greatest extent possible.

[0087] This invention constructs a clutter suppression method based on physical model prediction and adaptive cancellation, which can track environmental changes in real time and dynamically filter out specific clutter spectra that vary with the environment, significantly improving the signal-to-noise ratio. Compared with traditional fixed bandpass / bandstop filters, it is more intelligent and precise. It can suppress environmental noise to the greatest extent while preserving weak fire signals, greatly reducing the difficulty of subsequent processing and the false alarm rate.

[0088] The prediction process for the vegetation clutter spectrum also includes:

[0089] An interactive model between radar beams and three-dimensional vegetation canopy is constructed. The digital elevation model, average tree height, and canopy density from the forest vegetation data are input, and the predicted vegetation clutter spectrum is calibrated and spatially corrected.

[0090] A mapping model between leaf type and microDoppler spectrum is constructed. The tree species and leaf type in the forest vegetation data are input, and the spectral morphology of the predicted vegetation clutter spectrum is adjusted.

[0091] A radar beam-to-3D vegetation canopy interaction model solves the problem of predicting clutter power and spatial distribution. In implementation, a 3D forest canopy surface is first constructed using a digital elevation model and average tree height data. Then, the radar beam's transmission path, beamwidth, and energy distribution are simulated to calculate the intersection volume (i.e., illumination volume) between the beam and this 3D canopy. Combined with canopy density data, the model can accurately calibrate the total power of vegetation clutter under the beam's direction and correct its spatial distribution in terms of distance and azimuth, ensuring that the predicted clutter power spectrum matches the actual situation.

[0092] The model mapping relationship between blade type and micro-Doppler spectrum solves the problem of predicting the spectral morphology of clutter spectra. It is a knowledge base or machine learning model built upon a large amount of simulation or measured data, storing the characteristic micro-Doppler spectra generated by different blade types (such as pine needles and broadleaf blades) under different wind fields. For example, a large number of small pine needles in the wind will produce a wider spectrum that is closer to a Gaussian distribution, while larger broadleaf blades may produce a non-Gaussian spectrum with specific periodic modulation components due to flipping and twisting. During prediction, based on the input tree species and blade type, the corresponding spectral morphology function is called from the model library or generated to finely adjust the basic clutter spectrum.

[0093] This invention introduces two high-order physical models, elevating the prediction of vegetation clutter from a coarse empirical model to a refined physical simulation level. This enables the adaptive filtering algorithm to obtain a reference signal that is highly consistent with the real clutter in terms of power, spatial distribution, and spectral morphology, thereby achieving more thorough and accurate clutter cancellation. It also greatly enhances the ability to understand the environment, providing a foundation for separating extremely weak early fire signals from a strong clutter background.

[0094] The signal adaptive suppression model includes a fire identification and early warning module, which includes a probabilistic inference model based on machine learning.

[0095] The dynamic spectrum characteristics after noise suppression processing are used as the main input, and environmental factors related to fire risk level in the radar environmental data are fused as auxiliary input.

[0096] Based on the decision function and network weights learned through data training within the probabilistic reasoning model, the input multidimensional feature data is nonlinearly mapped into a quantitative confidence probability value representing the existence of an early fire in the target area.

[0097] In implementation, the probabilistic inference model based on machine learning can be a gradient boosting decision tree, support vector machine or deep neural network, etc. The model is designed to have two types of inputs: the main input receives the dynamic spectrum feature vector after clutter suppression processing; the auxiliary input receives environmental factors directly related to the fire hazard level, such as relative humidity, temperature and wind speed.

[0098] Training and Inference: Before deployment, the model requires supervised learning training on a labeled dataset containing a large number of known "fire" and "non-fire" samples. The training process adjusts the model's internal parameters (such as the structure of the decision tree and the weights of the network layers) to enable it to learn the complex nonlinear mapping relationship between input features and output results. During actual runtime, the model receives real-time extracted features and environmental factors, performs forward propagation calculations using its internally fixed decision functions or network weights, and ultimately outputs a continuous value between 0 and 1, i.e., the quantized confidence probability value.

[0099] This invention introduces machine learning technology to automate the complex feature fusion and judgment process, avoiding the difficulties and one-sidedness of manually setting complex thresholds and rules. It can discover potential correlations that are difficult for human experts to detect. Furthermore, it integrates environmental factors as auxiliary inputs to provide background knowledge for the model, making its judgments more logical. This significantly improves the accuracy and robustness of fire identification and enables the system to make high-performance decisions.

[0100] Furthermore, the machine learning-based probabilistic reasoning model can also be a spatiotemporal graph neural network model;

[0101] The construction and inference process of the spatiotemporal graph neural network model includes: dividing the three-dimensional space detected by radar into a series of gridded nodes to construct a spatial graph structure, where each node represents a monitoring unit, and the initial feature of the node is the dynamic spectral feature extracted by the unit; constructing a time series graph sequence based on snapshots of multiple consecutive detection cycles; aggregating the spatial feature information of neighboring nodes in each time step through graph convolutional network layers, and capturing the evolutionary dependency of each node's features in the time dimension through recurrent neural network layers or attention mechanisms; finally, the model outputs the confidence probability of each node having an early fire at the current moment, realizing end-to-end modeling of the spatiotemporal correlation of fires.

[0102] Specifically, firstly, based on the radar's range and azimuth resolution, the entire monitoring area is discretized into a set of nodes in a three-dimensional graph. Connections (edges) between nodes can be predefined based on spatial adjacency relationships (e.g., Euclidean distance less than a certain threshold). In each detection cycle T, the system calculates the dynamic spectral feature vector for each node, forming the graph signal at that moment. The graph signal sequence for M consecutive cycles (T, T-1, ..., T-M+1) is used as the model input. Inside the model, graph convolutional layers are responsible for feature propagation and fusion on the graph at each moment, enabling each node's decision to be aware of the state of its neighbors. Subsequently, a temporal evolution modeling layer (such as GRU or Transformer) receives the spatially fused feature sequence of each node and learns its temporal variation pattern. Finally, through a fully connected layer and a sigmoid activation function, the fire probability of each node at time T is output.

[0103] This invention elevates fire identification from a series of isolated, single-point classification tasks to a holistic and interconnected understanding of a dynamically evolving system. It can explicitly learn the spatiotemporal patterns of fire source spread, growth, and wind diffusion, enabling the model not only to identify fires but also to understand their development trends. It has a strong ability to suppress spatially correlated interference (such as moving rain clouds) and is more sensitive to emerging and emerging weak fire clusters, thus detecting fires at an earlier stage with lower false alarm and false alarm rates.

[0104] A spatiotemporal consistency joint verification is performed on the confidence probability, and the confirmed fire source is vertically located and the fire type is determined by combining the forest vegetation data, and an early warning information is issued.

[0105] The process of performing spatiotemporal consistency joint verification on the confidence probability includes:

[0106] Perform a time-domain continuity check to determine whether the confidence probability is consistently higher than the main threshold within a preset continuous detection period;

[0107] Perform spatial domain clustering verification, using a spatial clustering algorithm to determine whether monitoring units above the main threshold form connected clusters of a preset size in space;

[0108] Perform a physical-logical consistency check to verify whether the spatiotemporal evolution trend of the connected clusters matches the wind field data in the environmental interference data.

[0109] Furthermore, the time-domain continuity check is performed within a sliding time window. For example, the continuous detection period is set to N (e.g., N=3), and the main threshold is P. When the confidence probability of a monitoring unit first exceeds P, an alarm is not immediately triggered; instead, it is marked as "suspected." Over the next N-1 detection periods, the system continuously tracks the unit, and only when its confidence probability is higher than P N times consecutively does it pass the time check. Any instantaneous, impulsive high-probability signals (such as electromagnetic interference) are filtered out by this step.

[0110] Spatial domain clustering verification is performed on 3D spatial data from a single scan. All monitoring units with confidence probabilities higher than the main threshold P are extracted, and then a spatial clustering algorithm is applied to find spatially adjacent units and group them into clusters. It is checked whether the "suspected" units that pass the time verification belong to a connected cluster whose size (i.e., the number of units contained) and shape meet the preset conditions. Isolated, single high-probability points are likely caused by noise or model misjudgment, and this verification can effectively eliminate them.

[0111] Physical-logical consistency verification: For high-confidence connected clusters that have passed both temporal and spatial verifications, their spatiotemporal evolution trends over the past few detection periods are analyzed (e.g., the direction and velocity of the centroid's movement). Then, wind field data (wind direction and speed) from the environmental database for the same period are retrieved to verify whether the evolution trend of the event cluster is physically and logically consistent with the blowing action of the wind (e.g., the direction of plume diffusion should roughly follow the wind direction). If the movement trajectory of a connected cluster significantly contradicts the wind field data, the system will reduce its confidence level or even classify it as a false alarm.

[0112] This invention constructs a multi-dimensional, multi-logic verification and filtering system, which effectively filters out false alarms caused by transient interference, spatially isolated noise points, and abnormal signals that do not conform to physical laws, ensuring that the warning information output by the system has a very high degree of confidence; it not only enhances the practical value of the system, but also strengthens the user's trust in the system.

[0113] Before issuing the early warning information, a fire type identification module based on vertical evolution analysis is also included. After confirming a high-confidence fire event, this module performs the following steps: extracting the spatial location sequence of the three-dimensional connected clusters corresponding to the event within a continuous observation period; calculating its four-dimensional spatiotemporal centroid and tracing the altitude of its origin point; comparing the altitude of the origin point with the digital elevation model and the average height map of the tree canopy in the forest vegetation data; if the origin point is located near the ground surface and shows a propagation characteristic of penetrating the tree canopy from bottom to top in time, it is identified as a "surface fire"; if the origin point is located at the height of the tree canopy and the outbreak is sudden, it is identified as a "crown fire", and a differentiated early warning information level is generated based on the identification results.

[0114] Example 2:

[0115] This invention also proposes an adaptive noise suppression system for radar echo signal filtering, used to implement the aforementioned adaptive noise suppression method for radar echo signal filtering. The structure of the system is as follows: Figure 3 As shown, it includes:

[0116] The parameter decision module acquires radar station location and beam pointing data, and simultaneously collects forest vegetation data and environmental interference data of the target area pointed to by the beam to form radar environmental data; the radar environmental data is identified through a multi-band decision model to obtain multi-modal waveform parameters, including X-band parameters and W-band parameters.

[0117] This module is responsible for dynamically optimizing and generating multi-mode waveform parameters for radar transmission based on the acquired environmental data.

[0118] First, obtain the precise location of the radar station and its current beam pointing data; simultaneously, collect radar environmental data of the target area.

[0119] The radar environmental data includes forest vegetation data and environmental interference data. The forest vegetation data is obtained by calling the geographic information system database, specifically the digital elevation model of the target area, vegetation coverage data, and vegetation distribution data, such as the tree species being Scots pine, the average tree height being 15 meters, the leaf type being needle-like leaves, and the canopy density being 0.7.

[0120] Environmental disturbance data is obtained from local weather stations, such as current wind speed of 3 meters per second, wind direction of northwest, atmospheric temperature of 25 degrees Celsius, relative humidity of 60%, and no rainfall.

[0121] After acquiring the data, the multi-band decision model within the module begins to run. This model consists of three decision modules:

[0122] The first decision-making module is responsible for optimizing X-band parameters.

[0123] This module aims to maximize the residual signal-to-noise ratio (SNR) after the signal penetrates vegetation. The residual SNR is estimated using a physical model, which includes vegetation two-way attenuation. This vegetation two-way attenuation is calculated using empirical formulas based on input canopy density and average tree height data. Furthermore, a particle swarm optimization algorithm is used to iteratively calculate within a selectable pulse width and peak power range to ultimately determine the optimal combination.

[0124] The second decision-making module is responsible for optimizing W-band parameters.

[0125] This module is designed to maximize the detection capability of smoke and dust particles. The "standard smoke particle" model it relies on is specifically defined in this embodiment as follows: assuming spherical carbonaceous particles with a log-normal particle size distribution, based on Mie scattering theory and combined with an atmospheric attenuation coefficient corrected for real-time atmospheric humidity, scanning calculations are performed within the radar's selectable center frequency range to find the optimal center frequency and bandwidth that maximizes the average backscattering cross-section of this particle model.

[0126] In this embodiment, the standard smoke particulate matter model is defined as follows: the particulate matter is spherical carbonaceous particles with a particle size distribution following a log-normal distribution, a median particle size of 0.5 micrometers, and a geometric standard deviation of 1.5. Based on these specific parameters, the backscattering cross-section at different center frequencies is calculated using Mie scattering theory, thereby determining the optimal frequency.

[0127] The third decision-making module is responsible for optimizing the pulse repetition frequency.

[0128] Based on the current wind speed, the module calculates that the maximum Doppler frequency shift caused by the swaying vegetation is about 300 Hz. To ensure that the speed is not blurred, the radar pulse repetition frequency is dynamically selected to be 1000 Hz.

[0129] Finally, the parameter decision module combines the X-band parameters, W-band parameters, and pulse repetition frequency to form multimodal waveform parameters, which are then sent to the radar identification module.

[0130] The radar identification module, based on multi-modal waveform parameters, controls the radar equipment to periodically detect the target area and obtain radar echo data.

[0131] This module receives multimodal waveform parameters generated by the parameter decision module and controls the radar hardware accordingly to perform periodic scanning detection on the target area, thereby obtaining raw radar echo data. In this embodiment, the radar system adopts an alternating transmission mode, alternately transmitting optimized X-band pulses and W-band pulses within a coherent processing interval to obtain near-synchronous dual-band observation data.

[0132] The fire detection module constructs a signal adaptive suppression model to identify radar echo data and radar environmental data, extracts dynamic spectral features from the radar echo data, and uses the radar environmental data to perform clutter suppression processing on the dynamic spectral features to calculate the confidence probability of an early fire in the target area.

[0133] An adaptive signal suppression model is constructed to perform in-depth processing and analysis of radar echo data, ultimately calculating the confidence probability of early-stage fires in the target area. The internal processing flow is as follows:

[0134] Clutter suppression processing: First, the module uses radar environmental data to suppress clutter in dynamic spectrum characteristics.

[0135] Vegetation clutter spectrum prediction: Based on environmental disturbance data (wind speed 3 meters per second) and forest vegetation data (Pinus sylvestris, coniferous trees), the predicted vegetation clutter spectrum is generated; this process is completed through two refined physical models.

[0136] The first is the interaction model between the radar beam and the three-dimensional vegetation canopy. In this embodiment, a volume scattering model is adopted, which regards the canopy voxels illuminated by the radar beam as a set of independent scattering bodies. Their radar cross-section is related to the canopy density. The total power of clutter is calibrated and predicted by integrating the scattering cross-section of all voxels within the beam illumination volume.

[0137] The second is the mapping relationship model between blade type and micro-Doppler spectrum. In this embodiment, a pre-trained three-layer feedforward neural network is used. Its input is wind speed and blade type (needle blade), and its output is the shape and scale parameters of the Weibull distribution describing the morphology of the Doppler spectrum, thereby finely adjusting the morphology of the predicted clutter spectrum.

[0138] Adaptive filtering: Since there is currently no rainfall, the power of atmospheric precipitation particle interference spectrum is zero; the predicted vegetation clutter spectrum is used as a reference signal for environmental background noise, and an adaptive filtering algorithm (such as the minimum mean square error algorithm) is used to process the actual received radar echo data to suppress clutter components related to vegetation movement to the greatest extent.

[0139] Dynamic spectral feature extraction: Two key dynamic spectral features are extracted from the echo data after clutter suppression.

[0140] Cross-band structural features: Using X-band echo data, a three-dimensional structural background mask representing the tree trunk, main branches, and ground surface is generated through a back projection three-dimensional imaging algorithm; at the same time, using W-band echo, the spatial location and motion intensity of dynamic scatterers (swaying leaves, possible dust particles) are identified, and a dynamic scatterer map is obtained.

[0141] Next, spatial registration was performed on the three-dimensional data of the two bands. Specifically, using a high-resolution three-dimensional grid of the W-band as a reference, the background mask data of the X-band was resampled using a trilinear interpolation algorithm to ensure voxel alignment. After registration, a specific difference operation was performed to generate features: first, the W-band spectral intensity values ​​and the X-band mask intensity values ​​were normalized; then, the feature value was calculated, which was equal to the difference between the normalized W-band intensity value and the normalized X-band intensity value multiplied by a weighting coefficient, and the larger of the result and zero was taken. This method effectively highlights dynamic targets with non-vegetation structures suspended in space.

[0142] Turbulent Chaotic Characteristics: For the suspicious dynamic target region identified in the W-band echo, high-resolution time-frequency analysis is performed using continuous wavelet transform; then, the phase space is reconstructed from its Doppler time series, and parameters used to characterize the thermal turbulent chaotic characteristics above the fire source are calculated, including the energy entropy of the time-frequency distribution, the spectral fractal dimension, and the maximum Lyapunov exponent characterizing the nonlinear evolution of the spectrum.

[0143] Fire probability calculation: The extracted cross-band structural features and turbulent chaotic features are used as the main inputs, and environmental factors related to the fire risk level (such as relative humidity of 60% and temperature of 25 degrees Celsius) from radar environmental data are fused as auxiliary inputs. These multidimensional feature data are fed into a machine learning-based probabilistic inference model, which in this embodiment is a gradient boosting decision tree model. Based on the decision function and network weights learned by the model through training on a large amount of data, the input multidimensional feature data is nonlinearly mapped into a quantitative confidence probability value representing the existence of an early fire in the target area.

[0144] The fire early warning module performs a joint spatiotemporal consistency check on the confidence probability, combines the forest vegetation data to vertically locate the confirmed fire source and determine the fire type, and issues early warning information.

[0145] This module performs a joint spatiotemporal consistency check on the confidence probability output by the fire detection module to filter out false alarms and issue early warning information after confirming the fire.

[0146] Time domain continuity verification: Determine whether the confidence probability of a certain monitoring unit is continuously higher than the main threshold (e.g., 0.8) within a preset continuous detection period (e.g., three consecutive periods, each period being 30 seconds).

[0147] Spatial domain clustering verification: Using a spatial clustering algorithm, it is determined whether the monitoring units that have passed the time verification have formed a connected cluster in space that meets the preset size (e.g., contains at least five adjacent units).

[0148] Physical-logical consistency verification: For the formed connected clusters, analyze their spatiotemporal evolution trend (such as the direction of centroid movement) over the past few cycles, and verify whether the trend is physically and logically consistent with the wind field data (northwest wind, 3 meters per second) in the environmental data, that is, the direction of plume diffusion should roughly follow the wind direction.

[0149] Once a suspected incident passes the aforementioned triple verification, it is confirmed as a real fire. Subsequently, the module combines the digital elevation model and the average height map of the tree canopy in the forest vegetation data to vertically locate the confirmed fire source and determine the fire type; for example, it is determined to be a "surface fire" or a "crown fire," and finally issues detailed early warning information to the monitoring center, including the location of the fire point, the confidence level, and the fire type.

[0150] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An adaptive method for suppressing noise in radar echo signal filtering, characterized in that, include: The radar station location and beam pointing data are acquired, and forest vegetation data and environmental interference data of the target area pointed to by the beam are collected simultaneously to form radar environmental data. The radar environment data is identified by a multi-band decision model to obtain multi-modal waveform parameters, which include X-band parameters and W-band parameters. Based on multimodal waveform parameters, the radar equipment is controlled to periodically detect the target area and obtain radar echo data. An adaptive signal suppression model is constructed to identify radar echo data and radar environment data, extract dynamic spectral features from the radar echo data, and use the radar environment data to perform clutter suppression processing on the dynamic spectral features to calculate the confidence probability of early fire in the target area. The adaptive signal suppression model includes an echo feature extraction module, which extracts features from radar echo data to obtain dynamic spectrum features; the dynamic spectrum features include cross-band structural features and turbulent chaos features. The cross-band structural features utilize X-band echo signals to process and generate a three-dimensional structural background mask representing the tree trunk, main branches, and ground surface. By utilizing the high sensitivity of W-band signals to microparticles, the spatial location and motion intensity of dynamic scatterers are identified, and a dynamic scatterer map is obtained. The dynamic scatterers include vegetation leaves and dust particles. Cross-band structural features are generated by combining the W-band dynamic scatterer map with the X-band structural background mask. The turbulent chaotic characteristics are obtained by performing high-resolution time-frequency analysis on the W-band echo, calculating and extracting parameters to characterize the thermal turbulent chaotic characteristics above the fire source, including the energy entropy of the time-frequency distribution, the spectral fractal dimension, and the Lyapunov exponent characterizing the nonlinear evolution of the spectrum. The turbulent chaos feature aims to capture the physical properties of turbulence formed above the fire source by the rising hot air. Fire thermal turbulence exhibits randomness, complexity, and chaotic characteristics that are different from the periodic swaying of wind-induced vegetation. A spatiotemporal consistency joint verification is performed on the confidence probability, and the confirmed fire source is vertically located and the fire type is determined by combining the forest vegetation data, and an early warning information is issued.

2. The adaptive suppression method for radar echo signal filtering noise according to claim 1, characterized in that: The radar environmental data includes forest vegetation data and environmental interference data; The forest vegetation data includes: digital elevation models of the target area, vegetation cover data, and vegetation distribution data including tree species, average tree height, leaf type, and canopy density, obtained by calling and parsing data from a geographic information system; the environmental disturbance data includes: wind speed data, wind direction data, atmospheric temperature, relative humidity, and rainfall data.

3. The adaptive suppression method for radar echo signal filtering noise according to claim 1, characterized in that: The multi-band decision model includes a first decision module, a second decision module, and a third decision module; The first decision module, based on the digital elevation model, vegetation coverage and canopy density in the forest vegetation data, and the calculated X-band path attenuation coefficient, adaptively adjusts the pulse width and peak power of the X-band transmitted waveform with the criterion of maximizing the residual signal-to-noise ratio after the signal energy penetrates the vegetation canopy. The second decision module, based on the Mie scattering theory model of standard smoke particles and combined with real-time atmospheric humidity and W-band path attenuation coefficient, adaptively adjusts the center frequency and bandwidth of the W-band transmitted waveform with the criterion of maximizing smoke particle detection. The third decision module dynamically adjusts the pulse repetition frequency of the radar equipment based on the wind speed and wind direction data in the environmental interference data. The multimode waveform parameters are obtained based on the pulse width and peak power of the X-band transmitted waveform, the center frequency and bandwidth of the W-band transmitted waveform, and the pulse repetition frequency.

4. The adaptive suppression method for radar echo signal filtering noise according to claim 1, characterized in that: The signal adaptive suppression model includes an echo data processing module, which performs clutter suppression processing on dynamic spectral characteristics using radar environmental data, including: Based on the wind speed and wind direction data in the environmental interference data and the vegetation distribution data in the forest vegetation data, a predicted vegetation clutter spectrum is generated. Based on the environmental interference data, a predicted atmospheric precipitation particle interference spectrum is generated; based on the vegetation clutter spectrum and the atmospheric precipitation particle interference spectrum, environmental background noise is generated; and an adaptive filtering algorithm is used to process the radar echo data according to the environmental background noise.

5. The adaptive suppression method for radar echo signal filtering noise according to claim 4, characterized in that: The prediction process for the vegetation clutter spectrum also includes: An interactive model between radar beams and three-dimensional vegetation canopy is constructed. The digital elevation model, average tree height, and canopy density from the forest vegetation data are input, and the predicted vegetation clutter spectrum is calibrated and spatially corrected. A mapping model between leaf type and microDoppler spectrum is constructed. The tree species and leaf type in the forest vegetation data are input, and the spectral morphology of the predicted vegetation clutter spectrum is adjusted.

6. The adaptive suppression method for radar echo signal filtering noise according to claim 1, characterized in that: The signal adaptive suppression model includes a fire identification and early warning module, which includes a probabilistic inference model based on machine learning. The dynamic spectrum characteristics after noise suppression processing are used as the main input, and environmental factors related to fire risk level in the radar environmental data are fused as auxiliary input. Based on the decision function and network weights learned through data training within the probabilistic reasoning model, the input multidimensional feature data is nonlinearly mapped into a quantitative confidence probability value representing the existence of an early fire in the target area.

7. The adaptive suppression method for radar echo signal filtering noise according to claim 1, characterized in that: The process of performing spatiotemporal consistency joint verification on the confidence probability includes: Perform a time-domain continuity check to determine whether the confidence probability is consistently higher than the main threshold within a preset continuous detection period; Perform spatial domain clustering verification, using a spatial clustering algorithm to determine whether monitoring units above the main threshold form connected clusters of a preset size in space; Perform a physical-logical consistency check to verify whether the spatiotemporal evolution trend of the connected clusters matches the wind field data in the environmental interference data.

8. An adaptive noise suppression system for radar echo signals, characterized in that, An adaptive suppression method for filtering noise in radar echo signals according to any one of claims 1-7 includes: The parameter decision module acquires radar station location and beam pointing data, and simultaneously collects forest vegetation data and environmental interference data of the target area pointed to by the beam, forming radar environmental data; the radar environmental data is identified through a multi-band decision model to obtain multi-modal waveform parameters, including X-band parameters and W-band parameters. The radar identification module, based on multi-modal waveform parameters, controls the radar equipment to periodically detect the target area and obtain radar echo data. The fire detection module constructs a signal adaptive suppression model, identifies radar echo data and radar environment data, extracts dynamic spectrum features from the radar echo data, and uses the radar environment data to perform clutter suppression processing on the dynamic spectrum features, thereby calculating the confidence probability of an early fire in the target area. The fire early warning module performs a joint spatiotemporal consistency check on the confidence probability, combines the forest vegetation data to vertically locate the confirmed fire source and determine the fire type, and issues early warning information.

Images