Mountain fire monitoring method and system based on polarization radar

Abstract

The invention relates to the technical field of forest fire monitoring, and discloses a forest fire monitoring method based on a polarization radar, and the method comprises the steps: S10, obtaining the body scanning data of a monitoring point through the polarization radar, the body scanning data comprising polarization variables, and the polarization variables comprising a horizontal reflectivity, a differential reflectivity, a cross correlation coefficient and a radial speed standard deviation; s20, the polarization variables are preprocessed, the preprocessing comprises normalization and abnormal value cutting processing, and the polarization variables with effective pixel masks are constructed; s30, performing initial classification on the preprocessed polarization variables based on a fuzzy logic algorithm, performing confirmation by adopting multi-source cross validation, and generating forest fire segmentation labeling data of the monitoring points; s40, performing model training by using the mountain fire segmentation historical annotation data set to obtain a mountain fire monitoring model; and S50, inputting the mountain fire segmentation labeling data of the monitoring point into a mountain fire monitoring model to obtain mountain fire monitoring data.

Description

Technical Field

[0001] This invention relates to the field of wildfire monitoring technology, and in particular to a wildfire monitoring method and system based on polarization radar. Background Technology

[0002] Wildfires, as a highly destructive natural disaster, pose a continuous and increasingly severe threat to global ecosystems, public infrastructure, public health, and power supply networks. Under the intensifying influence of anthropogenic climate change, the frequency and intensity of "fire weather" (i.e., hot, dry, and windy conditions) that lead to wildfires have significantly increased, resulting in larger burned areas and more extreme fire behaviors. Against this backdrop, rapid and accurate detection and location of ignition points and fire lines have become a crucial link in achieving effective situational awareness and risk mitigation within emergency management systems.

[0003] Currently, wildfire monitoring primarily relies on satellite-based observation systems. However, these systems have inherent limitations in their core performance indicators. Specifically, while geostationary imagers (such as Himawari-8 and FY-4A) offer high temporal resolution and can provide frequent observations, their spatial resolution is generally coarse, making it difficult to accurately identify small-scale fires. In contrast, polar-orbiting sensors (such as VIIRS and MODIS) provide even higher spatial resolution, but their long revisit periods cannot meet the needs of continuous monitoring of rapidly changing fire conditions. More critically, both types of satellite sensors are susceptible to cloud cover, smoke interference, and forest canopy obstruction, leading to widespread missed detections of small-scale or low-intensity fires. This technical shortcoming has been repeatedly verified by relevant research and operational management agencies. It is precisely during the most critical period of fire occurrence and development that the lack of such monitoring information directly impacts the timeliness of early warnings and emergency responses.

[0004] Ground-based polarimetric weather radar provides a powerful, all-weather, and highly timely supplement to satellite monitoring, capable of resolving wildfire plumes from the S-band to the Ka-band. In operational observations of the WSR-88D polarimetric radar network deployed in the United States, wildfire plumes often exhibit lower cross-correlation coefficients (ρHV) and slightly greater than zero differential reflectivity (ZDR) compared to typical condensates; their amplitude is controlled by particle shape / orientation and wavelength. Existing results show that ZDR can be significantly positive when flaky ash is horizontally oriented. These "fingerprint" characteristics—along with horizontal reflectivity (Zh), Doppler velocity, and spectral width—have enabled the national WSR-88D network, compact X-band, and mobile Ka-band radars deployed in the United States to track fire boundaries and plume evolution in real time.

[0005] Radar-based wildfire detection is still under development. Existing methods rely on fuzzy logic expert systems with fixed membership functions and variable weights, which limits their flexibility. Currently, there is a lack of dedicated radar datasets for wildfire monitoring, as well as methods that can directly monitor wildfires from weather radar data. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides a wildfire monitoring method based on polarization radar, the method comprising: S10, acquire volume scan data of the monitoring point through polarization radar. The volume scan data includes polarization variables, including horizontal reflectivity, differential reflectivity, cross-correlation coefficient and radial velocity standard deviation. S20, preprocess the polarization variable, the preprocessing including normalization and outlier clipping, to construct a polarization variable with an effective pixel mask; S30, based on the fuzzy logic algorithm, the preprocessed polarization variables are initially classified and confirmed by multi-source cross-validation to generate wildfire segmentation and labeling data for monitoring points; S40, Use the historical annotation dataset of wildfire segmentation to train the model to obtain the wildfire monitoring model. The historical annotation dataset of wildfire segmentation is obtained by performing S10-S30 on multiple volume scan data of historical monitoring. S50, input the wildfire segmentation and labeling data of the monitoring points into the wildfire monitoring model to obtain wildfire monitoring data, which includes wildfire probability map, segmentation mask calculation of the centroid location of the fire zone and the burned area.

[0007] Furthermore, the method includes: S21. Fuzzy logic with customized membership functions and weights is used to perform initial classification of radar echoes, resulting in initial results including wildfire smoke categories. S22, perform spatiotemporal consistency post-processing on the initial results. The post-processing includes block filtering, connected component analysis, and temporal continuity constraints to remove isolated noise points and instantaneous false alarms. S23, the post-processed candidate regions are spatiotemporally matched and verified with satellite fire point products, and supplemented by manual review, to generate the final wildfire segmentation label.

[0008] Furthermore, the method includes: Image-level Z-score normalization is performed on each polarization variable channel, and the values ​​are symmetrically cropped to a preset range. The effective pixel mask is used to shield zero-echo and low signal-to-noise ratio regions during model training.

[0009] Furthermore, the initial classification is based on a hydrographic classification algorithm, which includes a membership function and weights for wildfire plumes, wherein the identification features of wildfire plumes include a cross-correlation coefficient ρHV of less than 0.85.

[0010] Furthermore, the preprocessing also includes: During use, the air conditioning system undergoes continuous post-processing to eliminate residual interference.

[0011] Furthermore, the method for training the wildfire monitoring model includes: The training end-to-end segmentation network adopts an encoder-bottleneck-decoder architecture, wherein the encoder uses fragmented blocks and depthwise separable convolutions, the bottleneck includes a spatial pyramid pooling module, and the decoder integrates an attention gating mechanism at the skip connections.

[0012] Furthermore, the incomplete block is an incomplete block using a Conv–GroupNorm–SiLU combination, and the attention mechanism includes an SE channel.

[0013] Furthermore, the spatiotemporal consistency post-processing specifically includes: Based on block filtering, connected component analysis, and time continuity constraints, isolated noise points and short-term false alarms are systematically eliminated.

[0014] According to another aspect of the present invention, a wildfire monitoring system based on polarization radar is provided, the system comprising: The data acquisition module is used to acquire volume scan data of the monitoring point through polarization radar. The volume scan data includes polarization variables, including horizontal reflectivity, differential reflectivity, cross-correlation coefficient, and radial velocity standard deviation. The preprocessing module is used to preprocess the polarization variables, including normalization and outlier clipping, to construct polarization variables with an effective pixel mask. The annotation generation module is used to perform initial classification of preprocessed polarization variables based on fuzzy logic algorithm and confirm it using multi-source cross-validation to generate wildfire segmentation annotation data for monitoring points. The model training module is used to train a wildfire monitoring model using a historically labeled wildfire segmentation dataset. The historically labeled wildfire segmentation dataset is obtained by processing multiple volume scan data from historical monitoring through a data acquisition module, a preprocessing module, and a generation module. The monitoring and inference module is used to input the wildfire segmentation and labeling data of the monitoring points into the wildfire monitoring model to obtain wildfire monitoring data, which includes wildfire probability maps, segmentation mask calculations of the centroid location of the fire zone, and burned area.

[0015] Compared with existing technologies, the wildfire monitoring method based on polarization radar of this invention has the following advantages: This invention, by employing polarization radar as the core data source, fundamentally overcomes the limitations of existing mainstream satellite monitoring methods, which are constrained by spatiotemporal resolution, revisit cycles, and cloud / smoke obstruction, resulting in a high rate of missed detection for small-scale or low-intensity fires. It is particularly suitable for early detection and real-time tracking of rapidly developing wildfires. By combining multivariate observations and high-quality annotation generation processes from polarization radar with a segmentation model, replacing the traditional fuzzy logic with fixed membership functions, and using a machine learning model to automatically learn feature weights, it adapts to the monitoring needs of different environments. Attached Figure Description

[0016] Figure 1 This is a flowchart of a wildfire monitoring method based on polarization radar provided in an embodiment of the present invention; Figure 2 This is a framework diagram of the wildfire monitoring model provided in an embodiment of the present invention; Figure 3 This is a comparative experimental result diagram provided by another embodiment of the present invention; Figure 4 This is an ablation experiment result diagram provided by another embodiment of the present invention; Figure 5 This is a schematic diagram of a wildfire monitoring system based on polarization radar provided in another embodiment of the present invention. Detailed Implementation

[0017] The following description, in conjunction with the accompanying drawings, illustrates exemplary embodiments of the present invention, including various details to aid understanding. These details should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope of the invention. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0018] like Figure 1-2 As shown, in an optional embodiment of the present invention, the wildfire monitoring method based on polarization radar includes: S10, acquire volume scan data of the monitoring point through polarization radar. The volume scan data includes polarization variables, including horizontal reflectivity, differential reflectivity, cross-correlation coefficient and radial velocity standard deviation. S20, preprocess the polarization variable, the preprocessing including normalization and outlier clipping, to construct a polarization variable with an effective pixel mask; S30, based on the fuzzy logic algorithm, the preprocessed polarization variables are initially classified and confirmed by multi-source cross-validation to generate wildfire segmentation and labeling data for monitoring points; S40, Use the historical annotation dataset of wildfire segmentation to train the model to obtain the wildfire monitoring model. The historical annotation dataset of wildfire segmentation is obtained by performing S10-S30 on multiple volume scan data of historical monitoring. S50, input the wildfire segmentation and labeling data of the monitoring points into the wildfire monitoring model to obtain wildfire monitoring data, which includes wildfire probability map, segmentation mask calculation of the centroid location of the fire zone and the burned area.

[0019] Among them, polarization radar volume scan data: data acquired by three-dimensional scanning of the monitoring area using polarization radar, containing polarization variable information in both spatial (distance, azimuth, elevation) and temporal dimensions, is the fundamental data source for wildfire monitoring. Horizontal reflectivity (Zh): reflects the backscattering intensity of radar waves by plume particles, and is positively correlated with plume concentration and particle size, serving as a basic indicator for identifying fire points. Differential reflectivity (ZDR): characterizes the difference between the reflectivity of horizontal and vertical polarized waves. For flaky ash with horizontal orientation, ZDR is significantly positive (e.g., 0.5-3 dB), unlike spherical condensate (ZDR≈0). Cross-correlation coefficient (ρHV): measures the correlation between horizontal and vertical polarized wave echoes. Wildfire plumes have lower ρHV (e.g., 0.7-0.9) due to irregular particle shapes, while condensate ρHV is close to 1. Radial velocity standard deviation: reflects the intensity of plume turbulence. In fire areas, increased velocity dispersion due to thermal convection and wind disturbances can be used as a dynamic feature to aid in classification.

[0020] To further illustrate this, let's take a specific example: In S10, data acquisition (study area and observation system) is conducted: The study area was selected in Guangdong Province, China (20.13°–25.31°N, 109.39°–117.19°E). Twelve S-band dual-polarization radars (compatible with WSR-88D) were deployed, employing a VCP21 volume scan strategy (9 elevation angles, 6 minutes / scan, radial resolution 250m) to collect four-dimensional polarization variables: horizontal reflectivity Zh, differential reflectivity ZDR, cross-correlation coefficient ρHV, and radial velocity standard deviation σv. The purpose was to construct a high spatiotemporal resolution (6 minutes / scan, 250m radial resolution) wildfire plume observation data source, covering a monitoring area of ​​1.2 million square kilometers, and supporting small-scale fire detection.

[0021] In S20, data preprocessing is performed (through normalization and mask construction). The four-dimensional polarization variable is z-score normalized (intensity clipped to [-6,6]), and an effective pixel mask is generated by thresholding the non-zero region across the channel to exclude ground clutter and low signal-to-noise ratio regions.

[0022] In S30, initial classification and cross-validation are performed (using a three-stage annotation process). Based on the improved hydrophyte classification (HCA) algorithm, a customized membership function was set (e.g., ZDR>1.0 and ρHV<0.8 for wildfire candidates). Combined with threshold rules, the echoes were initially classified into four categories: ground clutter, biological, hydrophytes, and wildfire smoke. Block filtering and temporal continuity checks (e.g., 3×3 neighborhood averaging + 5-frame time series smoothing) were applied to remove residual interference, such as transient noise or isolated pixels. Spatial-temporal matching was performed with VIIRS / MODIS satellite fire point data and ground camera images, retaining only dual-source confirmed samples, supplemented by manual verification and correction.

[0023] In S40, model training is performed (using the RadarFire network architecture). The architecture employs an encoder-bottleneck-decoder structure. The encoder uses Conv-GroupNorm-SiLU incomplete blocks, the bottleneck layer integrates ASPP (dilated convolutional dilation rates of 1 / 6 / 12 / 18) + global average pooling, and the decoder restores spatial resolution through bilinear upsampling + 1×1 convolution. Skip-connected features are weighted and fused using attention gates. It uses 12,000+ labeled samples (240 wildfire events), divided into training / validation / test sets in an 8:1:1 ratio. The loss function is Dice + Focal (supervised loss) + consistency loss (semi-supervised regularization). The optimizer is AdamW (learning rate 1e-3, weight decay 1e-4), trained for 60 epochs (batch size 1, gradient accumulation 8 steps). It supports mixed precision (AMP) and cuDNN acceleration.

[0024] Real-time monitoring is performed on the S50. Real-time volumetric scanning data is input into a pre-trained RadarFire model, which outputs a wildfire probability map. Points with a probability >0.7 are identified as fire points. A binary mask is generated after thresholding, and the centroid of the fire zone (geometric center method) and the burned area (pixel area accumulation) are calculated. The model has an inference latency of <1 minute on an NVIDIA V100 GPU and supports real-time dynamic updates. It outputs a wildfire probability map, segmentation mask, fire zone centroid location, and burned area to support emergency response decision-making.

[0025] This invention, by employing polarization radar as the core data source, fundamentally overcomes the limitations of existing mainstream satellite monitoring methods, which are constrained by spatiotemporal resolution, revisit cycles, and cloud / smoke obstruction, resulting in a high rate of missed detection for small-scale or low-intensity fires. It is particularly suitable for early detection and real-time tracking of rapidly developing wildfires. By combining multivariate observations and high-quality annotation generation processes from polarization radar with a segmentation model, replacing the traditional fuzzy logic with fixed membership functions, and using a machine learning model to automatically learn feature weights, it adapts to the monitoring needs of different environments.

[0026] In an optional embodiment of the present invention, the method includes: S21. Fuzzy logic with customized membership functions and weights is used to perform initial classification of radar echoes, resulting in initial results including wildfire smoke categories. S22, perform spatiotemporal consistency post-processing on the initial results. The post-processing includes block filtering, connected component analysis, and temporal continuity constraints to remove isolated noise points and instantaneous false alarms. S23, the post-processed candidate regions are spatiotemporally matched and verified with satellite fire point products, and supplemented by manual review, to generate the final wildfire segmentation label.

[0027] Specifically, based on the improved Hydrophoretic Classification (HCA) algorithm, the membership function and weighting rules for polarization variables (Zh / ZDR / ρHV / σv) are customized. For example, typical characteristics of wildfire plumes include: ZDR > 1.0 dB (horizontally oriented sheet-like ash), ρHV < 0.8 (irregular particle shape), and σv > 2 m / s (significant turbulent motion). "Candidate wildfire regions" are initially identified through threshold combinations (e.g., ZDR > 1.0 and ρHV < 0.8), simultaneously excluding interference from ground clutter (ρHV < 0.6) and biological echoes (e.g., flocks of birds). This achieves rapid initial screening of fire points / non-fire points, but carries the risk of misclassification of isolated noise points or low-intensity fire plumes. A 3×3 neighborhood average is applied to the initial classification results to suppress isolated pixel noise. Candidate regions are required to exist stably for 5 consecutive frames (approximately 30 minutes) to exclude transient interference such as fireworks and industrial emissions. The spatiotemporal continuity characteristics of real fire points are preserved, providing a high-quality candidate set for subsequent validation. Candidate regions were spatiotemporally matched with VIIRS / MODIS satellite thermal anomaly data (1km spatial resolution), retaining only regions consistent with both sources. For regions failing to match, experts manually interpreted the data using ground-based camera and meteorological station data (such as temperature, humidity, and wind speed) to correct boundary misjudgments or omissions, such as low-intensity fire plumes and forest canopy obstruction. This three-stage process preserved the ability of fuzzy logic to capture wildfire "fingerprint features" such as ZDR / ρHV anomalies, while spatiotemporal post-processing and multi-source verification improved the accuracy and robustness of the labeled data. This provided high-quality training samples for subsequent S40 model training, ultimately supporting high-precision output for S50 real-time monitoring. In an optional embodiment of the present invention, the method includes: Image-level Z-score normalization is performed on each polarization variable channel, and the values ​​are symmetrically cropped to a preset range. The effective pixel mask is used to shield zero-echo and low signal-to-noise ratio regions during model training.

[0028] In an optional embodiment of the present invention, the initial classification is based on a hydrographic classification algorithm, which includes a membership function and weights for wildfire plumes, wherein the identification features of wildfire plumes include a cross-correlation coefficient ρHV of less than 0.85.

[0029] In an optional embodiment of the present invention, the preprocessing further includes: During use, the air conditioning system undergoes continuous post-processing to eliminate residual interference.

[0030] In an optional embodiment of the present invention, the method for training a wildfire monitoring model includes: The training end-to-end segmentation network adopts an encoder-bottleneck-decoder architecture, wherein the encoder uses fragmented blocks and depthwise separable convolutions, the bottleneck includes a spatial pyramid pooling module, and the decoder integrates an attention gating mechanism at the skip connections.

[0031] In an optional embodiment of the invention, the incomplete block is an incomplete block using a Conv–GroupNorm–SiLU combination, and the attention mechanism includes an SE channel.

[0032] In an optional embodiment of the present invention, the spatiotemporal consistency post-processing specifically includes: Based on block filtering, connected component analysis, and time continuity constraints, isolated noise points and short-term false alarms are systematically eliminated.

[0033] like Figure 2-5 As shown, according to another aspect of the present invention, a wildfire monitoring system based on polarization radar is provided, the system comprising: The data acquisition module is used to acquire volume scan data of the monitoring point through polarization radar. The volume scan data includes polarization variables, including horizontal reflectivity, differential reflectivity, cross-correlation coefficient, and radial velocity standard deviation. The preprocessing module is used to preprocess the polarization variables, including normalization and outlier clipping, to construct polarization variables with an effective pixel mask. The annotation generation module is used to perform initial classification of preprocessed polarization variables based on fuzzy logic algorithm and confirm it using multi-source cross-validation to generate wildfire segmentation annotation data for monitoring points. The model training module is used to train a wildfire monitoring model using a historically labeled wildfire segmentation dataset. The historically labeled wildfire segmentation dataset is obtained by processing multiple volume scan data from historical monitoring through a data acquisition module, a preprocessing module, and a generation module. The monitoring and inference module is used to input the wildfire segmentation and labeling data of the monitoring points into the wildfire monitoring model to obtain wildfire monitoring data, which includes wildfire probability maps, segmentation mask calculations of the centroid location of the fire zone, and burned area.

[0034] Specifically, let's illustrate this further with a concrete example: The system consists of a data acquisition module, a preprocessing module, a label generation module, a model training module, and a monitoring and inference module. These modules interact via message queues or shared storage; offline training and online inference can be deployed on the same or different computing nodes.

[0035] Data source: Primarily S-band dual-polarization weather radar volume scan data (Example: VCP21, 9 elevation angles, volume scan period of approximately 6 minutes, radial resolution of approximately 250m). There are a total of 12 S-band dual-polarization radars in Guangdong Province. The input channels include four variables: ZH, ZDR, ρHV, and radial velocity standard deviation (σv). It also supports extension to X / Ka bands and more polarization / spectral width characteristic variables.

[0036] Normalization and cropping: Normalize each channel using the image-level z-score method, and symmetrically crop strong outliers according to the threshold range (example: [-6, 6]); construct an effective pixel mask to shield zero-echo or low signal-to-noise ratio regions and reduce the interference of invalid gradients on training.

[0037] (3) High-quality annotation generation The proposed "three-stage annotation pipeline" has the following specific process: The volume scan echo is initially screened based on an improved fuzzy logic classification system, which includes categories such as ground clutter, biological objects, meteorological echoes, and wildfire smoke. The membership degree and weight can be adaptively adjusted according to the local electromagnetic environment and terrain conditions.

[0038] Spatiotemporal consistency post-processing: Based on block filtering, connected component analysis and temporal continuity constraints, it systematically removes isolated noise points and short-term false alarms.

[0039] Multi-source cross-validation and manual verification: Candidate echoes are spatiotemporally matched with satellite fire points (such as VIIRS / MODIS) to significantly reduce mislabeling; key samples are manually verified and corrected to form a high-confidence annotation set.

[0040] (4) End-to-end segmentation network An improved U-Net encoder-bottleneck-decoder framework is adopted: Encoder: The incomplete block (Conv–GroupNorm–SiLU) is downsampled and expanded in stages, and depthwise separable convolution is combined to reduce computational overhead; Squeeze-and-Excitation channel attention is introduced to achieve adaptive channel recalibration; Bottleneck: Set up Spatial Pyramid Pooling (ASPP) with typical expansion rates of {1, 6, 12, 18} and supplement it with global average pooling to aggregate multi-scale context in a lossless resolution manner; Decoder: Bilinear upsampling + 1×1 convolution is used to avoid checkerboard artifacts from deconvolution; Attention gate is introduced at the jump connection between encoder and decoder to spatially adaptively suppress clutter / biological echoes and highlight wildfire feature boundaries. Output header: Use 1×1 convolution to map to a single-channel probability map and set a threshold to generate binary segmentation.

[0041] Semi-supervised training objectives By combining supervised loss and unsupervised regularization, robustness in sparse annotation scenarios can be improved. Supervision components: Focal Loss (alleviates foreground / background imbalance) + Dice Loss (directly optimizes region overlap and boundaries); Consistency regularization: Ensures consistent output for weak / strong augmented view constraints on the same input, improving stability against disturbances and noise; Entropy minimization: Suppressing high-entropy outputs in unlabeled regions prompts the model to make more confident judgments.

[0042] Overall objective function: Ltotal = λsupLsup + λconsLcons + λentLent.

[0043] (6) Key points for training and deployment (example values, can be adjusted as needed) Framework and hardware: PyTorch 2.x; NVIDIA V100 / A100 GPU; AMP mixed precision and cuDNN acceleration enabled; Hyperparameters: base channel width B=16; AdamW optimizer, learning rate 1e-3, weight decay 1e-4; supervision part uses Focal (α=0.25, γ=2.0) and Dice in combination; consistency and entropy terms have weights of 0.5 and 0.2 respectively; training for 60 epochs; mini-batch combined with gradient accumulation (e.g., bs=1, accumulating 8 steps). Inference: Threshold segmentation + morphological refinement (optional); calculate the centroid and area of ​​the segmented region, and output km² quantization and latitude and longitude positioning; supports online processing time of volume scans at least once every 6 minutes.

[0044] (7) System Implementation Form Method: The processing flow is as described in the steps above; Device / System: An all-in-one hardware and software machine containing CPU / GPU computing power, storage, network and display / alarm interfaces; Computer-readable storage medium: storing program instructions and parameter weights for performing the method.

[0045] Under realistic and complex backgrounds and sparse annotation conditions, the method of this invention significantly outperforms the prediction results of traditional fuzzy logic and the baseline U-Net model in key metrics: Region area estimation accuracy: approximately 0.96, higher than U-Net's approximately 0.76; Centroid deviation: approximately 0.54 km, significantly better than U-Net's approximately 2.74 km; Ablation analysis shows that removing attention gating, ASPP, or residual structures all lead to a decrease in area accuracy and an increase in centroid deviation, indicating that each module contributes to robust recognition and fine boundary recovery. The system has the ability to suppress clutter / organism echoes and has better coherence and detail fidelity for small-scale, irregularly shaped plume boundaries; It has the ability to generalize across radar sites and events, and meets the online processing time limit for volume scan, making it easy to integrate into the business early warning link.

[0046] Combination Figure 2-4 The following comparative experimental results were obtained based on the experimental data: Figure 3 In the table, (a) represents the true value, (b) represents the prediction result of the fuzzy logic algorithm, (c) represents the prediction result of the Unet model, and (d) represents the prediction result of the RadarFire model.

[0047] (1) Qualitative comparison Compared to fuzzy logic algorithms and the classic U-Net, RadarFire's segmentation results are highly consistent with the ground truth, and it excels in boundary alignment, detail preservation, and noise suppression. Fuzzy logic produces numerous artifacts / outliers in scenes with strong weather clutter or low-intensity fire plumes. U-Net under-segments small targets and over-segments complex backgrounds. Figure 3 (a)–(d)).

[0048] (2) Quantitative ablation Using area accuracy (%) and centroid offset (km) as indicators, the full model achieved 0.96 and 0.54 km respectively; attention removal reduced it to 0.93 / 1.23 km; ASPP removal was 0.94 / 0.97 km; residual removal was 0.88 / 1.53 km; and baseline U-Net was 0.76 / 2.74 km. Figure 4 (a) and (b)).

[0049] RadarFire achieves precise segmentation of wildfire plumes while maintaining high timeliness through modular innovations (such as attention gating and ASPP) and optimized training strategies (semi-supervised loss functions). Its advantages are particularly evident in: strong anti-interference capabilities, effectively distinguishing plumes from meteorological clutter and biological echoes; high detail fidelity, providing high-fidelity recovery of small-scale fire points and irregular boundaries; and robust generalization, exhibiting stable performance across radar sites and events, meeting operational deployment requirements. These experimental results validate the groundbreaking value of this invention in wildfire monitoring, providing more reliable technical support for early warning and emergency response.

[0050] It should be understood that the various forms of processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this invention can be achieved, and this is not limited herein.

[0051] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the principles of this invention should be included within the scope of protection of this invention.

Claims

1. A wildfire monitoring method based on polarization radar, characterized in that, The method includes: S10, acquire volume scan data of the monitoring point through polarization radar. The volume scan data includes polarization variables, including horizontal reflectivity, differential reflectivity, cross-correlation coefficient and radial velocity standard deviation. S20, preprocess the polarization variable, the preprocessing includes normalization and outlier clipping, and construct a polarization variable with an effective pixel mask; S30, based on the fuzzy logic algorithm, the preprocessed polarization variables are initially classified and confirmed by multi-source cross-validation to generate wildfire segmentation and labeling data for monitoring points; S40, Use the historical annotation dataset of wildfire segmentation to train the model to obtain the wildfire monitoring model. The historical annotation dataset of wildfire segmentation is obtained by performing S10-S30 on multiple volume scan data of historical monitoring. S50, input the wildfire segmentation and labeling data of the monitoring points into the wildfire monitoring model to obtain wildfire monitoring data, which includes wildfire probability map, segmentation mask calculation of the centroid location of the fire zone and the burned area.

2. The wildfire monitoring method based on polarization radar according to claim 1, characterized in that, The method includes: S21. Fuzzy logic with customized membership functions and weights is used to perform initial classification of radar echoes, resulting in initial results including wildfire smoke categories. S22, perform spatiotemporal consistency post-processing on the initial results. The post-processing includes block filtering, connected component analysis, and temporal continuity constraints to remove isolated noise points and instantaneous false alarms. S23, the post-processed candidate regions are spatiotemporally matched and verified with satellite fire point products, and supplemented by manual review, to generate the final wildfire segmentation label.

3. The wildfire monitoring method based on polarization radar according to claim 1, characterized in that, The method includes: Image-level Z-score normalization is performed on each polarization variable channel, and the values ​​are symmetrically cropped to a preset range. The effective pixel mask is used to shield zero-echo and low signal-to-noise ratio regions during model training.

4. The wildfire monitoring method based on polarization radar according to claim 1, characterized in that, The initial classification is based on a hydrographic classification algorithm, which includes a membership function and weights for wildfire plumes. The identification features of wildfire plumes include a cross-correlation coefficient ρHV of less than 0.

85.

5. The wildfire monitoring method based on polarization radar according to claim 1, characterized in that, The preprocessing also includes: During use, the air conditioning system undergoes continuous post-processing to eliminate residual interference.

6. The wildfire monitoring method based on polarization radar according to claim 1, characterized in that, The methods used in training wildfire monitoring models include: The training end-to-end segmentation network adopts an encoder-bottleneck-decoder architecture, wherein the encoder uses fragmented blocks and depthwise separable convolutions, the bottleneck includes a spatial pyramid pooling module, and the decoder integrates an attention gating mechanism at the skip connections.

7. The wildfire monitoring method based on polarization radar according to claim 6, characterized in that, The incomplete block is an incomplete block using a Conv–GroupNorm–SiLU combination, and the attention mechanism includes an SE channel.

8. The wildfire monitoring method based on polarization radar according to claim 7, characterized in that, The spatiotemporal consistency post-processing specifically includes: Based on block filtering, connected component analysis, and time continuity constraints, isolated noise points and short-term false alarms are systematically eliminated.

9. A wildfire monitoring system based on polarization radar, characterized in that, The system includes: The data acquisition module is used to acquire volume scan data of the monitoring point through polarization radar. The volume scan data includes polarization variables, including horizontal reflectivity, differential reflectivity, cross-correlation coefficient, and radial velocity standard deviation. The preprocessing module is used to preprocess the polarization variables, including normalization and outlier clipping, to construct polarization variables with an effective pixel mask. The annotation generation module is used to perform initial classification of preprocessed polarization variables based on fuzzy logic algorithm and confirm it using multi-source cross-validation to generate wildfire segmentation annotation data for monitoring points. The model training module is used to train a wildfire monitoring model using a historically labeled wildfire segmentation dataset. The historically labeled wildfire segmentation dataset is obtained by processing multiple volume scan data from historical monitoring through a data acquisition module, a preprocessing module, and a generation module. The monitoring and inference module is used to input the wildfire segmentation and labeling data of the monitoring points into the wildfire monitoring model to obtain wildfire monitoring data, which includes wildfire probability maps, segmentation mask calculations of the centroid location of the fire zone, and burned area.

Images