Wind speed measurement method based on monitoring sensor network
By using a monitoring sensor network to measure wind speed, and leveraging video from monitoring cameras and deep learning technology, high spatiotemporal resolution ground wind speed measurement was achieved. This solves the problem of insufficient spatiotemporal resolution in wind speed measurement in urban built environments and supports various application scenarios.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NANJING INST OF TECH
- Filing Date
- 2025-12-25
- Publication Date
- 2026-04-17
AI Technical Summary
The density and spatial heterogeneity of urban built environments make it difficult for existing wind speed measurement systems to characterize near-surface wind fields with high spatiotemporal resolution.
A wind speed measurement method based on a monitoring sensor network is adopted, including monitoring camera registration and video data acquisition, wind speed video and target detection dataset production, two-stage deep learning wind speed estimation model construction, multi-camera collaborative observation and ground wind field construction. The method realizes single-camera wind speed point observation through monitoring camera video and fuses multi-camera observations to construct a ground wind observation network with high spatiotemporal resolution.
It enables low-cost, high spatiotemporal resolution ground wind observation, supporting refined urban weather forecasting, issuance of severe wind warnings, microclimate analysis and modeling, and air pollution control.
Smart Images
Figure CN121878253A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of wind speed measurement and relates to a wind speed measurement method based on a monitoring sensor network. Background Technology
[0002] The density and spatial heterogeneity of urban built environments make it difficult for existing wind speed measurement systems to characterize the dynamic evolution of near-surface wind fields with high spatiotemporal resolution. Specifically:
[0003] Automatic weather stations are equipped with various anemometers (such as cup anemometers and ultrasonic anemometers), which can accurately obtain wind data at specific locations, but due to the low density of station distribution, the spatial resolution is insufficient.
[0004] Ground-based weather radar has a wide observation range, but due to the limitation of the observation elevation angle, the wind field data it acquires deviates significantly from the near-surface wind field.
[0005] Satellite remote sensing measures wind speed by analyzing cloud layers and atmospheric composition. Although it has high spatial resolution, it is limited by the revisit period and has insufficient temporal resolution, making it difficult to capture short-term changes in wind speed near the ground.
[0006] In addition, while numerical model-based wind field forecasts can accurately depict wind field information at multiple scales, they have extremely high requirements for initial field parameters. Furthermore, for urban areas with complex terrain features, the near-surface wind field simulation accuracy is insufficient, and the computational cost is relatively high.
[0007] With the continuous advancement of smart cities, surveillance cameras, as a crucial cornerstone of traffic management and public safety, have been widely deployed in urban streets, transportation hubs, and public places. These cameras continuously capture the movement of common objects in these environments, such as trees and flags, providing vital clues for wind speed measurement. This is thanks to the large number, high density, and high transmission speed of surveillance cameras. Summary of the Invention
[0008] 1. The technical problem to be solved:
[0009] The density and spatial heterogeneity of urban built environments make it difficult for existing wind speed measurement systems to characterize near-surface wind fields with high spatiotemporal resolution.
[0010] 2. Technical Solution:
[0011] To address the above problems, this invention provides a wind speed measurement method based on a monitoring sensor network, comprising the following steps:
[0012] Step 1: Register surveillance cameras and acquire video data;
[0013] Step 2: Creating the wind speed video and target detection dataset;
[0014] Step 3: Construction and implementation of a two-stage deep learning wind speed estimation model for surveillance videos; including a two-stage model.
[0015] Step 3-1: In the first stage, the model detects and identifies wind-induced reference objects through the wind-induced target detection model, and outputs their category and location information as priors;
[0016] Step 3-2: The second stage of model construction is a wind speed estimation model based on surveillance video, consisting of a multi-branch network composed of an image branch, a motion branch, an instance branch, and a frequency domain branch. The image branch extracts fine-grained appearance features from the original image; the motion branch supplements explicit dynamic cues; the instance branch uses the target mask and category prior output by the first stage model to enhance the focus on wind-induced areas; and the frequency domain branch extracts the frequency response of blade / flag vibration to mitigate morphological differences caused by uncertainties in viewing angle and wind direction.
[0017] Step 4: Multi-camera collaborative observation and ground wind field construction, and wind speed measurement.
[0018] 3. Beneficial effects:
[0019] This invention relies on existing monitoring resources, starting with urban surveillance camera videos, and uses computer vision and deep learning methods to achieve point observation of wind speed from a single camera. On this basis, it further integrates multi-camera observations to achieve collaborative perception from "point" to "area", thereby constructing a low-cost ground wind observation network with high spatiotemporal resolution, providing support for applications such as refined urban weather forecasting, issuance of severe wind warnings, microclimate analysis and modeling, and air pollution control. Attached Figure Description
[0020] Figure 1 This is the overall flowchart of the present invention.
[0021] Figure 2 This is the indoor scene verification diagram in Example 2.
[0022] Figure 3 This is a diagram discussing the performance of a real-world scenario test in Example 2.
[0023] Figure 4 This is a local ground wind speed grid map from Example 3. Detailed Implementation
[0024] This invention provides a wind speed measurement method based on a monitoring sensor network, such as... Figure 1 As shown, it includes the following steps:
[0025] Step 1: Register surveillance cameras and acquire video data.
[0026] Step 2: Creating the wind speed video and target detection dataset.
[0027] Step 3: Construction and implementation of a two-stage deep learning wind speed estimation model for surveillance videos; including a two-stage model.
[0028] Step 3-1: In the first stage, the model detects and identifies wind-induced reference objects through the wind-induced target detection model, and outputs their category and location information as priors.
[0029] Step 3-2: The second stage of model construction is a wind speed estimation model based on surveillance video, consisting of a multi-branch network composed of an image branch, a motion branch, an instance branch, and a frequency domain branch. The image branch extracts fine-grained appearance features from the original image; the motion branch supplements explicit dynamic cues; the instance branch uses the target mask and category prior output by the first stage model to enhance attention to wind-induced areas; and the frequency domain branch extracts the frequency response of blade / flag vibration to alleviate morphological differences caused by uncertainties in viewing angle and wind direction.
[0030] Step 4: Multi-camera collaborative observation and ground wind field construction, and wind speed measurement.
[0031] In one embodiment, step 1-1: Select diverse monitoring scenarios to cover different environmental conditions and include multiple tree types;
[0032] Step 1-2: Deploy fixed monitoring cameras in the monitoring scenario described in Step 1-1 to ensure that the monitoring screen simultaneously includes the tree crown and trunk.
[0033] Steps 1-3: Each surveillance camera connects to the unified data acquisition and management system via the local area network to complete registration, record the camera number, model, focal length, and installation coordinates, and synchronize with Beijing time.
[0034] Steps 1-4: Deploy a cup-type anemometer or portable anemometer at a distance of approximately 5-10 meters from the camera, with a height difference of 0-5 meters from the camera lens. Ensure that the anemometer corresponds spatially to the camera's observation area and does not interfere with the monitoring image, and synchronize it to Beijing time.
[0035] Steps 1-5: Synchronously acquire video streams, camera parameters, and anemometer time-series wind speed and direction data through the data acquisition system, store them in the local database, and create an index based on scene, camera number, and timestamp to achieve retrieval and traceable management.
[0036] In one embodiment, in step 1-1, the monitoring scenarios include, but are not limited to, parks, city streets, railway lines, and industrial parks.
[0037] In one embodiment, in steps 1-2, the camera is installed at a height of 8 to 12 meters above the ground, and the lens is kept horizontally forward-looking.
[0038] In one embodiment, step 2 includes the following steps.
[0039] Step 2-1: Based on the instantaneous maximum wind speed and wind direction labels recorded by the anemometer at 1 minute resolution, the monitoring video for the corresponding time period is cropped to align the video clips with the anemometer observations. The cropped video samples are then labeled with synchronized wind speed and wind direction to construct a wind speed video dataset.
[0040] Step 2-2: Select monitoring segments from the dataset mentioned in Step 2-1 that cover different seasons, different wind speed ranges, and different meteorological conditions, and extract video frames containing representative trees and flag targets from them. The meteorological conditions include sunny, cloudy, rainy, and foggy.
[0041] Steps 2-3: Use the LabelImg tool to manually and accurately annotate the video frames described in step 22.
[0042] In one embodiment, in step 2-1, the dataset covers the full wind speed range of 0–16 m / s, including the wind-induced response characteristics of trees and flag references at various wind speed levels.
[0043] In one embodiment, in steps 2-3, the method for precise annotation is as follows: select the wind-induced reference object area such as trees and flags by rectangular box, and label its type, including four categories: evergreen broad-leaved trees, deciduous broad-leaved trees, coniferous trees and flags, to construct a multi-category wind-induced reference object dataset.
[0044] In one embodiment, the method for building the wind-induced target detection model includes the following steps:
[0045] Step 3-1-1: Construct a multi-resolution convolutional encoder to extract hierarchical features ,
[0046]
[0047] in This indicates max pooling. Indicates the first There are 1, and the kernel size is 1. convolutional layers, This represents a single frame of the input video. This represents the height, width, and number of channels of the output feature maps of different levels of convolutional layers.
[0048] Step 3-1-2: Use residual connections and normalization to fuse multi-scale features, preserving low-level details while enhancing high-level semantic perception capabilities.
[0049]
[0050] in This indicates a normalization operation. This indicates an upsampling operation, with the height and width upsampling factors being respectively... .
[0051] Step 3-1-3: Introduce the gradient saliency map as a structural guidance signal, so that the attention of the convolutional features focuses on the contour edges of the dynamic target.
[0052] ;
[0053] in This indicates the calculation of the gradient significance plot. , This represents the Sobel directional gradient operator.
[0054] Step 3-1-4: Concatenate the convolutional features with the gradient saliency map and input the convolutional layer to generate a structural attention weight map.
[0055]
[0056] in This is the Sigmoid function.
[0057] Step 3-1-5: Apply attention weights to adjust the original feature response, forming structurally enhanced features.
[0058] .
[0059] Step 3-1-6: Fuse the multi-scale structural enhancement features to obtain the fused features.
[0060] .
[0061] Step 3-1-7: Apply the decoder output to determine the location of the wind-induced target. and categories ,
[0062]
[0063] in Indicates decoder, This represents the coordinates of the top-left corner of the target bounding box. , They respectively represent based on The width and height offsets, category C includes evergreen broad-leaved trees, deciduous broad-leaved trees, conifers and flags.
[0064] Step 3-1-8: The WDNet model finally generates instance masks, with the target box marked in red and the category marked in white in the upper left corner.
[0065] Step 3-1-9: Train the WDNet model by jointly constraining the bounding box regression loss and the class cross-entropy loss.
[0066]
[0067] in Indicates the total number of frames. Indicates smoothed L1 loss, Indicates the first The predicted coordinates of each bounding box. Indicates the first The actual coordinates of each box Representing 4 categories, Indicates the first One-hot encoding of the true label of each sample, Indicates the first One-hot encoding of the predicted label for each sample.
[0068] In one embodiment, the construction of the wind speed estimation model based on surveillance video includes the following steps:
[0069] Step 3-2-1: Extract 60 frames of images from a single surveillance video sample at equal intervals and input them into the image branch. Through residual blocks based on an attention mechanism, the residual structure allows the details of wind-induced targets to be preserved, and attention enables the model to focus on wind-induced targets.
[0070]
[0071] in This indicates a max pooling operation. This indicates the CBAM attention module. Represents the standard residual block. For a single frame image, by applying 4 layers The output channel dimensions are 64, 128, 256, and 512, respectively, and the final feature is obtained. .
[0072] Step 3-2-2: Use a Gaussian mixture model to separate moving targets in the original surveillance video, obtain a foreground mask sequence, and take 60 equally spaced frames as input to the motion branch. Then, apply 4 convolutional blocks to extract features. , As an effective supplement to motion information and to suppress background changes under complex lighting conditions, a single convolutional block is defined as follows:
[0073]
[0074] in This indicates the dropout operation, with the number of channels in the four convolutional blocks being 64, 128, 256, and 512, respectively.
[0075] Step 3-2-3: Input the surveillance video into WDNet to obtain the instance mask frame sequence, and take 60 frames at equal intervals to input into the instance branch, and apply the four convolutional blocks to extract features. ; It can help the model quickly locate targets and provide category priors, solving the problem of consistency in tree color, height, canopy width, and flag color and shape under different seasons and scenarios.
[0076] Step 3-2-4: Extract the amplitude spectrum of each frame of the surveillance video using Fast Fourier Transform, and take 60 equally spaced frames as input to the frequency domain branch. Apply the four convolutional blocks to extract features. ; It can effectively overcome the differences in the morphological response of reference objects caused by different camera orientations and wind speeds.
[0077] Step 3-2-5: Fuse the features extracted in steps 3-2-1 to 3-2-3 and apply global average pooling to reduce spatial domain redundancy.
[0078]
[0079] GAP stands for Global Average Pooling.
[0080] Step 3-2-6: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full context.] and The features are then input into two GRU layers for time series modeling to obtain the output features. and Set time step The input is The previous hidden state was The calculation process of GRU is as follows:
[0081]
[0082] in, It is the Sigmoid activation function. It is the hyperbolic tangent function. This represents element-wise multiplication. For learnable parameters, Indicates the candidate hidden state. Indicates the final hidden state. Indicates an update to the door. This indicates a reset gate. The hidden layer dimensions of the GRU are 1024 and 512. The above temporal modeling steps can be described as follows:
[0083] .
[0084] Step 3-2-7: Fuse the features obtained in Step 3-2-6 and input them into the fully connected layer. Apply the Kan layer to output the wind speed. The Kan layer is a kernel function-based method that improves the nonlinear fitting ability and enhances the sensitivity of the wind speed output to changes.
[0085] .
[0086] Step 3-2-8: Apply mean squared error loss constraint model training.
[0087] In one embodiment, in step 4, a local ground wind speed grid is constructed at 100 m intervals. The specific method includes the following steps:
[0088] Step 4-1: Obtain data through the data acquisition system A camera in The set of observations at any given time
[0089]
[0090] in For the first Camera coordinates Indicates the first The predicted wind speed from the camera.
[0091] Step 4-2: Calculate the median wind speed in the neighborhood. and median absolute deviation Filter outliers,
[0092] For the first Using a camera, select the neighborhood set within a spatial radius of 100 m:
[0093]
[0094] in This indicates that the median value is taken. If the wind speed value output by a single camera satisfies the following formula, it is determined to be an outlier and is removed.
[0095] ;
[0096] Step 4-3: Apply inverse distance weighting to calculate the estimated wind speed at any grid point.
[0097]
[0098] in express The camera and the first Taiwanese camera distance, This represents a very small positive number (1e-6 in this article).
[0099] Step 4-4: Generate a wind speed intensity distribution map using color mapping or contour lines, and draw a wind vector map based on the wind speed gradient direction of adjacent grid points to form a visualized regional wind field, ultimately generating a city ground wind speed raster map.
[0100] Example 1: Indoor scene verification.
[0101] This invention patent was first verified through experiments in a controlled indoor environment. In the experiment, a flag, a monitoring camera, and a portable anemometer were fixed on a stable support, with their relative positions remaining constant. By adjusting the fan's output speed and direction, wind-induced response scenarios were constructed under different combinations of wind speed and oncoming wind direction. Continuous video data of the flag's movement under various wind conditions was collected. Figure 2 As shown.
[0102] On the aforementioned video samples, the proposed two-stage deep learning model for wind speed measurement was used for wind speed estimation, and the results were compared and evaluated with those from synchronous observations using a portable anemometer. Indoor experimental results show that the proposed method achieves a mean absolute error (MAE) of 0.11 m / s, a mean absolute percentage error (MAPE) of 0.52%, and a root mean square error (RMSE) of 0.36 m / s in this scenario, verifying the high accuracy and stability of the proposed model in an ideal controlled environment.
[0103] 2. Real-world scenario verification.
[0104] Building upon indoor verification, this invention was further tested in a real-world observation environment. A total of 21 fixed monitoring cameras were deployed at the Nanjing National Reference Climate Station and the Ningde National Reference Climate Station. The camera installation height and other on-site configurations are detailed below. Figure 3 The parameters such as focal length and frame rate are shown in Table 1.
[0105] Table 1. Surveillance Camera Configuration
[0106]
[0107] At the Nanjing National Baseline Climate Station, monitoring video and synchronous wind measurement data from 10:00–15:00 on July 8, 2025, were selected as test data. The experimental results show that during this period, the MAE between the camera-estimated wind speed and the anemometer observation was 0.38 m / s, the MAPE was 16.4%, and the RMSE was 0.49 m / s.
[0108] At the Ningde National Baseline Climate Station, monitoring video and wind measurement data from 8:00–13:00 on February 11, 2025, were used for the same tests and evaluations. Under this scenario, the method of this invention achieved a MAE of 0.09 m / s, a MAPE of 0.8%, and an RMSE of 0.11 m / s. The comparison results at the two stations demonstrate that the proposed dual-stage wind speed measurement deep learning model maintains high measurement accuracy and good generalization ability under different station environments and time periods, effectively meeting the requirements for wind speed measurement accuracy and stability in real-world observations and engineering applications.
[0109] Example 3. Construction of Local Surface Wind Speed Grid Map
[0110] This invention patent constructs an urban surveillance sensor network covering approximately 100 square kilometers (longitude 118.86-118.94, latitude 31.90-31.98), comprising 114 fixed surveillance cameras, including 83 4mm and 31 6mm focal length cameras, with an average camera density of approximately 1.1 cameras per square kilometer. The cameras are mainly distributed in urban roads, parks and green spaces, residential areas, and railway lines, with priority given to locations within the field of view that simultaneously contain wind-induced reference objects such as trees or flags. Using the method proposed in this patent, a city ground wind speed grid for 2 PM on July 9, 2025, is finally generated, as shown below. Figure 4 As shown.
Claims
1. A method for measuring wind speed based on a monitoring sensor network, characterized in that: Includes the following steps: Step 1: Register surveillance cameras and acquire video data; Step 2: Creating the wind speed video and target detection dataset; Step 3: Construction and implementation of a two-stage deep learning wind speed estimation model for surveillance videos; Includes a two-stage model, Step 3-1: In the first stage, the model detects and identifies wind-induced reference objects through the wind-induced target detection model, and outputs their category and location information as priors; Step 3-2: The second-stage model construction consists of a multi-branch network composed of an image branch, a motion branch, an instance branch, and a frequency domain branch, forming a wind speed estimation model based on surveillance video. The image branch extracts fine-grained appearance features from the original image; the motion branch supplements explicit dynamic cues; the instance branch uses the target mask and category prior output by the first-stage model to enhance the focus on wind-induced areas; and the frequency domain branch extracts the frequency response of blade / flag vibration to mitigate morphological differences caused by uncertainties in viewing angle and wind direction. Step 4: Multi-camera collaborative observation and ground wind field construction, and wind speed measurement.
2. The monitoring sensor network based wind speed measurement method of claim 1, wherein: Step 1 includes the following steps: Step 1-1: Select diverse monitoring scenarios to cover different environmental conditions and include various tree types; Step 1-2: Deploy fixed monitoring cameras in the monitoring scenario described in Step 1-1 to ensure that the monitoring image simultaneously includes both the tree crown and the tree trunk; Steps 1-3: Each surveillance camera connects to the unified data acquisition and management system via the local area network to complete registration, record the camera number, model, focal length and installation coordinates, and synchronize with Beijing time; Steps 1-4: Deploy a cup-type anemometer or portable anemometer at a distance of about 5-10 meters from the camera, with a height difference of 0-5 meters from the camera lens. Ensure that the anemometer corresponds to the spatial area observed by the camera and does not interfere with the monitoring screen, and synchronize it to Beijing time. Steps 1-5: Synchronously acquire video streams, camera parameters, and anemometer time-series wind speed and direction data through the data acquisition system, store them in the local database, and create an index based on scene, camera number, and timestamp to achieve retrieval and traceable management.
3. The wind speed measurement method based on a monitoring sensor network as described in claim 2, characterized in that: In step 1-1, the monitoring scenarios include, but are not limited to, parks, city streets, railway lines, and industrial parks.
4. The wind speed measurement method based on a monitoring sensor network as described in claim 2, characterized in that: In steps 1-2, the camera is installed at a height of 8 to 12 meters above the ground, with the lens kept horizontal and looking forward.
5. The wind speed measurement method based on a monitoring sensor network as described in any one of claims 2-4, characterized in that: Step 2 includes the following steps: Step 2-1: Based on the instantaneous maximum wind speed and wind direction labels recorded by the anemometer at 1 minute resolution, the monitoring video of the corresponding time period is cropped to align the video clips with the anemometer observations. The cropped video samples are then attached with synchronized wind speed and wind direction labels to construct a wind speed video dataset. Step 2-2: Select monitoring segments from the dataset mentioned in Step 2-1 that cover different seasons, different wind speed ranges, and different meteorological conditions, and extract video frames containing representative trees and flag targets from them. The meteorological conditions include sunny, cloudy, rainy, and foggy. Step 2-3: Use the LabelImg tool to manually and accurately annotate the video frames described in Step 2-2.
6. The wind speed measurement method based on a monitoring sensor network as described in claim 5, characterized in that: In step 2-1, the dataset covers the full wind speed range of 0–16 m / s, including the wind-induced response characteristics of trees and flag references at various wind speed levels.
7. The wind speed measurement method based on a monitoring sensor network as described in claim 5, characterized in that: In steps 2-3, the precise annotation method is as follows: select the wind-induced reference object area such as trees and flags by rectangular box, and label its type, including four categories: evergreen broad-leaved trees, deciduous broad-leaved trees, coniferous trees and flags, to construct a multi-category wind-induced reference object dataset.
8. The wind speed measurement method based on a monitoring sensor network as described in claim 1, characterized in that: The method for building the wind-induced target detection model includes the following steps: Step 3-1-1: Construct a multi-resolution convolutional encoder to extract hierarchical features , ， in This indicates max pooling. Indicates the first There are 1, and the kernel size is 1. convolutional layers, This represents a single frame of the input video. This represents the height, width, and number of channels of the output feature maps of different levels of convolutional layers; Step 3-1-2: Use residual connections and normalization to fuse multi-scale features, preserving low-level details while enhancing high-level semantic perception capabilities. ， in This indicates a normalization operation. This indicates an upsampling operation, with the height and width upsampling factors being respectively... ; Step 3-1-3: Introduce the gradient saliency map as a structural guidance signal, so that the attention of the convolutional features focuses on the contour edges of the dynamic target. ； in This indicates the calculation of the gradient significance plot. , This represents the Sobel directional gradient operator; Step 3-1-4: Concatenate the convolutional features with the gradient saliency map and input the convolutional layer to generate a structural attention weight map. ， in For the Sigmoid function; Step 3-1-5: Apply attention weights to adjust the original feature response, forming structurally enhanced features. ； Step 3-1-6: Fuse the multi-scale structural enhancement features to obtain the fused features. ； Step 3-1-7: Apply the decoder output to determine the location of the wind-induced target. and categories , ， in Indicates decoder, This represents the coordinates of the top-left corner of the target bounding box. , They respectively represent based on Width and height offsets, Category C includes evergreen broad-leaved trees, deciduous broad-leaved trees, conifers and flags; Step 3-1-8: The WDNet model finally generates instance masks, with the target box marked in red and the category marked in white in the upper left corner; Step 3-1-9: Train the WDNet model by jointly constraining the bounding box regression loss and the class cross-entropy loss. ， in Indicates the total number of frames. Indicates smoothed L1 loss, Indicates the first The predicted coordinates of each bounding box. Indicates the first The actual coordinates of each box Representing 4 categories, Indicates the first One-hot encoding of the true label of each sample, Indicates the first One-hot encoding of the predicted label for each sample.
9. The wind speed measurement method based on a monitoring sensor network as described in claim 8, characterized in that: The construction of the wind speed estimation model based on surveillance video includes the following steps: Step 3-2-1: Extract 60 frames of images from a single surveillance video sample at equal intervals and input them into the image branch. Through residual blocks based on an attention mechanism, the residual structure allows the details of wind-induced targets to be preserved, and attention enables the model to focus on wind-induced targets. ， in This indicates a max pooling operation. This indicates the CBAM attention module. Represents the standard residual block. For a single frame image, by applying 4 layers The output channel dimensions are 64, 128, 256, and 512, respectively, and the final feature is obtained. ; Step 3-2-2: Use a Gaussian mixture model to separate moving targets in the original surveillance video, obtain a foreground mask sequence, and take 60 equally spaced frames as input to the motion branch. Then, apply 4 convolutional blocks to extract features. , As an effective supplement to motion information and to suppress background changes under complex lighting conditions, a single convolutional block is defined as follows: ， in This indicates the dropout operation, with the number of channels in the four convolutional blocks being 64, 128, 256, and 512, respectively. Step 3-2-3: Input the surveillance video into WDNet to obtain the instance mask frame sequence, and take 60 frames at equal intervals to input into the instance branch, and apply the four convolutional blocks to extract features. ; Step 3-2-4: Extract the amplitude spectrum of each frame of the surveillance video using Fast Fourier Transform, and take 60 equally spaced frames as input to the frequency domain branch. Apply the four convolutional blocks to extract features. ; Step 3-2-5: Fuse the features extracted in steps 3-2-1 to 3-2-3 and apply global average pooling to reduce spatial domain redundancy. ， Where GAP represents global average pooling; Step 3-2-6: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full context.] and The features are then input into two GRU layers for time series modeling to obtain the output features. and Set time step The input is The previous hidden state was The calculation process of GRU is as follows: ， in, It is the Sigmoid activation function. It is the hyperbolic tangent function. This represents element-wise multiplication. For learnable parameters, Indicates the candidate hidden state. Indicates the final hidden state. Indicates an update to the door. This indicates a reset gate. The hidden layer dimensions of the GRU are 1024 and 512. The above temporal modeling steps can be described as follows: ； Step 3-2-7: Fuse the features obtained in Step 3-2-6 and input them into the fully connected layer. Apply the Kan layer to output the wind speed. The Kan layer is a kernel function-based method that improves the nonlinear fitting ability and enhances the sensitivity of the wind speed output to changes. ； Step 3-2-8: Apply mean squared error loss constraint model training.
10. The wind speed measurement method based on a monitoring sensor network as described in claim 9, characterized in that: In step 4, a local ground wind speed grid is constructed at 100m intervals. The specific method includes the following steps: Step 4-1: Obtain data through the data acquisition system A camera in The set of observations at any given time ， in For the first Camera coordinates Indicates the first Predicted wind speed using a camera; Step 4-2: Calculate the median wind speed in the neighborhood. and median absolute deviation Filter outliers, For the first Using a camera, select the neighborhood set within a spatial radius of 100 m: ， in This indicates that the median value is taken. If the wind speed value output by a single camera satisfies the following formula, it is considered an outlier and is removed: ； Step 4-3: Apply inverse distance weighting to calculate the estimated wind speed at any grid point. ，， in express The camera and the first Taiwanese camera distance, This represents a very small positive number (1e-6 in this article). Step 4-4: Generate a wind speed intensity distribution map using color mapping or contour lines, and draw a wind vector map based on the wind speed gradient direction of adjacent grid points to form a visualized regional wind field, ultimately generating a city ground wind speed raster map.