A method for vegetation identification and interference elimination based on slope radar

Abstract

This invention discloses a vegetation identification and interference elimination method based on slope radar, belonging to the field of slope monitoring technology. It aims to solve the problems of signal misinterpretation, high false alarm rate, poor recognition generalization, and high hardware cost in existing slope radar monitoring systems, which are prone to vegetation disturbance. This invention acquires radar echo data and environmental factor data of the target slope, constructs the vegetation as a forced damping harmonic oscillator system to quantify the radar phase interference caused by vegetation disturbance, and achieves pixel-level vegetation area identification through a cross-modal identification model extracted by intensity and phase feature demultiplexing. The deformation warning threshold is dynamically adjusted and warnings are executed based on the identification results and phase interference. This invention significantly improves vegetation identification accuracy, greatly reduces false alarms caused by vegetation, while also considering the sensitivity of real slope deformation monitoring. It requires no additional high-cost hardware, has strong adaptability, and high engineering practicality.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, specifically relating to a method for vegetation identification and interference elimination based on slope radar. Background Technology

[0002] As a non-contact remote sensing monitoring device, slope radar has the advantages of millimeter-level deformation monitoring accuracy, all-weather continuous operation capability, and no need for on-site manual deployment of measuring points, and has become the mainstream core equipment for high-risk slope deformation monitoring.

[0003] In practical engineering applications, most slope monitoring areas have varying degrees of vegetation cover. Under the influence of natural environmental factors such as wind and rainfall, vegetation undergoes continuous swaying and displacement, which directly causes disturbances in the intensity and phase signals of radar echoes. Because the signal fluctuation characteristics caused by vegetation disturbance are highly similar to those generated by actual deformation of the slope's soil and rock, existing monitoring systems cannot effectively distinguish between the two. This easily leads to misidentification of vegetation disturbance as slope deformation displacement, triggering numerous false alarms. Frequent false alarms severely reduce the reliability of monitoring results, decrease the sensitivity of maintenance personnel to warning signals, and even result in missed detection of actual landslide risks, posing a significant threat to slope safety management.

[0004] To address the issue of vegetation interference in slope radar, the mainstream solution in the industry is currently the coherence screening method. This method is based on the characteristic of low radar echo coherence in vegetation-disturbed areas. By manually setting a coherence threshold, areas below the threshold are identified as vegetation interference areas and eliminated. However, the setting of the coherence threshold in this solution is highly dependent on the experience of engineers, resulting in strong subjectivity and uncertainty. It has poor adaptability to complex terrain and variable weather conditions, and cannot meet the identification needs of different vegetation types and different growth stages. Not only does it lack generalization ability, but it is also prone to problems such as over-elimination leading to the loss of true deformation signals, or under-elimination leading to residual interference. Therefore, its effect on reducing the false alarm rate is very limited.

[0005] To improve vegetation identification accuracy, some existing solutions employ multi-polarization slope radar, or a monitoring scheme that integrates slope radar with high-resolution optical cameras and drone aerial photography equipment, identifying vegetation coverage areas through multi-dimensional data cross-verification. However, such solutions require the additional purchase of high-cost multi-polarization radar and high-precision optical monitoring equipment, significantly increasing hardware investment and maintenance costs, making it difficult to widely promote in small- to medium-scale slope monitoring projects. Furthermore, optical monitoring equipment is susceptible to environmental conditions such as rain, fog, and darkness at night, failing to match the all-weather monitoring characteristics of slope radar. It also suffers from insufficient spatial registration accuracy of multi-source data and data temporal synchronization issues, making it difficult to guarantee the effectiveness of vegetation interference elimination.

[0006] Other existing solutions employ data processing methods such as fixed filtering and time smoothing to eliminate vegetation interference by filtering out high-frequency short-term fluctuations in radar signals. While these solutions have simple processing logic, they cannot handle long-term vegetation disturbances under conditions of strong winds and continuous rainfall. Furthermore, they are prone to filtering out minute deformation signals at the initial stage of slope landslides, leading to missed detections of real risks. Satellite-based vegetation index-assisted identification schemes suffer from low spatial resolution and long revisit periods, failing to meet the application requirements of millimeter-level, real-time continuous monitoring of slopes using radar, and thus unable to achieve accurate quantification and elimination of dynamic vegetation disturbances. In summary, existing technologies cannot simultaneously balance vegetation identification accuracy, interference elimination effectiveness, early warning reliability, and engineering implementation costs, exhibiting significant technical limitations. Summary of the Invention

[0007] To address the technical problems in existing slope radar deformation monitoring technologies, such as signal misjudgment easily caused by vegetation cover areas, high false alarm rate of early warning, poor generalization of vegetation identification, and high hardware cost of interference cancellation schemes, this invention provides a vegetation identification and interference cancellation method based on slope radar.

[0008] The technical solution adopted in this invention is as follows:

[0009] This invention provides a method for vegetation identification and interference elimination based on slope radar, applicable to slope deformation monitoring scenarios, including the following steps:

[0010] S1. Obtain slope radar echo data of the target slope monitoring area, as well as environmental factor data corresponding to the monitoring area;

[0011] S2. Construct the vegetation in the monitoring area into a forced damped harmonic oscillator system, and quantify the radar phase interference caused by vegetation disturbance in combination with the environmental factor data.

[0012] S3. Based on the slope radar echo data and the radar phase interference, identify the vegetation interference area within the monitoring area;

[0013] S4. Based on the identification results of the vegetation interference area and the corresponding radar phase interference, dynamically adjust the early warning threshold for slope deformation monitoring, and execute slope deformation early warning based on the adjusted early warning threshold.

[0014] Preferably, the slope radar echo data includes radar intensity data and phase time series data; the environmental factor data includes meteorological monitoring data and topographic data, wherein the meteorological monitoring data includes wind speed and rainfall, and the topographic data includes the elevation, slope, and curvature of the monitoring area.

[0015] Preferably, in step S2, when constructing the forced damped harmonic oscillator system, a vegetation dynamics differential equation is constructed by combining vegetation type, vegetation growth stage, wind speed, rainfall, and slope data, and the vegetation displacement is obtained by solving the equation. Based on the vegetation displacement, the corresponding radar phase interference is mapped to be obtained. The vegetation dynamics differential equation is:

[0016]

[0017] Where x represents the instantaneous vibration displacement of the vegetation canopy. The instantaneous vibration velocity of the vegetation canopy. Let be the instantaneous vibration acceleration of the vegetation canopy, m be the vegetation mass, c be the damping coefficient, and k be the vegetation stiffness. δ is the wind load, g is the slope correction factor, θ is the gravitational acceleration, and θ is the slope angle of the monitoring area.

[0018] Preferably, the vegetation mass m is determined by the formula:

[0019]

[0020] The calculation yielded, where Where α is the dry mass of vegetation, R is the rainfall interception coefficient, R is the rainfall amount, and A is the vegetation projected area; the vegetation stiffness k is calculated by the formula k=k0βγ, where k0 is the reference stiffness, β is the vegetation type coefficient, and γ is the vegetation growth stage factor.

[0021] Preferably, in step S3, spatial texture features are extracted from the radar intensity data to obtain intensity features, dynamic disturbance features are extracted from the phase time series data combined with environmental factor data to obtain phase features, cross-modal fusion is performed on the intensity features and phase features, and pixel-level vegetation category probability distribution of the monitoring area is generated by combining the radar phase interference amount, and vegetation interference area is identified based on the vegetation category probability distribution.

[0022] Preferably, in step S3, the prediction result of the vegetation category probability distribution is optimized through an environmental adaptive physical constraint joint loss function. The joint loss function includes a vegetation dynamic physical consistency loss term, which is calculated using the following formula:

[0023]

[0024] in, The term represents the loss of vegetation dynamic physical consistency, where W and H are the width and height of the radar image of the monitoring area. Let (i,j) be the predicted probability of the vegetation category k corresponding to pixel (i,j). The standard deviation of phase perturbation for vegetation of category k, calculated using a forced damped harmonic oscillator system. is the reference perturbation standard deviation for vegetation category k.

[0025] Preferably, in step S4, based on the radar phase interference, vegetation category probability, and environmental factor data corresponding to the vegetation interference area, a threshold adjustment coefficient for each pixel location within the monitoring area is calculated using the following formula. The preset fixed warning threshold is dynamically adjusted pixel by pixel, whereby... The adjusted warning threshold for pixel (i,j) The preset fixed warning threshold, This is the threshold adjustment coefficient corresponding to pixel (i,j).

[0026] Preferably, after acquiring the data in step S1, preprocessing operations are performed on the radar echo data and environmental factor data. The preprocessing operations include: performing environmental-guided non-uniform spatial filtering on the radar intensity data in combination with terrain data; performing meteorological-driven outlier removal and periodic disturbance extraction on the phase time series data in combination with meteorological monitoring data; and performing spatial interpolation and timestamp alignment on the environmental factor data to match the spatial grid and time scale corresponding to the radar data.

[0027] Preferably, when performing cross-modal fusion of intensity features and phase features, vegetation type and growth stage are encoded as category embedding vectors, environmental factor data are encoded as environmental feature vectors, class condition query vectors are generated based on the category embedding vectors and environmental feature vectors, and the fusion weights of intensity features and phase features are dynamically adjusted through the class condition query vectors.

[0028] The beneficial effects of this invention are as follows:

[0029] This invention constructs the vegetation in the monitoring area as a forced damping harmonic oscillator system, and for the first time incorporates slope topography parameters, meteorological environment parameters, vegetation type and growth stage parameters into the mechanical quantification system of vegetation disturbance. It can accurately solve the displacement caused by vegetation swaying under different scenarios and map the corresponding radar phase interference, providing clear physical mechanism support for the identification and elimination of vegetation interference, and solving the core defects of existing technologies that rely on empirical statistics and lack physical interpretability.

[0030] This invention employs a radar intensity and phase feature separation extraction mechanism. It extracts the texture and reflection features of radar intensity data through spatial convolutional coding and extracts the dynamic disturbance features of phase time series data through spatiotemporal coding. Combined with vegetation category coding and dynamic modulation of environmental factors, it achieves pixel-level accurate identification of vegetation type and growth stage in the monitoring area. It can effectively distinguish between vegetation interference areas and bare rock and soil areas on slopes, and significantly improves the accuracy and generalization adaptability of vegetation identification in complex terrain and variable weather conditions.

[0031] Based on the quantification results of vegetation phase interference and the results of vegetation identification, this invention realizes the pixel-by-pixel dynamic adjustment of the warning threshold for slope deformation monitoring. It can reasonably optimize the warning threshold in high vegetation interference scenarios to reduce the false alarm rate, and maintain the original monitoring sensitivity in the main area of ​​the slope with low interference to avoid missing the true deformation. Without weakening the monitoring capability of the true deformation of the slope, it significantly reduces the number of false alarms in vegetation-covered areas.

[0032] This invention constructs a technical framework that integrates physical mechanisms and data-driven approaches. By constructing a physical consistency loss term through a vegetation dynamics physical model, it constrains the training and prediction process of the deep learning model, solving the problems of black box nature and insufficient generalization in traditional AI recognition schemes. At the same time, it can optimize the input parameters of the physical model through data recognition results, forming a closed-loop optimization system, which further improves the stability and reliability of the method.

[0033] This invention does not require replacement or modification of existing slope radar main equipment. It only needs to be combined with conventional meteorological sensor data such as wind speed and rainfall and slope topography data to achieve vegetation identification and interference elimination throughout the entire process. The solution has low deployment cost, strong adaptability, and can be directly embedded into various existing slope radar monitoring systems, and has extremely high engineering practicality and promotion value. Attached Figure Description

[0034] Figure 1 This is a schematic diagram of vegetation being covered by a film when treating a certain original slope in an embodiment of the present invention;

[0035] Figure 2 This is a diagram showing the slope deformation before vegetation suppression when treating a certain original slope in an embodiment of the present invention.

[0036] Figure 3 This is a diagram showing the slope deformation after vegetation suppression when treating a certain original slope in an embodiment of the present invention;

[0037] Figure 4 This is a flowchart of the method in an embodiment of the present invention. Detailed Implementation

[0038] The present invention will be further explained below with reference to the accompanying drawings and specific embodiments.

[0039] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0040] Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0041] Example 1:

[0042] This exemplary embodiment provides a method for vegetation identification and interference cancellation based on slope radar, referring to... Figures 3-4 This method is applied to slope deformation monitoring scenarios such as open-pit mine slopes, highway subgrade slopes, and water conservancy hub slopes. The execution entity in this embodiment is the data processing unit of the slope radar deformation monitoring system. The method includes the following core steps:

[0043] S1. Obtain slope radar echo data of the target slope monitoring area, as well as environmental factor data corresponding to the monitoring area;

[0044] S2. Construct the vegetation in the monitoring area into a forced damped harmonic oscillator system, and quantify the radar phase interference caused by vegetation disturbance in combination with the environmental factor data.

[0045] S3. Based on the slope radar echo data and the radar phase interference, identify the vegetation interference area within the monitoring area;

[0046] S4. Based on the identification results of the vegetation interference area and the corresponding radar phase interference, dynamically adjust the early warning threshold for slope deformation monitoring, and execute slope deformation early warning based on the adjusted early warning threshold.

[0047] The above basic embodiments correspond to the superior technical solutions of the independent claims of this invention and are the core implementation logic of the method. Based on this core logic, the specific implementation methods of each step are described in detail below through hierarchical expansion.

[0048] Regarding step S1, in one possible implementation, the slope radar echo data includes radar intensity data S, phase time series data P, intensity texture grayscale co-occurrence matrix (GLCM), and phase difference ΔP; the environmental factor data includes meteorological monitoring data and topographic data, wherein the meteorological monitoring data includes rainfall R and wind speed V during the monitoring period, and the topographic data is a digital terrain model (DTM) of the monitoring area, including the elevation, slope, and curvature of the corresponding radar pixel location.

[0049] For example, the slope radar is a ground-based synthetic aperture radar (GB-SAR) with a Ku-band operating frequency, a wavelength λ of 2 cm, a range resolution of 0.5 m, an azimuth resolution of 0.3 m, and a data sampling interval of 5 minutes, enabling continuous monitoring of millimeter-level deformation in the monitored area. The meteorological monitoring data is collected by anemometers and rain gauges deployed at the radar site, with the sampling frequency aligned with the radar data sampling frequency. The terrain data is generated through UAV aerial surveying, with spatial resolution matching the pixel resolution of the radar monitoring image.

[0050] Optionally, in some optional embodiments provided in this disclosure, after acquiring the data in step S1, preprocessing operations are performed on the radar echo data and environmental factor data to improve the data signal-to-noise ratio and feature consistency.

[0051] In one possible implementation, the preprocessing operation includes:

[0052] Terrain correction and intensity feature enhancement: Combine terrain data to perform environment-guided non-uniform spatial filtering on radar intensity data. Based on the slope, curvature and vegetation distribution priors of the pixel location, the size of the spatial filtering window is dynamically adjusted. For example, the filtering window for steep slope areas is 3×3 pixels, and the filtering window for gentle slopes with dense vegetation is 5×5 pixels, so as to realize terrain correction and texture enhancement of radar intensity data, highlighting the reflection differences between vegetation and bare land and rock.

[0053] Phase time series data preprocessing: Combine meteorological monitoring data to perform meteorological-driven outlier removal and periodic disturbance extraction on the phase time series data. The time smoothing window parameters are dynamically adjusted according to real-time wind speed and rainfall. For example, when the wind speed is ≥5m / s or the daily rainfall is ≥10mm, the time smoothing window is adjusted to 3 sampling periods. Under windless weather, the smoothing window is 10 sampling periods. While suppressing random noise, the short-term phase fluctuation characteristics of vegetation under windy and rainy conditions are preserved.

[0054] Multi-source data spatiotemporal registration: Point-collected environmental factor data such as wind speed and rainfall are mapped to the same spatial raster as radar data through inverse distance weighted spatial interpolation. At the same time, timestamp alignment is used to match all data to a unified time scale, ensuring that subsequent analysis can accurately correlate the impact of environmental factors on vegetation disturbance.

[0055] In one possible implementation, in step S2, when constructing the forced damped harmonic oscillator system, a vegetation dynamics differential equation is constructed by combining vegetation type, vegetation growth stage, wind speed, rainfall, and slope data. The vegetation displacement x(t) is then solved, and the corresponding radar phase interference ΔP is mapped based on this vegetation displacement. In one possible implementation, the vegetation dynamics differential equation is:

[0056]

[0057] Where m is the vegetation mass, c is the damping coefficient, and k is the vegetation stiffness. Let be the wind load, δ be the slope correction factor, and g be the gravitational acceleration. The slope angle of the corresponding pixel position in the monitoring area.

[0058] In one possible implementation, the vegetation mass m is expressed by the formula... The calculation yields α, where α is the rainfall interception coefficient, R is the rainfall amount, and A is the projected area of ​​the vegetation canopy; the vegetation stiffness k is obtained through the formula... The calculations show that k0 is the baseline stiffness, β is the vegetation type coefficient, and γ is the vegetation growth stage factor.

[0059] For example, for common vegetation types on slopes, the vegetation type coefficient β is set as follows: β=1.5 for trees, β=1.0 for shrubs, and β=0.5 for herbs; for different growth stages of vegetation, the growth stage factor γ is set as follows: γ=0.8 for budding stage, γ=1.2 for mature stage, and γ=0.9 for declining stage; the rainfall interception coefficient α ranges from 0.2 to 0.8, for example, α=0.8 for trees, α=0.5 for shrubs, and α=0.2 for herbs; the damping coefficient c is determined by the formula... Calculations are performed, where the damping ratio ζ = 0.1; wind load is calculated using the formula... The calculation shows that, among which ρ is the drag coefficient, with a value of 0.8; ρ is the air density, with a value of 1.2 kg / m³. The turbulent wind speed is calculated as follows:

[0060] ( =0.2V)

[0061] in, The average wind speed, The standard deviation of wind speed fluctuation, This is noise. The slope correction factor δ=cosθ is used to correct the effect of slope on the vegetation gravity component.

[0062] Optionally, in some optional embodiments provided in this disclosure, the vegetation dynamics differential equation is solved using the fourth-order Runge-Kutta method to obtain the vegetation dynamic displacement time series x(t) corresponding to each radar pixel position during the monitoring period. Based on the mapping relationship between radar phase and displacement, the radar phase interference ΔP caused by vegetation disturbance is obtained. The mapping relationship is as follows: , where λ is the radar operating wavelength.

[0063] In one possible implementation, in step S3, spatial texture features are extracted from the radar intensity data to obtain intensity features, dynamic disturbance features are extracted from the phase time series data combined with environmental factor data to obtain phase features, cross-modal fusion is performed on the intensity features and phase features, and pixel-level vegetation category probability distribution of the monitoring area is generated by combining the radar phase interference amount, and vegetation interference area is identified based on the vegetation category probability distribution.

[0064] In one possible implementation, the feature extraction, fusion, and vegetation recognition are achieved through a cross-modal adaptive radar vegetation recognition model, which includes an intensity split encoder, a phase split encoder, a vegetation coding and environment modulation module, and a cross-modal fusion decoding module.

[0065] In one possible implementation, the intensity split encoder takes radar intensity data S and gray-level co-occurrence matrices (GLCMs) in four directions (0°, 45°, 90°, and 135°) as input. A spatial feature tensor is generated using a multi-channel, multi-scale convolutional encoder to capture the reflectivity and texture differences between vegetated and non-vegetated areas. The intensity features are calculated as follows:

[0066]

[0067] in, For multi-channel, multi-scale spatial convolution, These are the learnable parameters of the encoder. Let be the intensity feature tensor corresponding to pixel (i,j).

[0068] For example, the multi-scale convolution uses three convolution kernels of different sizes: 3×3, 5×5, and 7×7, to capture vegetation texture features of different canopy sizes, and the number of output feature channels is 64.

[0069] In one possible implementation, the inputs of the phase split encoder are radar phase time-series data P, rainfall R, and wind speed V. A spatiotemporal attention encoder extracts the spatial distribution features at the same moment and the temporal dynamic features at different time steps, generating a phase spatiotemporal feature tensor to capture the dynamic fluctuation features of vegetation under wind and rain conditions. The phase features are calculated as follows:

[0070]

[0071] in, For spatiotemporal attention encoders, These are the learnable parameters of the encoder. Let be the phase feature tensor corresponding to pixel (i,j).

[0072] For example, the spatiotemporal attention encoder adopts a Transformer structure, with a time dimension input length of 20 sampling periods, 8 attention heads, and 64 output feature channels.

[0073] In one possible implementation, when performing cross-modal fusion of intensity features and phase features, vegetation type and growth stage are encoded as category embedding vectors, environmental factor data are encoded as environmental feature vectors, class condition query vectors are generated based on the category embedding vectors and environmental feature vectors, and the fusion weights of intensity features and phase features are dynamically adjusted through the class condition query vectors.

[0074] In one possible implementation, vegetation type (trees, shrubs, herbs) is combined with growth stage (budding, mature, decaying), and bare land category is superimposed to obtain 10 preset categories. These categories are encoded as discrete labels k∈{0,1,…,9}, and the category labels are encoded into category embedding vectors through an embedding layer. ∈ ,Right now

[0075]

[0076] in, For learnable parameters of the embedding layer, Let be the embedding vector for vegetation category k. In one possible implementation, environmental factors such as rainfall R, wind speed V, slope, and curvature corresponding to pixel (i,j) are encoded into an environmental feature vector. ,Right now

[0077]

[0078] Where ϕ(⋅) is the nonlinear activation function of a multilayer perceptron or GELU. Let (i,j) be the terrain feature corresponding to pixel (i,j). , , This is the weight matrix. This is a bias term.

[0079] In one possible implementation, a class-conditional query vector is generated based on the class embedding vector and the environmental feature vector. The intensity feature weight and phase feature weight corresponding to each class are calculated using a class attention mechanism to achieve dynamic feature modulation.

[0080]

[0081] in, W is the class condition query vector corresponding to category k. q This is the weight matrix;

[0082] Linearly project the intensity and phase features into the attention key / value space:

[0083]

[0084]

[0085]

[0086] in, Let be the weight matrix projected from the intensity features onto the attention key space. The weight matrix is ​​the projection of the intensity features onto the attention value space. Let be the weight matrix projected onto the attention key space from the phase features. The weight matrix is ​​the projection of the phase features onto the attention value space. For the attention dimension (scaling factor), through promote The changes in phase and intensity are adjusted to reflect the contributions of different vegetation categories, resulting in modulated vegetation category feature representations: [This is a partial translation of the original text, which is incomplete and requires further context.] By varying and adjusting the contributions of phase and intensity to different vegetation categories, we obtain the modulated vegetation category feature representations:

[0087]

[0088] The intensity feature weight α is calculated based on the query vector. k,s (i,j) and phase feature weight α k,p (i,j), by using softmax normalization to ensure that the sum of the weights of the two types of features is 1, finally obtaining the modulated vegetation category feature F. k (i,j).

[0089] In one possible implementation, a slope vegetation environment interaction integration module is used to achieve complementary fusion of intensity features and phase features, generating cross-modal fused features. Based on the fusion feature, the vegetation category embedding vector and the environmental feature vector are combined and mapped to the pixel-level log probability of vegetation category. The vegetation category probability distribution P(i,j,k) is obtained by normalization through the Softmax function, which is the predicted probability that pixel (i,j) belongs to vegetation category k.

[0090] For example, the category with the highest predicted probability is used as the recognition result of the pixel. When the recognition result is tree, shrub, or herb, the pixel is marked as a vegetation interference area, and a vegetation interference area mask is generated for the monitoring area.

[0091] In one possible implementation, the training process of the cross-modal adaptive radar vegetation recognition model is optimized using an environment-adaptive physical constraint joint loss function, wherein the joint loss function is:

[0092]

[0093] in, Cross-entropy loss is used to optimize pixel-level classification accuracy. Dice loss is used to suppress class imbalance and enhance the sensitivity to the identification of sparse vegetation areas. This is a vegetation dynamic physical consistency loss term, used to constrain the model prediction results to be consistent with the output of the vegetation dynamic physical model, thus solving the problems of black box nature and insufficient generalization of deep learning models.

[0094] In one possible implementation, the vegetation dynamic physical consistency loss term is calculated using the following formula:

[0095]

[0096] Where W and H represent the width and height of the radar image of the monitored area, Let (i,j) be the predicted probability of the vegetation category k corresponding to pixel (i,j). The standard deviation of phase perturbation for vegetation of category k, calculated using a forced damped harmonic oscillator system. The reference perturbation standard deviation for vegetation category k is pre-calibrated using a labeled dataset.

[0097] For example, during training, the dynamic weight α ranges from 0.5 to 0.7, and the hyperparameter λ ranges from 0.2 to 0.4, which strengthens the constraint of the physical model on the model's prediction while ensuring classification accuracy.

[0098] In one possible implementation, in step S4, based on the radar phase interference, vegetation category probability, and environmental factor data corresponding to the vegetation interference area, a threshold adjustment coefficient for each pixel location within the monitoring area is calculated, and the preset fixed warning threshold is dynamically adjusted pixel by pixel using the following formula:

[0099]

[0100] in, The adjusted warning threshold for pixel (i,j) The preset fixed warning threshold, This is the threshold adjustment coefficient corresponding to pixel (i,j).

[0101] In one possible implementation, the threshold adjustment coefficient Calculated using the following formula:

[0102]

[0103] in, pixel coordinates The corresponding real-time rainfall data, pixel coordinates The corresponding real-time wind speed data, Pixel coordinates within the slope radar monitoring area The predicted probability value of belonging to the k-th type of vegetation. Pixel coordinates within the slope radar monitoring area The corresponding combination of terrain features, For learnable parameters, The standard deviation of the phase disturbance at this pixel location is calculated by taking the standard deviation of the phase disturbance ΔP from the time series data during the monitoring period; while ϕ(・) is a nonlinear fusion function, whose inputs are the vegetation category probability, rainfall, wind speed, and terrain data at this pixel location.

[0104] For example, the fixed early warning threshold The basic threshold for slope deformation early warning is set at 5mm; when the pixel location is a mature tree and the wind speed is ≥8m / s, the threshold adjustment coefficient is... The value is set to 2.0~3.0, corresponding to an adjusted warning threshold of 10mm-15mm, to reduce false alarms caused by vegetation disturbance; when the pixel location is bare ground or withering herbaceous plants, the threshold adjustment coefficient is... The value should be set between 1.0 and 1.2 to maintain the original monitoring sensitivity and avoid underreporting of true deformation.

[0105] In one possible implementation, slope deformation early warning is performed based on an adjusted early warning threshold, and the early warning logic is as follows:

[0106]

[0107] in, This represents the radar deformation monitoring inversion result corresponding to pixel (i,j).

[0108] Optionally, in some optional embodiments provided in this disclosure, in order to suppress false warnings of isolated pixels, neighborhood consistency and temporal continuity filtering are performed on the pixels that trigger alarms. Only when the alarm of pixel (i,j) has no less than 3 consecutive alarm pixels in its 3×3 spatial neighborhood and the alarm is triggered in 3 consecutive sampling periods is it considered as a valid deformation warning. Otherwise, it is judged as an isolated anomaly and filtered out.

[0109] In some optional embodiments provided in this disclosure, to avoid missing the small deformation signals in the early stage of landslides during the vegetation disturbance elimination process, a vegetation disturbance attribute discrimination step is added before step S3 of the above basic embodiment completes the identification of vegetation disturbance areas and step S4 executes the early warning threshold adjustment. This step distinguishes whether the displacement signal of the vegetation area is environmentally driven random disturbance or landslide-driven overall displacement, and a differentiated threshold adaptation strategy is executed. The specific implementation is as follows:

[0110] For each pixel marked as a vegetation disturbance area in step S3, four types of core discriminative features are extracted to provide a quantitative basis for the classification of disturbance attributes:

[0111] (1) Spatial continuity characteristics: Calculate the consistency of deformation displacement direction and standard deviation of displacement between the target pixel and pixels of the same vegetation category in the surrounding 3×3 and 5×5 neighborhoods. The consistency of direction is characterized by the cosine of the angle between the displacement vectors of adjacent pixels. The calculation formula is as follows:

[0112]

[0113] in, Let N be the neighborhood range of the target pixel (i,j), and let N be the number of vegetation pixels in the neighborhood. The deformation displacement vector of the target pixel. The directional consistency coefficient has a value range of [-1, 1]. The closer the value is to 1, the higher the consistency of vegetation displacement direction in the neighborhood, and the more it conforms to the overall displacement characteristics driven by landslides.

[0114] (2) Temporal correlation characteristics: The displacement time series of the target pixel within 20 consecutive sampling periods is calculated, and the Pearson correlation coefficient Corr is calculated with the synchronously collected wind speed and rainfall time series. env The closer the absolute value of the correlation coefficient is to 0, the lower the correlation between the displacement and the random disturbance driven by wind and rain, and the greater the probability that it is a landslide-driven displacement.

[0115] (3) Physical model residual characteristics: Based on the forced damped harmonic oscillator system in step S2, calculate the residual Res(i,j) between the measured displacement time series of the target pixel and the environmental driving disturbance time series predicted by the model. The larger the residual, the more the measured displacement contains a landslide driving displacement component that exceeds the random disturbance.

[0116] (4) Vegetation coupling characteristics: Based on the probability distribution of vegetation categories obtained in step S3, the vegetation type and growth stage of the target pixel are extracted. The root system of trees and mature shrubs has a higher coupling degree with the slope rock and soil, and the transferability of landslide deformation is stronger. They are given higher weights in the discrimination process.

[0117] The extracted discriminative features are input into a pre-trained gradient boosting tree classifier, which outputs the probability P that the displacement of the target pixel belongs to the overall displacement driven by the landslide. slide (i,j), based on the aforementioned probability, a differentiated threshold adjustment strategy is executed, which complements the original dynamic threshold adjustment scheme:

[0118] (1) When P slide (i,j)≥P th When the vegetation displacement of a pixel is determined to be a landslide-driven overall displacement, regardless of the magnitude of the vegetation disturbance, the warning threshold for that pixel will not be increased, and the basic fixed warning threshold τ0 will remain unchanged. Simultaneously, the pixel will be marked as a deformation-concealed pixel and included in the slope deformation time-series tracking sequence to continuously monitor its displacement changes. For example, the discrimination threshold P... th The value is 0.75.

[0119] (2) When P slide (i,j) <P th When the vegetation displacement of the pixel is determined to be an environment-driven random disturbance, the dynamic threshold adjustment strategy in the above basic embodiment is executed to increase the warning threshold based on the amount of vegetation phase interference, thereby suppressing false warnings.

[0120] To avoid misjudgment of single pixels, a region contiguousness check is performed on pixels initially identified as landslide-driven displacement: only when there is a contiguous landslide-driven displacement region with an area of ​​not less than 9 consecutive pixels within the monitoring area in 3 consecutive sampling periods is it finally identified as a valid landslide-driven vegetation displacement region and its monitoring sensitivity is retained; otherwise, it is identified as a scattered anomaly point and a threshold adjustment strategy of random disturbance is implemented.

[0121] In one possible implementation, the training samples of the gradient boosting tree classifier include labeled environmentally driven random vegetation perturbation samples, artificially simulated slope soil and rock mass driving vegetation displacement samples, and vegetation displacement samples in the early stage of landslides in historical landslide events. During training, the F1 score is used as the optimization target to ensure a balance between classification accuracy and recall.

[0122] To verify the technical effect of this supplementary embodiment, an artificial landslide simulation test was added to the test scenario of an open-pit mine slope in Southwest China. In the mature shrub-covered area in the lower part of the slope, the slope soil and rock were slowly pushed with jacks to simulate the small deformation of 0.5 mm / day in the early stage of a landslide, causing the overlying shrubs to undergo continuous overall displacement. The supplementary embodiment, the original basic embodiment, and the existing mainstream coherence screening method were used for synchronous monitoring. The test results are as follows:

[0123] Existing coherence screening methods identify the shrub area of ​​the simulated landslide as a low-coherence vegetation disturbance area, directly eliminating all deformation signals in the area and failing to identify the small deformations in the early stage of the simulated landslide, resulting in serious underreporting.

[0124] Original basic implementation scheme: Because the area is covered by shrubs, the warning threshold was raised based on the amount of vegetation disturbance during windy weather. The warning was not triggered for the small deformation of 0.5 mm / day, which poses a risk of landslide underreporting.

[0125] This supplementary embodiment identifies a directional consistency coefficient of 0.92 for shrub displacement in the area by perturbation attribute discrimination, and a correlation coefficient of only 0.13 between displacement time series and wind speed and rainfall, which is consistent with the overall displacement characteristics driven by landslide. Therefore, the basic warning threshold remains unchanged, and a tiny deformation of 0.5 mm / day is accurately identified and a warning is triggered with no missed reports. At the same time, the false alarm rate of 4.6% is still maintained for random vegetation disturbance in other surrounding areas.

[0126] The above test results prove that this supplementary embodiment perfectly solves the dilemma of eliminating vegetation interference and preserving early landslide signals. Without increasing the false alarm rate of early warning, it achieves 100% detection of small deformations in the early stage of landslides in vegetated areas, further improving the monitoring reliability and engineering practical value of the present invention.

[0127] Furthermore, to intuitively verify the technical effects of this embodiment, this embodiment uses a steep slope in an open-pit mine in southwestern my country as the test monitoring object to conduct on-site comparative verification tests, combined with the attached... Figure 1-4 The effects of implementing the present invention will be described in detail.

[0128] The monitoring target slope in this embodiment is the north slope of an open-pit mine. The overall height of the slope is 216m, the average slope is 38°, and the total monitoring area is 0.72km². The vegetation coverage rate in the area reaches 67%. The upper and middle parts of the slope are densely covered with trees, the lower and middle gentle slopes are covered with a mixture of shrubs and herbs, and the toe and platform areas are bare rock / bare soil areas. Figure 1 The original radar intensity image of the slope shown corresponds perfectly.

[0129] The monitoring equipment used in this embodiment is a Ku-band ground-based synthetic aperture slope radar with a radar wavelength of 2 cm, a range resolution of 0.4 m, an azimuth resolution of 0.3 m, a data sampling interval of 5 minutes, and a continuous monitoring period of 60 days, which fully covers typical meteorological conditions that are prone to vegetation disturbance, such as heavy rainfall during the rainy season and strong winds in spring. Simultaneously, real-time meteorological data are collected through anemometers and rain gauges deployed on the slope top, and a 1:1000 digital terrain model of the slope is obtained through UAV aerial surveying, providing basic data support for the implementation of the method.

[0130] The original radar intensity image of the target slope in this embodiment can clearly distinguish the topographic contour of the slope and the differences in radar reflection of different ground features, but it cannot directly identify the type and spatial distribution of vegetation cover area. Figure 1 This is a pixel-level vegetation type mask output by the method of this embodiment. The mask is generated based on the cross-modal adaptive radar vegetation recognition model of this embodiment. Different colors correspond to four core areas: trees, shrubs, herbs, and bare land, and are fully registered with the original slope spatial location.

[0131] from Figure 1 As can be seen intuitively, the method of this embodiment can accurately match the spatial distribution of vegetation on the original slope: the dense tree area in the upper and middle part of the slope and the mixed shrub-herb area in the lower and middle part are completely identified, and the working platform at the foot of the slope and the exposed rock mass area of ​​the slope are accurately identified as bare land areas, with no large-scale omissions or misidentifications.

[0132] Verified through on-site manual survey and labeling, the method in this embodiment achieved an overall identification accuracy of 96.4% for vegetation areas on slopes, with an accuracy rate of 97.1% for trees, 95.3% for shrubs, and 94.8% for herbs. It can achieve pixel-level accurate identification of different vegetation types and areas with different coverage levels, providing a reliable foundation for subsequent vegetation interference elimination.

[0133] Figure 2 The attached image shows the slope deformation monitoring results obtained using a traditional fixed threshold early warning mechanism but without employing the method described in this embodiment. Figure 3 The two images show the slope deformation monitoring results after vegetation disturbance elimination and dynamic threshold adjustment using the method of this embodiment. The deformation color scale ranges of the two images are completely consistent. The red area is the alarm area where the deformation exceeds the warning threshold, the yellow area is the deformation concern area, and the blue area is the stable area.

[0134] from Figure 2 It can be clearly seen that in the traditional monitoring scheme, large areas of continuous red alarm zones appeared in the dense tree area in the middle and upper part of the slope and the shrub-herb cover area in the middle and lower part. After on-site verification, there were no signs of slope rock deformation in the above areas. All alarms were false warnings caused by vegetation swaying under the action of wind and rainfall. False alarm areas accounted for 92.3% of the total alarm areas, with a false alarm rate as high as 31.7%, which seriously interfered with the normal judgment of slope safety monitoring.

[0135] from Figure 3 It can be clearly seen that after adopting the method of this embodiment, the original Figure 2The large-scale false red alert areas caused by vegetation disturbance have been completely eliminated, with only two continuous deformation alert areas remaining in the jointed rock mass area in the middle of the slope. On-site verification confirmed that these two areas were indeed areas of genuine minor deformation in the slope rock mass, with deformation values ​​perfectly matching the radar inversion results. The method in this embodiment achieved a 100% recognition rate for genuine deformation, with no missed alarms. Statistics show that after adopting this method in this embodiment, the false alarm rate for slope deformation warnings decreased to 4.6%, a reduction of 85.5% compared to the traditional fixed threshold scheme, effectively eliminating vegetation disturbance and accurately suppressing false alarms.

[0136] To further verify the technical advantages of this embodiment, parallel comparative tests were conducted simultaneously using mainstream industry methods such as coherence screening and fixed-time smoothing filtering. The test results are shown in the table below:

[0137] This embodiment 96.40% 4.60% 100% 7.80% Traditional coherence screening method 72.10% 30.80% 82.40% 63.20% Fixed-time smoothing filtering method No vegetation identification ability 26.90% 76.50% 58.70%

[0138] The comparison results clearly show that the method in this embodiment is significantly superior to the existing mainstream technologies in three core dimensions: vegetation identification accuracy, false alarm rate suppression, and real deformation detection capability. Especially under extreme weather conditions such as strong winds and heavy rainfall, which cause severe vegetation disturbance, it can still maintain an extremely low false alarm rate and stable monitoring performance, solving the problems of high false alarm rate, easy omission of real deformation, and poor environmental adaptability of existing technologies in vegetation coverage scenarios.

[0139] This invention is not limited to the optional embodiments described above, and anyone can derive other various forms of products based on the inspiration of this invention. The specific embodiments described above should not be construed as limiting the scope of protection of this invention; the scope of protection of this invention should be determined by the claims, and the specification can be used to interpret the claims.

Claims

1. A method for vegetation identification and interference cancellation based on slope radar, applied to slope deformation monitoring scenarios, characterized in that, Includes the following steps: S1. Obtain slope radar echo data of the target slope monitoring area, as well as environmental factor data corresponding to the monitoring area; S2. Construct the vegetation in the monitoring area into a forced damped harmonic oscillator system, and quantify the radar phase interference caused by vegetation disturbance in combination with the environmental factor data. S3. Based on the slope radar echo data and the radar phase interference, identify the vegetation interference area within the monitoring area; S4. Based on the identification results of the vegetation interference area and the corresponding radar phase interference, dynamically adjust the early warning threshold of the slope deformation monitoring, and execute the slope deformation early warning based on the adjusted early warning threshold. The slope radar echo data includes radar intensity data and phase time series data; the environmental factor data includes meteorological monitoring data and topographic data, wherein the meteorological monitoring data includes wind speed and rainfall, and the topographic data includes the elevation, slope, and curvature of the monitoring area. In step S2, when constructing the forced damped harmonic oscillator system, a vegetation dynamics differential equation is constructed by combining vegetation type, vegetation growth stage, wind speed, rainfall, and slope data. The vegetation displacement is then solved to obtain the vegetation displacement, and the corresponding radar phase interference is mapped based on the vegetation displacement. The vegetation dynamics differential equation is as follows: ; Where x represents the instantaneous vibration displacement of the vegetation canopy. The instantaneous vibration velocity of the vegetation canopy. Let be the instantaneous vibration acceleration of the vegetation canopy, m be the vegetation mass, c be the damping coefficient, and k be the vegetation stiffness. δ is the wind load, g is the slope correction factor, θ is the gravitational acceleration, and θ is the slope angle of the monitoring area.

2. The vegetation identification and interference cancellation method based on slope radar according to claim 1, characterized in that, The vegetation mass m is expressed by the formula: ; The calculation yielded, where α is the dry mass of vegetation, R is the rainfall interception coefficient, R is the rainfall amount, and A is the projected area of ​​vegetation. The vegetation stiffness k is calculated by the formula k=k0βγ, where k0 is the reference stiffness, β is the vegetation type coefficient, and γ is the vegetation growth stage factor.

3. The vegetation identification and interference cancellation method based on slope radar according to claim 1, characterized in that, In step S3, spatial texture features are extracted from the radar intensity data to obtain intensity features, dynamic disturbance features are extracted from the phase time series data combined with environmental factor data to obtain phase features, cross-modal fusion is performed on the intensity features and phase features, and pixel-level vegetation category probability distribution of the monitoring area is generated by combining the radar phase interference amount, and vegetation interference areas are identified based on the vegetation category probability distribution.

4. The vegetation identification and interference cancellation method based on slope radar according to claim 3, characterized in that, In step S3, the prediction result of the vegetation category probability distribution is optimized through an environmentally adaptive physical constraint joint loss function. The joint loss function includes a vegetation dynamic physical consistency loss term, which is calculated using the following formula: ; in, The term represents the loss of vegetation dynamic physical consistency, where W and H are the width and height of the radar image of the monitoring area. The vegetation category corresponding to pixel (i,j) The predicted probability, The category calculated using the forced damped harmonic oscillator system Standard deviation of phase perturbation of vegetation Vegetation category The reference perturbation standard deviation.

5. The method for vegetation identification and interference cancellation based on slope radar according to claim 1, characterized in that, In step S4, based on the radar phase interference, vegetation category probability, and environmental factor data corresponding to the vegetation interference area, the threshold adjustment coefficient for each pixel location within the monitoring area is calculated using the following formula. The preset fixed warning threshold is dynamically adjusted pixel by pixel, whereby... The adjusted warning threshold for pixel (i,j) The preset fixed warning threshold, This is the threshold adjustment coefficient corresponding to pixel (i,j).

6. The vegetation identification and interference cancellation method based on slope radar according to claim 1, characterized in that, After acquiring the data in step S1, preprocessing operations are performed on the radar echo data and environmental factor data. The preprocessing operations include: performing environmental-guided non-uniform spatial filtering on the radar intensity data in combination with terrain data; performing meteorological-driven outlier removal and periodic disturbance extraction on the phase time series data in combination with meteorological monitoring data; and performing spatial interpolation and timestamp alignment on the environmental factor data to match the spatial grid and time scale corresponding to the radar data.

7. The method for vegetation identification and interference cancellation based on slope radar according to claim 3, characterized in that, When performing cross-modal fusion of intensity and phase features, vegetation type and growth stage are encoded as category embedding vectors, and environmental factor data are encoded as environmental feature vectors. Based on the category embedding vectors and environmental feature vectors, class condition query vectors are generated, and the fusion weights of intensity and phase features are dynamically adjusted through the class condition query vectors.

Images