A high-precision soil three-dimensional mapping method based on radar detection

Abstract

The present invention provides a high-precision three-dimensional soil mapping method based on radar detection, which collects radar data of the soil in the target area, wherein the radar data includes three-dimensional laser point cloud data collected by laser radar and two-dimensional ground radar data collected by ground detection radar, separates ground points from the three-dimensional laser point cloud data through an end-to-end trained semantic segmentation network, and models a digital ground model, matches and aligns the two-dimensional ground radar data with the three-dimensional laser point cloud data through a filtering algorithm, generates a three-dimensional soil visualization model, extracts soil attribute parameters from the three-dimensional soil visualization model, and generates a thematic map for reflecting the spatial distribution of the physical and chemical properties of the soil in the target area. The present invention can improve the accuracy, reliability and comprehensive application value of soil mapping through the precise acquisition and fusion expression of multi-dimensional, multi-scale and multi-temporal soil information.

Description

Technical Field

[0001] The invention belongs to the technical field of soil mapping, and in particular relates to a high-precision three-dimensional soil mapping method based on radar detection. Background Art

[0002] In modern precision agriculture management, high-precision soil mapping is the key to improving agricultural production efficiency and optimizing resource allocation. The physical and chemical properties of soil, such as water content, dielectric constant, porosity, etc., directly affect the growth and yield of crops. With the development of agricultural technology, more and more research and applications rely on efficient and accurate soil information acquisition technology to meet the needs of refined management. Traditional soil mapping methods rely on manual sampling and laboratory analysis, which have problems such as complex operation, low efficiency, and limited coverage, and are difficult to cope with large-area, heterogeneous soil distribution characteristics. In recent years, with the development of technologies such as ground probing radar (GPR) and three-dimensional laser point cloud, rapid acquisition of soil information and high-resolution modeling have become possible. These technologies can realize the simultaneous measurement of soil surface morphology and internal properties, providing rich data support for the construction of soil models. The current existing soil mapping schemes have the following problems:

[0003] (1) The soil data obtained are relatively single-dimensional and can only reflect some of the physical or chemical properties of the soil. The property changes of soil layers at different depths are often simplified or ignored, resulting in the inability of the surveying and mapping results to truly reflect the multi-dimensional characteristics of the soil, making it difficult to meet the needs of precision agriculture for multi-attribute and multi-scale soil data.

[0004] (2) The spatial and temporal resolution is insufficient, requiring a long sampling and analysis time, making it difficult to achieve real-time updates of soil properties;

[0005] (3) The quantitative inversion capability is poor. Due to differences in scale, accuracy, and temporal resolution, it is difficult to effectively integrate data from different sources into a unified model, and it is impossible to comprehensively and accurately characterize the spatiotemporal distribution of soil properties.

[0006] Therefore, we need to develop a high-precision three-dimensional soil mapping method based on radar detection, which can fuse multi-source data, combine multi-phase GPR data and real-time monitoring methods, dynamically update the spatiotemporal distribution of soil properties, and realize accurate modeling and visualization of the spatiotemporal distribution of soil properties. Summary of the invention

[0007] The purpose of the present invention is to provide a high-precision three-dimensional soil mapping method based on radar detection to solve the problems of the existing soil mapping schemes mentioned in the above background technology, such as single soil data dimension, insufficient temporal and spatial resolution, and poor quantitative inversion capability.

[0008] To achieve the above object, the present invention provides a high-precision three-dimensional soil mapping method based on radar detection, and the method is specifically as follows:

[0009] Collecting radar data of the soil in the target area, wherein the radar data includes three-dimensional laser point cloud data collected by laser radar and two-dimensional ground radar data collected by ground detection radar;

[0010] Through an end-to-end trained semantic segmentation network, ground points are separated from the three-dimensional laser point cloud data, and a digital ground model is modeled. The end-to-end training adopts a stochastic gradient descent optimization algorithm, and prevents overfitting through strategies including but not limited to learning rate decay and early stopping. The end-to-end training includes the following stages: training data preparation stage, training stage, forward reasoning stage, and prediction result post-processing stage;

[0011] Matching and registering the two-dimensional ground radar data with the three-dimensional laser point cloud data through a filtering algorithm to generate a three-dimensional soil visualization model, wherein the matching and registration process includes three-dimensional projection, state estimation, and depth association;

[0012] Extracting soil attribute parameters from the three-dimensional soil visualization model and generating a thematic map reflecting the spatial distribution of soil physical and chemical properties in the target area;

[0013] The three-dimensional projection is to convert the two-dimensional ground radar data into a three-dimensional coordinate system corresponding to the three-dimensional laser point cloud data, and for each ground detection radar survey line, establish a mapping relationship between the sampling points of the ground detection radar and the digital ground model through an extended Kalman filter, estimate the real position of the sampling point as the estimated position, and use the estimated position to update the three-dimensional coordinates.

[0014] Based on the above solution, the radar data collection specifically includes:

[0015] For the collection of the three-dimensional laser point cloud data, the serpentine route of the laser radar is planned according to the soil area of ​​the target area, ensuring that the route spacing does not exceed 50m, the heading overlap rate is not less than 60%, and the lateral overlap rate is not less than 30%; the serpentine route planning is calculated by the following formula: route spacing = 2×(relative ground height×tan(laser radar vertical field of view angle / 2)×(1-lateral overlap rate)), during the laser radar aerial photography process, the ground target is continuously scanned, and the information of each scanning point is recorded, and the information of the scanning point includes but is not limited to the three-dimensional coordinates (x, y, z), reflection intensity, and echo order;

[0016] The ground detection radar is specifically a GPRover vehicle-mounted ground penetrating radar. For the collection of the two-dimensional ground radar data, a high-frequency electromagnetic pulse signal is emitted through the transmitting antenna of the GPRover vehicle-mounted ground penetrating radar. When it penetrates the underground and encounters an interface with electrical differences, it generates reflection. The receiving antenna records the two-way travel time, amplitude and other information of the reflected signal, and stores it synchronously with the vehicle positioning information.

[0017] Based on the above scheme, in the training data preparation stage, a large number of high-density point clouds are collected under different terrain conditions using a vehicle-mounted mobile measurement system. The high-density point clouds are used as original point clouds, and point-level semantic labels are formed through manual annotation. The original point clouds are subjected to data enhancement operations such as random cropping, rotation, and scaling, and are divided into training sets, validation sets, and test sets.

[0018] The forward reasoning stage is to use the trained semantic segmentation network to perform forward reasoning on the collected three-dimensional laser point cloud data after the training stage is completed to obtain a prediction result, that is, a ground point prediction probability or a non-ground point prediction probability of each point;

[0019] In the prediction result post-processing stage, the prediction result is post-processed using a graph-based smoothing optimization method, the ground point cloud data predicted as ground points in the three-dimensional laser point cloud data are extracted, exported into .las or .xyz format, and input into DTM software for modeling of a digital ground model.

[0020] Based on the above scheme, after the radar data is collected, the radar data is preprocessed, including but not limited to outlier removal, time zero point correction, distance calibration, background denoising, and after the preprocessing, the radar data is converted into SEG-Y format and recorded as the first radar data;

[0021] The end-to-end training adopts a stochastic gradient descent optimization algorithm, and prevents overfitting by using strategies including but not limited to learning rate decay and early stopping. The cross entropy loss function is used as the main optimization target to measure the difference between the predicted probability of the semantic segmentation network and the true label: , where N is the total number of training samples, C is the number of categories (ground points and non-ground points), is the true label of the i-th point in the c-th category, is the predicted probability of the i-th point in the c-th category;

[0022] The training set is divided into small batches, each batch is calculated in parallel by a GPU, the training set is input, and the category prediction probability of each point is calculated by the semantic segmentation network; based on the cross entropy loss function, the gradient is calculated and the parameters of the semantic segmentation network are updated.

[0023] Based on the above scheme, the semantic segmentation network is specifically a U-Net style encoder-decoder architecture, the network input is the point cloud block of the first radar data, and the network output is the ground or non-ground classification label of each point in the point cloud; the encoder downsamples layer by layer to extract the multi-scale features of the point cloud, and obtains a series of multi-scale feature maps; the decoder upsamples layer by layer to restore the point cloud resolution, and obtains the resolution output corresponding to the input point cloud; the local features of each layer of the encoder and the corresponding layer of the decoder are fused through jump connections to accurately classify the point cloud in the first radar data as ground points or non-ground points.

[0024] Based on the above solution, the two-dimensional ground radar data is converted into a three-dimensional coordinate system corresponding to the three-dimensional laser point cloud data, as follows:

[0025] Set up The coordinates of the starting and ending points of the ground detection radar survey line are and , then any sampling point on the measuring line The three-dimensional coordinates of are expressed as:

[0026]

[0027]

[0028]

[0029] in, is the sampling point number, ; For the The total number of sampling points for each survey line; is obtained from the digital terrain model by bilinear interpolation. The ground elevation value at the location; the three-dimensional coordinates and amplitude values ​​of all sampling points are recorded as ,in is the ground detection radar signal amplitude, Respectively The three-dimensional coordinates of the sampling points.

[0030] Based on the above scheme, the state estimation is completed through the state space model, and the state space model is defined as follows:

[0031] State vector: , , , Respectively represent The offset of the sampling point of the ground detection radar relative to the initial three-dimensional projection position;

[0032] Observation vector: , is obtained from the digital terrain model. The ground elevation value at the location, It is The signal amplitude of each sampling point of the ground detection radar;

[0033] State transfer equation , used to describe how the system state changes from time step Transfer to time step , where the state transfer matrix is the identity matrix, is process noise, with mean zero and covariance Normal distribution, covariance matrix Control the smoothness of state changes;

[0034] Observation equation: , where the observation function ,in Indicates the elevation value in the digital terrain model, depending on the current Location, Indicates the depth The signal amplitude of the ground detection radar at

[0035] Initial conditions: initial state estimate Set to zero vector, initial state covariance Let it be a diagonal matrix with diagonal elements , reflecting the uncertainty of the initial position.

[0036] Based on the above scheme, the depth association includes: sampling points of the ground detection radar after estimating the position Voxelization is performed to divide the three-dimensional space into three-dimensional grids of 0.1m×0.1m×0.02m, recorded as three-dimensional voxel grids, and each of the three-dimensional voxel grids takes the weighted average of the sampling point coordinates as the center coordinate of the three-dimensional voxel grid, and takes the normalized amplitude of the ground detection radar of the sampling point as the weight. The center coordinates of are calculated as:

[0037] ,in Indicates The three-dimensional voxel grid, weight By sampling point The signal amplitude of the ground detection radar Decide, define , in summary, substituting the weight into the central coordinate calculation formula, the central coordinate of the three-dimensional voxel grid is finally expressed as:

[0038] .

[0039] Based on the above scheme, the voxelized three-dimensional voxel grid is smoothed by a spherical linear interpolation algorithm, and the isosurface is extracted from the three-dimensional voxel grid by the Marching Cubes algorithm to generate a three-dimensional soil grid model. The amplitude value of the ground detection radar is mapped to RGB pseudo-color and rendered on the three-dimensional soil grid model to obtain the final three-dimensional soil visualization model.

[0040] Based on the above scheme, the generating of the thematic map reflecting the spatial distribution of soil physical and chemical properties in the target area specifically includes:

[0041] Based on the propagation characteristics of the ground detection radar signal in the soil medium, the three-dimensional soil visualization model is divided into soil layers through an image processing algorithm, and the geometric boundaries of different soil layers are automatically identified;

[0042] Based on the physical relationship between the GPR amplitude value and the dielectric constant and water content, combined with the soil dielectric model and the Topp equation, a soil attribute parameter inversion model is established; by substituting the GPR amplitude value into the soil attribute parameter inversion model, the dielectric constant and water content of each three-dimensional voxel grid are estimated;

[0043] Based on the Archie equation and Maxwell equation, the relationship between conductivity, dielectric constant and pore structure is described, the soil volume conductivity is estimated by the wave velocity of the ground detection radar, and the soil porosity is inverted by substituting it into the Archie equation, and other physical parameters are converted;

[0044] The soil property parameters obtained by inversion, including but not limited to dielectric constant, water content, soil volume conductivity, soil texture, and soil porosity, are subjected to Kriging interpolation calculation in the voxelized three-dimensional voxel grid to form regular three-dimensional voxel grid data;

[0045] The regular three-dimensional voxel grid data is visualized through a GIS platform to generate a spatial distribution map reflecting the soil attribute parameters, namely the thematic map. The thematic map overlays and analyzes monitoring information, which includes but is not limited to crop yields and growth conditions, reveals the spatial differences in soil fertility, identifies low-yield fields and salinized areas, and provides a scientific basis for precision agriculture management measures.

[0046] Compared with the prior art, the present invention has the following advantages and effects:

[0047] (1) By combining ground-based probing radar (GPR) signals, laser point clouds, and other measured data, a comprehensive soil information model with multiple attributes and scales is constructed, which significantly improves the multi-dimensional integrity of soil mapping results and achieves a comprehensive description of soil physical, chemical, and spatial properties;

[0048] (2) We use an end-to-end trained semantic segmentation network to accurately extract ground information from complex scene point clouds. We design a fusion filtering algorithm by introducing category weight balancing and graph optimization post-processing. We achieve accurate matching and registration of radar profiles and 3D point clouds through state estimation and deep correlation, and construct a high-resolution soil visualization 3D model. This effectively solves problems such as sample imbalance and discrete misclassification, and improves ground segmentation accuracy.

[0049] (3) By introducing physical models such as the Archie equation and the Topp equation, a quantitative soil parameter inversion method based on the propagation mechanism of ground penetrating radar signals was constructed. The coupled effects of multiple factors such as soil texture, temperature, and salinity were comprehensively considered, which significantly improved the reliability and applicability of the estimation of key attribute parameters such as soil moisture content, porosity, and bulk density. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following is a brief introduction to the drawings required for the specific embodiments or the description of the prior art. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn according to the actual scale.

[0051] Figure 1 The present invention provides a flow chart of a high-precision three-dimensional soil mapping method based on radar detection. DETAILED DESCRIPTION

[0052] In order to more clearly explain the purpose, technical solutions and advantages of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. The example implementation methods can be implemented in various forms and should not be understood as being limited to the examples described herein. On the contrary, these implementation methods are provided to make the present invention more comprehensive and complete, and to fully convey the concepts of the example implementation methods to those skilled in the art.

[0053] In addition, the described features, structures or characteristics may be combined in one or more embodiments in any suitable manner. In the following description, many specific details are provided to provide a full understanding of the embodiments of the present invention. However, those skilled in the art will appreciate that the technical solution of the present invention can be practiced without one or more of the specific details, or other methods, components, devices, steps, etc. may be adopted. In other cases, known methods, devices, implementations or operations are not shown or described in detail to avoid blurring various aspects of the present invention.

[0054] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities may be implemented in software form, or in one or more hardware modules or integrated circuits, or in different networks and / or processor devices and / or microcontroller devices.

[0055] The flowcharts shown in the accompanying drawings are only exemplary and do not necessarily include all the contents and operations / steps, nor must they be executed in the order described. For example, some operations / steps can be decomposed, and some operations / steps can be combined or partially combined, so the actual execution order may change according to actual conditions.

[0056] The present invention is described in detail below in conjunction with specific embodiments:

[0057] As attached Figure 1 As shown, the embodiment of the present invention provides a high-precision soil three-dimensional mapping method based on radar detection, and the specific steps of the method are as follows:

[0058] Step S1, collecting radar data of the soil in the target area, wherein the radar data includes three-dimensional laser point cloud data collected by a laser radar and two-dimensional ground radar data collected by a ground detection radar;

[0059] Step S2: separating ground points from the three-dimensional laser point cloud data through an end-to-end trained semantic segmentation network, and building a digital ground model;

[0060] Step S3, matching and registering the two-dimensional ground radar data with the three-dimensional laser point cloud data through a filtering algorithm to generate a three-dimensional soil visualization model;

[0061] Step S4: extracting soil attribute parameters from the three-dimensional soil visualization model, and generating a thematic map reflecting the spatial distribution of soil physical and chemical properties in the target area.

[0062] Preferably, in step S1, the radar data collection specifically includes:

[0063] For the collection of the three-dimensional laser point cloud data, the serpentine route of the laser radar is planned according to the soil area of ​​the target area, ensuring that the route spacing does not exceed 50m, the heading overlap rate is not less than 60%, and the lateral overlap rate is not less than 30%; the serpentine route planning is calculated by the following formula: route spacing = 2×(relative ground height×tan(laser radar vertical field of view angle / 2)×(1-lateral overlap rate)), during the laser radar aerial photography process, the ground target is continuously scanned, and the information of each scanning point is recorded, and the information of the scanning point includes but is not limited to the three-dimensional coordinates (x, y, z), reflection intensity, and echo order;

[0064] The ground detection radar is specifically a GPRover vehicle-mounted ground penetrating radar. For the collection of the two-dimensional ground radar data, a high-frequency electromagnetic pulse signal is emitted through the transmitting antenna of the GPRover vehicle-mounted ground penetrating radar. When it penetrates the underground and encounters an interface with electrical differences, it generates reflection. The receiving antenna records the two-way travel time, amplitude and other information of the reflected signal, and stores it synchronously with the vehicle positioning information.

[0065] Preferably, after the radar data acquisition in step S1 is completed, the radar data is preprocessed, including but not limited to outlier removal, time zero point correction, distance calibration, background denoising, and after the preprocessing, the radar data is converted into SEG-Y format and recorded as the first radar data.

[0066] Preferably, in step S2, the end-to-end training adopts a stochastic gradient descent optimization algorithm, and prevents overfitting by using strategies including but not limited to learning rate decay and early stopping. The end-to-end training includes the following stages: training data preparation stage, training stage, forward reasoning stage, and prediction result post-processing stage, as follows:

[0067] Step S201, in the training data preparation stage, a large number of high-density point clouds are collected under different terrain conditions by using a vehicle-mounted mobile measurement system. The high-density point clouds are used as original point clouds, and point-level semantic labels are formed by manual annotation. The original point clouds are subjected to data enhancement operations such as random cropping, rotation, and scaling, and are divided into training sets, validation sets, and test sets;

[0068] Step S202, the training phase, includes initialization, loss function definition, training process control and model verification, specifically:

[0069] Initialize the parameters of the semantic segmentation network, randomly distribute and set the initial weights or use pre-trained weights for transfer learning, and define the input and output of the semantic segmentation network: Input: the training set, including multi-dimensional features such as spatial coordinates and intensity values ​​of the point cloud; Output: the category prediction probability of each point (ground point or non-ground point);

[0070] The cross entropy loss function is used as the main optimization objective to measure the difference between the predicted probability of the semantic segmentation network and the true label: , where N is the total number of training samples, C is the number of categories (ground points and non-ground points), is the true label of the i-th point in the c-th category, is the predicted probability of the i-th point in the c-th category;

[0071] It should be noted that if the categories are unbalanced, that is, the number of ground points and non-ground points is significantly different, weighted cross entropy or focal loss can be introduced to enhance the learning ability of minority categories;

[0072] Use stochastic gradient descent or Adam optimizer to train the semantic segmentation network and set the learning rate scheduling strategy: the initial learning rate is , adopt a learning rate decay strategy, including but not limited to cosine annealing scheduling and step-by-step descent, to gradually reduce the learning rate to improve the convergence effect; for example, such as:

[0073] , where t is the current iteration number, is the decay coefficient; when the validation set loss does not improve for multiple consecutive times, the training is terminated early to avoid overfitting;

[0074] The training set is divided into small batches, each batch is calculated in parallel by a GPU, the training set is input, and the category prediction probability of each point is calculated by a semantic segmentation network; based on the cross entropy loss function, the gradient is calculated and the parameters of the semantic segmentation network are updated;

[0075] After each training cycle, the performance of the semantic segmentation network is evaluated using the validation set, and the classification ability of the model is evaluated using the following validation set indicators:

[0076] Accuracy: ;

[0077] Precision: ;

[0078] Recall: ;

[0079] F1 score: ;

[0080] Where TP is the number of correctly classified ground points, FP is the number of misclassified non-ground points, and FN is the number of missed ground points;

[0081] In each training cycle, the current best semantic segmentation network is saved according to the validation set indicator to prevent overfitting in subsequent training.

[0082] Step S203, the forward reasoning stage is to use the trained semantic segmentation network to perform forward reasoning on the collected three-dimensional laser point cloud data after the training stage is completed, to obtain a prediction result, that is, a ground point prediction probability or a non-ground point prediction probability of each point;

[0083] Step S204, in the prediction result post-processing stage, the prediction result is post-processed using a graph-based smoothing optimization method, the ground point cloud data predicted as ground points in the three-dimensional laser point cloud data is extracted, exported to .las or .xyz format, and input into DTM software for modeling of a digital ground model.

[0084] Preferably, in step S2, the semantic segmentation network is specifically a U-Net style encoder-decoder architecture, the network input is the point cloud block of the first radar data, and the network output is the ground or non-ground classification label of each point in the point cloud; the encoder downsamples layer by layer to extract multi-scale features of the point cloud to obtain a series of multi-scale feature maps; the decoder upsamples layer by layer to restore the point cloud resolution to obtain a resolution output corresponding to the input point cloud; the local features of each layer of the encoder and the corresponding layer of the decoder are fused through jump connections to accurately classify the point cloud in the first radar data as ground points or non-ground points.

[0085] Preferably, the matching and registration process in step S3 includes three-dimensional projection, state estimation and depth association, specifically:

[0086] Step S301, the three-dimensional projection is to convert the two-dimensional ground radar data into a three-dimensional coordinate system corresponding to the three-dimensional laser point cloud data, as follows:

[0087] Set up The coordinates of the starting and ending points of the ground detection radar survey line are and , then any sampling point on the measuring line The three-dimensional coordinates of are expressed as:

[0088]

[0089]

[0090]

[0091] in, is the sampling point number, ; For the The total number of sampling points for each survey line; is obtained from the digital terrain model by bilinear interpolation. The ground elevation value at the location; the three-dimensional coordinates and amplitude values ​​of all sampling points are recorded as ,in is the ground detection radar signal amplitude, Respectively The three-dimensional coordinates of the sampling points;

[0092] For each of the ground detection radar survey lines, a mapping relationship between the sampling points of the ground detection radar and the digital ground model is established through extended Kalman filtering, the real position of the sampling point is estimated as the estimated position, and the three-dimensional coordinates are updated through the estimated position.

[0093] Step S302: complete the state estimation through a state space model, and the state space model is defined as follows:

[0094] State vector: , , , Respectively represent The offset of the sampling point of the ground detection radar relative to the initial three-dimensional projection position;

[0095] Observation vector: , is obtained from the digital terrain model. The ground elevation value at the location, It is The signal amplitude of each sampling point of the ground detection radar;

[0096] State transfer equation , used to describe how the system state changes from time step Transfer to time step , where the state transfer matrix is the identity matrix, is process noise, with mean zero and covariance Normal distribution, covariance matrix Control the smoothness of state changes;

[0097] Observation equation: , where the observation function ,in Indicates the elevation value in the digital terrain model, depending on the current Location, Indicates the depth The signal amplitude of the ground detection radar at

[0098] Initial conditions: initial state estimate Set to zero vector, initial state covariance Let it be a diagonal matrix with diagonal elements , reflecting the uncertainty of the initial position.

[0099] Step S303: Based on the state space model and the observation equation, the Kalman filter is extended, specifically as follows: First, state prediction is performed and covariance prediction ; Secondly, calculate the Kalman gain ,in is the observation function Jacobian matrix, state update ,in is the observation value, covariance update ; By continuously iterating the prediction step and the update step until the filter converges, the optimal estimate of the position of all sampling points is obtained:

[0100]

[0101] in, It is the sampling point The optimal estimated 3D coordinates, is the corresponding GPR signal amplitude.

[0102] Step S304: the depth association includes: sampling points of the ground detection radar after the estimated position Voxelization is performed to divide the three-dimensional space into three-dimensional grids of 0.1m×0.1m×0.02m, recorded as three-dimensional voxel grids, and each of the three-dimensional voxel grids takes the weighted average of the sampling point coordinates as the center coordinate of the three-dimensional voxel grid, and takes the normalized amplitude of the ground detection radar of the sampling point as the weight. The center coordinates of are calculated as:

[0103] ,in Indicates The three-dimensional voxel grid, weight By sampling point The signal amplitude of the ground detection radar Decide, define , in summary, substituting the weight into the central coordinate calculation formula, the central coordinate of the three-dimensional voxel grid is finally expressed as:

[0104] .

[0105] The voxelized three-dimensional voxel grid is smoothed by a spherical linear interpolation algorithm, and isosurfaces are extracted from the three-dimensional voxel grid by a MarchingCubes algorithm to generate a three-dimensional soil grid model. The amplitude value of the ground detection radar is mapped into RGB pseudo-color and rendered onto the three-dimensional soil grid model to obtain the final three-dimensional soil visualization model.

[0106] Preferably, in step S4, generating a thematic map for reflecting the spatial distribution of soil physical and chemical properties in the target area specifically includes:

[0107] Step S401: based on the propagation characteristics of the ground detection radar signal in the soil medium, the three-dimensional soil visualization model is divided into soil layers through an image processing algorithm, and the geometric boundaries of different soil layers are automatically identified;

[0108] Specifically, the image processing algorithm includes threshold segmentation, which combines soil expert knowledge and field drilling data to determine the selection of the threshold to ensure the geological rationality of the layer division results, and statistics the geometric and physical attribute parameters such as volume, thickness, average GPR amplitude, etc. of each soil layer to preliminarily quantify the spatial heterogeneity of the soil layer.

[0109] Step S402: Based on the physical relationship between the GPR amplitude value and the dielectric constant and water content, a soil property parameter inversion model is established in combination with the soil dielectric model and the Topp equation; by substituting the GPR amplitude value into the soil property parameter inversion model, the dielectric constant and water content of each three-dimensional voxel grid can be estimated;

[0110] Specifically, since the moisture characteristic curves of soils with different textures are different, the empirical relationship between water content and sand and silt content is constructed and calibrated with measured data, the soil texture composition can be further inferred from the water content.

[0111] Step S403: Based on the Archie equation and the Maxwell equation, the relationship between conductivity, dielectric constant and pore structure is described, the soil volume conductivity is estimated by the wave velocity of the ground detection radar, and the soil porosity is inverted by substituting it into the Archie equation, and then other physical parameters such as bulk density are calculated;

[0112] It should be noted that the electromagnetic parameters measured by the ground detection radar are often the apparent values ​​of multi-parameter coupling, and the influence of factors such as soil texture, temperature, and salinity need to be corrected during inversion.

[0113] Step S404: performing Kriging interpolation calculation on the soil property parameters obtained by inversion, including but not limited to dielectric constant, water content, soil volume conductivity, soil texture, soil porosity and bulk density, in the voxelized three-dimensional voxel grid to form regular three-dimensional voxel grid data. The Kriging interpolation calculation needs to select a suitable theoretical model according to the spatial distribution characteristics of the actual data, and optimize the parameters of the theoretical model;

[0114] It should be noted that compared with traditional deterministic interpolation methods such as inverse distance weighting, Kriging interpolation takes into account the spatial autocorrelation of soil properties. By constructing a variogram to fit the semi-variogram, it characterizes the similarity between samples at different distances and can estimate the attribute values ​​of unknown points based on this. It is widely used in the field of spatial statistics, and is particularly suitable for mapping soil attribute parameters with spatial correlation.

[0115] Step S405: Visualize the regular three-dimensional voxel grid data through a GIS platform to generate a thematic map reflecting the plane distribution of properties such as soil texture, bulk density, and porosity. The thematic map reveals the spatial difference law of soil fertility by superimposing and analyzing monitoring information such as crop yield and growth status, identifies low-yield fields and salinized areas, and can provide a scientific basis for precision agricultural management measures such as soil improvement and targeted fertilization;

[0116] It should be pointed out that the soil property parameters inverted based on the ground detection radar may have certain uncertainties, such as being affected by the instantaneous conditions of soil moisture, temperature, etc. It is possible to build a three-dimensional soil information database with multiple parameters, multiple scales, and multiple phases by strengthening the mutual verification and complementarity of soil data from different sources and combining it with other geophysical exploration, remote sensing, near-ground sensing and other means.

[0117] In this embodiment, radar data is collected on the soil in the target area, and the radar data includes three-dimensional laser point cloud data collected by laser radar and two-dimensional ground radar data collected by ground detection radar. The ground points are separated from the three-dimensional laser point cloud data through an end-to-end trained semantic segmentation network, and a digital ground model is modeled. The two-dimensional ground radar data is matched and registered with the three-dimensional laser point cloud data through a filtering algorithm to generate a three-dimensional soil visualization model, and soil attribute parameters are extracted from the three-dimensional soil visualization model, and a thematic map is generated to reflect the spatial distribution of the physical and chemical properties of the soil in the target area. The accurate acquisition and fusion expression of multi-dimensional, multi-scale, and multi-temporal soil information can improve the accuracy, reliability, and comprehensive application value of soil mapping.

[0118] Those skilled in the art will readily appreciate other embodiments of the present invention after considering the specification and practicing the invention disclosed herein. This application is intended to cover any modification, use or adaptation of the present invention, which follows the general principles of the present invention and includes common knowledge or customary techniques in the art that are not disclosed by the present invention. The specification and examples are intended to be exemplary only, and the true scope and spirit of the present invention are indicated by the claims. It should be understood that the present invention is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from the scope thereof. The scope of the present invention is limited only by the appended claims.

Claims

1. A high-precision three-dimensional soil mapping method based on radar detection, characterized in that: include: Collecting radar data of the soil in the target area, wherein the radar data includes three-dimensional laser point cloud data collected by a laser radar and two-dimensional ground radar data collected by a ground detection radar; Using an end-to-end trained semantic segmentation network, ground points are separated from the three-dimensional laser point cloud data and a digital ground model is modeled. The end-to-end training uses a stochastic gradient descent optimization algorithm and prevents overfitting by including a learning rate decay and early stopping strategy. The end-to-end training includes the following stages: training data preparation, training, forward reasoning, and prediction result post-processing. Matching and registering the two-dimensional ground radar data with the three-dimensional laser point cloud data through a filtering algorithm to generate a three-dimensional soil visualization model, wherein the matching and registration process includes three-dimensional projection, state estimation, and depth correlation; Extracting soil attribute parameters from the three-dimensional soil visualization model and generating a thematic map reflecting the spatial distribution of soil physical and chemical properties in the target area; The three-dimensional projection is to convert the two-dimensional ground radar data into a three-dimensional coordinate system corresponding to the three-dimensional laser point cloud data, establish a mapping relationship between the sampling points of the ground detection radar and the digital terrain model for each ground detection radar survey line through an extended Kalman filter, estimate the actual position of the sampling point as the estimated position, and use the estimated position to update the three-dimensional coordinates; The conversion of the two-dimensional ground radar data into a three-dimensional coordinate system corresponding to the three-dimensional laser point cloud data is specifically as follows: Set up the first The coordinates of the starting and ending points of the ground detection radar survey line are and , then any sampling point on the survey line The three-dimensional coordinates of are expressed as: , , ,in, is the sampling point number, ; For the The total number of sampling points on the survey line; is obtained from the digital terrain model by bilinear interpolation. The ground elevation value at the location; the three-dimensional coordinates and amplitude values ​​of all sampling points are recorded as ,in is the ground detection radar signal amplitude, Respectively The three-dimensional coordinates of the sampling points; The sampling points of the ground detection radar after estimated position are voxelized, and the dielectric constant and water content of each three-dimensional voxel grid are estimated by substituting the GPR amplitude value into the soil property parameter inversion model. The inverted soil property parameters are subjected to Kriging interpolation calculation in the voxelized three-dimensional voxel grid to form regular three-dimensional voxel grid data. The regular three-dimensional voxel grid data is visualized through a GIS platform to generate a thematic map reflecting the planar distribution of soil texture, bulk density, and porosity properties.

2. The high-precision three-dimensional soil mapping method based on radar detection according to claim 1 is characterized in that: The radar data collection specifically includes: For the collection of the three-dimensional laser point cloud data, the laser radar's serpentine route is planned based on the soil area of ​​the target area, ensuring that the route spacing does not exceed 50 meters, the heading overlap rate is not less than 60%, and the lateral overlap rate is not less than 30%. The serpentine route planning is calculated using the following formula: route spacing = 2 × (relative ground height × tan (laser radar vertical field of view angle / 2) × (1-lateral overlap rate)). During the laser radar aerial photography process, the ground target is continuously scanned and the information of each scanning point is recorded. The scanning point information includes the three-dimensional coordinates (x, y, z), reflection intensity, and echo order. The ground detection radar is specifically a GPRover vehicle-mounted ground penetrating radar. For the collection of the two-dimensional ground radar data, a high-frequency electromagnetic pulse signal is emitted through the transmitting antenna of the GPRover vehicle-mounted ground penetrating radar. When it penetrates the underground and encounters an interface with electrical differences, it is reflected. The receiving antenna records the two-way travel time and amplitude information of the reflected signal, and stores it synchronously with the vehicle positioning information.

3. The high-precision three-dimensional soil mapping method based on radar detection according to claim 1 is characterized in that: In the training data preparation stage, a vehicle-mounted mobile measurement system is used to collect a large number of high-density point clouds under different terrain conditions. The high-density point clouds are used as original point clouds, and point-level semantic labels are formed through manual annotation. The original point clouds are randomly cropped, rotated, and scaled for data augmentation, and divided into training sets, validation sets, and test sets; The forward reasoning stage is to use the trained semantic segmentation network to perform forward reasoning on the collected three-dimensional laser point cloud data after the training stage is completed to obtain a prediction result, that is, a ground point prediction probability or a non-ground point prediction probability of each point; In the prediction result post-processing stage, a graph-based smoothing optimization method is used to post-process the prediction results, and the ground point cloud data predicted as ground points in the three-dimensional laser point cloud data are extracted, exported into .las or .xyz format, and input into DTM software for modeling of a digital ground model.

4. The high-precision three-dimensional soil mapping method based on radar detection according to claim 3 is characterized in that: After the radar data is collected, the radar data is preprocessed, including outlier removal, time zero correction, distance calibration, and background denoising. After the preprocessing, the radar data is converted into a SEG-Y format and recorded as first radar data. The end-to-end training adopts the stochastic gradient descent optimization algorithm and prevents overfitting by including learning rate decay and early stopping strategies. The cross entropy loss function is used as the main optimization objective to measure the difference between the predicted probability of the semantic segmentation network and the true label: , where N is the total number of training samples, C is the number of categories of ground points and non-ground points, is the true label of the i-th point in the c-th category, is the predicted probability of the i-th point in the c-th category; The training set is divided into small batches, each batch is calculated in parallel by a GPU, the training set is input, and the category prediction probability of each point is calculated by the semantic segmentation network; based on the cross entropy loss function, the gradient is calculated and the parameters of the semantic segmentation network are updated.

5. The high-precision three-dimensional soil mapping method based on radar detection according to claim 4 is characterized in that: The semantic segmentation network is specifically a U-Net-style encoder-decoder architecture. The network input is the point cloud block of the first radar data, and the network output is the ground or non-ground classification label of each point in the point cloud. The encoder extracts multi-scale features of the point cloud through layer-by-layer downsampling to obtain a series of multi-scale feature maps. The decoder restores the point cloud resolution through layer-by-layer upsampling to obtain a resolution output corresponding to the input point cloud. The local features of each layer of the encoder are fused with the corresponding layers of the decoder through jump connections to accurately classify the point cloud in the first radar data as ground points or non-ground points.

6. The high-precision three-dimensional soil mapping method based on radar detection according to claim 1, characterized in that: The state estimation is accomplished through a state-space model, which is defined as follows: State vector: , 、 、 Respectively represent An offset of a sampling point of the ground detection radar relative to the initial three-dimensional projection position; Observation vector: , is obtained from the digital terrain model. The ground elevation value at the location, It is The signal amplitude of each sampling point of the ground detection radar; State transition equation , used to describe how the system state changes from time step Transfer to time step , where the state transfer matrix is the identity matrix, is process noise, with mean zero and covariance Normal distribution, covariance matrix Control the smoothness of state changes; Observation equation: , where the observation function ,in Indicates the elevation value in the digital terrain model, depending on the current Location, Indicates depth The signal amplitude of the ground detection radar at Initial conditions: initial state estimate Set to zero vector, initial state covariance Let it be a diagonal matrix with diagonal elements , reflecting the uncertainty of the initial position.

7. The high-precision three-dimensional soil mapping method based on radar detection according to claim 1, characterized in that: The depth correlation includes: sampling points of the ground detection radar after estimating the position Perform voxelization and divide the three-dimensional space into three-dimensional grids of 0.1m×0.1m×0.02m, which are recorded as three-dimensional voxel grids. The weighted average of the sampling point coordinates in each three-dimensional voxel grid is taken as the center coordinate of the three-dimensional voxel grid, and the normalized amplitude of the ground detection radar of the sampling point is used as the weight. The center coordinates of are calculated as: ,in Indicates the The three-dimensional voxel grid, weight By sampling point The signal amplitude of the ground detection radar Decision, definition , in summary, substituting the weights into the center coordinate calculation formula, the center coordinates of the three-dimensional voxel grid are finally expressed as: 。 8. The high-precision three-dimensional soil mapping method based on radar detection according to claim 7, characterized in that: The voxelized three-dimensional voxel grid is smoothed using a spherical linear interpolation algorithm, and an isosurface is extracted from the three-dimensional voxel grid using a Marching Cubes algorithm to generate a three-dimensional soil grid model. The amplitude value of the ground detection radar is mapped to RGB pseudo-color and rendered onto the three-dimensional soil grid model to obtain the final three-dimensional soil visualization model.

9. The high-precision three-dimensional soil mapping method based on radar detection according to claim 8, characterized in that: The generating of the thematic map reflecting the spatial distribution of soil physical and chemical properties in the target area specifically includes: Based on the propagation characteristics of the ground detection radar signal in the soil medium, the three-dimensional soil visualization model is divided into soil layers through an image processing algorithm, and the geometric boundaries of different soil layers are automatically identified; Based on the physical relationship between GPR amplitude values ​​and dielectric constant and water content, combined with the soil dielectric model and Topp equation, a soil property parameter inversion model is established; by substituting the GPR amplitude values ​​into the soil property parameter inversion model, the dielectric constant and water content of each three-dimensional voxel grid are estimated; Based on the Archie equation and Maxwell equation describing the relationship between conductivity, dielectric constant and pore structure, the soil volume conductivity is estimated by the wave velocity of the ground detection radar, and the soil porosity is inverted by substituting it into the Archie equation, and other physical parameters are converted; The inverted soil property parameters including dielectric constant, water content, soil bulk conductivity, soil texture, and soil porosity are subjected to Kriging interpolation calculation in the voxelized three-dimensional voxel grid to form regular three-dimensional voxel grid data; The regular three-dimensional voxel grid data is visualized through a GIS platform to generate a spatial distribution map reflecting the soil attribute parameters, namely the thematic map. The thematic map overlays and analyzes monitoring information, including crop yield and growth status, to reveal the spatial differences in soil fertility, identify low-yield fields and salinized areas, and provide a scientific basis for precision agricultural management measures.

Images