An intelligent detection method and system for subgrade filling compaction degree

Abstract

The application relates to the technical field of road engineering quality detection, and discloses a subgrade filling compaction degree intelligent detection method and system, which comprises the following steps: a multispectral remote sensing subsystem is controlled to collect images and generate orthographic images, and initial compaction degrees are calculated after interference masks are removed by vegetation and water body indexes; data heterogeneity and spectral characteristics are analyzed to generate sampling instructions containing suggestion coordinates; measured values fed back by a dynamic sampling subsystem are acquired, residual error vectors are calculated, and residual error correction surfaces covering the whole field are constructed; initial data and residual error surfaces are weighted and superimposed by using a dynamic correction coefficient to generate a final compaction degree distribution map. Through the combination of a multivariate regression model and a ground measured residual error correction mechanism, an adaptive sampling strategy based on spatial heterogeneity is used to guide accurate detection, the problems that a single remote sensing inversion is greatly disturbed by the environment and traditional sampling is insufficient in representativeness are effectively solved, and the comprehensiveness and accuracy of subgrade compaction degree detection are improved.

Description

Technical Field

[0001] This invention relates to the field of road engineering quality testing technology, specifically to an intelligent testing method and system for the compaction degree of roadbed fill. Background Technology

[0002] The compaction degree of subgrade fill is a key indicator for evaluating the quality of road construction, directly affecting the stability, load-bearing capacity, and service life of the completed road. In highway and railway engineering construction, real-time, accurate, and comprehensive testing of subgrade compaction degree is a crucial step in ensuring project quality.

[0003] Traditional methods for testing roadbed compaction primarily rely on physical contact techniques such as sand cone testing, ring cutter testing, or nuclear density metering. While these methods provide relatively accurate point data, they are destructive or semi-destructive tests, involving cumbersome procedures, long processing times, and high labor intensity. More importantly, traditional methods only acquire discrete, single-point data, using a limited number of sample points to infer the compaction quality of the entire work surface. This point-to-surface assessment approach is ill-suited for long-distance, large-area roadbed construction sites, failing to comprehensively reflect the spatial distribution of roadbed compaction and easily overlooking insufficient compaction in unsampled areas.

[0004] With the development of remote sensing technology, using UAVs equipped with spectral sensors for roadbed quality inspection has become a research hotspot. This technology utilizes the correlation between soil spectral reflectance and its physical properties to establish an inversion model for rapid large-area inspection. However, in practical engineering applications, relying solely on spectral remote sensing to invert compaction has limitations. On the one hand, the complex environment of construction sites, fluctuations in soil moisture content, differences in surface roughness, and changes in lighting conditions can all cause nonlinear interference to spectral reflectance, leading to unstable prediction accuracy of general inversion models and making it difficult to directly meet engineering acceptance standards. On the other hand, existing remote sensing inspections and ground measurements are often disconnected. Ground sampling typically uses random or fixed grid sampling methods, lacking utilization of the surface heterogeneity characteristics reflected in remote sensing images, resulting in a lack of targeted sampling and an inability to accurately capture weak areas with questionable quality. How to effectively integrate the large-area coverage advantage of remote sensing technology with the high-precision characteristics of ground inspection, achieving intelligent and accurate sampling and full-field data calibration while eliminating environmental interference, is a problem that current roadbed inspection technology needs to solve. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an intelligent detection method and system for roadbed fill compaction. It solves the problems of traditional physical contact detection methods being unable to achieve large-area full coverage and having low operational efficiency, single-spectral remote sensing inversion technology being easily affected by soil moisture and lighting conditions leading to insufficient detection accuracy, and the lack of data support for on-site sampling points resulting in poor representativeness of detection results and easy omission of key hidden danger areas.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] The first aspect of this invention provides an intelligent detection method for the compaction degree of roadbed fill. This method combines the advantages of large-area coverage of multispectral remote sensing with the precise characteristics of ground physical detection through an air-ground collaborative operation mode, thereby realizing the digital inversion and calibration of the compaction degree of roadbed.

[0008] The method first controls the multispectral remote sensing subsystem to acquire multiband spectral image sequences of the area to be detected, and then processes the multiband spectral image sequences to generate a multispectral digital orthophoto map. The data processing unit performs non-filled area masking and removal processing on the multispectral digital orthophoto map. Specifically, it calculates the normalized difference vegetation index and normalized water index of each pixel in the image, and identifies areas with indices exceeding preset thresholds as either vegetation-covered areas or waterlogged areas and marks them as invalid, retaining only the spectral data of bare fill areas.

[0009] Subsequently, based on the spectral data of the retained effective fill area, the initial compaction distribution data were calculated using a pre-constructed multivariate basic regression model. This multivariate basic regression model characterizes the linear mapping relationship between spectral reflectance and subgrade compaction. The basic model prediction value is obtained by multiplying the spectral reflectance values ​​of each band of each pixel by the corresponding weight coefficients and summing them, and adding a constant intercept term.

[0010] Building upon this foundation, the method introduces a dynamic sampling decision mechanism based on spatial heterogeneity. The system analyzes the spatial texture and spectral characteristics of the initial compaction distribution data to generate sampling instructions. Specifically, a sliding window is used to traverse the initial compaction distribution data, calculating the coefficient of variation of the spectral data within the window, and statistically analyzing the mean and standard deviation of the overall coefficient of variation to determine an adaptive variation threshold. Regions with coefficients of variation higher than the adaptive variation threshold are identified as highly heterogeneous regions, and their connected geometric centers are extracted as the first type of sampling points. Simultaneously, cluster analysis is performed on the stable regions after removing highly heterogeneous regions, dividing them into multiple sub-regions with similar spectral characteristics. The geometric center of each sub-region is selected as the second type of sampling point. The geographic coordinates of these two types of sampling points are merged and written into the sampling instructions.

[0011] After receiving the sampling command, the dynamic sampling subsystem collects measured compaction values ​​at the specified geographic coordinates and feeds them back to the data processing unit via a wireless communication link. The data processing unit calculates the difference between the measured compaction values ​​and the initial compaction distribution data at the corresponding locations, generating a discrete residual vector. To achieve full-field calibration, the system uses a spatial interpolation algorithm to map the discrete residual vector into a residual correction surface covering the area to be detected. During the interpolation process, for unsampled pixels, the Euclidean distance between them and each retained sampling point is calculated, and an inverse distance weighted algorithm is used to perform a weighted average of the residual vector with the reciprocal of the power of the Euclidean distance as the weight, thereby estimating the residual value of the unsampled points.

[0012] Finally, the method performs a dual calibration calculation based on dynamic correction coefficients. Dynamic correction coefficients are constructed as variables in the distance function, decreasing as the distance between a pixel and its nearest sampled point increases. The base model predictions for unsampled pixels are added to the weighted estimated residuals using the dynamic correction coefficients to generate the final compaction degree prediction. After performing the above calculations across all pixels, the final compaction degree distribution map of the area to be inspected is obtained. Furthermore, the method sets acceptable thresholds according to engineering construction specifications, performs graded rendering of the final compaction degree distribution map, identifies unacceptable areas, extracts their vector contours, areas, and center coordinates, and outputs an inspection report.

[0013] A second aspect of the present invention provides an intelligent detection system for the compaction degree of roadbed fill, which is used to perform the above-described intelligent detection method for the compaction degree of roadbed fill.

[0014] The system mainly comprises a multispectral remote sensing subsystem and a dynamic sampling subsystem connected via a wireless communication link. The multispectral remote sensing subsystem includes an unmanned aerial vehicle (UAV) platform, a multispectral sensor, and a data processing unit. The UAV platform, acting as the flight carrier, carries the multispectral sensor and performs flight operations above the area to be detected. The multispectral sensor is responsible for acquiring multi-band spectral image sequences containing geographic information. The data processing unit, configured as the core computing hub of the system, is responsible for receiving spectral data and generating digital orthophoto maps, performing mask removal of non-filled areas and calculating initial compaction degree, and generating sampling instructions containing suggested coordinates based on the spatial heterogeneity of the data. Simultaneously, the data processing unit is also responsible for receiving ground feedback data, calculating residual vectors, constructing residual correction surfaces, and using dynamic correction coefficients to generate the final compaction degree distribution map.

[0015] The dynamic sampling subsystem comprises ground sampling equipment, a positioning module, and a wireless communication module. The ground sampling equipment responds to sampling commands from the data processing unit, acquiring measured soil compaction values ​​at designated locations. The positioning module simultaneously acquires the precise geographic coordinates of the current sampling point. The wireless communication module packages the measured values ​​and coordinate data and feeds them back to the multispectral remote sensing subsystem, forming a closed-loop data transmission between the air and ground. The system also features path planning capabilities, capable of determining the optimal operational path based on the sampling point coordinates and current location, guiding the ground equipment to efficiently complete the sampling task.

[0016] In summary, the present invention has at least one of the following beneficial technical effects:

[0017] 1. This invention constructs a multivariate basic regression model combined with a residual correction mechanism based on ground-measured data. By using dynamic correction coefficients to weight and fuse trend information from large-area remote sensing inversion with calibration information from local precision measurements, it effectively solves the accuracy fluctuation problem caused by interference from ambient light and soil moisture in single-spectral remote sensing. At the same time, it makes up for the lack of spatial resolution in traditional point sampling inspections, so that the final generated compaction distribution map has both macroscopic coverage and microscopic accuracy, improving the overall reliability of roadbed compaction detection results.

[0018] 2. This invention employs an adaptive sampling strategy based on spatial heterogeneity and spectral distribution characteristics. It uses the coefficient of variation to identify highly heterogeneous regions and extract their centers. Simultaneously, it performs cluster analysis on stable regions to select representative sample points. This mechanism can automatically guide the sampling equipment to prioritize the detection of key locations with drastic fluctuations in compaction quality or representative characteristics. It avoids the blindness and potential missed detections caused by traditional uniform grid methods or random sampling. While ensuring that the test samples are sufficiently representative, it optimizes the sampling path and reduces the workload and time cost of on-site operations.

[0019] 3. This invention introduces a masking and removal process for non-filled areas based on normalized difference vegetation index and normalized water index. This process can automatically identify and filter invalid interference sources such as vegetation-covered areas and waterlogged areas in the early stages of data processing. By modeling and analyzing only the effective filled areas and combining the closed-loop interaction of air and ground data between UAVs and ground equipment, the entire process from environmental perception and invalid area removal to final mapping is automated, effectively improving the anti-interference capability and the intelligence level of detection operations in complex construction environments. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the intelligent detection system for roadbed fill compaction degree of the present invention;

[0021] Figure 2 This is a flowchart illustrating the intelligent detection method for roadbed fill compaction degree according to the present invention.

[0022] Figure 3 This is a schematic diagram illustrating the principle of the multispectral data acquisition and preprocessing workflow of the present invention;

[0023] Figure 4 This is a schematic diagram of the dynamic sampling decision logic based on spectral heterogeneity of the present invention;

[0024] Figure 5 This is a schematic diagram illustrating the calculation principle of the dual calibration model based on residual correction of the present invention.

[0025] Among them, 10. Multispectral remote sensing subsystem; 11. UAV platform; 12. Multispectral sensor; 13. Data processing unit; 20. Dynamic sampling subsystem; 21. Ground sampling equipment; 22. Positioning module; 23. Wireless communication module. Detailed Implementation

[0026] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] See attached document Figure 1 This invention provides an intelligent detection system for the compaction degree of roadbed fill, which includes a multispectral remote sensing subsystem 10 and a dynamic sampling subsystem 20. The multispectral remote sensing subsystem 10 and the dynamic sampling subsystem 20 interact with each other via a wireless communication link, and work together to complete the inversion and calibration of the compaction degree of roadbed fill in the area to be detected.

[0028] The multispectral remote sensing subsystem 10 includes an unmanned aerial vehicle (UAV) platform 11, a multispectral sensor 12, and a data processing unit 13. The UAV platform 11, acting as an aerial flight carrier, is equipped with a flight control module. This module controls the UAV platform 11 to perform flight operations above the area to be detected based on preset waypoint coordinates. The UAV platform 11 can be a multi-rotor aircraft or a fixed-wing aircraft, possessing low-altitude flight capabilities in complex terrain environments.

[0029] A multispectral sensor 12 is mounted on an unmanned aerial vehicle (UAV) platform 11 to collect spectral reflectance signals from the surface of the area to be detected. The multispectral sensor 12 includes multiple independent optical channels, each equipped with a filter for a specific wavelength. In this embodiment, the multispectral sensor 12 is configured to collect spectral image data including, but not limited to, blue, green, red, red-edge, and near-infrared bands. The multispectral sensor 12 has a communication interface for outputting the collected raw spectral image data.

[0030] The data processing unit 13 is communicatively connected to the multispectral sensor 12 to receive raw spectral image data and perform data processing, sampling decisions, and compaction degree inversion calculations. The data processing unit 13 can be an onboard computing board integrated on the UAV platform 11, or a high-performance computer workstation or cloud server located on the ground. When the data processing unit 13 is located on the ground, the UAV platform 11 transmits data back to the ground in real time via a data transmission link.

[0031] The dynamic sampling subsystem 20 includes a ground sampling device 21, a positioning module 22, and a wireless communication module 23. The ground sampling device 21, upon receiving a sampling command, performs physical testing on the subgrade fill at a specific location within the testing area to obtain the measured compaction degree value at that location. The ground sampling device 21 can be an automatic penetration tester, a lightweight dynamic penetrometer, or a nuclear density meter. The ground sampling device 21 has a data output interface for outputting the measured compaction degree value.

[0032] The positioning module 22 is mounted on the ground sampling device 21 and is used to acquire the geospatial coordinates of the current ground sampling point in real time. The geospatial coordinates include longitude, latitude, and elevation information. The positioning module 22 uses a Global Navigation Satellite System receiver. In this embodiment, the positioning module 22 supports real-time dynamic differential technology to ensure that the accuracy of the coordinate data matches the spatial resolution of the multispectral remote sensing subsystem 10. Furthermore, the data processing unit 13, the multispectral sensor 12, and the ground sampling device 21 are all equipped with a unified clock synchronization module based on Network Time Protocol (NAT) or GPS timing. Each frame of image acquired by the multispectral sensor 12 and each sampling data packet uploaded by the ground sampling device 21 has a unified timestamp with microsecond-level precision written in the data header file to facilitate subsequent data fusion processing.

[0033] The wireless communication module 23 is connected to both the ground sampling device 21 and the positioning module 22. It packages the measured compaction degree values ​​and their corresponding geospatial coordinates into sampling feedback data and sends this data to the data processing unit 13. The data processing unit 13 then performs real-time corrections to the compaction degree detection model based on the received sampling feedback data.

[0034] The data processing unit 13 is internally equipped with a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement a dual calibration logic based on spectral regression and residual correction. Specifically, the data processing unit 13 calculates the final predicted compaction degree value at any target location within the detection area based on the following formula:

[0035] ;

[0036] in: This represents the predicted compaction degree of the roadbed fill calculated at the target location; This indicates the total number of spectral bands acquired by the multispectral sensor 12; Indicates the first The weight coefficients of each spectral band in the basic regression model represent the degree of response of the spectral reflectance of that band to the degree of compaction. Indicates the target position is at the 1st position. Spectral reflectance values ​​for each spectral band; This represents the constant intercept term of the basic regression model; This represents the dynamic correction coefficient, which ranges from 0 to 1 and is used to adjust the correction strength of the ground-measured data to the overall model. This represents the residual vector at the sampling point, which is obtained by subtracting the predicted value of the basic regression model at that location from the measured compaction value obtained by the ground sampling device 21. This represents a spatial interpolation function used to map the residual vector at discrete sampling points into an error correction surface covering the entire field. The spatial interpolation function employs either the inverse distance weighted algorithm or the kriging interpolation algorithm.

[0037] The data processing unit 13 is also configured to generate sampling instructions based on the spatial heterogeneity characteristics of the spectral image data. The sampling instructions include the geographic coordinates of suggested sampling points and are sent to the dynamic sampling subsystem 20 via a wireless network to guide operators or automated equipment to the designated location for operation.

[0038] The intelligent detection system for roadbed fill compaction also includes a power supply module, which provides power to the electrical components in the multispectral remote sensing subsystem 10 and the dynamic sampling subsystem 20. The UAV platform 11 and the ground sampling equipment 21 are each equipped with independent battery packs.

[0039] See attached document Figure 2 This invention provides an intelligent detection method for the compaction degree of roadbed fill. This method is based on the aforementioned intelligent detection system for the compaction degree of roadbed fill and mainly includes the following steps:

[0040] Initialization and sensor radiometric calibration of the intelligent detection system for roadbed fill compaction degree.

[0041] Before performing flight operations, the multispectral remote sensing subsystem 10 is in standby mode. The operator places a standard gray card with known reflectance within the field of view of the multispectral sensor 12 and controls the multispectral sensor 12 to capture multiple sets of images of the standard gray card. The data processing unit 13 calculates the radiometric calibration coefficients based on the known reflectance data of the standard gray card and the digital quantization values ​​(DN values) in the images.

[0042] The radiometric calibration coefficient is used to eliminate the interference of ambient light intensity variations and the sensor's own dark current on the spectral data, ensuring the physical consistency of the subsequently acquired spectral reflectance data. While acquiring standard grayscale images, the multispectral sensor 12 acquires multiple frames of dark current noise images with the shutter closed. The data processing unit 13 calculates the average value matrix of the dark current noise images and subtracts this average value matrix from each frame of observed images before converting the original DN values ​​to radiance values, in order to eliminate thermal noise interference from the sensor's photosensitive element.

[0043] Spectral data acquisition and orthophoto construction.

[0044] The drone platform 11 flies above the area to be detected according to preset flight path planning parameters. The flight path planning parameters include flight altitude, forward overlap rate, and lateral overlap rate. In this embodiment, the forward overlap rate is set to be no less than 70%, and the lateral overlap rate is set to be no less than 60% to ensure the geometric accuracy of image stitching.

[0045] The multispectral sensor 12 continuously triggers exposures at preset time intervals or distance intervals to acquire a sequence of multi-band spectral images containing geographic location information. The data processing unit 13 receives the image sequence and uses a motion reconstruction algorithm or a preset stitching algorithm to generate a multispectral digital orthophoto map covering the entire area to be detected from multiple independent spectral images.

[0046] Masking and initial screening of non-filled areas.

[0047] The data processing unit 13 performs pixel-by-pixel classification processing on the generated multispectral digital orthophoto image. The data processing unit 13 calculates the normalized vegetation index or a specific soil index for each pixel in the image and compares it with a preset threshold. For non-effective fill areas identified as vegetation, large construction machinery, temporary buildings, or waterlogged ponds, the data processing unit 13 generates a binarized mask to block them, retaining only the spectral data of the exposed fill areas for subsequent calculations. Subsequently, the data processing unit 13 uses a basic regression model to calculate the initial compaction distribution data of the entire fill area.

[0048] Dynamic sampling decision based on spatial heterogeneity.

[0049] Data processing unit 13 performs spatial texture analysis on the initial compaction distribution data to determine the specific location of ground sampling. This step specifically includes: data processing unit 13 uses a sliding window to traverse the entire field data, calculates the standard deviation of the spectral data within the window, marks areas with standard deviations higher than a preset variation threshold as high heterogeneity areas, and generates the coordinates of a first-class sampling point at the geometric center of the area;

[0050] Simultaneously, the data processing unit 13 performs K-Means clustering analysis on the full-field spectral data, dividing the fill area into several sub-regions with similar spectral characteristics, and selecting the geometric center of each sub-region as the coordinates of the second type of sampling point. The data processing unit 13 merges the coordinates of the first type of sampling point and the coordinates of the second type of sampling point to generate a sampling command, which is then sent to the dynamic sampling subsystem 20 via a wireless communication link.

[0051] Ground data acquisition and real-time feedback.

[0052] The dynamic sampling subsystem 20 receives a sampling command, and the ground sampling device 21 moves to the geographic coordinates specified in the command. The ground sampling device 21 performs a physical detection operation to obtain the true value of soil compaction at that location. The positioning module 22 simultaneously records the precise three-dimensional coordinates during sampling. The wireless communication module 23 correlates the true compaction value with the three-dimensional coordinates and transmits it back to the data processing unit 13. This step achieves spatiotemporal alignment between macroscopic remote sensing data and microscopic physical property data.

[0053] Model residual calculation and dual calibration.

[0054] Data processing unit 13 extracts the predicted values ​​from the basic regression model at the coordinates of the sampling points, calculates the difference between these predicted values ​​and the true compaction values ​​transmitted from the ground, and obtains the residual vector. Data processing unit 13 then uses a spatial interpolation algorithm to expand the discrete residual vector into a residual correction surface covering the entire field. Subsequently, data processing unit 13 weights and superimposes the initial compaction distribution data with the residual correction surface to generate a calibrated final compaction distribution map.

[0055] Results output and engineering early warning.

[0056] The data processing unit 13 stores the final compaction distribution map in the form of raster data and sets a qualified threshold according to the engineering design standards. For areas with compaction values ​​lower than the qualified threshold, the system marks them with color on the distribution map and outputs a test report containing the area, location coordinates, and average compaction value of the unqualified areas.

[0057] See attached document Figure 3 This embodiment details the technical process from UAV flight operations to the generation of pure spectral reflectance data, which is controlled and executed by the data processing unit 13.

[0058] Multispectral sensor 12 band response setting and data acquisition. During the data acquisition process, the multispectral sensor 12 does not acquire the full-band spectrum, but rather uses a specific narrow-band filter configured for the physical characteristics of the roadbed fill.

[0059] Specifically, the multispectral sensor 12 is configured to respond to the following center wavelengths:

[0060] The multispectral sensor 12 simultaneously records attitude angle data (pitch, roll, yaw) and GPS timestamps provided by the inertial measurement unit at the moment of exposure, and writes this metadata into the image file header for subsequent geometric correction. The wavelengths are 450 nm (blue light band, corresponding to soil mineral absorption characteristics), 560 nm (green light band), 650 nm (red light band, corresponding to soil color and iron oxide content), 730 nm (red edge band, used to distinguish sparse vegetation from soil boundaries), and 840 nm (near-infrared band, corresponding to soil moisture content characteristics).

[0061] Radiometric correction based on the bidirectional reflectance distribution function. Since the surface of the roadbed fill is not a Lambertian surface, variations in the solar incidence angle and observation angle can lead to differences in the spectral brightness of the same material. After converting the image DN values ​​into radiance values, the data processing unit 13 introduces an anisotropic correction algorithm. Specifically, based on the acquired solar altitude angle, azimuth angle, and observation geometry angle during UAV shooting, and combined with preset bidirectional reflectance distribution function model parameters for the roadbed fill, the data processing unit 13 compensates for uneven illumination in the image edge areas, eliminating hotspot effects and vignetting, ensuring consistent color and brightness in the stitched image.

[0062] High-precision geometric registration and band alignment. Since the multispectral sensor 12 typically employs a multi-lens array, the optical centers of lenses in different bands have physical positional deviations, resulting in spatial non-overlapping of the original images. The data processing unit 13 reads the internal parameters of the multispectral sensor 12 (including focal length, principal point coordinates, and lens distortion coefficients) and the relative position transformation matrix between lenses. Using a specific band (e.g., the green band) as a reference image, the data processing unit 13 calculates the correspondence matrix of other band images relative to the reference image using a feature point matching algorithm (such as SIFT or SURF), performing sub-pixel-level band registration to ensure that the same pixel position in all bands corresponds to the same ground feature point.

[0063] Multispectral orthophoto mosaicking. Data processing unit 13 performs aerial triangulation on the registered single-frame multispectral images, calculates the precise exterior orientation elements of each image in the survey area coordinate system, generates sparse and dense point clouds, and then constructs a digital surface model. Data processing unit 13 uses this digital surface model to orthorectify the single-frame images, eliminating projection distortion caused by terrain undulations. Subsequently, data processing unit 13 performs image mosaicking, automatically finding the optimal stitching line in overlapping areas, and smoothing the pixel values ​​at the stitching gaps using a feathering fusion algorithm, finally outputting a multi-band digital orthophoto map with geographic reference coordinates.

[0064] Spectral threshold removal of non-fill interference. To improve the purity of compaction degree inversion, data processing unit 13 needs to remove interference on the roadbed surface. Data processing unit 13 calculates the normalized difference vegetation index and normalized water index of the image. Pixels with a normalized difference vegetation index value greater than 0.2 are identified as vegetation-covered areas; pixels with a normalized water index value greater than 0 are identified as waterlogged areas; areas with reflectance higher than 0.4 across all bands and extremely small variance are identified as concrete or light-colored stones.

[0065] The data processing unit 13 generates a binary mask matrix, marking the identified interfering pixel locations as 0 (invalid) and the pixel locations that match the spectral characteristics of bare soil as 1 (valid). Subsequent compaction inversion calculations are performed only on the regions marked as 1 in the mask matrix, thereby avoiding contamination of the calibration model parameters by non-fill objects.

[0066] See attached document Figure 4 This embodiment details how the multispectral remote sensing subsystem 10 intelligently generates ground sampling instructions based on the spatial distribution characteristics of spectral data, achieving air-ground collaborative operation. This process is executed by the data processing unit 13, aiming to solve the technical problem that traditional fixed grid sampling cannot balance efficiency and representativeness.

[0067] Construction of soil compaction sensitivity index. Before making sampling decisions, data processing unit 13 first performs feature transformation on the preprocessed multispectral image to construct a spectral feature map sensitive to compaction changes. Data processing unit 13 calculates the ratio index of red band to near-infrared band and the texture feature component of red band.

[0068] Specifically, data processing unit 13 uses the gray-level co-occurrence matrix algorithm to extract four second-order statistics from the red band image: energy, entropy, contrast, and correlation. These features reflect the particle roughness and moisture content distribution of the soil surface, thus indirectly characterizing the uniformity of compaction. Data processing unit 13 normalizes the above spectral indices and texture features, and weights them to synthesize a single-channel compaction response feature map.

[0069] Local spatial variability calculation and high-risk area identification. Data processing unit 13 uses a sliding window algorithm to traverse the compaction response feature map to quantify the data dispersion of local areas. The size of the sliding window is set to a pixel matrix corresponding to the actual ground size of 3 meters × 3 meters. For each window position, data processing unit 13 calculates the coefficient of variation (i.e., standard deviation divided by mean) of the pixel values ​​within the window.

[0070] The data processing unit 13 sets a threshold for the coefficient of variation (e.g., 0.15). This threshold is not a fixed value, but is adaptively calculated by the data processing unit 13 based on the histogram statistical distribution of the coefficient of variation across the entire field. Specifically, the data processing unit 13 calculates the coefficient of variation for all sliding windows across the entire field and then calculates its mean. and standard deviation Set the threshold to (in It is a preset adjustable scaling factor, usually with a value of 1.5 to 2.0, so as to dynamically adapt to the natural background differences of different fill materials.

[0071] When the coefficient of variation within a certain window exceeds the threshold, it indicates that the soil properties in that area fluctuate drastically, suggesting either insufficient or excessive compaction. This represents a high-uncertainty region that is difficult to predict with a single model. Data processing unit 13 extracts the geometric centroids of the connected components in this high-variability region and marks them as anomaly verification sampling points. The purpose of these sampling points is to capture extreme values ​​and prevent the model from failing locally.

[0072] Representative regions were selected based on spectral clustering. To correct for overall model bias across the entire field, data processing unit 13 performed K-Means clustering analysis on the remaining stable regions after removing highly variable areas. Data processing unit 13 mapped the multi-band spectral vector of each pixel to a high-dimensional feature space and divided all valid pixels into K categories based on Euclidean distance (the K value was dynamically set according to the area of ​​the survey region, for example, one category per 1000 square meters). After the clustering process iteratively converged, data processing unit 13 obtained K soil categories with spectral differences (e.g., high-moisture clay, dry sand, graded gravel). Data processing unit 13 calculated the spatial coordinates of the pixel closest to the cluster center in each category and marked it as a calibration baseline sampling point. The purpose of these sampling points is to obtain the average compaction level of different soil types for calibrating the global parameters of the basic regression model.

[0073] Path planning and sampling instruction generation. The data processing unit 13 merges the generated set of coordinates of abnormal verification sampling points with the set of coordinates of calibration benchmark sampling points to form the final set of ground sampling target points. Subsequently, based on the current GPS position of the ground sampling device 21, the data processing unit 13 uses an ant colony algorithm or a genetic algorithm to solve for the shortest path to traverse all target point sets, generating the optimal sampling route. The data processing unit 13 packages the target point's latitude and longitude coordinates, the recommended path order, and the sampling type label (verification point or benchmark point) corresponding to each point into a structured sampling instruction data packet, and sends it to the terminal display screen of the dynamic sampling subsystem 20 via a wireless communication link to guide ground personnel or equipment to perform precise operations according to the map.

[0074] See attached document Figure 5 This embodiment details the algorithm logic of how the data processing unit 13 uses ground-measured data to perform in-depth correction on the remote sensing inversion results, which is the core step in achieving high-precision detection. The multivariate basic regression model consists of two parts: a basic linear regression module and a nonlinear spatial residual compensation module.

[0075] Training and application of a multivariate basic regression model. Data processing unit 13 first establishes the basic mapping relationship between spectral reflectance and subgrade compaction. The basic mapping relationship is expressed using a multiple linear regression model, with the formula as follows: .in, The degree of compaction predicted by the basic model; Indicates the first The weight coefficients of each spectral band in the basic regression model represent the degree of response of the spectral reflectance of that band to the degree of compaction. Indicates the target position is at the 1st position. Spectral reflectance values ​​for each spectral band; This represents the constant intercept term of the basic regression model.

[0076] In the initial stage of the project, the coefficient and The data can be obtained through pre-training based on the historical sample database of the region, or by using the data from the calibration benchmark sampling points in this dynamic sampling for least squares fitting and updating. Considering the multicollinearity among multispectral bands, partial least squares regression or ridge regression algorithms are preferred for the fitting and updating process.

[0077] Data processing unit 13 determines the optimal number of principal components or regularization parameters through cross-validation to improve the generalization ability of the basic model to unsampled areas. The role of the basic model is to extract the macroscopic trend of the overall compaction distribution and reflect the general law between soil compaction and spectral reflectance (e.g., the higher the compaction, the smaller the soil porosity, the smoother the surface, and the higher the reflectance in a specific band).

[0078] Calculation and extraction of residuals from the basic model. This involves obtaining the residuals of the entire field. After distribution, the data processing unit 13 introduces ground dynamic sampling data for error verification. For each ground sampling point... The system reads the true value of its measured ground compaction. And extract the base model prediction value corresponding to the coordinates of that point. The data processing unit 13 calculates the difference between the two values ​​to obtain the residual value at that point. The residual value Physically, this represents the spectral inversion error caused by factors such as differences in soil microstructure, local anomalies in moisture content, or surface shading.

[0079] if A positive value indicates that the basic model underestimated the compaction at that location; conversely, a negative value indicates an overestimation. Before performing interpolation calculations, the data processing unit 13 will process the residual set... The Grubbs criterion or the Laida criterion is applied. If the residual value of a sampling point is determined to be a statistical outlier, the data processing unit 13 will automatically remove the abnormal residual, exclude it from the spatial interpolation calculation, and generate a retest instruction to send back to the dynamic sampling subsystem 20, prompting the operator to resample at that location.

[0080] Full-field residual surface reconstruction based on spatial interpolation. Since ground sampling points are discrete and finite (e.g., only 10-20 points), while remote sensing images are continuous pixel matrices, data processing unit 13 needs to calculate residual values ​​for unknown points to correct non-sampling areas. This embodiment uses an inverse distance-weighted interpolation algorithm to construct a residual correction surface covering the entire field. For any unsampled pixel within the detection area. The estimated residual value From the surrounding The weighted average of the residual values ​​from known sampling points is used to obtain the result, with the weights proportional to the distance. The power is inversely proportional to the exponent. The specific calculation logic is as follows:

[0081] ;

[0082] in: For pixels With the Euclidean distance between sampling points; The exponent (usually 2) is used to control the local influence range. Through this step, the data processing unit 13 generates an error distribution map with the same size as the original image, which quantifies the potential bias of the base model at various locations across the entire field.

[0083] Adaptive weighted fusion and final compaction degree calculation. Finally, data processing unit 13 fuses the results of the basic regression model with the residual correction surface to calculate the final compaction degree. To avoid introducing erroneous correction values ​​in areas too far from the sampling point, this embodiment introduces a dynamic correction coefficient. This coefficient is not a fixed constant, but rather designed as a function of distance: the closer to the nearest sampling point, the higher the coefficient. The closer to 1 (full correction); the farther away, Gradually decay to 0 (without correction, relying only on the base model).

[0084] Data processing unit 13 performs pixel-level addition operations: Through this dual calibration mechanism, the multispectral remote sensing subsystem 10 retains the consistency advantage of multispectral remote sensing in large-area detection, while effectively correcting nonlinear deviations in local areas using dynamic sampling data, thereby improving the overall detection accuracy.

[0085] This embodiment details how the data after dual calibration is transformed into visualized engineering management information and generates a guiding inspection report. This process is completed by the data processing unit 13, marking the end of a single inspection process.

[0086] Hierarchical rendering and visualization of compaction raster data. Data processing unit 13 obtains the final calculated compaction prediction value. Each element in the predicted value corresponds to an actual spatial unit on the ground (e.g., 0.1 m × 0.1 m). To visually represent the distribution of subgrade quality, the data processing unit 13 establishes a color-mapped lookup table based on the compaction standard values ​​(e.g., 93%, 94%, 96%) set by engineering construction specifications (e.g., the "Technical Specification for Highway Subgrade Construction").

[0087] Data processing unit 13 maps the values ​​in the compaction matrix to RGB color values: areas with compaction below the standard lower limit (e.g., <93%) are rendered in red, indicating serious non-compliance; areas within the critical range (e.g., 93%-94%) are rendered in yellow, indicating warning zones; and areas above the standard value (e.g., >96%) are rendered in green, indicating acceptable zones. The system overlays the generated pseudo-color layer onto the original orthophoto or digital line map to form a thematic distribution map of roadbed compaction with geographic coordinate reference. This distribution map supports zooming and panning operations, allowing engineers to intuitively view the compaction status at any location.

[0088] Automatic extraction and vectorization of non-conforming areas. To guide subsequent compaction operations, data processing unit 13 performs connected component analysis on the red-rendered area (non-conforming area). The system automatically identifies and extracts all continuous non-conforming polygons, uses edge detection algorithms to trace their contour boundaries, and converts them into vector polygon data (such as Shapefile or GeoJSON format). For each non-conforming vector polygon, data processing unit 13 calculates its geometric attributes, including the polygon area, perimeter, and center point coordinates, and calculates the minimum and average compaction values ​​within the area. This attribute information is written into the attribute table of the vector data as a digital basis for engineering rectification.

[0089] Intelligent generation of multi-dimensional inspection reports. Data processing unit 13 summarizes all data from this inspection and automatically generates a standardized inspection report file (PDF or HTML format). The report content includes, but is not limited to: basic project information (project name, chainage range, inspection time), inspection summary (average compaction degree, pass rate statistics, standard deviation), list of unqualified areas (number, location, area, recommended number of recompaction passes), sampling point data comparison table (error analysis between remote sensing inversion values ​​and ground measured values), and screenshots of the above-mentioned compaction degree thematic distribution map. After the report is generated, the system pushes it to the project management platform or the handheld terminal of the supervisor via network interface.

[0090] Spatiotemporal archiving and trend analysis of historical data. Data processing unit 13 stores the final compaction matrix, vectorized non-compliant patches, and raw spectral data from this test into a spatiotemporal database. The database is indexed using location and time as dual primary keys. When multiple layers of filling or multiple tests are conducted on the same subgrade section, the intelligent subgrade fill compaction detection system can call historical data for comparative analysis and generate compaction change trend curves. This allows engineering managers to trace the evolution of compaction quality at a specific location during different construction stages. For example, it can determine whether the compaction measures are effective or whether there is a compaction attenuation due to moisture evaporation, thereby achieving full lifecycle digital management of subgrade construction quality.

Claims

1. A method for intelligent detection of the compaction degree of roadbed fill, characterized in that, Includes the following steps: The multispectral remote sensing subsystem is controlled to acquire multiband spectral image sequences of the area to be detected, and the multiband spectral image sequences are processed to generate a multispectral digital orthophoto map. The multispectral digital orthophoto image is masked to remove non-filled areas, and the initial compaction distribution data is calculated based on the retained spectral data of the filled areas. A sampling instruction is generated based on the spatial heterogeneity and spectral characteristics of the initial compaction distribution data, and the sampling instruction includes the geographic coordinates of the suggested sampling points; The dynamic sampling subsystem acquires the measured compaction value at the geographic coordinates according to the sampling command. Calculate the difference between the measured compaction value and the initial compaction distribution data at the corresponding location to generate a residual vector; The residual vector is mapped to a residual correction surface covering the region to be detected using a spatial interpolation algorithm; The initial compaction distribution data and the residual correction surface are weighted and superimposed according to the dynamic correction coefficient to obtain the final compaction distribution map of the area to be detected.

2. The intelligent detection method for roadbed fill compaction degree according to claim 1, characterized in that, The steps for mask removal in non-filled areas include: Calculate the normalized difference vegetation index and normalized water index for each pixel in the multispectral digital orthophoto image. The area where the normalized difference vegetation index is greater than the first preset threshold is determined to be a vegetation-covered area, and the area where the normalized water index is greater than the second preset threshold is determined to be a waterlogged area. Generate a binary mask matrix, mark the vegetation-covered area and the waterlogged area as invalid areas, and mark the remaining area as valid fill areas; The subsequent initial compaction distribution data calculation and sampling command generation are performed only for the effective fill area.

3. The intelligent detection method for roadbed fill compaction degree according to claim 1, characterized in that, The steps for generating sampling instructions include: The coefficient of variation of the spectral data within the sliding window is calculated by traversing the initial compaction distribution data using a sliding window. The mean and standard deviation of the coefficient of variation of all sliding windows in the entire field are statistically analyzed, and an adaptive variation threshold is calculated based on the mean and standard deviation. Regions with a coefficient of variation higher than the adaptive variation threshold are marked as high heterogeneity regions, and the geometric center of the connected domain of the high heterogeneity region is extracted as the coordinates of the first type of sampling point. Write the coordinates of the first type of sampling points into the sampling instruction.

4. The intelligent detection method for roadbed fill compaction degree according to claim 3, characterized in that, The step of generating sampling instructions further includes: After removing the highly heterogeneous regions, cluster analysis is performed on the spectral data of the remaining stable regions. The stable region is divided into multiple sub-regions with similar spectral characteristics, and the geometric center of each sub-region is selected as the coordinates of the second type of sampling point; The coordinates of the first type of sampling point and the coordinates of the second type of sampling point are combined and written into the sampling instruction as the geographical coordinates of the suggested sampling point.

5. The intelligent detection method for roadbed fill compaction degree according to claim 1, characterized in that, The step of calculating the initial compaction distribution data based on the retained spectral data of the fill area includes: A multivariate basic regression model is constructed, which represents the linear mapping relationship between spectral reflectance and subgrade compaction degree; Multiply the spectral reflectance values ​​of each band of each pixel in the fill area by the corresponding weighting coefficients and sum them up. Add the constant intercept term to obtain the basic model prediction value of the pixel position. The initial compaction distribution data is constructed by combining the base model predictions for all pixel locations.

6. The intelligent detection method for roadbed fill compaction degree according to claim 5, characterized in that, The step of mapping the residual vector to a residual correction surface covering the region to be detected using a spatial interpolation algorithm includes: Extract the base model prediction values ​​at the geographic coordinates of the suggested sampling points; Subtracting the corresponding predicted value from the basic model from the measured compaction value yields the discrete residual vector. Perform statistical outlier detection on the residual vector to remove abnormal residuals; For unsampled pixels within the region to be detected, calculate the Euclidean distance between the unsampled pixel and the sampling points corresponding to all retained residual vectors; Using the inverse distance weighting algorithm, the weights are calculated by weighting all the retained residual vectors with the reciprocal of the power of the Euclidean distance as the weights, and the estimated residual values ​​of the unsampled pixels are obtained.

7. The intelligent detection method for roadbed fill compaction degree according to claim 6, characterized in that, The step of weighted superposition of the initial compaction distribution data and the residual correction surface according to the dynamic correction coefficient includes: The dynamic correction coefficient is constructed and set as a variable of the distance function, such that the dynamic correction coefficient decreases as the distance between the pixel and the nearest sampling point increases; The final compaction prediction value of the unsampled pixel is calculated by adding the basic model prediction value of the unsampled pixel to the estimated residual value after weighting by the dynamic correction coefficient. The above addition operation is performed on all pixels in the entire field to generate the final compaction distribution map.

8. The intelligent detection method for roadbed fill compaction degree according to claim 4, characterized in that, The method further includes: Obtain the current position coordinates of the dynamic sampling subsystem; Based on the current location coordinates and the geographic coordinates of the suggested sampling points, a path planning algorithm is used to find the shortest path to traverse all sampling points. The data packet containing the geographic coordinates of the suggested sampling points and the recommended path order is sent to the dynamic sampling subsystem via a wireless network.

9. The intelligent detection method for roadbed fill compaction degree according to claim 1, characterized in that, The method further includes: The acceptable compaction threshold is set according to the engineering construction specifications; The final compaction degree distribution map is rendered using raster data hierarchical rendering, and areas with values ​​lower than the compaction degree qualification threshold are marked as unqualified areas. Connectivity analysis and edge detection are performed on the non-conforming areas to extract the vector contours of the non-conforming patches; Calculate the area and center point coordinates of the defective patches, and output an inspection report.

10. An intelligent detection system for the compaction degree of roadbed fill, characterized in that, The method for performing the intelligent detection method for compaction degree of subgrade fill as described in any one of claims 1 to 9 includes a multispectral remote sensing subsystem, including an unmanned aerial vehicle platform, a multispectral sensor, and a data processing unit; The unmanned aerial vehicle platform is used to carry the multispectral sensor and fly above the area to be detected. The multispectral sensor is used to acquire multi-band spectral image sequences. The data processing unit is configured as follows: Receive the multi-band spectral image sequence and generate a multispectral digital orthophoto map; The multispectral digital orthophoto image is subjected to masking and removal of non-filled areas, and the initial compaction distribution data is calculated. Based on the initial compaction distribution data, a sampling instruction containing the geographic coordinates of the suggested sampling points is generated; The residual vector is calculated based on the measured compaction values ​​fed back by the dynamic sampling subsystem; Residual correction surfaces are constructed using spatial interpolation algorithms; And generate the final compaction distribution map of the area to be tested based on the dynamic correction coefficient; The dynamic sampling subsystem includes ground sampling equipment, a positioning module, and a wireless communication module; The ground sampling device is used to collect measured compaction values ​​according to the sampling instructions sent by the data processing unit; the positioning module is used to obtain the geographic coordinates of the sampling location; and the wireless communication module is used to feed back the measured compaction values ​​and the geographic coordinates to the data processing unit. The multispectral remote sensing subsystem is connected to the dynamic sampling subsystem via a wireless communication link.

Images