Multispectral vegetation root system identification system and method based on differentiable physical engine

Abstract

The invention relates to the technical field of vegetation root system nondestructive testing, in particular to a multispectral vegetation root system identification system and method based on a differentiable physical engine, and the system comprises a multispectral LiDAR module which is used for transmitting near-infrared laser and short-wave infrared laser, scanning a target area, obtaining dual-wavelength reflectivity data, and determining the vegetation root system according to the dual-wavelength reflectivity data; generating a reflectivity ratio matrix by adopting a reflectivity ratio classification model; the differentiable physical engine is used for quantizing the dynamic relationship between the vegetation root growth rate and the soil resistance by adopting a coupling model according to the reflectivity ratio matrix, generating a physical model and updating the physical model at a preset time interval; and the data processing and three-dimensional modeling module is used for generating a three-dimensional root system model and a health assessment result according to the reflectivity ratio matrix and the physical model, and outputting the three-dimensional root system model and the health assessment result. According to the scheme, soil penetration and root system and foreign matter distinguishing can be carried out, the vegetation native environment is not damaged, dynamic monitoring analysis is carried out, and the accuracy of recognition analysis is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of non-destructive testing of vegetation root systems, in particular to a multi-spectral vegetation root system identification system and method based on a differentiable physics engine. BACKGROUND

[0002] Vegetation root system identification refers to obtaining and analyzing the morphology, distribution and biomass information of underground roots using technical means, which can assist in ecological research, precision agriculture and ecological restoration, and is an important basis for ecological research, agricultural management and engineering protection. Current vegetation root system identification technologies include: Ground Penetrating Radar (GPR): By emitting high-frequency electromagnetic waves and receiving reflected signals, it analyzes the differences in underground dielectric constant to identify roots, but its resolution is limited by the wavelength (usually only centimeters), and changes in soil moisture can significantly interfere with the signal, resulting in a misjudgment rate of more than 30%, for example, in clay environments, GPR has difficulty distinguishing between root systems and water-bearing layer interfaces; X-ray computed tomography (CT): By three-dimensional imaging technology, it realizes root system visualization, with a resolution of sub-millimeter level, but it needs to dig soil samples and place them in laboratory equipment, which destroys the original environment of the vegetation, and the cost of a single detection is more than 5000 yuan, which cannot be used for large-scale dynamic monitoring; Electrical impedance tomography (EIT): By measuring the current distribution in the soil to reconstruct the root structure, however, uneven soil composition can cause current path deviation, resulting in low reconstruction accuracy, and it cannot reflect the biomechanical properties of the root system; Single-wavelength LiDAR: Some studies attempt to use near-infrared LiDAR (such as 905nm) to scan the ground vegetation, but only the canopy height information can be obtained, and it cannot penetrate the soil layer to detect the roots.

[0003] Although there are many methods for existing vegetation root system identification, there are still some problems: 1. Conflict between physical limitations and precision: GPR and CT are limited by electromagnetic wave attenuation and destructive sampling, respectively, making it difficult to balance non-destructive testing and high-precision requirements, for example, GPR has a resolution of less than 1cm in sandy soil, while CT has high accuracy but cannot be used in situ; while single-wavelength LiDAR, although it can be non-destructively detected and has high accuracy, it cannot penetrate the soil layer to detect roots, nor can it distinguish between roots and foreign objects; 2. Lack of dynamic modeling capability: Existing technologies focus on static structure analysis, lacking quantitative models for dynamic coupling of root growth rate and soil resistance, for example, traditional finite element analysis relies on offline calculation and cannot update parameters in real time, resulting in a deviation of more than 40% between simulation results and real environment; 3. Poor environmental adaptability: The existing system model parameters (such as soil elastic modulus) are usually fixed, and the influence of temperature and humidity changes on root-soil interaction is not considered. For example, the soil resistance decreases by 50% in the rainy season, but the traditional model still calculates according to the parameters in the dry season, resulting in prediction failure.

[0004] Therefore, there is an urgent need for a multi-spectral vegetation root recognition system and method based on a differentiable physics engine, which can penetrate soil and distinguish roots and foreign objects, without damaging the original environment of the vegetation, and dynamically monitor and analyze. SUMMARY

[0005] One of the purposes of the present application is to provide a multi-spectral vegetation root recognition system based on a differentiable physics engine, which can penetrate soil and distinguish roots and foreign objects, without damaging the original environment of the vegetation, and dynamically monitor and analyze, to improve the accuracy of recognition and analysis.

[0006] The basic scheme provided by the present application is: a multi-spectral vegetation root recognition system based on a differentiable physics engine, comprising: a multi-spectral LiDAR module, a differentiable physics engine, a data processing and three-dimensional modeling module; The multi-spectral LiDAR module is used to emit near-infrared laser and short-wave infrared laser, scan the target area, obtain dual-wavelength reflectivity data, and generate a reflectivity ratio matrix according to the dual-wavelength reflectivity data using a reflectivity ratio classification model; The differentiable physics engine is used to quantize the dynamic relationship between vegetation root growth rate and soil resistance according to the reflectivity ratio matrix using a coupling model, generate a physical model, and update at intervals of a preset time; The data processing and three-dimensional modeling module is used to generate a three-dimensional root model and a health assessment result according to the reflectivity ratio matrix and the physical model, and output.

[0007] The beneficial effects of the basic scheme are: the multi-spectral LiDAR module emits near-infrared laser and short-wave infrared laser, wherein the attenuation coefficient of near-infrared laser in soil is less than 0.15 dB / cm, which can penetrate soil and detect roots, and the absorption rate of short-wave infrared laser differs for different objects, with an absorption rate difference of more than 45%, especially the presence of root lignin absorption peak, so that roots and foreign objects can be distinguished; by emitting near-infrared laser and short-wave infrared laser, scanning the target area, obtaining dual-wavelength reflectivity data, and generating a reflectivity ratio matrix according to the dual-wavelength reflectivity data using a reflectivity ratio classification model, compared with the existing technology that LiDAR cannot detect underground structures, the present scheme can penetrate soil and distinguish roots and foreign objects; compared with the existing technology of GPR, the present scheme is less disturbed and has higher accuracy; compared with the existing technology of CT, the present scheme does not damage the original environment of the vegetation; The scheme also provides a differentiable physical engine, which adopts a coupling model according to the reflectivity ratio matrix, quantifies the dynamic relationship between the growth rate of the vegetation root system and the soil resistance, generates a physical model, and updates at intervals of a preset time, so that the three-dimensional root system model and the health assessment result generated by the data processing and three-dimensional modeling module according to the reflectivity ratio matrix and the physical model are also updated, the limitation of a traditional static model is broken through, errors caused by the static model are optimized, and the accuracy of identification and analysis is improved.

[0008] In summary, the scheme can distinguish the soil penetration, the root system and foreign matters, does not damage the original environment of the vegetation, and performs dynamic monitoring and analysis, so as to improve the accuracy of identification and analysis.

[0009] Further, the multispectral LiDAR module adopts a dual-wavelength laser to scan a target region and obtain dual-wavelength reflectivity data. The emission end of the dual-wavelength laser alternately emits near-infrared laser and short-wave infrared laser. The receiving end of the dual-wavelength laser detects echo signals of the near-infrared laser pulse and the short-wave infrared laser pulse respectively, and records the reflection intensity and .

[0010] The dual-wavelength laser alternately emits near-infrared laser and short-wave infrared laser, which can avoid spectral aliasing.

[0011] Further, the wavelength of the near-infrared laser is 1550 nm, and the wavelength of the short-wave infrared laser is 1064 nm. The combination of 1550 nm near-infrared laser and 1064 nm short-wave infrared laser is adopted, the former has a penetration depth of 30 cm, and the latter is directed to the absorption peak of root lignin.

[0012] Further, the reflectivity ratio matrix is generated by using a reflectivity ratio classification model according to the dual-wavelength reflectivity data, including: The reflectivity of the dual-wavelength laser is calculated for each scanning point of the target region: ; ; wherein and are the initial intensities of the near-infrared laser and the short-wave infrared laser respectively, and are calibration values; According to the reflectivity, the reflectivity ratio of all scanning points in the target region is calculated by using a reflectivity ratio model, and a reflectivity ratio matrix is formed; The reflectivity ratio model is: ; wherein is the reflectivity of the near-infrared laser, is the reflectivity of the short-wave infrared laser; The reflectivity matrix formed is: ; Wherein the matrix element represents the reflectivity ratio of the row, the column scanning point.

[0013] Further, the principal component analysis dimension reduction algorithm is used to extract the first several principal components. Reduce noise interference.

[0014] Further, the coupling model adopts a Verhulst-Pennes coupling model; Wherein the Verhulst-Pennes coupling model includes a root growth rate model and a soil resistance model; The root growth rate model includes: Wherein, is the root density, indicating the mass of the root in unit volume of soil, is the time, is the variable of root density with time; is the maximum bearing density, is the root growth coefficient, is the resistance sensitivity coefficient, is the soil resistance; The soil resistance model includes: Wherein is the soil density, is the humidity correction factor, is the humidity influence coefficient, is the temperature variable, is the variable of temperature with time. Quantify the dynamic relationship between root growth rate and soil resistance, and break through the limitations of traditional static models.

[0015] Further, a differentiable computation graph is constructed using PyTorch, and the , , , coefficient is optimized by automatic differentiation. Improve parameter tuning efficiency and adapt to complex environments such as drought and flood.

[0016] Further, the data processing and three-dimensional modeling module is used for reflectivity ratio matrix and physical model, to generate three-dimensional root system model and health assessment results, and output, including: Identify vegetation roots based on the reflectivity ratio matrix, match the reflectivity ratio matrix with the location coordinates in real time, partition the model into grids of preset sizes, and construct and generate a three-dimensional root system model. Update the three-dimensional root system model according to the physical model; If any indicator in the 3D root system model exceeds a preset threshold, a corresponding health warning or health recommendation is triggered as a health assessment result. The indicators include: 、 and . Dynamic monitoring improves monitoring accuracy.

[0017] Further, if If the value is greater than or equal to the first preset ratio, the reflected object is a root system; like If the ratio is less than or equal to the second preset value, the reflected object is a foreign object, thereby distinguishing roots from foreign objects such as rocks or plastic.

[0018] The second purpose of the present invention is to provide a multispectral vegetation root identification method based on a differentiable physics engine, which can penetrate the soil and distinguish between roots and foreign objects without destroying the native environment of the vegetation, and perform dynamic monitoring and analysis to improve the accuracy of identification and analysis.

[0019] The present invention provides a second basic solution: a multispectral vegetation root identification method based on a differentiable physics engine, which adopts the above-mentioned multispectral vegetation root identification system based on a differentiable physics engine.

[0020] Beneficial effects of this solution: This solution can penetrate the soil and distinguish between roots and foreign objects without destroying the native environment of vegetation, and conduct dynamic monitoring and analysis to improve the accuracy of identification and analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 This is a logic block diagram of an embodiment of a multispectral vegetation root recognition system based on a differentiable physics engine according to the present invention. DETAILED DESCRIPTION

[0022] The following is further described in detail through specific implementation methods: The symbols in the drawings of the specification include: Example 1 This embodiment is basically as shown in the attached Figure 1 As shown: A multispectral vegetation root recognition system based on a differentiable physics engine, including: a multispectral LiDAR module, a differentiable physics engine, a data processing and 3D modeling module; The multispectral LiDAR module is used to scan the target area, obtain dual-wavelength reflectance data, and generate a reflectance ratio matrix based on the dual-wavelength reflectance data using a reflectance ratio classification model; Specifically, the multi-spectral LiDAR module adopts a dual-wavelength laser to scan a target region and obtain dual-wavelength reflectivity data. The transmitting end of the dual-wavelength laser synchronously or alternately emits near-infrared laser and short-wave infrared laser through pulse code modulation (PCM) with a pulse interval ≤1 μs to avoid spectral interference; in this embodiment, the wavelength of the near-infrared laser is 1550 nm, and the wavelength of the short-wave infrared laser is 1064 nm; the penetration depth of the near-infrared laser reaches 30 cm (soil attenuation coefficient <0.15 dB / cm), and the short-wave infrared laser is directed against the root lignin absorption peak (absorption rate difference >45%); the dual-wavelength laser is alternately emitted to avoid spectral aliasing, and the sampling rate is increased to 1.2 MHz. The 1550 nm can penetrate deeper, and the 1064 nm can exclude foreign matter interference, and the two cooperate with each other to have a higher vegetation penetration ratio. The receiving end of the dual-wavelength laser detects the echo signals of the near-infrared laser pulses and the short-wave infrared laser pulses through avalanche photodiodes (APDs) respectively, and records the reflection intensities and .

[0023] For each scanning point (pixel) of the target region, the reflectivity of the dual-wavelength laser is calculated as follows: ; ; wherein and are the initial emission intensities of the near-infrared laser and the short-wave infrared laser respectively, and are calibration values; According to the reflectivity, a reflectivity ratio model is used to calculate the reflectivity ratio of all scanning points (i.e. pixel points) in the scanning region to form a reflectivity ratio matrix. The reflectivity ratio model is as follows: ; wherein is the reflectivity of the near-infrared laser, is the reflectivity of the short-wave infrared laser; The formed reflectivity matrix is as follows: ; wherein the matrix element represents the reflectivity ratio of the scanning point in the th row and the th column; Anomalies in the reflectivity matrix are filtered as follows: >3.0 or <0.1, it is a noise point, and is filtered, such as specular reflection or shielding; the rows and columns of the reflectivity ratio matrix correspond to the physical coordinates of the scanning region, such as each scanning point representing a 0.5 cm×0.5 cm ground area; If greater than or equal to a first preset ratio, in the embodiment , the reflected object is a root system; because the lignin enrichment of the root system causes the reflection ratio to increase; If less than or equal to a second preset ratio, in the embodiment , the reflected object is a rock or a plastic foreign object; the classification of the first preset ratio and the second preset ratio can cover a 95% confidence interval, and the classification accuracy is 96%; the second preset ratio is 0.75, which can further exclude soil interference, and in other embodiments, the first preset ratio and the second preset ratio can be fine-tuned according to the environment parameters, such as: 1.25 in the clay area can be reduced to 1.22, and 1.25 in the acidic soil (pH < 5.5) can be increased to 1.3; In addition, a principal component analysis (PCA) dimension reduction algorithm is used to extract the first several principal components (in the embodiment, the cumulative contribution rate is > 85%) to reduce noise interference; in the embodiment, three principal components are extracted, including: soil-root system overall contrast (reflecting the intensity of the 1550 nm / 1064 nm reflectance); root system density distribution characteristics (high value corresponding to root system dense area); foreign object and noise response (high value indicating rock / plastic interference). In the three-dimensional modeling process, the original data have wavelength reflectance corresponding to the spatial information of 1550 nm (near-infrared laser), wavelength reflectance corresponding to the spatial information of 1064 nm (short-wave infrared laser), spatial information corresponding to Rratio, spatial information corresponding to ≥1.25 (≥1.25 is assigned a value of 1, and others are 0), spatial information corresponding to ≤0.75 (≤0.75 is assigned a value of 1, and others are 0), after dimension reduction, the first three principal components are: soil-root system overall contrast (reflecting the intensity of the 1550 nm / 1064 nm reflectance) - corresponding to the spatial information corresponding to Rratio; root system density distribution characteristics (high value corresponding to root system dense area) - corresponding to spatial information corresponding to ≥1.25 (≥1.25 is assigned a value of 1, and others are 0); foreign object and noise response (high value indicating rock / plastic interference) - corresponding to R spatial information corresponding to ≤0.75 (≤0.75 is assigned a value of 1, and others are 0); only three dimensions are considered in the five-dimensional data, and the selection of three dimensions is the optimal balance between accuracy and efficiency. The amount of principal component data and the calculation time (ms / pixel) are: the original 5-dimensional data amount is retained by 100%, and the calculation time is 0.85; the first 3-dimensional data amount is retained by 87%, and the calculation time is 0.12; the first 2-dimensional data amount is retained by 80%, and the calculation time is 0.08.

[0024] The wavelength of the near-infrared laser pulse is 1550 nm, which belongs to the second window of the near-infrared, and the water absorption rate is the lowest (only 0.1%), while the water in the soil is the main attenuation source; the wavelength of the short-wave infrared laser pulse is 1064 nm, and according to the specific absorption of root lignin: the C-H bond of lignin has a strong frequency doubling absorption peak at 1064 nm (the absorption coefficient is 2.3 times that of 1550 nm), and the soil minerals have weak absorption at this waveband, thereby facilitating the detection of plant roots. In addition, the cost of 1550 nm laser pulses is 30% higher than that of 1064 nm laser pulses, but it is more suitable for ground equipment due to its eye safety level (Class I), while 1064 nm requires protection (Class IV), but the response efficiency of silicon-based detectors to it reaches 80%, and the cost performance is higher.

[0025] In the present scheme, the combination of 1550 nm and 1064 nm in the multispectral LiDAR module is the optimal balance of penetration ability, biochemical specificity, and engineering cost: 1550 nm laser pulses break through the soil barrier and directly reach the deep roots; 1064 nm laser pulses lock lignin and exclude foreign matter interference; and the ratio method of constructing the two provides high-purity input data for the physical engine by combining the detection advantages of the two laser pulses.

[0026] The multispectral LiDAR module is used to penetrate the soil and distinguish between roots and foreign matter, solving the defect that traditional LiDAR cannot detect underground structures.

[0027] The differentiable physical engine is used to adopt a coupling model to quantify the dynamic relationship between the growth rate of plant roots and soil resistance according to the reflectance ratio matrix, generate a physical model, and update it at intervals of a preset time; Specifically, the coupling model adopts a Verhulst-Pennes coupling model. The Verhulst-Pennes coupling model includes a root growth rate model and a soil resistance model. The root growth rate model includes: wherein, is the root density, indicating the mass of roots per unit volume of soil, is the time, is the variable of the root density over time; is the maximum bearing density, which is defined by a plant species database constructed by planting experiments, is the root growth coefficient, quantifying the inhibition intensity of soil resistance on root growth (range: 0.1~0.3), is the resistance sensitivity coefficient, is the soil resistance; The soil resistance model includes: wherein is the soil density, in g / cm³, is the humidity correction factor, and is a 0-1 normalized value, is the humidity influence coefficient, is the temperature variable, is the temperature variable over time.

[0028] In addition, a differentiable computational graph is constructed using PyTorch, and the coefficients are optimized by automatic differentiation 、 、 、 The training set covers 100,000 sets of simulation data, including extreme scenarios such as drought and flood; The simulation results are compared in COMSOL Multiphysics, and the model prediction error is less than 8% (double-index verification of root density and stress distribution).

[0029] The data processing and three-dimensional modeling module is used to reflect the ratio matrix and the physical model, generate a three-dimensional root system model and health assessment results, and output.

[0030] Specifically, According to the reflectance ratio matrix, the vegetation root system is identified, and real-time matching is performed according to the reflectance ratio matrix and the position coordinates, and the three-dimensional root system model is constructed and generated by dividing the modeling according to the preset size grid; In this embodiment, the position coordinates use GPS coordinates and RTK positioning, with an accuracy of ±1 cm; the preset size is 10m x 10m; If is greater than or equal to the first preset ratio, in this embodiment , the reflected object is a root system; because the lignin enrichment of the root system will cause the reflectance ratio to rise; If is less than or equal to the second preset ratio, in this embodiment , the reflected object is rock or plastic foreign matter; According to the physical model, the three-dimensional root system model is updated, specifically, the physical engine updates the distribution of root density once every preset time, and the data processing and three-dimensional modeling model updates the three-dimensional root system model according to the distribution of root density; If any index in the three-dimensional root system model exceeds the preset threshold, the corresponding health warning or health suggestion is triggered as the health assessment result; the indexes include: 、 and ; Correspondingly, if greater than a preset soil resistance threshold, a root decay early warning is triggered; in the embodiment, the preset soil resistance threshold is 25 kPa; If less than a preset reflectivity ratio threshold, a slope reinforcement suggestion is generated; in the embodiment, the preset reflectivity ratio threshold is 0.3.

[0031] The specific implementation process is as follows: taking the ancient tree root health monitoring using the scheme as an example, the specific process is as follows: First, the system deployment and parameter setting are performed: The multispectral LiDAR module is fixed on a tripod (height 1.2 m, inclination 20°), the scanning radius is 3 m, and the resolution is 0.5 cm; The differentiable physics engine adopts a physics engine solving software and is pre-installed in the data processing and three-dimensional modeling module.

[0032] Initialization: set the initial parameters according to the tree species database: ginkgo tree 、 =0.12、 =1.05、 =0.8; Maximum bearing density =2.5 g / cm3 (based on historical growth data).

[0033] Secondly, data acquisition and processing are performed: Obtain the dual-wavelength reflectivity data, that is, the spectral data, the multispectral LiDAR module emits laser at a frequency of 15 Hz, generates a reflectivity ratio matrix, and then transmits it to the differentiable physics engine through Wi-Fi 6; after PCA dimension reduction, the first three principal components are retained, and more than 80% of soil scattering noise is filtered out; Solve the physical model, and the differentiable physics engine updates the root density distribution every 0.5 seconds, calculates the soil resistance gradient ; And when local >25 kPa is detected, a root decay early warning is triggered; Finally, the result output and application are performed: The data processing and three-dimensional modeling module generates a three-dimensional root model, and the red highlighted area represents a high-risk area >25 kPa.

[0034] Taking the slope vegetation root coverage rate evaluation using the scheme as an example, the specific process is as follows: First, the system deployment and parameter setting are performed: Hardware configuration: multispectral LiDAR mounted on a drone, flight height 60 m, scanning bandwidth 120 m, daily detection area 25 acres; onboard computer transmits data to cloud-based differentiable physics engine in real time; Model initialization: set , = 0.18, = 1.3, = 0.6 (for grass vegetation); maximum bearing density = 1.8 g / cm3; Second, dynamic monitoring and optimization: Spectral data processing: real-time matching of reflectance matrix and GPS coordinates (RTK positioning, accuracy ±1 cm), modeling by 10m x 10m grid partitioning; for <0.8 area automatically marked as "foreign matter interference area" and excluded from analysis.

[0035] Model solution: differentiable physics engine updates root density distribution in the entire area every 2 seconds; when root coverage is detected to be <30%, slope reinforcement recommendations are generated.

[0036] Finally, engineering verification: In the slope project, the system evaluated root coverage at 82%, guiding accurate seeding, and vegetation survival rate increased by 25%; the entire process took 30 minutes / acre, a 15-fold improvement in efficiency over manual reconnaissance.

[0037] This scheme has made significant improvements in technical performance, economic effect, and environmental and social benefits, as follows: Technical performance: This scheme can perform high-precision root identification, with a root and foreign matter differentiation accuracy of 95% through the reflectance ratio model of the multispectral LiDAR module, a significant improvement over the 70% of traditional ground penetrating radar (GPR); principal component analysis (PCA) dimensionality reduction algorithm filters out more than 80% of soil noise, with a signal-to-noise ratio of 25dB, ensuring the purity of spectral data; This scheme performs dynamic modeling, with the Verhulst-Pennes coupled model quantifying the dynamic relationship between root growth and soil resistance, with a model prediction error of <8% (COMSOL simulation verification), a significant optimization over traditional static model error (>30%); differentiable programming supports automatic optimization , , , coefficients, with a 5-fold improvement in parameter tuning efficiency, suitable for complex environments such as drought and flooding; The whole process from data acquisition to three-dimensional model output takes only 2.5 minutes (in the scene of ancient tree monitoring), which is 40 times more efficient than X-ray CT detection (which takes several hours); the root system density distribution resolution reaches 0.5 cm, which can accurately locate abnormal areas such as decay and insect damage (error < 0.4 cm).

[0038] In terms of economic benefits: The cost is reduced, the multispectral LiDAR module supports in-situ non-destructive testing, and the single cost is reduced from 5000 yuan of CT to 200 yuan, which is suitable for large-scale engineering applications; the cloud physical engine is combined with edge computing, and the hardware investment is reduced by 40% (compared with the traditional GPU server solution).

[0039] The engineering application value of the present scheme is improved, in the ecological restoration of slope, the system early warning weak root area, reduce the reinforcement cost 30%; ancient tree health monitoring guide accurate repair, maintenance cost reduction 40%; In terms of environmental and social benefits: The present scheme carries out ecological protection, non-destructive testing technology avoids damage to ancient tree roots, prolongs the life expectancy of famous trees by more than 10 years; the slope vegetation coverage rate evaluation accuracy reaches 95%, and helps to improve the ecological restoration standard rate by 25%; The present scheme efficiently utilizes resources, precise irrigation recommendation (based on root density distribution) saves water by 15%, and corn yield in agricultural test fields is increased by 8%; the dynamic model adapts to changes in soil moisture, reducing the pollution risk caused by excessive fertilization; In addition, the present scheme carries out multidisciplinary integration, for the first time combining multispectral LiDAR with biomechanical equations to overcome the difficulty of non-destructive detection of underground roots; the differentiable physical engine realizes self-optimization of model parameters, forming a "data-model" closed-loop iteration technology barrier; the present scheme supports unmanned aerial vehicle carrying (daily detection of 25 hectares) and fixed deployment (millimeter-level precision), covering slope repair, agricultural monitoring and other scenes; the specific setting of standardized data interface (JSON protocol) is compatible with mainstream engineering management platforms, and has strong adaptability.

[0040] The embodiment also provides a multispectral vegetation root identification method based on a differentiable physical engine, which adopts the multispectral vegetation root identification system based on the differentiable physical engine.

[0041] The above-mentioned are only embodiments of the present application, and the common knowledge of specific structures and characteristics in the scheme is not described too much herein. The ordinary skilled person in the art knows all the ordinary technical knowledge in the field of the present application before the application date or the priority date, can know all the prior art in the field, and has the ability to apply conventional experimental means before that date. The ordinary skilled person in the art can perfect and implement the present scheme under the guidance of the present application, combined with their own ability. Some typical known structures or known methods should not be an obstacle for the ordinary skilled person in the art to implement the present application. It should be noted that, for those skilled in the art, without departing from the structure of the present application, a number of modifications and improvements can be made, which should also be considered as the protection scope of the present application. These will not affect the effect and practicality of the patent. The protection scope of the present application should be subject to the content of its claims, and the specific implementation mode and the like in the specification can be used to explain the content of the claims.

Claims

1. A multispectral vegetation root recognition system based on a differentiable physics engine, characterized by: include: Multispectral LiDAR module, differentiable physics engine, data processing and 3D modeling module; A multispectral LiDAR module is used to emit near-infrared laser and short-wave infrared laser to scan the target area, obtain dual-wavelength reflectivity data, and generate a reflectivity ratio matrix based on the dual-wavelength reflectivity data using a reflectivity ratio classification model; A differentiable physics engine is used to quantify the dynamic relationship between vegetation root growth rate and soil resistance using a coupled model based on the reflectivity ratio matrix, generate a physical model, and update it at preset intervals; The data processing and three-dimensional modeling module is used to generate and output a three-dimensional root system model and health assessment results based on the reflectivity ratio matrix and physical model.

2. The multispectral vegetation root recognition system based on a differentiable physics engine according to claim 1 is characterized in that: The multispectral LiDAR module uses a dual-wavelength laser to scan the target area and obtain dual-wavelength reflectivity data; The emitting end of the dual-wavelength laser emits near-infrared laser and short-wave infrared laser alternately; The receiving end of the dual-wavelength laser detects the echo signals of the near-infrared laser pulse and the short-wave infrared laser pulse respectively, and records the reflection intensity and .

3. The multispectral vegetation root recognition system based on a differentiable physics engine according to claim 1 is characterized in that: The wavelength of the near-infrared laser is 1550 nm, and the wavelength of the short-wave infrared laser is 1064 nm.

4. The multispectral vegetation root recognition system based on a differentiable physics engine according to claim 1 is characterized in that: The reflectivity ratio matrix is ​​generated based on the dual-wavelength reflectivity data using a reflectivity ratio classification model, including: For each scanning point in the target area, calculate the dual-wavelength reflectivity: ； ； in and are the initial intensity of near-infrared laser emission and the initial intensity of short-wave infrared laser emission, which are calibration values; According to the reflectivity, the reflectivity ratio model is used to calculate the reflectivity ratio of all scanning points in the target area to form a reflectivity ratio matrix; The reflectivity ratio model is: ;in is the reflectivity of the near-infrared laser, is the reflectivity of short-wave infrared laser; The resulting reflectivity matrix is: ； The matrix elements Indicates the Row, No. The reflectivity ratio of the scanning points.

5. The multispectral vegetation root recognition system based on a differentiable physics engine according to claim 1 is characterized in that: The principal component analysis dimensionality reduction algorithm is used to extract the first several principal components.

6. The multispectral vegetation root recognition system based on a differentiable physics engine according to claim 1 is characterized in that: The coupling model adopts the Verhulst-Pennes coupling model; The Verhulst-Pennes coupling model includes: root growth rate model and soil resistance model; Root growth rate model, including: in, is the root density, which represents the mass of roots per unit volume of soil. For time, is the root density variable over time; is the maximum load density, is the root growth coefficient, is the resistance sensitivity coefficient, is soil resistance; Soil resistance models, including: in is the soil density, is the humidity correction factor, is the humidity influence coefficient, is the temperature variable, is the temperature variation over time.

7. The multispectral vegetation root recognition system based on a differentiable physics engine according to claim 6 is characterized in that: Use PyTorch to build a differentiable computational graph and optimize it through automatic differentiation 、 、 、 coefficient.

8. The multispectral vegetation root recognition system based on a differentiable physics engine according to claim 1 is characterized in that: The data processing and 3D modeling module is used to generate a 3D root system model and health assessment results using the reflectivity ratio matrix and physical model, and outputs the following: Identify vegetation roots based on the reflectivity ratio matrix, match the reflectivity ratio matrix with the location coordinates in real time, partition the model into grids of preset sizes, and construct and generate a three-dimensional root system model. Update the three-dimensional root system model according to the physical model; If any indicator in the three-dimensional root system model exceeds the preset threshold, the corresponding health warning or health advice will be triggered as the health assessment result.

9. The multispectral vegetation root recognition system based on a differentiable physics engine according to claim 8, characterized in that: If the reflectivity ratio is greater than or equal to the first preset ratio, the reflected object is a root system; If the reflectivity ratio is less than or equal to the second preset ratio, the reflected object is a foreign object.

10. A multispectral vegetation root recognition method based on a differentiable physics engine, characterized by: A multispectral vegetation root recognition system based on a differentiable physics engine as described in any one of claims 1 to 9 is used.

Images