A Multispectral Vegetation Root System Identification System and Method Based on Differentiable Physics Engine

Abstract

This invention relates to the field of non-destructive testing technology for vegetation roots, specifically a multispectral vegetation root identification system and method based on a differentiable physics engine. The system includes: a multispectral LiDAR module for emitting near-infrared and short-wave infrared lasers to scan the target area, acquire dual-wavelength reflectance data, and generate a reflectance ratio matrix using a reflectance ratio classification model based on the dual-wavelength reflectance data; a differentiable physics engine for quantifying the dynamic relationship between vegetation root growth rate and soil resistance using a coupled model based on the reflectance ratio matrix, generating a physical model, and updating it at preset intervals; and a data processing and 3D modeling module for generating and outputting a 3D root model and health assessment results based on the reflectance ratio matrix and the physical model. This solution can distinguish between soil penetration and roots and foreign objects without damaging the original vegetation environment, and performs dynamic monitoring and analysis to improve the accuracy of identification and analysis.

Description

Technical Field

[0001] This invention relates to the field of non-destructive testing technology for vegetation roots, specifically a multispectral vegetation root identification system and method based on a differentiable physics engine. Background Technology

[0002] Vegetation root identification refers to the use of technical means to obtain and analyze information on the morphology, distribution and biomass of underground roots. It can assist in ecological research, precision agriculture and ecological restoration, and is an important foundation for ecological research, agricultural management and engineering protection.

[0003] The mainstream technologies currently used for vegetation root identification include:

[0004] Ground penetrating radar (GPR): It identifies roots by analyzing differences in the dielectric constant of the soil by emitting high-frequency electromagnetic waves and receiving reflected signals. However, its resolution is limited by wavelength (usually only at the centimeter level), and changes in soil moisture can significantly interfere with the signal, resulting in a false positive rate of more than 30%. For example, in clay environments, GPR has difficulty distinguishing the root system from the aquifer interface.

[0005] X-ray computed tomography (CT): It enables the visualization of root systems through three-dimensional imaging technology with a resolution of up to sub-millimeter level. However, it requires digging up soil samples and placing them in laboratory equipment, which damages the original environment of vegetation. Moreover, the cost of a single test exceeds 5,000 yuan, making it unsuitable for large-scale dynamic monitoring.

[0006] Electrical impedance imaging (EIT): Root structure is inverted by measuring the current distribution in the soil. However, uneven soil composition can cause current path deviation, resulting in low reconstruction accuracy and failure to reflect the biomechanical properties of the roots.

[0007] Single-wavelength LiDAR: Some studies have attempted to use near-infrared LiDAR (such as 905nm) to scan surface vegetation, but they can only obtain information on canopy height and cannot penetrate the soil layer to detect roots.

[0008] While there are various existing methods for identifying vegetation root systems, some problems still exist:

[0009] 1. Physical limitations and precision contradictions exist: GPR and CT are limited by electromagnetic wave attenuation and destructive sampling, respectively, making it difficult to meet the requirements of non-destructive testing and high precision. For example, GPR has a resolution of less than 1 cm in sandy soil, while CT, although highly accurate, cannot be applied in situ; while single-wavelength LiDAR, although capable of non-destructive testing with high precision, cannot penetrate the soil layer to detect roots and cannot distinguish between roots and foreign objects.

[0010] 2. Lack of dynamic modeling capabilities: Existing technologies mostly focus on static structural analysis and lack quantitative models for the dynamic coupling of root growth rate and soil resistance. For example, traditional finite element analysis relies on offline calculations and cannot update parameters in real time, resulting in simulation results deviating from the real environment by more than 40%.

[0011] 3. Poor environmental adaptability: Existing system model parameters (such as soil elastic modulus) are usually fixed and do not consider the impact of temperature and humidity changes on root-soil interactions. For example, soil resistance decreases by 50% during the rainy season, but traditional models still calculate based on dry season parameters, leading to prediction failure.

[0012] Therefore, there is an urgent need for a multispectral vegetation root identification system and method based on a differentiable physics engine, which can perform soil penetration and distinguish between roots and foreign objects, without damaging the original vegetation environment, and can perform dynamic monitoring and analysis. Summary of the Invention

[0013] One of the objectives of this invention is to provide a multispectral vegetation root identification system based on a differentiable physics engine, which can penetrate the soil and distinguish between roots and foreign objects without damaging the original vegetation environment, and perform dynamic monitoring and analysis to improve the accuracy of identification and analysis.

[0014] The basic solution provided by this invention is a multispectral vegetation root identification system based on a differentiable physics engine, comprising: a multispectral LiDAR module, a differentiable physics engine, and a data processing and 3D modeling module;

[0015] The multispectral LiDAR module is used to emit near-infrared laser and short-wave infrared laser to scan the target area, acquire dual-wavelength reflectivity data, and generate a reflectivity ratio matrix based on the dual-wavelength reflectivity data using a reflectivity ratio classification model.

[0016] A differentiable physics engine is used to quantify the dynamic relationship between vegetation root growth rate and soil resistance based on the reflectivity ratio matrix and a coupled model, generate a physical model, and update it at preset intervals.

[0017] The data processing and 3D modeling module is used to generate and output a 3D root system model and health assessment results based on the reflectivity ratio matrix and the physical model.

[0018] Beneficial effects of the basic scheme: This scheme emits near-infrared and short-wave infrared lasers through a multispectral LiDAR module. Near-infrared lasers have an attenuation coefficient of less than 0.15 dB / cm in soil, allowing them to penetrate the soil and detect roots. Short-wave infrared lasers exhibit varying absorption rates for different objects, with differences exceeding 45%, particularly showing a lignin absorption peak in roots, thus distinguishing roots from foreign objects. By emitting near-infrared and short-wave infrared lasers, the target area is scanned to acquire dual-wavelength reflectance data. Based on this data, a reflectance ratio classification model is used to generate a reflectance ratio matrix. Compared to existing LiDAR technologies that cannot detect underground structures, this scheme can penetrate soil and distinguish between roots and foreign objects. Compared to existing GPR technologies, this scheme is less susceptible to interference and has higher accuracy. Compared to existing CT technologies, it does not damage the original vegetation environment.

[0019] This solution also includes a differentiable physics engine. Based on the reflectance ratio matrix, it uses a coupled model to quantify the dynamic relationship between vegetation root growth rate and soil resistance, generating a physical model that is updated at preset intervals. Consequently, the data processing and 3D modeling modules also update the 3D root model and health assessment results generated based on the reflectance ratio matrix and the physical model. This overcomes the limitations of traditional static models, optimizes the errors introduced by static models, and improves the accuracy of identification and analysis.

[0020] In summary, this solution can penetrate the soil and distinguish between roots and foreign objects without damaging the original vegetation environment, and can perform dynamic monitoring and analysis to improve the accuracy of identification and analysis.

[0021] Furthermore, the multispectral LiDAR module uses a dual-wavelength laser to scan the target area and acquire dual-wavelength reflectivity data;

[0022] The dual-wavelength laser emits near-infrared laser and short-wave infrared laser alternately at its emitting end.

[0023] The receiver of the dual-wavelength laser detects the echo signals of near-infrared and short-wave infrared laser pulses respectively, and records the reflection intensity. and .

[0024] Dual-wavelength lasers alternately emit near-infrared and short-wave infrared lasers, which can avoid spectral aliasing.

[0025] Furthermore, the near-infrared laser has a wavelength of 1550nm, and the short-wave infrared laser has a wavelength of 1064nm. By combining the 1550nm near-infrared laser with the 1064nm short-wave infrared laser, the former achieves a penetration depth of 30cm, while the latter targets the lignin absorption peak in the root system.

[0026] Furthermore, the step of generating a reflectance ratio matrix based on dual-wavelength reflectance data using a reflectance ratio classification model includes:

[0027] For each scan point in the target area, calculate the reflectivity of both wavelengths:

[0028] ;

[0029] ;

[0030] in and These are the initial emission intensities of near-infrared laser and short-wave infrared laser, respectively, and are calibration values;

[0031] Based on reflectance, a reflectance ratio model is used to calculate the reflectance ratio of all scanning points in the target area, forming a reflectance ratio matrix.

[0032] The reflectivity ratio model is as follows: ;in The reflectivity of near-infrared laser, The reflectivity of short-wave infrared laser;

[0033] The resulting reflectivity matrix is:

[0034] ;

[0035] matrix elements Indicates the first Okay, number The reflectance ratio of the scan points.

[0036] Furthermore, principal component analysis (PCA) is used to reduce dimensionality by extracting the first few principal components, thereby reducing noise interference.

[0037] Furthermore, the coupling model adopted is the Verhulst-Pennes coupling model;

[0038] The Verhulst-Pennes coupled model includes: a root growth rate model and a soil resistance model;

[0039] Root growth rate models include:

[0040]

[0041] in, Root density represents the mass of roots per unit volume of soil. For time, Root density is a variable over time; For maximum load-bearing density, Root growth coefficient For drag sensitivity coefficient, For soil resistance;

[0042] Soil resistance models, including:

[0043]

[0044] in For soil density, Humidity correction factor Humidity influence coefficient For temperature variables, Temperature is a variable over time. This method quantifies the dynamic relationship between root growth rate and soil resistance, overcoming the limitations of traditional static models.

[0045] Furthermore, a differentiable computational graph is constructed using PyTorch, and automatic differentiation optimization is performed. , , , Coefficients. Improve parameter tuning efficiency and adapt to complex environments such as drought and floods.

[0046] Furthermore, the data processing and 3D modeling module is used to generate a 3D root system model and health assessment results from the reflectivity ratio matrix and physical model, and outputs them, including:

[0047] Based on the reflectance ratio matrix, the vegetation root system is identified, and the reflectance ratio matrix is ​​matched with the location coordinates in real time. The model is then constructed and generated by dividing the model into grid sections of a preset size.

[0048] Update the three-dimensional root system model based on the physical model;

[0049] In the three-dimensional root system model, if any indicator exceeds a preset threshold, a corresponding health warning or health suggestion is triggered as a health assessment result; the indicators include: , and Dynamic monitoring improves monitoring accuracy.

[0050] Furthermore, if If the ratio is greater than or equal to the first preset value, then the reflected object is a root system;

[0051] like If the ratio is less than or equal to the second preset value, the reflected object is considered a foreign object. This distinguishes between root systems and foreign objects such as rocks or plastic.

[0052] The second objective of this invention is to provide a multispectral vegetation root identification method based on a differentiable physics engine, which can perform soil penetration and distinguish between roots and foreign objects, without damaging the original vegetation environment, and can perform dynamic monitoring and analysis to improve the accuracy of identification and analysis.

[0053] 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.

[0054] The beneficial effects of this solution are: This solution can penetrate the soil and distinguish between roots and foreign objects without damaging the original vegetation environment, and can perform dynamic monitoring and analysis to improve the accuracy of identification and analysis. Attached Figure Description

[0055] Figure 1 This is a logic block diagram of an embodiment of the multispectral vegetation root identification system based on a differentiable physics engine of the present invention. Detailed Implementation

[0056] The following detailed description illustrates the specific implementation method:

[0057] The markings in the accompanying drawings include:

[0058] Example 1

[0059] This embodiment is basically as shown in the appendix. Figure 1 As shown: A multispectral vegetation root identification system based on a differentiable physics engine, including: a multispectral LiDAR module, a differentiable physics engine, and a data processing and 3D modeling module;

[0060] The multispectral LiDAR module is used to scan the target area, acquire dual-wavelength reflectance data, and generate a reflectance ratio matrix based on the dual-wavelength reflectance data using a reflectance ratio classification model.

[0061] Specifically, the multispectral LiDAR module uses a dual-wavelength laser to scan the target area and acquire dual-wavelength reflectivity data;

[0062] The dual-wavelength laser emitter synchronously or alternately emits near-infrared and short-wave infrared lasers via pulse code modulation (PCM), with a pulse interval ≤1μs to avoid spectral interference. In this embodiment, the near-infrared laser wavelength is 1550nm and the short-wave infrared laser wavelength is 1064nm. The near-infrared laser has a penetration depth of 30cm (soil attenuation coefficient <0.15dB / cm), while the short-wave infrared laser targets the lignin absorption peak in roots (absorption rate difference >45%). Alternating emission of the dual-wavelength laser avoids spectral aliasing, increasing the sampling rate to 1.2MHz. The 1550nm wavelength allows for deeper penetration, while the 1064nm wavelength eliminates interference from foreign objects; their combined effect results in a higher penetration rate through vegetation.

[0063] At the receiver of the dual-wavelength laser, an avalanche photodiode (APD) is used to detect the echo signals of the near-infrared laser pulse and the short-wave infrared laser pulse, respectively, and the reflection intensity is recorded. and .

[0064] For each scan point (pixel) in the target area, calculate the reflectivity using two wavelengths:

[0065] ;

[0066] ;

[0067] in and These are the initial emission intensities of near-infrared laser and short-wave infrared laser, respectively, and are calibration values;

[0068] Based on reflectance, a reflectance ratio model is used to calculate the reflectance ratio of all scan points (i.e., pixels) within the scan area, forming a reflectance ratio matrix;

[0069] The reflectivity ratio model is as follows: ;in The reflectivity of near-infrared laser, The reflectivity of short-wave infrared laser;

[0070] The resulting reflectivity matrix is:

[0071] ;

[0072] matrix elements Indicates the first Okay, number The reflectance ratio of the scan points;

[0073] Filtering outliers in the reflectivity matrix: Points with a reflectivity >3.0 or <0.1 are considered noise points and are filtered out, such as those caused by specular reflection or occlusion. The rows and columns of the reflectivity ratio matrix correspond to the physical coordinates of the scanned area, such as each scan point representing a 0.5cm × 0.5cm ground area.

[0074] like Greater than or equal to the first preset ratio, in this embodiment If the reflection is positive, then the object reflecting the light is the root system; because the lignin enrichment in the root system leads to an increased reflectance.

[0075] like Less than or equal to the second preset ratio, in this embodiment If the reflected object is rock or plastic foreign object, the classification of the first preset ratio and the second preset ratio can cover the 95% confidence interval and the classification accuracy reaches 96%. The second preset ratio of 0.75 can also exclude soil interference. In other embodiments, the first preset ratio and the second preset ratio can be adaptively fine-tuned according to environmental parameters. For example, 1.25 in clay areas can be reduced to 1.22, and 1.25 can be increased to 1.3 for acidic soil (pH<5.5).

[0076] Furthermore, Principal Component Analysis (PCA) dimensionality reduction algorithm was employed to extract the first few principal components (cumulative contribution rate > 85% in this embodiment) to reduce noise interference. In this embodiment, three principal components were extracted: overall soil-root contrast (reflecting 1550nm / 1064nm reflectance intensity); root density distribution characteristics (high values ​​correspond to dense root areas); and foreign object and noise response (high values ​​indicate interference from rocks / plastics, etc.). The original data used in the 3D modeling process included spatial information corresponding to the 1550nm (near-infrared laser) wavelength reflectance, the 1064nm (short-wave infrared laser) wavelength reflectance, and the spatial information corresponding to Rratio. Spatial information corresponding to ≥1.25 (≥1.25 is assigned a value of 1, others are assigned a value of 0), Spatial information corresponding to values ​​≤0.75 (≤0.75 is assigned a value of 1, others are assigned 0). After dimensionality reduction, the first three principal components are: overall soil-root contrast (reflecting 1550nm / 1064nm reflectance intensity) - spatial information corresponding to the Rratio; root density distribution characteristics (high values ​​correspond to dense root areas) - spatial information corresponding to the Rratio. Spatial information corresponding to ≥1.25 (≥1.25 is assigned a value of 1, others are 0); Foreign object and noise response (high values ​​indicate interference such as rocks / plastics) - corresponding R Spatial information corresponding to values ​​≤0.75 (≤0.75 is assigned a value of 1, others are 0); for five-dimensional data, only three dimensions are considered, and choosing three dimensions is the optimal balance between accuracy and efficiency. Principal component quantity, data volume retention, and computation time (ms / pixel): 100% of the original 5-dimensional data volume is retained, with a computation time of 0.85; 87% of the first 3-dimensional data volume is retained, with a computation time of 0.12; 80% of the first 2-dimensional data volume is retained, with a computation time of 0.08.

[0077] Near-infrared laser pulses, with a wavelength of 1550nm, fall within the second near-infrared window, where water absorption is lowest (only 0.1%), and soil moisture is the primary source of attenuation. Short-wave infrared laser pulses, with a wavelength of 1064nm, benefit from the specific absorption of lignin in roots: the CH bonds in lignin exhibit a strong overtone absorption peak at 1064nm (absorption coefficient 2.3 times that of 1550nm), while soil minerals show weak absorption in this band, making it advantageous for detecting plant roots. Furthermore, while 1550nm laser pulses cost 30% more than 1064nm laser pulses, their eye safety level (Class I) makes them more suitable for ground-based equipment. Although 1064nm requires protection (Class IV), silicon-based detectors achieve an 80% response efficiency, making them more cost-effective.

[0078] In this scheme, the multispectral LiDAR module uses a combination of 1550nm and 1064nm, which is the optimal balance between penetration capability, biochemical specificity, and engineering cost: the 1550nm laser pulse breaks through the soil barrier and reaches deep roots; the 1064nm laser pulse locks onto lignin and eliminates foreign interference; and the ratio method of the two is used to construct a noise-resistant reflectivity ratio matrix, combining the advantages of the two laser pulse detection methods to provide high-purity input data for the physics engine.

[0079] By using a multispectral LiDAR module to penetrate soil and distinguish between roots and foreign objects, the limitations of traditional LiDAR in detecting underground structures are overcome.

[0080] A differentiable physics engine is used to quantify the dynamic relationship between vegetation root growth rate and soil resistance based on the reflectivity ratio matrix and a coupled model, generate a physical model, and update it at preset intervals.

[0081] Specifically, the coupling model adopted is the Verhulst-Pennes coupling model;

[0082] The Verhulst-Pennes coupled model includes: a root growth rate model and a soil resistance model;

[0083] Root growth rate models include:

[0084]

[0085] in, Root density represents the mass of roots per unit volume of soil. For time, Root density is a variable over time; The maximum carrying capacity is defined by a database of plant species constructed from planting trials. The root growth coefficient quantifies the inhibitory effect of soil resistance on root growth (range: 0.1~0.3). For drag sensitivity coefficient, For soil resistance;

[0086] Soil resistance models, including:

[0087]

[0088] in Soil density, in g / cm³. This is a humidity correction factor, and it is a 0-1 normalized value. Humidity influence coefficient For temperature variables, Temperature is a variable that changes with time.

[0089] Furthermore, a differentiable computation graph is constructed using PyTorch, and automatic differentiation optimization is performed. , , , The coefficients, the training set covers 100,000 sets of simulated data, including extreme scenarios such as drought and flood;

[0090] The simulation results in COMSOL Multiphysics show that the model prediction error is less than 8% (verified by both root density and stress distribution).

[0091] The data processing and 3D modeling module is used for reflectivity ratio matrix and physical model, generating 3D root system model and health assessment results, and outputting them.

[0092] Specifically,

[0093] Based on the reflectance ratio matrix, the vegetation root system is identified, and the reflectance ratio matrix is ​​matched with the location coordinates in real time. The model is then constructed and generated by dividing the model into grid sections of a preset size.

[0094] In this embodiment, the location coordinates are GPS coordinates, RTK positioning, with an accuracy of ±1cm; the preset size is 10m×10m.

[0095] like Greater than or equal to the first preset ratio, in this embodiment If the reflection is positive, then the object reflecting the light is the root system; because the lignin enrichment in the root system leads to an increased reflectance.

[0096] like Less than or equal to the second preset ratio, in this embodiment If so, the reflected object is a rock or a plastic foreign object;

[0097] Based on the physical model, the 3D root system model is updated. Specifically, the physics engine updates the root density distribution once at a preset time interval. The data processing and 3D modeling model update the 3D root system model based on the root density distribution.

[0098] In the three-dimensional root system model, if any indicator exceeds a preset threshold, a corresponding health warning or health suggestion is triggered as a health assessment result; the indicators include: , and ;

[0099] Correspondingly, if If the soil resistance exceeds a preset soil resistance threshold, a root rot warning is triggered; in this embodiment, the preset soil resistance threshold is 25 kPa.

[0100] like If the reflectance ratio is less than a preset threshold, a slope reinforcement suggestion is generated; in this embodiment, the preset reflectance ratio threshold is 0.3.

[0101] The specific implementation process is as follows: Taking the root health monitoring of ancient and famous trees using this scheme as an example, the specific process is as follows:

[0102] First, deploy and configure the system parameters:

[0103] The multispectral LiDAR module was fixed on a tripod (height 1.2m, tilt angle 20°), with a scanning radius of 3m and a resolution of 0.5cm;

[0104] The differentiable physics engine uses physics engine solution software and is pre-installed in the data processing and 3D modeling modules.

[0105] Initialization: Set initial parameters based on the tree species database: Ginkgo tree , =0.12、 =1.05、 =0.8;

[0106] Maximum bearing density =2.5g / cm3 (based on historical growth data).

[0107] Secondly, data collection and processing are carried out:

[0108] Dual-wavelength reflectance data, i.e. spectral data, is acquired. The multispectral LiDAR module emits laser at a frequency of 15Hz to generate a reflectance ratio matrix, which is then transmitted to the differentiable physics engine via Wi-Fi 6. After PCA dimensionality reduction, the first three principal components are retained, and more than 80% of soil scattering noise is filtered out.

[0109] The physical model is solved using a differentiable physics engine that updates the root density distribution every 0.5 seconds to calculate the soil resistance gradient. ;

[0110] And local detection When the pressure is >25 kPa, a root rot warning is triggered;

[0111] Finally, output and apply the results:

[0112] The data processing and 3D modeling module generates a 3D root system model, indicated by the red highlighted area. High-risk areas >25kPa.

[0113] Taking the assessment of slope vegetation root coverage using this scheme as an example, the specific process is as follows:

[0114] First, deploy the system and configure the parameters:

[0115] Hardware configuration: Multispectral LiDAR is mounted on a drone, with a flight altitude of 60m, a scanning bandwidth of 120m, and a daily detection area of ​​25 hectares; the onboard computer transmits data to the cloud-based differentiable physics engine in real time.

[0116] Model initialization: settings , =0.18、 =1.3、 =0.6 (applicable to grassy vegetation); maximum carrying capacity =1.8g / cm3;

[0117] Secondly, dynamic monitoring and optimization should be carried out:

[0118] Spectral data processing: Real-time matching of reflectance matrix with GPS coordinates (RTK positioning, accuracy ±1cm), and modeling by 10m×10m grid partitioning; Areas with a value <0.8 are automatically marked as "foreign object interference areas" and excluded from analysis.

[0119] Model solution: The differentiable physics engine updates the root density distribution across the entire region every 2 seconds; when the root coverage is detected to be less than 30%, slope reinforcement suggestions are generated.

[0120] Finally, engineering verification was conducted:

[0121] In slope projects, this system assesses root coverage at 82%, guides precise sowing, and increases vegetation survival rate by 25%; the entire process takes 30 minutes per hectare, which is 15 times more efficient than manual surveying.

[0122] This solution represents a significant improvement in technical performance, economic efficiency, and environmental and social benefits, as detailed below:

[0123] In terms of technical performance:

[0124] This solution can perform high-precision root identification. Through the reflectance ratio model of the multispectral LiDAR module, the accuracy of distinguishing roots from foreign objects reaches 95%, which is a significant improvement over the 70% of traditional ground-penetrating radar (GPR). The principal component analysis (PCA) dimensionality reduction algorithm filters out more than 80% of soil noise, improving the signal-to-noise ratio to 25dB and ensuring the purity of spectral data.

[0125] This scheme employs dynamic modeling, using a Verhulst-Pennes coupled model to quantify the dynamic relationship between root growth and soil resistance. The model's prediction error is <8% (verified by COMSOL simulation), significantly improving upon the error of traditional static models (>30%). Differentiable programming supports automatic optimization. , , , The coefficient and parameter optimization efficiency is improved by 5 times, making it adaptable to complex environments such as drought and floods;

[0126] This solution enables efficient 3D reconstruction, with the entire process from data acquisition to 3D model output taking only 2.5 minutes (in the ancient tree monitoring scenario), which is 40 times more efficient than X-ray CT detection (which takes several hours); the root density distribution resolution reaches 0.5cm, which can accurately locate abnormal areas such as rot and pests (error <0.4cm).

[0127] In terms of economic benefits:

[0128] This solution reduces costs; the multispectral LiDAR module supports in-situ non-destructive testing, reducing the cost per test from 5,000 yuan for CT to 200 yuan, making it suitable for a wide range of engineering applications; the combination of cloud-based physics engine and edge computing reduces hardware investment by 40% (compared to traditional GPU server solutions).

[0129] This solution enhances engineering application value. In slope ecological restoration, the system provides early warnings of areas with weak root systems, reducing reinforcement costs by 30%; and the health monitoring of ancient trees guides precise restoration, reducing maintenance costs by 40%.

[0130] Regarding environmental and social benefits:

[0131] This plan prioritizes ecological protection, employing non-destructive testing techniques to avoid damage to the root systems of ancient trees and extending their lifespan by an estimated 10 years or more. The accuracy of slope vegetation coverage assessment reaches 95%, contributing to a 25% increase in the ecological restoration compliance rate.

[0132] This solution utilizes resources efficiently, with precise irrigation recommendations (based on root density distribution) saving 15% of water and increasing corn yield in agricultural experimental fields by 8%; the dynamic model adapts to changes in soil moisture, reducing the risk of pollution caused by excessive fertilization;

[0133] Furthermore, this solution deeply integrates multiple disciplines, combining multispectral LiDAR with biomechanical equations for the first time to overcome the challenge of non-destructive testing of underground root systems; the differentiable physics engine enables self-optimization of model parameters, forming a closed-loop iterative technology barrier between data and model; this solution supports drone-mounted (25 hectares per day) and fixed deployment (millimeter-level accuracy), covering multiple scenarios such as slope restoration and agricultural monitoring; the specific settings include a standardized data interface (JSON protocol) compatible with mainstream engineering management platforms, making it highly adaptable to implementation.

[0134] This embodiment also provides a multispectral vegetation root identification method based on a differentiable physics engine, using the aforementioned multispectral vegetation root identification system based on a differentiable physics engine.

[0135] The above descriptions are merely embodiments of the present invention. Commonly known structures and characteristics are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are aware of all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, under the guidance of this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A multispectral vegetation root identification system based on a differentiable physics engine, characterized in that, include: Multispectral LiDAR module, differentiable physics engine, data processing and 3D modeling module; The multispectral LiDAR module is used to emit near-infrared laser and short-wave infrared laser to scan the target area, acquire 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 based on the reflectivity ratio matrix and a coupled model, generate a physical model, and update it at preset intervals. The data processing and 3D modeling module is used to generate and output a 3D root system model and health assessment results based on the reflectivity ratio matrix and the physical model. The coupling model adopted is the Verhulst-Pennes coupling model; The Verhulst-Pennes coupled model includes: a root growth rate model and a soil resistance model; Root growth rate models include: in, Root density represents the mass of roots per unit volume of soil. For time, Root density is a variable over time; For maximum load-bearing density, Root growth coefficient For drag sensitivity coefficient, For soil resistance; Soil resistance models, including: in For soil density, Humidity correction factor Humidity influence coefficient For temperature variables, Temperature is a variable over time.

2. The multispectral vegetation root identification system based on a differentiable physics engine according to claim 1, characterized in that, The multispectral LiDAR module uses a dual-wavelength laser to scan the target area and acquire dual-wavelength reflectivity data. The dual-wavelength laser emits near-infrared laser and short-wave infrared laser alternately at its emitting end. The receiver of the dual-wavelength laser detects the echo signals of near-infrared and short-wave infrared laser pulses respectively, and records the reflection intensity. and .

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

4. The multispectral vegetation root identification system based on a differentiable physics engine according to claim 1, characterized in that, The step of generating a reflectivity ratio matrix based on dual-wavelength reflectivity data using a reflectivity ratio classification model includes: For each scan point in the target area, calculate the reflectivity of both wavelengths: ； ； in and These are the initial emission intensities of near-infrared laser and short-wave infrared laser, respectively; they are calibration values. Based on reflectance, a reflectance ratio model is used to calculate the reflectance ratio of all scanning points in the target area, forming a reflectance ratio matrix. The reflectivity ratio model is as follows: ;in The reflectivity of near-infrared laser, The reflectivity of short-wave infrared laser; The resulting reflectivity matrix is: ； matrix elements Indicates the first Okay, number The reflectance ratio of the scan points.

5. The multispectral vegetation root identification system based on a differentiable physics engine according to claim 1, characterized in that, Principal component analysis (PCA) dimensionality reduction algorithm is used to extract the first few principal components.

6. The multispectral vegetation root identification system based on a differentiable physics engine according to claim 1, characterized in that, A differentiable computation graph is constructed using PyTorch, and automatic differentiation optimization is performed. , , , coefficient.

7. The multispectral vegetation root identification system based on a differentiable physics engine according to claim 1, characterized in that, The data processing and 3D modeling module is used to generate a 3D root system model and health assessment results from the reflectivity ratio matrix and physical model, and outputs the following: Based on the reflectance ratio matrix, the vegetation root system is identified, and the reflectance ratio matrix is ​​matched with the location coordinates in real time. The model is then constructed and generated by dividing the model into grid sections of a preset size. Update the three-dimensional root system model based on the physical model; If any indicator in the three-dimensional root system model exceeds a preset threshold, a corresponding health warning or health suggestion will be triggered as a health assessment result.

8. The multispectral vegetation root identification system based on a differentiable physics engine according to claim 7, characterized in that, If the reflectivity ratio is greater than or equal to the first preset ratio, then 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.

9. A multispectral vegetation root identification method based on a differentiable physics engine, characterized in that, The multispectral vegetation root identification system based on a differentiable physics engine as described in any one of claims 1-8 is adopted.

Images