Field day lily phenotype measurement system and method based on multi-view three-dimensional reconstruction

Abstract

The application discloses a field day lily phenotype measurement system and method based on multi-view three-dimensional reconstruction, and belongs to the technical field of three-dimensional phenotype detection. The system comprises a multi-view image acquisition module, a foreground mask segmentation module, a three-dimensional reconstruction module, a point cloud preprocessing module, a ground point cloud stripping module, a three-dimensional phenotype analysis module and a time sequence phenotype state mining module. The multi-view image acquisition module is used for acquiring multi-view image data of the whole day lily before bolting and pre-processing the images. The foreground mask segmentation module is used for foreground mask extraction on the pre-processed images to construct a pure multi-view image data set. The three-dimensional reconstruction module is used for three-dimensional reconstruction based on the pure multi-view image data set to generate a dense point cloud model of the whole day lily. The application solves the problems that the existing phenotype method based on crops with main stems has poor adaptability on day lily, inaccurate measurement and cannot dynamically evaluate the ideal plant type.

Description

Technical Field

[0001] This invention relates to the field of three-dimensional phenotypic detection technology, specifically to a field daylily phenotypic measurement system and method based on multi-view three-dimensional reconstruction. Background Technology

[0002] As an important specialty economic crop, daylily's phenotypic characteristics, such as plant height, canopy width, projected canopy area, and leaf morphology, are crucial for variety selection, evaluation, and field cultivation management. Currently, field phenotypic data collection still relies primarily on traditional methods such as manual measurement, visual assessment, and destructive sampling. These methods suffer from significant drawbacks, including high labor intensity, low efficiency, large subjective errors, and poor repeatability, failing to meet the urgent needs of modern agriculture for large-scale, rapid, and accurate phenotypic measurements. With the rapid development of digital agriculture and sensor technology, vision-based non-contact phenotypic measurement techniques are gaining increasing attention. Existing methods based on 3D point cloud processing have made some progress in plant phenotypic analysis, but still face numerous challenges.

[0003] Single-view images only provide two-dimensional projection information, making it difficult to reflect the true three-dimensional spatial structural features of plants. Depth cameras are prone to noise and data loss under conditions of strong light and shadow in the field. The motion-reconstruction-multi-view stereo vision method is prone to problems such as missing point clouds and surface discontinuities when faced with interference from farmland weeds and soil, and densely overlapping leaves. Although neural radiation fields have strong three-dimensional detail representation capabilities, they are slow to train, consume large computational resources, and are sensitive to dynamic field backgrounds. In recent years, 3D Gaussian splashing has emerged as an emerging 3D representation and reconstruction method, showing certain advantages in detail preservation and reconstruction efficiency. However, due to the slender and overlapping leaves of daylilies, traditional 3D Gaussian splashing tends to parametrically model irrelevant background areas in complex field backgrounds, leading to decreased accuracy in plant body reconstruction and increased free noise.

[0004] Most existing crop phenotypic analysis algorithms are based on the prior assumption of "vertical uniaxial growth with a distinct main stem," and are mainly applicable to gramineous crops such as corn and wheat. Before bolting, daylilies exhibit a typical rosette-like, ground-hugging clump growth pattern without a distinct main stem, with leaves radiating outwards from the base of the shortened stem. This unique biological structure makes traditional algorithms poorly adaptable when directly applied to daylilies: two-dimensional convex hull algorithms struggle to characterize the irregular gaps between leaves, leading to a significant overestimation of the projected area; methods based on main stem cylinder fitting or conventional principal component analysis are prone to biases in determining the principal direction, failing to accurately quantitatively describe the "upright" or "spreading" posture of the daylily leaf cluster. Existing methods also largely remain at the level of single, static spatial geometric measurements, making it difficult to describe the dynamic evolution of the three-dimensional structural state of the same plant over continuous growth periods.

[0005] Therefore, the present invention aims to provide a field daylily phenotypic measurement system and method based on multi-view three-dimensional reconstruction to solve the defects existing in the prior art. Summary of the Invention

[0006] The purpose of this invention is to provide a field daylily phenotypic measurement system and method based on multi-view 3D reconstruction, in order to solve the problems of missing single-view image information, the large amount of noise in traditional 3D reconstruction methods in farmland background and difficulty in measuring irregular volumes, the difficulty in customized 3D phenotypic analysis of daylily without a main stem and rosette structure, and the lack of dynamic state mining and candidate ideal plant type evaluation methods for continuous multi-week 3D phenotypic data.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution, specifically including: A field phenotypic measurement system for daylilies based on multi-view 3D reconstruction, comprising: The multi-view image acquisition module is used to acquire multi-view image data of the whole daylily plant before bolting, and to preprocess the images; The foreground mask segmentation module is used to extract the foreground mask from the preprocessed image, filter out field background interference, and construct a clean multi-view image dataset. The 3D reconstruction module is used to perform 3D reconstruction based on the clean multi-view image dataset to generate a dense point cloud model of the whole daylily plant. The point cloud preprocessing module is used to denoise the dense point cloud and extract the main body of the plant to obtain the main point cloud of the daylily plant. The ground point cloud stripping module is used to strip the ground point cloud from the main point cloud of the plant to obtain a pure plant point cloud. The three-dimensional phenotypic analysis module is used to automatically analyze the three-dimensional phenotypic traits of daylily based on the point cloud of the pure plant. The three-dimensional phenotypic traits include plant height, main effective crown width, projected crown area, three-dimensional crown volume, spatial light interception potential index and leaf posture morphology parameters. The time-series phenotypic state mining module is used to construct the three-dimensional phenotypic parameters within a continuous monitoring period into a plant-week state vector, identify the three-dimensional phenotypic state categories at different growth stages and generate a single plant state transition trajectory, construct a candidate ideal plant type evaluation model and output the candidate ideal plant type ranking results.

[0008] A method for measuring the phenotypic characteristics of daylilies in the field based on multi-view 3D reconstruction includes the following steps: S1. Multi-view image acquisition and preprocessing: Acquire multi-view image data of the whole daylily plant before bolting and perform distortion correction and photometric consistency compensation. S2. Foreground mask segmentation: The U²-Net salient object detection network is used to extract the foreground mask of the preprocessed image, and the field background interference is filtered out by pixel-by-pixel Hadamard product operation. S3, 3D reconstruction: Based on the structure-of-motion algorithm, the camera spatial pose is restored and sparse point cloud is generated. A differentiable 3D Gaussian splash rendering mechanism is introduced to guide joint optimization with foreground mask to generate a dense point cloud model of the whole daylily plant. S4. Point cloud preprocessing: Statistical outlier filtering for noise reduction and density-based spatial clustering for subject extraction are performed sequentially on dense point clouds. S5. Ground point cloud stripping: The ground reference equation is obtained by adaptive voxel downsampling and random sampling consistency plane fitting, and then mapped back to the original point cloud for point-by-point physical orthogonal distance determination to achieve in-situ stripping of ground point cloud. S6. Three-dimensional phenotypic analysis: Based on pure plant point cloud, the plant height, main effective crown width, projected crown area, three-dimensional crown volume, spatial light interception potential index and leaf posture morphology parameters are automatically analyzed. S7. Time-series phenotypic state mining and candidate ideal plant type evaluation: The three-dimensional phenotypic parameters within the continuous monitoring period are constructed into plant-week state vectors. Unsupervised clustering is used to identify the three-dimensional phenotypic state categories and generate state transition trajectories. A candidate ideal plant type evaluation model is constructed and the ranking results are output.

[0009] In steps S1 and S2, the image acquisition device uses a motion camera to construct a continuous sequence of images based on a surrounding viewpoint trajectory of the daylily plant's geometric center. To balance the efficiency of high-throughput data acquisition in the field with the quality of 3D reconstruction, a single-ring track acquisition strategy is adopted in the image acquisition process. By setting a single fixed downward tilt angle, a single 360° surround shot is performed to efficiently acquire a complete multi-view image sequence that can characterize the main canopy morphology and the spatial interlacing structure of the leaves. The acquired video sequence data undergoes preprocessing operations, including equal-interval video frame extraction, optical distortion correction, and photometric consistency compensation. Based on this, a U²-Net salient object detection network is introduced to extract a high-resolution foreground mask of the target plant. Based on the unique double-layer nested U-shaped network topology of this model, it has excellent multi-scale feature extraction and fusion capabilities. It can accurately segment and retain the complex leaf boundaries of daylilies that are long and severely overlapping. Thus, it can actively strip and effectively filter out complex background interference such as weeds, soil and light and shadow in natural farmland environment in the two-dimensional image domain, and then construct a clean multi-view image dataset with high signal-to-noise ratio for three-dimensional reconstruction.

[0010] In step S3, the multi-view image data acquired in step S1 is subjected to 3D reconstruction processing to generate a 3D point cloud model of the entire daylily plant. During the reconstruction process, the camera spatial pose of the multi-view image sequence is first recovered based on a feature matching algorithm. Then, based on the generated sparse point cloud, a differentiable 3D Gaussian rendering mechanism is introduced to perform model densification processing. Thanks to the high-precision mask output by the previous U²-Net model, which effectively strips away the complex background in the field, the differentiable Gaussian rendering process can concentrate the parameterized representation of anisotropic 3D Gaussian primitives primarily on the daylily plant itself. This mechanism effectively avoids excessive consumption of optimization computing power by invalid background regions, fundamentally suppressing spatial levitation artifacts and detached noise caused by background mismatch, thereby reconstructing a 3D geometric structure of the daylily with high spatial continuity and rich detail, accurately restoring the spatial morphological features of key organs such as its shortened stem and basal radial overlapping leaves. This step achieves automated conversion from 2D multi-view images to a 3D point cloud model, significantly improving the high-quality reconstruction accuracy of plant structural features.

[0011] In step S4, the generated dense point cloud undergoes cascaded denoising and subject separation preprocessing to significantly improve the signal-to-noise ratio and spatial continuity of the 3D data. Specifically, a statistical outlier filtering mechanism is first introduced to effectively remove sparse, free noise points that deviate from the global statistical distribution, addressing multi-view reconstruction errors introduced by drastic changes in field lighting and wind-induced leaf swaying. Subsequently, for the dense, blocky flying points and isolated suspended matter remaining after preliminary denoising, a density-based spatial clustering algorithm is used to perform point cloud connectivity topology analysis, extracting the largest connected cluster containing the most points as the daylily plant body. This cascaded processing strategy can accurately remove complex farmland environmental disturbances, achieving high-quality purification of the target plant point cloud while preserving the fine geometric topology of daylilies without loss, providing a reliable data benchmark for subsequent high-precision extraction of phenotypic parameters.

[0012] In step S5, a more accurate ground point cloud stripping mechanism based on a cascaded adaptive scale downsampling and random sampling consistency algorithm is introduced. This mechanism first extracts the bounding box scale of the point cloud in 3D space, adaptively calculates the voxel size, and performs downsampling to obtain a low-resolution point cloud that accelerates fitting and has good noise resistance. Then, robust fitting of the globally optimal ground reference plane is performed in the low-resolution space to extract the ground plane equation. Finally, this reference equation is losslessly mapped back to the original high-resolution dense point cloud, and point-by-point segmentation is performed by calculating a physical orthogonal distance threshold. This scheme balances the processing efficiency and accuracy of large-scale point clouds, achieving efficient in-situ stripping of the natural environment's ground background while strictly preserving the original spatial pose and high-resolution detail features of the target point cloud, thus obtaining pure 3D data of the daylily plant.

[0013] In step S6, after the accurate stripping of the ground point cloud in step S5, the system extracts a high-purity three-dimensional point cloud model of the daylily plant. Based on this high-quality point cloud data, the system can automatically analyze and quantify the key macroscopic phenotypic traits of the entire daylily plant, specifically covering plant height, crown width, projected crown area, three-dimensional crown volume, spatial light interception potential index, and leaf morphological parameters. Specifically, plant height is measured by extracting the orthogonal distance between the maximum extreme value of the target point cloud in the vertical Z-axis direction and the reference plane; crown width measurement introduces a crown truncation enhancement model, which strips away interference from scattered leaves at the bottom layer by setting a height truncation ratio coefficient, thereby accurately quantifying the maximum horizontal span of the plant in the upper part of the point cloud subset. For the analysis of spatial area and volume, the system employs a high-resolution two-dimensional spatial grid mapping algorithm to statistically determine the occupied grid and obtain the projected canopy area. Simultaneously, it introduces a three-dimensional Alpha-Shape edge wrapping algorithm to extract the three-dimensional concave geometric shell of the plant body to calculate the three-dimensional canopy volume, thereby constructing a spatial light interception potential index based on the ratio of projected area to volume. This mechanism effectively overcomes the inherent spatial overestimation defect of traditional convex hull fitting algorithms when dealing with complex pores. For the quantitative assessment of leaf posture morphology, principal component analysis is introduced to perform spatial structure tensor analysis on the pure point cloud. Given the rosette-shaped distribution of daylilies before bolting, the system defines the ratio of the third eigenvalue to the sum of all eigenvalues ​​as the uprightness index. Combined with parameters such as the vertical skew angle of the principal axis, this avoids the dimensional mismatch of conventional crop phenotyping algorithms, achieving a quantitative characterization of the leaf spread and upright morphology of daylilies.

[0014] In step S7, the system introduces an unsupervised machine learning mining method for multi-period three-dimensional point cloud phenotypic data to construct a three-dimensional phenotypic state transition analysis and candidate ideal plant type evaluation model for daylilies. Specifically, the system standardizes the parameters output from step S5 within the continuous monitoring period, such as plant height, effective canopy width, projected canopy area, three-dimensional canopy volume, spatial light interception potential index, uprightness index, principal axis vertical skew angle, and anisotropy. It then constructs a plant-week three-dimensional phenotypic state vector using a single plant and a single week as the basic analysis unit, thus forming a state dataset containing plant number, monitoring week, and multi-dimensional phenotypic parameters. Subsequently, the system employs unsupervised machine learning algorithms such as principal component analysis, Gaussian mixture model, K-means clustering, or hierarchical clustering to identify different three-dimensional phenotypic states in the reduced-dimensional phenotypic space. These states characterize the spatial structural features of daylilies, such as upright structure, high canopy volume, wide canopy expansion, or high light interception potential. Furthermore, the system generates a state transition sequence based on the temporal changes in the phenotypic state of the same plant within consecutive monitoring weeks. This sequence characterizes the dynamic developmental pattern of daylily evolving from an early phenotypic state to a dominant spatial structure state in the middle and later stages. Based on the state identification results and the state transition sequence, the system establishes a candidate ideal plant type evaluation index system, including plant height, crown width, projected crown area, three-dimensional crown volume, uprightness index, main axis posture stability, and spatial light interception potential index, according to the three-dimensional phenotypic parameters of the final monitoring week and the dynamic change characteristics throughout the entire monitoring period. The system then calculates the comprehensive evaluation value of each plant according to preset weights or similarity measurement rules to determine the scoring and ranking results of the candidate ideal plant types. This step reduces the reliance on a large number of manually labeled samples and enables the comprehensive mining of three-dimensional plant structure, dynamic developmental patterns, and candidate ideal plant types in daylilies under multi-stage and weakly labeled conditions. This provides interpretable quantitative evidence for early screening of daylily varieties, construction of ideal plant types, and breeding decisions.

[0015] Compared with the prior art, the beneficial effects achieved by the present invention are: This invention extracts the foreground mask through the U²-Net salient object detection network and guides the joint optimization process of 3D Gaussian splashing, so that the Gaussian meta-parameterization is strictly constrained to the plant body region, effectively avoiding the erroneous modeling of the field background region by traditional 3D Gaussian splashing, significantly improving the accuracy and purity of 3D reconstruction in densely overlapping leaf scenes, and suppressing the generation of suspension artifacts and free noise from the source.

[0016] To address the unique biological structure of daylilies before bolting, characterized by a rosette-like, ground-hugging growth pattern without a distinct main stem, this study abandons conventional algorithms based on the prior assumption of "vertical uniaxial growth" and proposes a customized three-dimensional phenotypic analytical model. Through a canopy truncation enhancement model, a two-dimensional spatial mesh mapping method, and a three-dimensional Alpha-Shape concave hull algorithm, the model achieves accurate measurements of the effective canopy width, the true projected canopy area, and the three-dimensional canopy volume, effectively overcoming the inherent defect of traditional convex hull algorithms in spatial overestimation under complex pore structures.

[0017] In leaf posture morphology assessment, the ratio of the third eigenvalue of principal component analysis to the sum of all eigenvalues ​​is defined as the uprightness index. Combined with the vertical deviation angle of the principal axis, a three-dimensional non-destructive quantitative characterization of the uprightness and openness of daylily leaves is achieved. This breaks through the technical bottleneck of conventional algorithms based on the first principal component direction in determining the principal direction in crops without a main stem, and provides a customized quantitative index for the leaf posture evaluation of rosette-shaped crops.

[0018] A complete closed loop was constructed, from 3D reconstruction to temporal state mining to ideal plant type evaluation. The 3D phenotypic parameters within the continuous monitoring period were organized into plant-week state vectors. Typical 3D phenotypic state categories were automatically identified and single plant state transition trajectories were generated through unsupervised clustering. Under weak labeling conditions, the automated analysis of the dynamic development pattern of the 3D plant type of daylily and the quantitative ranking of candidate ideal plant types were realized. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart illustrating the overall architecture of the field daylily phenotypic measurement system based on multi-view three-dimensional reconstruction of the present invention. Figure 2 This is a schematic diagram of the data acquisition device and spatial viewpoint of the present invention; Figure 3 These are comparison images of the background stripping effect based on U²-Net in this invention; Figure 4 This is a flowchart of the Gaussian splash reconstruction process based on mask guidance of the present invention; Figure 5 This is a flowchart of the dense point cloud cascaded noise reduction and main body fine separation process of the present invention; Figure 6This is a schematic diagram of the automatic extraction of three-dimensional phenotypic parameters in this invention; wherein, (a) is an analytical schematic diagram of plant height and maximum canopy span, (b) is a schematic diagram of spatial grid discretization covering method, (c) is a schematic diagram of three-dimensional spatial covariance matrix and eigenvalue decomposition, (d) is a schematic diagram of three-dimensional canopy volume, (e) is a schematic diagram of spatial light interception potential index, (f) is a correlation analysis diagram of plant height measurement, (g) is a correlation analysis diagram of canopy width measurement, and (h) is a correlation analysis diagram of projected canopy area measurement; Figure 7 This is a flowchart of the plant-week three-dimensional phenotypic state mining process of the present invention; Figure 8 This is a schematic diagram of the three-dimensional phenotypic state transition trajectory and candidate ideal plant type scoring and ranking of the present invention; Figure 9 This is a flowchart of the steps of the field daylily phenotypic measurement method based on multi-view three-dimensional reconstruction of the present invention. Detailed Implementation

[0020] The technical solutions of 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.

[0021] Please see Figures 1-9 The present invention provides the following technical solution: like Figure 1 As shown, the multi-view image acquisition module is responsible for acquiring multi-view image data of daylily plants in the field. The hardware components include: a motion camera, an adjustable aluminum alloy rotating bracket, a stepper motor and its driver, an STM32F407 microcontroller, and a laser alignment device. During acquisition, the bracket is fixed next to the plant, and the laser alignment device is used to adjust the rotation center to coincide with the center of the plant's base. The camera's pitch angle is set to 35°, and the stepper motor rotates at a uniform angular velocity of 6° / s to complete a single 360° surround acquisition. The acquisition frame rate is 30fps, and the acquisition time per plant is 60 seconds, acquiring approximately 1800 frames of raw video images. The acquired data is uploaded to a cloud server in real time via a 5G wireless router.

[0022] The foreground mask segmentation module is deployed on a cloud server and receives the raw image sequence from the acquisition module. The module internally loads a pre-trained U²-Net salient object detection network. After image input, the network performs forward inference and outputs a single-channel foreground mask image with the same resolution as the input image, with pixel values ​​ranging from [0,1]. Subsequently, a Hadamard product operation is performed, multiplying the mask pixel-by-pixel with the original RGB image to generate a clean image sequence after background removal, which is then stored in the cloud database.

[0023] The 3D reconstruction module runs on the same cloud server, performing motion-based structure recovery and implementing differentiable 3D Gaussian splash rendering based on the PyTorch framework. The module first extracts SIFT feature points from the clean image sequence, performs feature matching and geometric verification, and then optimizes the camera pose and sparse point cloud using bundle adjustment. Subsequently, using the sparse point cloud as initialization, it constructs an anisotropic 3D Gaussian primitive set, projects the Gaussian primitives onto a 2D plane using a differentiable renderer, calculates mask-guided photometric loss and structural similarity loss, and generates a dense point cloud through 300 iterations using the Adam optimizer.

[0024] The point cloud preprocessing module performs cascade purification on the reconstructed dense point cloud, the ground point cloud stripping module performs ground stripping on the main point cloud, the three-dimensional phenotypic analysis module automatically calculates phenotypic parameters based on the pure plant point cloud, and the time-series phenotypic state mining module organizes the phenotypic parameters of each plant per week within the continuous monitoring period.

[0025] The signal flow between the modules is as follows: multi-view image acquisition module, foreground mask segmentation module, 3D reconstruction module, point cloud preprocessing module, ground point cloud stripping module, 3D phenotypic analysis module, and time-series phenotypic state mining module, forming a complete data processing pipeline.

[0026] like Figure 2 As shown, this embodiment uses daylily plants before bolting as the research object, systematically acquiring whole-plant image data and the true values ​​of key phenotypic traits such as plant height and crown width. The specific implementation method is as follows: Step S1: Refined acquisition of multi-view image data of daylilies and obtaining phenotypic true values. A stratified random sampling design was adopted, following the spatial pattern of daylily planting in the field, dividing the area into several independent sample areas in a checkerboard or strip layout, and strictly controlling the spatial interval of the sample areas to eliminate edge effects. A number of daylily plants were randomly selected from each sample area as measured samples. The data acquisition device and spatial viewpoint distribution are shown below. Figure 2 As shown, during the image acquisition phase, a motion camera mounted on an adjustable aluminum alloy rotating bracket was used as the visual sensor. During acquisition, the rotation center of the bracket was precisely aligned with the geometric center of the target daylily plant, constructing a horizontal circular track with a radius of approximately 0.5 meters. To balance high-throughput phenotypic measurement efficiency in the field with the accuracy of 3D feature point calculation, the camera was set to perform single-loop track acquisition with a single fixed tilt angle. Under this fixed perspective, the camera was driven at a constant speed by a stepper motor to complete a single 360° continuous circular shot. This single tilt angle spatial configuration maximizes the balance between the visual visibility of the daylily's top canopy projection and the lateral leaf edge features. While ensuring that adjacent frames have sufficient common feature points to support subsequent motion-based structural reconstruction, it significantly shortens the physical acquisition cycle of single-plant data and reduces the computational cost of subsequent calculations.

[0027] After acquiring continuous video sequence data for a single concentric layer, the FFmpeg algorithm library is first used to perform equally spaced sampling on the acquired video stream, extracting approximately 200-300 still images per plant to form an initial multi-view image sequence. Subsequently, to address the edge perspective distortion problem caused by wide-angle lens shooting, optical distortion correction is performed using the OpenCV algorithm library with pre-calibrated camera intrinsics. Furthermore, photometric consistency compensation processing is performed on the extracted image sequence, specifically employing algorithms such as adaptive histogram equalization and white balance correction to eliminate local overexposure and underexposure caused by highlights, shadows, or backlighting.

[0028] Step S2: The preprocessed image sequence still contains a large amount of farmland background interference, so a complex background active stripping mechanism based on U²-Net needs to be introduced. The processed image is input into a pre-trained U²-Net salient object detection network for forward inference. The network inference finally outputs a high-precision single-channel foreground mask image with the same resolution as the input image, where the foreground pixels belonging to daylily plants are close to 1, while the background pixels such as weeds, soil, and light and shadow are close to 0.

[0029] Finally, the mask image is multiplied pixel by pixel using the Hadamard product operation, which is mathematically expressed as follows: , where I original (x,y) represents the original pixel value at coordinates (x,y), Mask(x,y) represents the mask weights output by the network, and I pure (x, y) represents the clean pixel values ​​after background stripping, and ⊙ represents the Hadamard product. This operation sets complex background pixels from the natural environment as background values, thereby constructing a clean multi-view image dataset with extremely high signal-to-noise ratio specifically for 3D reconstruction. The background stripping process and effect comparison are shown in the figure below. Figure 3 As shown, this eliminates the reprojection error caused by background noise during feature matching and Gaussian parameter optimization at the source.

[0030] Plant height and crown width were measured: For plant height, the linear distance from the base of the plant to the highest point of the crown was measured vertically; for crown width, the maximum lateral span was measured on the horizontal projection surface of the plant. To reduce random observation errors, both indicators were measured n times for each plant, and the arithmetic mean was taken as the final benchmark verification data. The calculation formula is as follows: ,in V is the arithmetic mean of the target phenotypic traits of a single daylily plant. i Let n be the physical measurement value of the i-th measurement. In this embodiment, the number of measurements is n=3.

[0031] Preliminary observations of projected canopy area and leaf posture parameters were conducted by acquiring preliminary projected images through vertical overhead photography to record the plant's coverage area on the horizontal plane. Simultaneously, morphological records were made of the spatial tilt angle and growth potential of typical leaves as a reference for subsequent algorithm comparisons.

[0032] Step S3: High-quality reconstruction of plants using COLMAP pose calculation and mask-guided 3D Gaussian splashing. The specific algorithm implementation process is as follows... Figure 4 As shown.

[0033] After obtaining the clean multi-view image sequence and its corresponding foreground mask output in step S2, this step, in conjunction with incremental motion recovery structure and differentiable 3D Gaussian rendering mechanism, achieves high-precision 3D geometric reconstruction of daylily plants before bolting. The clean multi-view image sequence is input into COLMAP software. The system first uses the scale-invariant feature transform algorithm to extract local key points of each frame image, and performs cross-view Euclidean distance matching based on feature descriptors. Based on the matched feature point pairs, bundle adjustment is used to jointly optimize camera parameters and 3D point coordinates. Let X... j Let x be the j-th point in three-dimensional space. ij Given the two-dimensional observation pixel coordinates of the i-th camera viewpoint, the system accurately solves for the camera's intrinsic parameter matrix K and rotation matrix R by minimizing the overall reprojection error E. i With translation vector t i :

[0034] Where π(·) represents the perspective projection function, and ρ(·) is the Huber robust loss function, used to suppress interference from mismatched points. After optimization convergence, the system outputs the initial sparse 3D point cloud of the daylily plant and the precise spatial pose of each frame, providing crucial geometric priors and initial centroids for subsequent dense reconstruction. Using the sparse point cloud output by COLMAP as the initial position set, the 3D continuous structure of the daylily before bolting is discretized into a set of anisotropic 3D Gaussian ellipsoids. For the Kth anisotropic 3D Gaussian element, its probability density distribution function G(x) in space is defined as:

[0035] Where x is the coordinate of the spatial sampling point, μ k Let ∑ be the three-dimensional center coordinates of the Gaussian sphere. k Let be the three-dimensional covariance matrix. To ensure that the covariance matrix maintains its positive semi-definite property throughout the differentiable optimization process and can accurately fit the anisotropic characteristics of the slender and specifically curved daylily leaves, ∑ k Decomposed into the product of the scaling matrix S and the rotation matrix R: In addition, each Gaussian sphere is assigned an opacity parameter α.k And the view-dependent color parameter c based on the spherical harmonic function. k To realistically reproduce the light and shadow reflection characteristics of plant leaf surfaces.

[0036] Given an arbitrary camera viewpoint, an anisotropic 3D Gaussian primitive is projected onto a 2D image plane using an affine approximation method. Its covariance matrix in the 2D pixel coordinate system is obtained. The calculation is as follows:

[0037] Where W is the transformation matrix from the world coordinate system to the camera view coordinate system, and J is the Jacobian matrix of the perspective projection transformation. Pixel colors are calculated using a point-based alpha blending rendering technique. For any pixel on the screen, depth-ordered cumulative blending is performed along the ray direction that passes through the Gaussian set to calculate the rendered color C of that pixel.

[0038]

[0039] Where N is the number of ordered Gaussian spheres participating in the rendering of this pixel, c i With α i Let U be the color and opacity of the i-th projected Gaussian sphere at that pixel. The predicted image generated by the above volume rendering process is compared with the actual acquired image obtained in step S1. The loss is calculated to compute the gradient and update the Gaussian parameters (μ, S, R, α, c) in reverse. To effectively reduce floating artifacts and incorrect optimizations caused by complex backgrounds such as weeds and soil in the field, this embodiment deeply integrates U from step S1 into the traditional loss function. 2 -Net outputs a high-precision foreground mask matrix M. Defines the mask-guided photometric consistency loss L. 1_mask Structural similarity loss L guided by mask SSM_mask : , , in The predicted image generated for volume rendering, I GT For realistic multi-view imagery, ⨀ denotes the Hadamard product of matrices. The overall objective function is L. total This is a weighted combination of the two.

[0040]

[0041] Here, λ is the balance coefficient. Since the mask M forces the values ​​of the background region to be 0, the objective function forces the gradient-based adaptive density control to be strictly constrained within the effective region of the daylily plant body (M=1). After a specified number of iterations of optimization, a high-density 3D reconstruction model of daylily with high-quality leaf edges and a relatively accurate spatial interlacing structure is finally generated.

[0042] Step S4: Cascaded denoising and refined separation of the plant body from the dense point cloud. After obtaining the initial dense point cloud model of daylily in step S3, due to variations in natural field lighting, slight wind-induced leaf swaying, and inherent computational biases in multi-view reconstruction, the point cloud often contains free noise and suspended specks. To remove noise, reduce redundant point cloud density, and optimize the continuity and structural integrity of daylily organ surfaces, the specific cascaded preprocessing flow is as follows: Figure 5 As shown, the following cascaded preprocessing operations are performed on the point cloud to perform statistical outlier filtering on the dense point cloud model, in order to more accurately remove isolated sparse noise points introduced by light reflection or blade vibration. For any point p in the point cloud set... i Calculate the average Euclidean distance d between it and its k nearest neighbors. i Calculate the mean distance μ and standard deviation μ of the global point cloud. Determine the point p. i The mathematical model for determining whether a point is an outlier or noise point is as follows: Where α is the distance threshold coefficient, used to control the intensity of outlier removal. If point p i Average neighborhood distance d i If the above conditions are met, it will be identified as an outlier and physically removed.

[0043] After removing sparse noise, a density-based spatial clustering algorithm is used to perform point cloud connectivity topology analysis on the remaining dense, blocky flying points and isolated suspended objects. The spatial neighborhood search radius R is set. eps Minimum nearest neighbor threshold N of the core point min Spatially connected points are grouped into the same cluster C. j In this process, clusters are formed. Count the number of point clouds contained within each cluster |C j | and extract the largest connected cluster C with the most points. target The extraction criteria for the main point cloud of a complete daylily plant are as follows: The remaining small, independent clusters were removed as environmental disturbances, thus achieving high-quality stripping of the target plant. To further focus on the effective canopy area of ​​the daylily plant and improve the computational efficiency of subsequent phenotypic extraction, C... target Perform spatial region of interest clipping and downsampling. By calculating the 3D bounding box, remove irrelevant areas such as the bottom ground reference plate. If the 3D coordinates of the point cloud are (x...i ,y i ,z i If the effective ROI is selected, then the following conditions must be met: Z ground The ground threshold is obtained statistically from the elevation histogram, and ∆h is the error buffer margin. Voxelized downsampling is performed while preserving the cross-geometric structure of daylily leaves before bolting. The three-dimensional space is divided into a voxel grid, and all points within the grid are replaced by the geometric centroids within the voxels. To ensure that the voxel size adaptively matches the structural scale of the slender daylily leaves, the voxel scale V... s Defined as: , where d min Let p be the minimum nearest neighbor distance of the local point cloud, and γ be the scaling factor. To eliminate the step effect caused by downsampling and improve the continuity and smoothness of the point cloud surface, a three-dimensional Gaussian filter is used to locally weight and adjust the position of the points. i The original coordinates are p i raw Smoothed optimized coordinates p i smooth The expression is:

[0044] Among them, Ω i For point p i The set of spatial neighborhood points. To strictly preserve the curvature variations and detailed topological structure of plant leaf edges while smoothing surface noise, the weight w... ij Using the Gaussian kernel function definition:

[0045] Where, σ w To smooth the bandwidth parameters, the above multi-step cascaded processing ultimately yields a complete point cloud model of the daylily plant, characterized by a clean background, sharp edges, and stable representation.

[0046] Step S5: More accurate stripping of ground point cloud based on adaptive downsampling and RANSAC mapping. After completing the denoising and preliminary extraction of the main body in step S4, in order to effectively reduce the interference of the soil reference surface in the natural farmland environment on the subsequent canopy phenotypic calculation, this embodiment introduces a more accurate stripping mechanism for ground point cloud that takes into account both computational efficiency and high-quality details. The specific implementation steps are as follows: Let the extreme value difference of the bounding box in the X, Y, and Z coordinate axes be L. x ,L y ,L z Then calculate its spatial diagonal scale L. diag : To avoid environmental incompatibility caused by manually setting voxel sizes, the system adaptively calculates the side length of the downsampled voxels based on this diagonal scale. , where k is a preset adaptive scaling factor.

[0047] Voxelization downsampling is performed on high-resolution dense point clouds, replacing all points within the grid with the centroids of the point clouds within voxels. This operation compresses the amount of point cloud data, accelerates the subsequent spatial fitting process, and endows the underlying point cloud with stronger resistance to local deformation and noise, thus outputting a low-resolution point cloud set for fitting. On the low-resolution point cloud obtained by the downsampling, a random sampling consensus algorithm is introduced for globally optimal plane fitting. In each iteration, the algorithm randomly selects three non-collinear spatial points to form an initial plane and calculates the distances from the remaining points to this plane; points whose distances are less than the fitting tolerance are classified as inliers. This process is repeated multiple times, with the number of iterations set to N. iter The system locks the optimal fitted plane model with the most local points, and extracts the normal vector features (A,B,C) and origin offset constant D of the plane model, and then constructs the ground reference plane equation of the target farmland: Ax+By+Cz+D=0.

[0048] After obtaining the global ground plane equation, to avoid the loss of plant root edge details due to downsampling, the system abandons direct segmentation on the low-resolution point cloud and instead maps the ground reference plane equation directly back to the original high-resolution dense point cloud space before downsampling in step S5. This involves traversing all spatial points p in the original dense point cloud. i (x i ,y i ,z i ), calculate the physical orthogonal absolute distance d from each point to the ground reference plane. i The analytical formula for calculating its distance is: The physical distance separation threshold T is set based on empirical statistical values ​​of farmland surface undulations. ground Based on this, a point cloud segmentation criterion is constructed, which satisfies the orthogonal distance d. i <T ground Points that are determined to be invalid ground background point clouds are completely removed; points that satisfy the distance d are... i ≥T ground Points identified as valid plant tissues were extracted in situ within the original 3D spatial coordinate system. This cascaded processing workflow, while strictly maintaining the original point cloud spatial pose and absolute scale without any shift, achieved relatively accurate stripping of farmland ground point clouds. The final output was a high-quality, pure 3D point cloud model of the main body of the daylily plant, containing only leaves and shortened stems, laying a data foundation for subsequent stable measurements of plant height, crown width, and projected area.

[0049] Step S6: Automated extraction of daylily phenotypic parameters based on high-quality 3D point clouds. After obtaining the pure daylily plant main body point cloud output in Step S5, the system will automatically extract plant height, crown width, projected crown area, 3D crown volume, spatial light interception potential index, and leaf posture morphological parameters. The overall phenotypic analysis and correlation analysis results are as follows: Figure 6 As shown. The specific measurement steps and underlying model are as follows: Plant height is calculated using the spatial coordinate distribution characteristics of point clouds in the vertical direction. Let the pure point cloud set be P={(x... i ,y i ,z i The tree height H of the reconstructed point cloud is calculated by extracting the maximum extreme value in the z-axis direction. max To address the issue of canopy width measurement bias caused by the random dispersion of basal leaves in daylilies, a canopy truncation enhancement model is introduced. A canopy height truncation ratio coefficient β is set, and the effective canopy height threshold is calculated. Extract the upper half of the point cloud subset P whose height is greater than the threshold using a Boolean mask. upper . In P upper In the subset, calculate the spatial span W in the X-axis and Y-axis directions respectively. x With W y :

[0050] The final crown width parameter W is defined based on the actual phenotypic pattern, taking the maximum span. This truncated model effectively avoids interference from scattered lower-level leaves, accurately focusing on the main effective crown width of the daylily. The point cloud analysis of plant height and maximum crown width span is shown below. Figure 6 As shown in (a). Abandoning the traditional two-dimensional convex hull calculation method, which carries the risk of overestimation, this embodiment employs a high-precision spatial grid discretization covering method to calculate the true canopy projection area. The spatial grid discretization covering process is as follows: Figure 6 As shown in (b), the 3D point cloud set is orthogonally projected onto a 2D horizontal reference plane to obtain a 2D point set containing only X and Y coordinates. Set the unit measurement grid resolution to Rgrid, mapping continuous two-dimensional coordinates to a discrete integer grid index matrix:

[0051] After mapping all projected points to the grid, count the total number N of independent grid cells containing at least one data point. unique The projected canopy area S of daylily projected Calculated as the product of the number of independent grid cells and the area of ​​a single grid cell: This mesh-covering analytical method can better adapt to irregular gaps and edge depressions between daylily leaves, significantly improving the fidelity of projected area calculation. To objectively describe the spatial distribution and posture of daylily leaves before bolting, principal component analysis was performed on the pure point cloud. First, the point cloud was centered with zero mean, and a 3×3 three-dimensional spatial covariance matrix Cov was constructed:

[0052] in Let be the coordinates of the geometric center of the point cloud. Singular value decomposition or eigenvalue solving is performed on the variance matrix to obtain three spatial eigenvalues ​​λ1, λ2, λ3 and their corresponding eigenvectors v1, v2, v3, arranged in descending order. Based on this, the following leaf pose morphology quantitative index, anisotropy index, is constructed: used to measure the asymmetry of the plant's spatial extension and the degree of ellipsoidal flattening, and its calculation formula is as follows: .

[0053] Uprightness index: used to quantify the degree of concentration of the overall vertical growth of daylily leaves. Given the biological characteristics of daylilies in a rosette shape before bolting, with the main crown extension located on the horizontal plane, the variance contribution in the vertical direction is characterized by the third principal component (λ3).

[0054] Its formula is defined as .

[0055] Vertical deflection of the principal axis ( ): Calculate the Z-axis vector that is absolutely perpendicular to the largest eigenvector v1. The spatial angle between them: .

[0056] Through the aforementioned PCA feature analysis, the macroscopic sensory leaf pose is transformed into a stable and computable multivariate numerical index. The physical mapping relationship between its three-dimensional spatial covariance matrix and eigenvalue decomposition is as follows: Figure 6 As shown in (c), core quantitative data were provided for the morphological screening and lodging resistance assessment of daylily field germplasm resources.

[0057] Furthermore, the system incorporates an accuracy verification and statistical analysis module, which automatically compares the phenotypic parameters extracted by the aforementioned algorithm with the baseline true values ​​and outputs comprehensive analysis charts. For the physical dimensional measurements of plant height, crown width, and projected crown area, the system automatically calculates the coefficient of determination (R²), root mean square error (RMSE), and mean error. Thanks to the aforementioned "crown truncation enhancement model" and "high-resolution two-dimensional spatial mesh mapping method," the system effectively avoids the overestimation effects of scattered leaves and irregular pores at the bottom layer. Correlation analysis between the automatically extracted values ​​of plant height, crown width, and projected crown area and the manually measured true values ​​is performed, respectively... Figure 6 (f) Figure 6 (g) and Figure 6 The plant height, crown width, and projected crown area measurements are shown in (h). These results include a scatter plot, a Bland-Altman consistency analysis plot, and an error distribution histogram. The results show that the algorithm-extracted values ​​have a very high linear correlation with the actual physical dimensions. The quantitative comparison statistics of specific phenotypic parameters are shown in Table 1. Table 1: Comparison of automatically extracted values ​​and manually measured true values ​​of macroscopic phenotypic parameters of daylily

[0058] For leaf posture morphology assessment, the system outputs the distribution statistics of the uprightness index, anisotropy index, and principal axis vertical skewing angle of the daylily germplasm population based on the analysis of the eigenvectors of the covariance matrix. The system automatically interprets morphological characteristics through quantitative logic: plants with higher uprightness indices have smaller calculated principal axis vertical skewing angles, exhibiting better uprightness and light interception potential in three-dimensional space. The specific leaf posture quantitative output statistics for the selected daylily samples are shown in Table 2.

[0059] For the accurate calculation of the three-dimensional canopy volume and spatial light interception potential index of daylily: Traditional two-dimensional images cannot obtain the true biomass volume of the crop. This embodiment uses a three-dimensional Alpha-Shape edge wrapping algorithm to extract a three-dimensional concave geometric shell from the clean plant point cloud after removing the ground. The point cloud envelope and volume calculation process of the three-dimensional canopy volume are as follows: Figure 6 As shown in (d).

[0060] Perform a 3D Delaunay tetrahedral partitioning on the pure plant point cloud set P to generate a set containing several tetrahedra T. k The set. Set an adaptive rolling ball radius parameter α, and calculate T for each tetrahedron. k circumsphere radius R k Extracting elements that satisfy R k <α tetrahedrons constitute the set of points V within the fine three-dimensional geometric envelope of daylily. alpha : For set V alpha By integrating and summing the volumes of all effective tetrahedrons within the canopy, the true three-dimensional canopy volume Vol above ground can be calculated: Based on this, and combined with the previously obtained projected area, a spatial light interception potential index for daylilies is constructed. .

[0061] This index characterizes the horizontal coverage capacity per unit three-dimensional canopy volume of the plant. A higher LIE value indicates that the daylily leaves are arranged more planarly in space, providing a novel three-dimensional quantitative perspective for evaluating its photosynthetic potential and spatial competitiveness. The geometric representation of the spatial light interception potential index is as follows: Figure 6 As shown in (e).

[0062] Step S7: Construction of candidate ideal plant type based on plant-week 3D phenotypic state mining. After obtaining the plant height, crown width, projected crown area, 3D crown volume, spatial light interception potential index, and leaf posture morphology parameters output in step S6, the system further performs temporal organization and unsupervised mining on the daylily 3D phenotypic data within the continuous monitoring period. Specifically, the multi-dimensional 3D phenotypic parameters of the same plant under different monitoring weeks are used as plant-week phenotypic state units to construct a 3D phenotypic state dataset that can reflect the dynamic growth process of daylily. The overall processing flow of plant-week 3D phenotypic state mining and candidate ideal plant type construction is as follows: Figure 7 As shown.

[0063] Construct a plant-week three-dimensional phenotypic state vector. Let the three-dimensional phenotypic feature vector of the i-th daylily plant in the t-th monitoring week be... H i,t Let W be the height of the i-th daylily plant in week t. i,t As the main effective crown width, A p,i,t V is the projected canopy area. c,i,t For the three-dimensional canopy volume, UI i,t For uprightness index, AI i,t θ is the anisotropy index. i,t Main axis vertical deflection angle, LIE A,i,t and LIE S,i,t These are indicators related to the spatial light interception potential index. Each x... i,t Each corresponds to a single plant's three-dimensional phenotypic state at a specific cycle. To eliminate differences in dimensions and numerical scales among different phenotypic indicators, the above three-dimensional phenotypic state vectors are standardized.

[0064] Let the mean of the k-th phenotypic index in all plant-week samples be μ. k The standard deviation is σ k Then the standardized k-th eigenvalue is Among them, Z i,t,k x represents the standardized phenotypic eigenvalues. i,t,k Here, represents the original phenotypic feature values, and ε is a minimal constant to prevent the denominator from being zero. After standardization, the plant-week three-dimensional phenotypic state vector can be obtained: Where m is the number of three-dimensional phenotypic indicators involved in state mining. The standardized vectors of all plants across all monitoring weeks are arranged row-wise to construct the three-dimensional phenotypic state matrix Z: , where n is the number of daylily plants, T is the number of consecutive monitoring weeks, and each row in Z represents a plant-week three-dimensional phenotypic state unit.

[0065] To reduce redundancy among multidimensional phenotypic indicators and extract the dominant comprehensive features characterizing the three-dimensional plant morphology differences of daylilies, the system employs principal component analysis to reduce the dimensionality of the state matrix Z. Let C be the covariance matrix of the standardized state matrix Z, then... , where N z This represents the total number of plant-week state samples. For the covariance matrix C... z Perform eigenvalue decomposition: Where λ1 is the l-th eigenvalue and u1 is its corresponding eigenvector. The eigenvalues ​​are sorted from largest to smallest, and the top d principal components whose cumulative contribution rate reaches a preset threshold are selected to form the projection matrix U. d : Then the low-dimensional three-dimensional phenotypic state of the i-th plant in week t is represented as: , where r i,t This is the dimensionality-reduced plant-week phenotypic state vector, used for subsequent unsupervised state recognition.

[0066] Unsupervised clustering of low-dimensional three-dimensional phenotypic state vectors using the K-means clustering algorithm was employed to identify typical three-dimensional phenotypic states exhibited by daylilies at different growth cycles. Let K be the number of clusters, and S be the set of states in the k-th cluster. k Its cluster center is C k The K-means clustering objective function is:

[0067] Cluster center c k The calculation formula is .

[0068] The status label S for each plant-week sample i,t Determined by its nearest cluster center:

[0069] To determine a reasonable number of clusters K, the system can use the silhouette coefficient to evaluate the clustering results. Let a(q) be the average distance between sample q and other samples in the same cluster, and b(q) be the average distance between sample q and the nearest sample in a different cluster. Then the silhouette coefficient of sample q is: .

[0070] The average silhouette coefficient of all samples is: When Sil reaches a large value and the number of samples in each cluster meets the stability requirements, the corresponding K can be used as the preferred number of categories for three-dimensional phenotypic state identification. Different state categories can respectively characterize the three-dimensional plant structure features of daylilies, such as upright structure type, high canopy volume type, wide canopy expansion type, or high light interception potential type. After obtaining the state label of each plant-week sample, the system constructs the single plant three-dimensional phenotypic state transition trajectory according to the state change sequence of the same plant in consecutive monitoring weeks.

[0071] The state transition sequence of the i-th daylily plant is represented as follows: Among them, T i This is used to describe the dynamic process of the i-th daylily plant evolving from an early growth state to a mid-to-late spatial structure state. Further, a population state transition matrix M is constructed, whose elements M... a,b This represents the number of times a plant transitions from state a to state b.

[0072] Where I(·) is an indicator function, taking a value of 1 when the condition within the parentheses is true, and a value of 0 otherwise. The state transition matrix M can be used to statistically analyze the main transition paths between different three-dimensional phenotypic states in the daylily population. Based on the three-dimensional phenotypic parameters of the final monitoring week and the dynamic change characteristics throughout the entire monitoring period, a candidate ideal plant type evaluation vector is constructed. Let g be the candidate ideal plant type evaluation vector for the i-th daylily plant. i Then there is H i,T W i,T A p,i,T V c,i,T U Ii,T and LIE S,i,T These are the plant height, main effective canopy width, projected canopy area, three-dimensional canopy volume, uprightness index, and spatial light interception potential index, respectively, under the final monitoring week; b i,H b i,A and b i,Vc These represent the slopes of change in plant height, projected canopy area, and three-dimensional canopy volume over a continuous monitoring period; PS i This is an attitude stability index. The slope of the k-th dynamic phenotypic index is obtained by least squares fitting: ,in To monitor the weekly average, is the average value of the kth phenotypic index of the i-th daylily plant across all monitoring weeks.

[0073] Attitude stability index PS i It can be calculated from the degree of variability of the uprightness index over a continuous monitoring period: std(UI)i,1:T Let be the standard deviation of the uprightness index of the i-th daylily plant across all monitoring weeks, and mean(UI) i,1:T (This is its average value.) i The higher the value, the more stable the leaf structure of the plant is within the continuous monitoring period.

[0074] Eliminate the dimensional differences of different indicators in the candidate ideal plant type evaluation vector, and apply g i Each indicator in the data is normalized to 0-1: ,in, Let ω be the normalized value of the k-th evaluation index. Let the evaluation weight vector be ω: The comprehensive score of the candidate ideal plant type of the i-th daylily plant is: i It can be represented as .

[0075] As another implementation method, a candidate ideal plant type reference vector g* can be constructed, and the distance D between the i-th daylily plant and the candidate ideal plant type reference vector can be calculated using weighted Euclidean distance. i :

[0076] According to distance D i Calculate the similarity score of candidate ideal plant type:

[0077] The system ultimately relies on the score. i Output the candidate ideal plant type ranking results from high to low, and simultaneously output the state transition trajectory T of each daylily plant. i Final week phenotypic state category S i,T Overall score of candidate ideal plant type i And the main contribution phenotypic indicators, their state transition trajectories and scoring ranking results are as follows: Figure 8 As shown.

[0078] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0079] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A field daylily phenotypic measurement system based on multi-view 3D reconstruction, characterized in that, include: The system includes a multi-view image acquisition module, a foreground mask segmentation module, a 3D reconstruction module, a point cloud preprocessing module, a ground point cloud stripping module, a 3D phenotypic analysis module, and a temporal phenotypic state mining module. The multi-view image acquisition module is used to acquire multi-view image data of the whole daylily plant before bolting, and to preprocess the images; The foreground mask segmentation module is used to extract the foreground mask from the preprocessed image, filter out field background interference, and construct a clean multi-view image dataset. The 3D reconstruction module is used to perform 3D reconstruction based on the clean multi-view image dataset to generate a dense point cloud model of the whole daylily plant. The point cloud preprocessing module is used to denoise the dense point cloud and extract the main body of the plant to obtain the main point cloud of the daylily plant. The ground point cloud stripping module is used to strip the ground point cloud from the main point cloud of the plant to obtain a pure plant point cloud. The three-dimensional phenotypic analysis module is used to automatically analyze the three-dimensional phenotypic traits of daylily based on the point cloud of the pure plant. The three-dimensional phenotypic traits include plant height, main effective crown width, projected crown area, three-dimensional crown volume, spatial light interception potential index and leaf posture morphology parameters. The leaf posture morphology parameters include at least the uprightness index and the vertical deviation angle of the main axis. The time-series phenotypic state mining module is used to construct the three-dimensional phenotypic parameters within a continuous monitoring period into a plant-week state vector, identify the three-dimensional phenotypic state categories at different growth stages and generate a single plant state transition trajectory, construct a candidate ideal plant type evaluation model and output the candidate ideal plant type ranking results.

2. The field daylily phenotypic measurement system based on multi-view three-dimensional reconstruction according to claim 1, characterized in that, The multi-view image acquisition module uses the geometric center of the target daylily plant as the rotation center, sets a single fixed top tilt angle, and performs a single-layer 360° continuous orbital acquisition of video sequences under the drive of a motor. Then, it performs equal-interval video frame sampling based on the FFmpeg tool, and uses OpenCV to realize camera distortion correction and photometric consistency compensation.

3. The field daylily phenotypic measurement system based on multi-view three-dimensional reconstruction according to claim 1, characterized in that, In the 3D reconstruction module, a motion recovery structure algorithm based on feature point matching is first used to jointly estimate the camera intrinsic and extrinsic parameters of each image in the multi-view image dataset to obtain the spatial pose information of the camera; then, based on the principle of triangulation, the sparse point cloud structure of the daylily plant in 3D space is restored. Finally, based on the sparse point cloud, a differentiable Gaussian splash rendering mechanism is introduced. Through foreground mask constraints, the 3D Gaussian optimization process is concentrated on the plant body region, and the plant structure is continuously parametrically modeled and iteratively optimized to generate a dense point cloud model of the whole daylily plant with high spatial continuity and detail expression capability.

4. The field daylily phenotypic measurement system based on multi-view three-dimensional reconstruction according to claim 1, characterized in that, In the point cloud preprocessing module, statistical outlier filtering determines and removes outliers by calculating the average Euclidean distance of each point within its nearest neighbor range and comparing it with the global statistical distribution; density-based spatial clustering performs connectivity clustering on the point cloud by setting the spatial neighborhood search radius and the minimum nearest neighbor threshold of the core point, and extracts the largest connected cluster with the most points as the main point cloud of the daylily plant.

5. The field daylily phenotypic measurement system based on multi-view three-dimensional reconstruction according to claim 1, characterized in that, In the ground point cloud stripping module, the ground reference plane equation obtained by the random sampling consistency algorithm is Ax + By + Cz + D = 0, and the point-by-point physical orthogonal distance calculation formula is... Points whose distance is less than a preset physical distance threshold are identified as ground point clouds and removed.

6. The field daylily phenotypic measurement system based on multi-view three-dimensional reconstruction according to claim 1, characterized in that, In the three-dimensional phenotypic analysis module, the effective canopy width of the subject is obtained through a canopy truncation enhancement model: a canopy height truncation ratio coefficient β is set, and the effective canopy height threshold is calculated. H max Plant height, extraction height greater than Z threshold After the upper half of the point cloud subset is calculated, the horizontal spatial span in the X-axis and Y-axis directions is calculated respectively, and the maximum span is taken as the effective crown width.

7. The field daylily phenotypic measurement system based on multi-view three-dimensional reconstruction according to claim 1, characterized in that, In the three-dimensional phenotypic analysis module, the uprightness index is defined as follows: Where λ1, λ2, and λ3 are three spatial feature values ​​arranged in descending order after principal component analysis of the pure plant point cloud, and λ1>λ2>λ3; the principal axis vertical deflection angle θ is defined as... Where v1 is the eigenvector corresponding to the largest eigenvalue λ1. It is the spatial direction vector perpendicular to the Z-axis.

8. The field daylily phenotypic measurement system based on multi-view three-dimensional reconstruction according to claim 1, characterized in that, In the three-dimensional phenotypic analysis module, the three-dimensional canopy volume is obtained by extracting the three-dimensional concave geometric shell from the pure plant point cloud using an algorithm; the spatial light interception potential index is defined as the ratio of the projected canopy area to the three-dimensional canopy volume. .

9. The field daylily phenotypic measurement system based on multi-view three-dimensional reconstruction according to claim 1, characterized in that, In the time-series phenotypic state mining module, the candidate ideal plant type evaluation model constructs an evaluation vector based on the plant height, main effective crown width, projected crown area, three-dimensional crown volume, uprightness index, and spatial light interception potential index of the final monitoring week, as well as the slope of change of plant height, projected crown area, and three-dimensional crown volume and posture stability index within the continuous monitoring period. After normalization, the comprehensive score of each plant is calculated according to the preset weight and the ranking results are output.

10. A method for measuring the phenotypic characteristics of daylilies in the field based on multi-view three-dimensional reconstruction, characterized in that, Includes the following steps: S1. Multi-view image acquisition and preprocessing: Acquire multi-view image data of the whole daylily plant before bolting and perform distortion correction and photometric consistency compensation. S2. Foreground mask segmentation: The U²-Net salient object detection network is used to extract the foreground mask of the preprocessed image, and the field background interference is filtered out by pixel-by-pixel Hadamard product operation. S3, 3D reconstruction: Based on the structure-of-motion algorithm, the camera spatial pose is restored and sparse point cloud is generated. A differentiable 3D Gaussian splash rendering mechanism is introduced to guide joint optimization with foreground mask to generate a dense point cloud model of the whole daylily plant. S4. Point cloud preprocessing: Statistical outlier filtering for noise reduction and density-based spatial clustering for subject extraction are performed sequentially on dense point clouds. S5. Ground point cloud stripping: The ground reference equation is obtained by adaptive voxel downsampling and random sampling consistency plane fitting, and then mapped back to the original point cloud for point-by-point physical orthogonal distance determination to achieve in-situ stripping of ground point cloud. S6. Three-dimensional phenotypic analysis: Based on pure plant point cloud, the plant height, main effective crown width, projected crown area, three-dimensional crown volume, spatial light interception potential index and leaf posture morphology parameters are automatically analyzed. S7. Time-series phenotypic state mining and candidate ideal plant type evaluation: The three-dimensional phenotypic parameters within the continuous monitoring period are constructed into plant-week state vectors. Unsupervised clustering is used to identify the three-dimensional phenotypic state categories and generate state transition trajectories. A candidate ideal plant type evaluation model is constructed and the ranking results are output.

Images