Banana plant phenotype extraction and health assessment method, system, equipment and medium

Abstract

The invention discloses a banana plant phenotype extraction and health assessment method, system, device and medium, and the method comprises the steps: determining a three-dimensional Gaussian splashing model of a banana plant, the three-dimensional Gaussian splashing model comprising a plurality of Gaussian units; based on the three-dimensional Gaussian splashing model, distinguishing organ instance segmentation results of cauloid, leaf and fruit comb by using anisotropic characteristics and spatial distribution relation of a Gaussian covariance matrix; and based on the organ instance segmentation result, extracting phenotypic parameters of the banana plant and evaluating health of the banana plant. According to the embodiment of the invention, accurate instance segmentation of each organ of a banana plant, nondestructive measurement of morphological parameters and robust health assessment based on multi-modal characteristics can be realized in a complex field environment.

Description

Technical Field

[0001] This invention relates to a method, system, equipment, and medium for extracting banana plant phenotypes and assessing health, belonging to the field of agricultural machine vision and crop phenotype analysis technology. Background Technology

[0002] Bananas are an important economic crop in tropical and subtropical regions worldwide. Phenotypic parameters such as plant height, pseudostem girth, leaf morphology, and fruit yield are key indicators for assessing growth vigor, guiding precise water and fertilizer management, and breeding selection. Meanwhile, vascular diseases, such as banana wilt, pose a serious threat to the industry, making early monitoring and precise intervention crucial. With the development of smart agriculture and machine vision technology, acquiring data using non-contact sensors and automatically analyzing high-throughput phenotypic parameters and health status has become an urgent industry need.

[0003] Current technologies for banana plant phenotyping and health assessment mainly fall into two categories: two-dimensional image processing and three-dimensional point cloud analysis. However, these technologies still have significant limitations when dealing with bananas, a tall, complex, and unstructured crop.

[0004] Traditional two-dimensional image analysis methods primarily rely on RGB cameras. While these methods are low-cost and have mature algorithms, their limitations in imaging principles make it difficult to recover the true three-dimensional scale of plants from a single perspective. Banana plants are typically 2 to 4 meters tall, with large, overlapping leaves that cause significant occlusion. The lack of depth information in two-dimensional images leads to severe perspective distortion, making it difficult to meet the accuracy requirements for measuring geometric parameters such as plant height and stem diameter in agronomic measurements. Furthermore, disease identification methods that rely solely on color features are highly susceptible to changes in field lighting, shadows, and leaf surface reflections, resulting in a high false detection rate in complex environments.

[0005] To acquire spatial geometric information, 3D point cloud reconstruction technology based on LiDAR or multi-view stereo vision is increasingly being applied in agriculture. However, point cloud data is essentially a collection of discrete coordinates, lacking continuous surface representation and rich physical properties, which limits its ability to perform refined phenotypic analysis. Specifically: First, in organ segmentation, banana leaves are spirally distributed and intertwined. Traditional point cloud clustering algorithms based on Euclidean distance or density struggle to distinguish physically adjacent but semantically independent leaves, often leading to multiple leaves being misclassified as the same leaf, failing to achieve accurate segmentation, and consequently affecting the accuracy of leaf area and leaf inclination angle calculations. Second, in morphological measurement, banana pseudostems are not standard cylinders, and their surfaces are often covered with dried leaf sheaths and fibers. Noise in discrete point clouds easily causes outlier interference in cylinder fitting algorithms, resulting in generally overestimated girth measurements. Third, existing yield prediction methods are mostly based on identifying fruit combs or fruit indexes, ignoring individual differences in fruit size and fullness, making it difficult to achieve high-confidence weight predictions through precise volume integration.

[0006] In recent years, 3D Gaussian Splatting, as an emerging explicit radiation field representation method, has attracted much attention in the field of computer vision due to its excellent rendering quality and training speed. However, its current application in agriculture is mostly limited to the 3D visualization and reproduction of scenes, lacking in-depth mining and agronomic interpretation of the physical properties of the Gaussian sphere. Existing methods usually treat the scene as a uniform whole for optimization, which has the following shortcomings: First, it fails to fully utilize the anisotropic characteristics of the covariance matrix of the Gaussian sphere to analyze the structure of plant organs, and it is difficult to solve the instance segmentation problem of complex canopies by leveraging geometric distribution features; Second, it lacks a dual-modal health diagnosis mechanism that integrates spectral features and geometric morphology. Existing disease monitoring often ignores the key geometric feature of leaf drooping that accompanies the early stage of banana wilt disease, resulting in the inability to achieve early warning before the leaves turn significantly yellow; Third, it lacks high-precision, non-destructive measurement algorithms for discrete Gaussian distributions, making it difficult to extract accurate biomass parameters from the Gaussian representation.

[0007] In summary, existing technologies struggle to simultaneously achieve precise instance segmentation of various banana plant organs, high-precision measurement of morphological parameters, and robust health assessment based on multimodal features in complex field environments. Therefore, there is an urgent need to develop a high-precision method for banana plant phenotypic extraction and health assessment that can deeply analyze three-dimensional Gaussian distribution characteristics and integrate geometric priors and spectral information. Summary of the Invention

[0008] In view of this, the present invention proposes a method, system, device and medium for banana plant phenotypic extraction and health assessment, which can achieve accurate instance segmentation of various organs of banana plants, non-destructive measurement of morphological parameters and robust health assessment based on multimodal features in complex field environments.

[0009] The first objective of this invention is to provide a method for extracting banana plant phenotypes and assessing their health.

[0010] The second objective of this invention is to provide a system for extracting banana plant phenotypes and assessing their health.

[0011] A third objective of this invention is to provide a computer device.

[0012] A fourth objective of this invention is to provide a computer-readable storage medium.

[0013] The first objective of this invention can be achieved by adopting the following technical solution: A method for banana plant phenotypic extraction and health assessment includes: A three-dimensional Gaussian splash model of a banana plant is determined, wherein the three-dimensional Gaussian splash model includes multiple Gaussian units; Based on the aforementioned three-dimensional Gaussian splash model, the anisotropic characteristics and spatial distribution relationships of the Gaussian covariance matrix are used to distinguish the organ instance segmentation results of pseudostem, leaf, and fruit comb. Based on the organ instance segmentation results, banana plant phenotypic parameters were extracted and banana plant health was assessed.

[0014] In some embodiments, the organ instance segmentation results based on the three-dimensional Gaussian splash model, utilizing the anisotropy characteristics and spatial distribution relationships of the Gaussian covariance matrix to distinguish pseudostems, leaves, and fruit combs, include: The covariance matrix of each Gaussian unit is decomposed into eigenvalues ​​to obtain multiple eigenvalues ​​and corresponding principal axis direction vectors. Gaussian units whose eigenvalue relationships reflect linear characteristics and are spatially located in the central region of the plant are classified as candidate pseudostems. Gaussian units whose eigenvalue relationships reflect areal features, whose principal plane is approximately horizontal or tilted at a preset angle, and whose spatial distribution is around the pseudostem, are classified as leaf candidate sets. Gaussian units whose eigenvalue relationships reflect spherical features, are spatially distributed above or on the side of the pseudostem, and exhibit a predetermined density clustering in local areas are classified as the fruit comb candidate set; Within their respective candidate sets, spatial clustering methods based on distance and density are used to perform instance-level clustering of Gaussian units, and each Gaussian unit is assigned a specific organ instance label.

[0015] In some embodiments, extracting banana plant phenotypic parameters based on the organ instance segmentation results includes: Measure pseudostem girth based on Gaussian soft slices; Based on the Gaussian volume integral, the yield of fruit combs is estimated.

[0016] In some embodiments, the measurement of pseudostem girth based on Gaussian soft slices includes: The main axis of the pseudostem is estimated based on the Gaussian central distribution, and a virtual horizontal slice plane is set at a preset height above the ground. Project the pseudo-stem Gaussian covariance matrix of the adjacent virtual horizontal slice plane onto the virtual horizontal slice plane, and perform weighted superposition of the two-dimensional probability density to form a continuous probability density function p(x,y); Set a density threshold τ, extract the contour lines of p(x,y)=τ as the effective cross-sectional profile of the pseudostem, calculate the cross-sectional area Area, and obtain the equivalent diameter based on D = 2 × √(Area / π), and then obtain the pseudostem girth.

[0017] In some embodiments, estimating comb yield based on Gaussian volume integral includes: For each fruit comb Gaussian element, estimate the equivalent ellipsoidal volume based on its eigenvalues; Based on the equivalent ellipsoidal volume, a weighted summation is performed using density parameters, and a fill factor is introduced to correct the volume deviation caused by Gaussian overlap, thus obtaining the effective volume of the fruit comb. Based on the effective volume of the fruit comb, the estimated weight of a single comb or the entire banana plant is obtained using a pre-calibrated density constant.

[0018] In some embodiments, the assessment of banana plant health includes: Color level: Based on Gaussian color parameters, the super green index ExG is calculated to quantify the fading of green; Geometric level: Using the blade skeleton line as a reference, calculate the downward angle θ of the blade tip relative to the base, and calculate the rate of change of the skeleton line curvature; The health of banana plants is comprehensively assessed based on the color level and the geometric level.

[0019] In some embodiments, the comprehensive assessment of banana plant health based on the color level and the geometric level includes: When ExG > 0.1 and θ < 45 degrees, the banana plant is in a healthy state; When ExG>0.1 but the increase in θ exceeds the preset angle within a preset time period, the banana plant is in an early physical wilting state; When ExG < 0.05 and θ > 75 degrees, the banana plant is in a state of wilt disease.

[0020] The second objective of this invention can be achieved by adopting the following technical solution: A banana plant phenotypic extraction and health assessment system includes: A determination module is used to determine a three-dimensional Gaussian splash model of a banana plant, wherein the three-dimensional Gaussian splash model includes multiple Gaussian units; The differentiation module is used to differentiate the organ instance segmentation results of pseudostem, leaf and fruit comb based on the three-dimensional Gaussian splash model and by using the anisotropic characteristics and spatial distribution relationship of the Gaussian covariance matrix. The extraction and evaluation module is used to extract banana plant phenotypic parameters and evaluate banana plant health based on the organ instance segmentation results.

[0021] The third objective of this invention can be achieved by adopting the following technical solution: A computer device includes a processor and a memory for storing a processor-executable program, wherein when the processor executes the program stored in the memory, it implements the above-described method for banana plant phenotypic extraction and health assessment.

[0022] The fourth objective of this invention can be achieved by adopting the following technical solution: A computer-readable storage medium storing a program that, when executed by a processor, implements the above-described method for banana plant phenotypic extraction and health assessment.

[0023] The present invention has the following advantages over the prior art: Advantages from the perspective of structural features and functional relationships: 1. It has a stronger ability to precisely segment organs with complex coronal layers.

[0024] Existing technologies often employ point cloud clustering algorithms based on Euclidean distance or point density. However, in complex canopies where banana leaves intertwine and physically overlap, this approach is prone to oversegmentation or undersegmentation, resulting in unclear boundaries between leaves, pseudostems, and fruit combs. This invention utilizes the covariance matrix of three-dimensional Gaussian units to construct a geometric classification mechanism based on eigenvalue decomposition. By analyzing the "flatness" (corresponding to planar structures like leaves), "linearity" (corresponding to linear structures like pseudostems), and "sphericity" (corresponding to near-isotropic structures like fruits) of the Gaussian units, it can effectively separate different organs at the semantic level, even when they are spatially very close or even locally adhered. This structural design significantly improves the completeness and accuracy of single-plant organ identification under heavily shaded environments, providing a reliable structural foundation for subsequent phenotypic parameter extraction.

[0025] 2. It has achieved a leap from "visualization only" to "quantifiable" phenotypic measurement.

[0026] Traditional 3D Gaussian splashing technology is mainly used for novel perspective image synthesis and high-quality rendering, but it lacks clear physical dimensions, making it difficult to directly apply to agronomic measurements. This invention proposes a soft-slice integration and volume integral method based on 3D Gaussian representation. By defining cross-sectional contour lines and effective volumes in a Gaussian probability density field and introducing density thresholds and fill factors, overlapping parts and noise regions are corrected, establishing a clear correspondence between the Gaussian distribution and key phenotypic parameters such as pseudostem girth, plant height, and comb volume. Therefore, the 3D Gaussian splashing model can not only be "clearly seen" but also "accurately calculated," expanding its application scope in precision agriculture phenotyping and yield assessment.

[0027] Advantages from a theoretical and principle perspective: 1. Geometric measurements based on probability density integrals have better noise resistance and stability.

[0028] Traditional pseudostem girth measurements typically rely on fitting cylindrical or elliptical shapes to point cloud cross-sections, which is highly sensitive to outliers and noise such as dried leaf sheaths, easily leading to systematic overestimation. The soft-section measurement method proposed in this invention is based on the principle of Gaussian probability density integration. It superimposes the two-dimensional probability densities of various Gaussians within the slice plane and automatically suppresses low-density surface burrs and isolated noise points by setting an effective density threshold, extracting cross-sections only from the high-confidence core fleshy region. Theoretically, this method performs geometric measurements in a continuous probability field, exhibiting higher robustness and repeatability accuracy compared to directly fitting a geometric model to a discrete point cloud.

[0029] 2. The spectral-geometric dual-modal coupling mechanism improves the timeliness and robustness of disease monitoring.

[0030] Traditional disease monitoring methods largely rely on RGB color characteristics, which are easily affected by changes in light intensity, shadows, and reflections, and struggle to capture early physical symptoms in leaves before significant color changes. This invention utilizes the spherical harmonic coefficients in a three-dimensional Gaussian spray to extract approximately light-independent albedo or greenness indices, fundamentally reducing the influence of ambient light on color assessment. Simultaneously, it employs geometric analysis to calculate leaf skeleton curvature and drooping angle, explicitly quantifying the leaf dehydration, drooping, and deformation characteristics caused by vascular diseases. The combined judgment of spectral and geometric information allows this invention to indicate disease risk before significant leaf chlorosis, theoretically resulting in a lower false negative rate and an earlier detection window. Attached Figure Description

[0031] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0032] Figure 1 This is a simplified flowchart of a banana plant phenotypic extraction and health assessment method according to an embodiment of the present invention.

[0033] Figure 2 This is a flowchart illustrating a method for extracting banana plant phenotypes and assessing health according to an embodiment of the present invention.

[0034] Figure 3 This is a flowchart of an organ-level Gaussian instance segmentation according to an embodiment of the present invention.

[0035] Figure 4 This is a flowchart illustrating the measurement of pseudostem girth and plant height according to an embodiment of the present invention.

[0036] Figure 5This is a flowchart illustrating the yield estimation of fruit combs according to an embodiment of the present invention.

[0037] Figure 6 This is a flowchart of a dual-modal health assessment according to an embodiment of the present invention.

[0038] Figure 7 This is a structural diagram of a banana plant phenotypic extraction and health assessment system according to an embodiment of the present invention. Detailed Implementation

[0039] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0040] To overcome the difficulties in organ segmentation, the significant noise interference affecting morphological measurement accuracy, the limited dimensionality of yield prediction, and the sensitivity and lag of disease monitoring to light in existing banana plant phenotypic extraction and health assessment technologies, this invention proposes a method for banana plant phenotypic extraction and health assessment based on three-dimensional Gaussian sputtering. This invention aims to solve the following specific technical problems: 1. Solving the problem of difficult banana organ instance segmentation under complex canopies. Existing 3D point cloud clustering algorithms struggle to effectively separate spatially intersecting and physically contacting banana leaves, leading to errors in calculating parameters such as leaf area and leaf inclination angle. This invention aims to utilize the covariance matrix characteristics of a 3D Gaussian sphere and analyze the anisotropic geometric properties of the Gaussian distribution to achieve accurate instance segmentation of banana leaves, pseudostems, and fruit combs, overcoming the semantic confusion problem in traditional methods.

[0041] 2. Addressing the problem of large measurement errors in the girth and plant height of non-standard cylindrical pseudostems. To address the shortcomings of traditional cylindrical fitting algorithms, which tend to overestimate diameter due to the presence of dried leaf sheaths and fibers on the surface of banana pseudostems leading to high noise in the discrete point cloud, this invention aims to propose a soft-slice measurement method based on a Gaussian probability density function. Through probability integration and edge threshold optimization, this method achieves non-destructive, high-precision measurement of the true fleshy part girth and plant height of the pseudostem without relying on the ideal cylinder assumption.

[0042] 3. Addressing the limitations of traditional yield estimation methods that rely on counting and have limited accuracy. Existing yield estimation methods only count fruit combs or fruit indexes, ignoring differences in individual fruit size and plumpness, leading to significant weight estimation errors. This invention aims to establish a yield assessment model based on Gaussian volume integrals. By calculating the effective cumulative volume of the Gaussian ellipsoid in the fruit comb region and introducing a fill factor correction, it achieves a leap from "quantitative statistics" to "volume weighing," thereby improving the accuracy of yield estimation.

[0043] 4. Addressing the issues of poor robustness and lag in early monitoring of banana wilt disease. To address the shortcomings of traditional two-dimensional image recognition, which is easily affected by field lighting and shadows and struggles to capture early physical drooping characteristics of leaves, this invention aims to construct a dual-modal health assessment mechanism that integrates spherical harmonic coefficient color information and geometric spatial posture. By jointly analyzing the degree of leaf chlorosis and geometric collapse characteristics, the influence of changes in environmental light can be effectively suppressed, achieving early and robust monitoring of vascular diseases such as banana wilt disease.

[0044] The technical solution of this invention consists of two parts: a "data acquisition and 3D reconstruction hardware system" and a "method for phenotypic parameter extraction and health assessment based on 3D Gaussian splashing". The hardware system is mainly used to acquire multi-source observation data in banana plantations and reconstruct a 3D Gaussian splashing model of banana plants; the method part, based on this model, completes organ-level segmentation, automatic calculation of phenotypic parameters, and health status assessment.

[0045] (1) Hardware device: The method of this invention operates on a 3D reconstruction system capable of acquiring 3D observation data of banana plantations and generating a 3D Gaussian splash model. The system typically includes a data acquisition platform, a multi-source sensor module, and a computing and storage unit. The data acquisition platform can be a four-wheeled mobile vehicle, a tracked platform, or a manually movable support, used to carry sensors and move between rows in the banana plantation; the sensor module can include one or more types of 3D sensors, such as mechanical or solid-state LiDAR, RGB cameras, multispectral cameras, structured light cameras, TOF cameras, or binocular depth cameras, used to acquire multi-view observation data of banana plants and their surrounding environment; the computing and storage unit can be an industrial control computer, a portable workstation, or a remote server, used to execute the 3D Gaussian splash reconstruction program and output a 3D Gaussian splash model of the banana plantation containing parameters such as the spatial location, covariance, color or spherical harmonic coefficients, and density of Gaussian elements.

[0046] This invention does not focus on specific hardware configurations. The form of the data acquisition platform, the model and combination of sensors, and the implementation of the computing unit can all be replaced or adjusted according to actual application needs. As long as a three-dimensional Gaussian splash model of a banana plantation that meets the accuracy requirements can be obtained, the automatic extraction of banana plant phenotypic parameters and health assessment method of this invention can be implemented on the basis of this model.

[0047] (2) Method: This invention proposes a method for banana plant phenotypic extraction and health assessment based on three-dimensional Gaussian splashing. This method processes data from a three-dimensional Gaussian splashing model of a banana plantation in a computer, aiming to solve specific technical problems such as difficulty in segmenting organ instances under complex canopies, large errors in pseudostem girth and plant height measurement, yield prediction relying solely on counting, and the sensitivity and lag of disease monitoring to light. Figure 1 and Figure 2 As shown, this method includes the following specific steps: Step 1: Model Input and Coordinate Normalization This invention first retrieves a 3D Gaussian splash model of a banana plantation from a 3D reconstruction system. This model consists of numerous Gaussian elements, each denoted as (x_i, Σ_i, c_i, α_i), where x_i is a 3D spatial position vector representing the coordinates of the Gaussian center in the scene; Σ_i is a 3×3 covariance matrix representing the shape and scale of the Gaussian element in 3D space; c_i is a color or spherical harmonic coefficient vector describing the albedo and view-dependent color at that location; and α_i is a density or opacity parameter representing the probability of an entity existing at that location. For ease of subsequent analysis, this invention obtains an approximately horizontal ground plane by fitting a ground Gaussian element and recalibrates the coordinate system to a local banana row coordinate system with the banana row direction as the X-axis and the gravity direction as the Z-axis. Simultaneously, the background area without banana plants is cropped, retaining only the model portion containing the banana plants.

[0048] Step 2: Organ-level Gaussian instance segmentation based on anisotropic features This method automatically distinguishes pseudostems, leaves, and fruit combs using the anisotropy and spatial distribution of the Gaussian covariance matrix. The specific physical process is as follows: Eigenvalue decomposition is performed on the covariance matrix Σi of each Gaussian unit to obtain three eigenvalues ​​λ1, λ2, and λ3 (with the convention λ1 ≥ λ2 ≥ λ3) and their corresponding principal axis direction vectors. Pseudostem determination: Gaussian units with approximately vertical principal axes, similar three eigenvalues, and located in the central region of the plant are classified as candidate pseudostems. Leaf determination: Gaussian units with two eigenvalues ​​significantly larger than the third eigenvalue (appearing flattened), approximately horizontal or slightly inclined principal planes, and located on the periphery of the pseudostem are classified as candidate leaves. Fruit comb determination: Gaussian units with elongated eigenvalues, distributed above or to the side of the pseudostem, and densely clustered in local space are classified as candidate fruit combs. Subsequently, within each candidate set, a distance- and density-based spatial clustering method (such as DBSCAN) is used to perform instance-level clustering of the Gaussian units, assigning each Gaussian unit a specific organ instance label.

[0049] Step 3: Measurement of pseudostem girth and plant height based on Gaussian soft slices A soft-slice measurement method based on Gaussian probability density is proposed to overcome the fitting error of non-standard cylinders. Circumference measurement: In the pseudostem example, the main axis of the pseudostem is estimated based on the Gaussian center distribution, and a virtual horizontal slice plane is set at a preset height above the ground (e.g., 1.0m). The Gaussian covariance matrix of the pseudostems adjacent to this plane is projected onto the plane, and the two-dimensional probability densities are weighted and superimposed to form a continuous probability density function p(x,y). A density threshold τ is set, and the contour lines of p(x,y)=τ are extracted as the effective cross-sectional profile of the pseudostem. The cross-sectional area Area is calculated, and the equivalent diameter is obtained according to D = 2 × √(Area / π). Plant height measurement: The highest reliable Gaussian center point in the Z-axis direction is selected from the pseudostem example, and the difference between its height and the ground plane height is calculated as the plant height.

[0050] Step 4: Prediction of comb yield based on Gaussian volume integral A weight estimation method based on Gaussian volume integrals is proposed, transforming the method from counting to volume measurement. For each fruited comb Gaussian, the equivalent ellipsoidal volume V_i is estimated based on its eigenvalues, and then weighted summation is performed using the density parameter α_i. A fill factor is introduced. (Values ​​range from 0.6 to 0.8) Correct for volume deviation caused by Gaussian overlap to obtain the effective volume of the fruit comb. Using a pre-calibrated density constant Through formula Convert the volume into a weight estimate for a single comb or the entire banana plant.

[0051] Step 5: Health assessment based on color-geometry bimodality A banana leaf health index was constructed by comprehensively utilizing Gaussian color parameters and geometric posture changes. At the color level: RGB albedo, approximately independent of light intensity, was extracted from the spherical harmonic coefficients to calculate the supergreen index ExG = 2G - R - B. The average greenness and spatial variability of leaf instances were statistically analyzed to quantify chlorosis and mottling anomalies. At the geometric level: Using the leaf skeleton line as a reference, the drooping angle of the leaf tip relative to the base and the rate of change of the skeleton line curvature were calculated to characterize abnormal bending or collapse. Comprehensive judgment: A judgment logic was set in the color-geometric two-dimensional index space: when greenness is low and mottling is high, the leaf is marked as a suspected diseased yellow leaf; when greenness is normal but the drooping angle and curvature are significantly increased, the leaf is marked as an early physical wilting or a high-risk leaf.

[0052] Step 6: Output and Application of Results The above segmentation results, phenotypic parameters, and health indices are organized into a structured data table, and the pseudostems, leaves, and fruit combs and their health status are marked with different colors in a 3D visualization interface. The data is then output to the field management terminal or decision-making system to complete the entire process.

[0053] Example 1: (I) System Environment and Data Acquisition Configuration: This embodiment uses a self-developed ground-based mobile agricultural phenotyping system.

[0054] Hardware platform configuration: A four-wheel differential drive chassis (model Scout-mini) is used as the mobile platform. The sensor suite includes a 16-line mechanical LiDAR (360-degree horizontal field of view, 30-degree vertical field of view) and an RGB-D depth camera (1280×720 resolution, 30fps). The computing unit uses an industrial PC equipped with an NVIDIA RTX 3090 graphics card.

[0055] Data acquisition process: During the banana budding stage, the mobile platform is controlled to travel at a constant speed of 0.6 m / s along the center line of the banana row. The LiDAR and camera are synchronized via a hardware clock, and the acquisition frequency is uniformly set to 10 Hz. During the acquisition process, it is ensured that each banana plant is covered by at least 3 keyframes from different perspectives to guarantee the integrity of the 3D reconstruction.

[0056] (II) Specific Implementation Steps: The processing flow of the method of the present invention is as follows: Figure 1 and Figure 2 As shown, the specific steps include: Step 1: Construction of a 3D Gaussian field model of the banana plant. This step utilizes collected multimodal data to construct a high-fidelity 3D scene model.

[0057] Data Preprocessing: The camera pose of each frame is calculated using the Structure for Motion Restoration (SfM) algorithm, and a sparse point cloud is generated as initialization input. Gaussian Model Training: The image and pose are input into a 3D Gaussian Splash (3DGS) network. During initialization, a 3D Gaussian sphere is generated at the location of the sparse point cloud. During training, an adaptive density control strategy is adopted, and Gaussian splitting and cloning are performed every 100 iterations. The total number of iterations is set to 30,000. To ensure the repeatability and stability of phenotypic parameter measurements, a fixed random seed is set during the model training initialization phase to ensure that the geometric structure of the same plant is consistent across multiple reconstructions. Model Output: The final generated model is represented by the set G = {gi}, where each Gaussian unit gi contains four parameters: position mean μ, covariance matrix Σ, spherical harmonic coefficients SH, and opacity α. Coordinate system transformation: The RANSAC algorithm is used to fit the ground point cloud to determine the ground plane normal vector. The model is rotated so that the Z-axis is perpendicular to the ground, and the coordinate origin is translated to the root position of the target plant.

[0058] Step 2: Organ-level instance segmentation based on anisotropic features. Figure 3 The flowchart illustrates how the geometric features of a Gaussian sphere can be used to solve the segmentation problem of intersecting leaves and adhered organs.

[0059] Eigenvalue calculation: Traverse all Gaussian spheres in the model with an opacity α greater than 0.1, and perform singular value decomposition (SVD) on their 3×3 covariance matrix Σ to obtain three eigenvalues ​​λ1, λ2, and λ3 (by convention, λ1 ≥ λ2 ≥ λ3). Geometric feature operator calculation: Formula 1 calculates three shape descriptors based on the eigenvalues. Formula 1 A larger L value indicates a shape closer to linear (such as a pseudostem), a larger P value indicates a shape closer to planar (such as a leaf), and a larger S value indicates a shape closer to spherical (such as a fruit).

[0060] Classification and Clustering: For leaf extraction, Gaussian points with P > 0.5 and L < 0.3 were selected, and DBSCAN density clustering (search radius 0.05m, minimum number of points 50) was used to separate independent leaf instances. For pseudostem extraction, Gaussian points with L > 0.6 and P < 0.3 were selected, and the angle between the principal axis and the Z-axis was less than 30 degrees, and clustering was used to obtain pseudostem instances. For fruit comb extraction, Gaussian points with S > 0.6 were selected in the region above the pseudostem for clustering to obtain fruit comb instances.

[0061] Step 3: Measurement of pseudostem girth and plant height based on Gaussian soft slices. This step achieves non-destructive measurement using the probability integration method.

[0062] Figure 4 Virtual slice plane construction: In the identified pseudostem instances, the measurement section is set at a height h = 1.0m above the ground. Probability density projection and integration: Search for all pseudostem Gaussian spheres within the Z-axis coordinate range of h ± 0.05m. Project them onto the 2D slice plane and calculate the cumulative opacity density D(x,y) at any point (x,y) on the plane. Effective cross-sectional area calculation: Set a density threshold τ = 0.5 to remove surface burr noise. Calculate the area Area satisfying D(x,y) > τ. Girth calculation formula: Formula 2 In Formula 2, D_stem is the equivalent diameter of the pseudostem, C_stem is the girth of the pseudostem, Area is the effective cross-sectional area obtained by integration, and π is pi. Next, the plant height is calculated: traversing all Gaussian spheres in the pseudostem instance, the 99th percentile of the Z-axis coordinate is taken as the plant top height, and the difference between this height and the ground height is the plant height.

[0063] Step 4: Prediction of comb yield based on volume integral. Figure 5 Estimate weight using volumetric characteristics.

[0064] Calculate the volume of a single Gaussian sphere. For each Gaussian sphere j in the fruit comb instance, calculate its geometric volume Vj: Formula 3 In Formula 3, λ1, λ2, and λ3 are the eigenvalues ​​of the covariance matrix of the Gaussian sphere, representing the squared lengths of the three semi-axises of the ellipsoid.

[0065] Total weight estimation formula: Formula 4 In Formula 4, Weight is the estimated weight; ρ is the average physical density of the banana fruit, which is 0.95 g / cm³ in this embodiment; η is the fill factor, used to correct the volume overhang caused by Gaussian sphere overlap, which is an empirical value of 0.7 in this embodiment. This value is a correction coefficient obtained after pre-experimental calibration based on multiple sets of standard sphere models with known volumes, used to eliminate volume redundancy caused by Gaussian overlap. αj is the opacity of the j-th Gaussian sphere, and ∑Vj is the sum of the volumes of all Gaussian spheres.

[0066] Step 5: Health assessment based on spectral and geometric dual modes. Figure 6 The flowchart illustrates the process of using color and posture to provide early disease warnings.

[0067] Color characteristics (spectral modes): The 0th order spherical harmonic coefficients of the Gaussian sphere of the leaf were extracted and converted into RGB color values. The supergreen index was calculated. Formula 5 In Formula 5, G, R, and B are the normalized pixel values ​​for the green, red, and blue channels, respectively. The lower the ExG value, the more severe the green fading.

[0068] Geometric features (attitude modes): Extract the blade skeleton line and calculate the angle θ (droop angle) between the blade tip tangent and the direction of gravity.

[0069] Comprehensive diagnostic criteria: Normal: ExG > 0.1 and θ < 45 degrees. Early physical wilting: ExG > 0.1 (normal color) but θ has increased by more than 15 degrees in the past week (abnormal drooping posture). Diagnosed wilt disease: ExG < 0.05 (severe chlorosis) and θ > 75 degrees.

[0070] (III) Experimental Data To verify the superiority of this invention in phenotypic parameter extraction, standard banana plants at different growth stages were selected as test samples. Precise manual measurements (using a measuring tape, vernier caliper, and high-precision electronic scale) were used as ground truth values. Evaluation indicators included pseudostem thickness (mean absolute error MAE), plant height (root mean square error RMSE), leaf segmentation (mIoU), and yield per plant (relative error MAPE). Table 1 selects five mainstream advanced methods currently used in the industry as a comparison group.

[0071] Table 1. Comparison of Measurement Accuracy of Banana Plant Phenotypic Parameters Method number Method type pseudostem thickness MAE / cm Plant height RMSE / cm Blade segmentation mIoU / % MAPE yield per plant / % Time consumption and computing cost Comparison Group 1 TLS + cylindrical fitting 1.2 9.0 68.5 15.6 Low Comparison Group 2 AdQSM skeletonization 1.0 7.8 66.9 14.2 middle Comparison Group 3 PointNet++ Deep Learning 0.9 6.5 81.5 10.8 High (requires training) Comparison Group 4 SfM–MVS Photogrammetry 1.4 10.2 74.2 18.5 high Comparison Group 5 NeRF Neural Radiation Field Network 1.1 8.8 70.1 16.7 Extremely high This invention 3DGS + Soft Slice Integration Method 0.5 3.5 91.4 5.2 middle Table 2 compares five mainstream monitoring technologies for early monitoring of banana wilt disease (yellow leaf disease).

[0072] Table 2. Comparison of early monitoring performance of banana Fusarium wilt disease Method number core features of the method Detection rate under strong light / % Detection rate of shaded areas / % Early physical wilting detection rate / % (Key point) False alarm rate / % Comparison Group 1 Traditional color features (thresholding method) 82.3 67.8 6.1 17.9 Comparison Group 2 CNN (ResNet, RGB) 90.7 75.6 14.8 9.8 Comparison Group 3 Transformer (ViT, RGB) 92.9 80.4 18.7 7.9 Comparison Group 4 Multispectral NDVI Index Method 89.6 84.2 21.5 6.7 Comparison Group 5 3D point cloud color features 88.5 78.3 19.4 8.8 This invention Color + Pose Bimodal Evaluation 97.1 95.3 86.4 3.1 Example 2: Figure 7 This is a structural diagram of a banana plant phenotypic extraction and health assessment system according to an embodiment of the present invention. Figure 7 As shown, the system includes: The determination module 701 is used to determine a three-dimensional Gaussian splash model of a banana plant, wherein the three-dimensional Gaussian splash model includes multiple Gaussian units.

[0073] The differentiation module 702 is used to differentiate the organ instance segmentation results of pseudostem, leaf and fruit comb based on the three-dimensional Gaussian splash model and by using the anisotropic characteristics and spatial distribution relationship of the Gaussian covariance matrix.

[0074] The extraction and evaluation module 703 is used to extract banana plant phenotypic parameters and evaluate banana plant health based on the organ instance segmentation results.

[0075] Example 3: Embodiments of this application also provide a computer device, including: a memory storing an executable program; and a processor for running the program, wherein the program executes the methods in various embodiments of the present invention during runtime.

[0076] The aforementioned memory can refer to devices inside a computer used to store data and programs, including RAM, hard disks, etc. RAM can be used to temporarily store running programs and data, while hard disks can be used to store programs and data long-term. Memory enables the computer to read and write data and execute programs. The aforementioned processor is responsible for executing instructions in computer programs and performing data processing. It can also be responsible for controlling and executing various operations, including arithmetic operations, logical operations, and data transmission.

[0077] Example 4: Embodiments of this application also provide a computer-readable storage medium including a stored executable program, wherein, when the executable program is executed, it controls the device where the computer-readable storage medium is located to perform the methods of various embodiments of the present invention.

[0078] The aforementioned computer storage media can refer to the media used in computer memory to store certain discontinuous physical quantities. Computer storage media mainly include semiconductors, magnetic cores, magnetic drums, magnetic tapes, laser discs, etc. Computer-readable storage media include stored programs, which can be a set of instructions that a computer can recognize and execute, running on an electronic computer to meet certain information needs.

[0079] Example 5: Embodiments of this application also provide a computer program product, including a computer program that, when executed by a processor, implements the methods of various embodiments of the present invention.

[0080] The aforementioned computer program products can refer to software programs that have been written, tested, and released, and can run on computers or other devices. Computer program products can include application programs, operating systems, utility software, etc., used to achieve specific functions or solve specific problems.

[0081] Example 6: Embodiments of this application also provide a computer program product, including a non-volatile computer-readable storage medium for storing a computer program that, when executed by a processor, implements the methods in various embodiments of the present invention.

[0082] The aforementioned non-volatile computer-readable storage medium can refer to a medium for storing data. Non-volatile computer-readable storage media can retain data without loss when power is off and can be used to store long-term data, such as operating systems, applications, and user files. Non-volatile storage media can include hard disk drives, solid-state drives, optical disks, and flash memory storage devices, etc.

[0083] Example 7: Embodiments of this application also provide a computer program that, when executed by a processor, implements the methods described in the various embodiments of the present invention.

[0084] The aforementioned computer program can refer to a set of instructions used to tell the computer to perform specific tasks or operations. Computer programs can be written by programmers using specific programming languages ​​and can include algorithms, data structures, logic, and control flow. Computer programs can be used for a variety of purposes, including application software, operating systems, etc.

[0085] In the above embodiments of the present invention, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0086] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For instance, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0087] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0088] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0089] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0090] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for extracting phenotypic characteristics and assessing the health of banana plants, characterized in that, include: A three-dimensional Gaussian splash model of a banana plant is determined, wherein the three-dimensional Gaussian splash model includes multiple Gaussian units; Based on the aforementioned three-dimensional Gaussian splash model, the anisotropic characteristics and spatial distribution relationships of the Gaussian covariance matrix are used to distinguish the organ instance segmentation results of pseudostem, leaf, and fruit comb. Based on the organ instance segmentation results, banana plant phenotypic parameters were extracted and banana plant health was assessed.

2. The method for banana plant phenotypic extraction and health assessment according to claim 1, characterized in that, The organ instance segmentation results based on the three-dimensional Gaussian splash model, which utilizes the anisotropy characteristics and spatial distribution relationships of the Gaussian covariance matrix to distinguish pseudostems, leaves, and fruit combs, include: The covariance matrix of each Gaussian unit is decomposed into eigenvalues ​​to obtain multiple eigenvalues ​​and corresponding principal axis direction vectors. Gaussian units whose eigenvalue relationships reflect linear characteristics and are spatially located in the central region of the plant are classified as candidate pseudostems. Gaussian units whose eigenvalue relationships reflect areal features, whose principal plane is approximately horizontal or tilted at a preset angle, and whose spatial distribution is around the pseudostem, are classified as leaf candidate sets. Gaussian units whose eigenvalue relationships reflect spherical features, are spatially distributed above or on the side of the pseudostem, and exhibit a predetermined density clustering in local areas are classified as the fruit comb candidate set; Within their respective candidate sets, spatial clustering methods based on distance and density are used to perform instance-level clustering of Gaussian units, and each Gaussian unit is assigned a specific organ instance label.

3. The method for banana plant phenotypic extraction and health assessment according to claim 1, characterized in that, The extraction of banana plant phenotypic parameters based on the organ instance segmentation results includes: Measure pseudostem girth based on Gaussian soft slices; Based on the Gaussian volume integral, the yield of fruit combs is estimated.

4. The method for banana plant phenotypic extraction and health assessment according to claim 3, characterized in that, The measurement of pseudostem girth based on Gaussian soft slices includes: The main axis of the pseudostem is estimated based on the Gaussian central distribution, and a virtual horizontal slice plane is set at a preset height above the ground. Project the pseudo-stem Gaussian covariance matrix of the adjacent virtual horizontal slice plane onto the virtual horizontal slice plane, and perform weighted superposition of the two-dimensional probability density to form a continuous probability density function p(x,y); Set a density threshold τ, extract the contour lines of p(x,y)=τ as the effective cross-sectional profile of the pseudostem, calculate the cross-sectional area Area, and obtain the equivalent diameter based on D = 2 × √(Area / π), and then obtain the pseudostem girth.

5. The method for banana plant phenotypic extraction and health assessment according to claim 3, characterized in that, The method for estimating comb yield based on Gaussian volume integral includes: For each fruit comb Gaussian element, estimate the equivalent ellipsoidal volume based on its eigenvalues; Based on the equivalent ellipsoidal volume, a weighted summation is performed using density parameters, and a fill factor is introduced to correct the volume deviation caused by Gaussian overlap, thus obtaining the effective volume of the fruit comb. Based on the effective volume of the fruit comb, the estimated weight of a single comb or the entire banana plant is obtained using a pre-calibrated density constant.

6. The method for banana plant phenotypic extraction and health assessment according to claim 1, characterized in that, The assessment of banana plant health includes: Color level: Based on Gaussian color parameters, the super green index ExG is calculated to quantify the fading of green; Geometric level: Using the blade skeleton line as a reference, calculate the downward angle θ of the blade tip relative to the base, and calculate the rate of change of the skeleton line curvature; The health of banana plants is comprehensively assessed based on the color level and the geometric level.

7. The method for banana plant phenotypic extraction and health assessment according to claim 6, characterized in that, The comprehensive assessment of banana plant health based on the color level and the geometric level includes: When ExG > 0.1 and θ < 45 degrees, the banana plant is in a healthy state; When ExG > 0.1 but the increase in θ exceeds the preset angle within a preset time period, the banana plant is in an early physical wilting state. When ExG < 0.05 and θ > 75 degrees, the banana plant is in a state of wilt disease.

8. A system for extracting banana plant phenotypes and assessing health, characterized in that, include: A determination module is used to determine a three-dimensional Gaussian splash model of a banana plant, wherein the three-dimensional Gaussian splash model includes multiple Gaussian units; The differentiation module is used to differentiate the organ instance segmentation results of pseudostem, leaf and fruit comb based on the three-dimensional Gaussian splash model and by using the anisotropic characteristics and spatial distribution relationship of the Gaussian covariance matrix. The extraction and evaluation module is used to extract banana plant phenotypic parameters and evaluate banana plant health based on the organ instance segmentation results.

9. A computer device comprising a processor and a memory for storing a processor-executable program, characterized in that, When the processor executes the program stored in the memory, it implements the banana plant phenotypic extraction and health assessment method according to any one of claims 1 to 7.

10. A computer-readable storage medium storing a program, characterized in that, When the program is executed by the processor, it implements the banana plant phenotypic extraction and health assessment method according to any one of claims 1 to 7.

Images