Alfalfa phenotype dynamic analysis and germplasm evaluation method based on few-sample time series segmentation and growth knowledge guidance

Abstract

The present application relates to a method for alfalfa phenotype dynamic analysis and germplasm evaluation based on few-shot temporal segmentation and growth knowledge guidance. The method comprises the following steps: collecting multi-temporal unmanned aerial vehicle remote sensing data of alfalfa in the field, constructing a DINO-XMem few-shot instance segmentation network, and using the network to perform single alfalfa instance segmentation on the alfalfa image in the field; high-throughput extraction of alfalfa multi-dimensional phenotype traits and inversion of key growth indicators, integration of single-scale and plot-scale temporal phenotype information, and construction of a multi-dimensional temporal high-throughput phenotype dataset; based on the prior knowledge of alfalfa growth process, the multi-dimensional temporal high-throughput phenotype dataset is decoupled based on the knowledge-guided temporal growth mechanism, and a plurality of mutually orthogonal biological growth dimensions are obtained by decoupling; fusing the characteristics of each biological growth dimension to construct a germplasm-specific representation vector, using a Gaussian mixture model to perform probability clustering on the germplasm-specific representation vector, and based on the clustering result, the functional typing of alfalfa germplasm resources is completed, and the alfalfa phenotype dynamic analysis and germplasm evaluation are realized.

Description

Technical Field

[0001] This invention relates to the field of agricultural information technology, specifically a method for dynamic analysis of alfalfa phenotypic traits and germplasm evaluation based on few-sample temporal segmentation and growth knowledge guidance. Background Technology

[0002] Alfalfa, the most widely cultivated legume forage crop globally, is a core crop ensuring the safety of herbivorous livestock farming. With the increasing demand for high-quality forage, developing breakthrough new alfalfa varieties that are high-yielding, highly regenerative, and widely adaptable has become a core objective of breeding efforts. However, traditional manual phenotyping methods suffer from low throughput, high destructiveness, and high labor intensity, making it difficult to meet the needs of large-scale, precise screening of germplasm resources and severely restricting the efficiency of discovering superior genetic resources.

[0003] In recent years, low-altitude remote sensing technology using unmanned aerial vehicles (UAVs) has provided a new opportunity to overcome phenotypic bottlenecks due to its advantages of high throughput and high spatiotemporal resolution. However, existing studies mostly adopt a "coarse-grained" monitoring strategy, that is, using vegetation indices (such as NDVI) or canopy height models (CHM) to perform mean-based statistics on the entire plot. This "population overview" mode faces significant challenges in the early stages of crop growth, when the crop has not yet closed the canopy, the proportion of exposed soil background is high, and weeds are randomly distributed, resulting in extracted phenotypic data that are often a "mixed signal" of crops, soil, and weeds, which seriously reduces the signal-to-noise ratio of the data. In order to obtain low-noise, high-purity germplasm characteristics, shifting the observation granularity from "plot" to "individual plant" is a good attempt. Through high-precision instance segmentation, soil background and weed interference can be accurately filtered out, thereby extracting pure individual plant morphological and textural features. Mainstream deep learning methods (such as YOLO series, Mask R-CNN, U-Net) have significantly improved the segmentation ability in agricultural scenarios. However, the fully supervised model used in this method heavily relies on massive amounts of manually labeled data for training. In agricultural scenarios, crop morphology changes drastically with the growth cycle, and the differences between different varieties are subtle. Constructing high-quality labeled datasets is not only costly, but the model's generalization ability is often significantly reduced due to changes in variety or environment. Recently, visual foundational models, represented by DINOv3, have demonstrated strong feature generalization capabilities, providing a new paradigm for achieving interference-resistant and high-precision single-plant phenotypic extraction under small sample conditions. However, translating advanced large-scale visual models into practical field phenotypic applications still faces the severe challenge of environmental noise. Canopy overlap caused by dense planting and complex and variable light heterogeneity often distort the feature discrimination power in the model's latent space. This instability of feature representation ultimately leads to poor segmentation results, specifically manifested as blurred edges, instance adhesion, and missed detection of targets in extreme environments.

[0004] Furthermore, existing phenotypic analysis studies still suffer from a disconnect in terms of resolution and continuity across the "spatiotemporal dimension," failing to establish a full-cycle analysis system adapted to the growth and development patterns of crops. This is mainly reflected in the following two aspects: First, insufficient mining of individual-scale characteristics in key developmental windows during early growth. Pre-canopy closure is a critical window period for alfalfa individual structure establishment and spatial resource competition. For example, microscopic traits such as canopy density, branching angle, and leaf texture contain extremely high genetic variation information, directly determining the plant's future light-harvesting efficiency and competitive advantage. However, existing studies are often limited by complex background noise interference, making it difficult to obtain high-purity individual-scale traits, often relying on coarse plot coverage indicators instead, resulting in the loss of a large amount of key information for early identification of superior germplasm. Second, current time-series analyses are mostly limited to discrete "static snapshot" monitoring or rely on machine learning algorithms to establish a "black box" mapping from spectrum to yield. While this method can predict yield, it ignores the continuity and dynamic mechanisms of crop growth and development, failing to explain the biological processes of yield formation. Roth et al.'s research shows that dynamic modeling is an effective alternative to black-box methods. By introducing prior information through parameterized or semi-parameterized growth models, discrete time series can be transformed into "intermediate traits" with clear biological significance.

[0005] For perennial, multi-crop crops like alfalfa, its core agronomical value depends not only on single-crop biomass but also on the regeneration rate after cutting and its growth stability across seasons. Ignoring these temporal dynamics can easily lead to misjudgments of germplasm potential. For example, simply pursuing plant height may result in selecting "leggy" but sparsely populated varieties, while overlooking specific germplasm that is short but dense and grazing-resistant. Therefore, to achieve the leap from simple "phenotypic monitoring" to in-depth "mechanism breeding," it is urgent to construct an analytical strategy that conforms to the developmental patterns of alfalfa: namely, using visual models in the early growth stage to accurately extract low-noise individual plant morphological and structural characteristics to capture individual developmental differences; and combining population temporal trajectories in the later growth stage to extract intermediate traits such as maximum growth rate, regeneration lag time, and key inflection points through dynamic modeling. Summary of the Invention

[0006] To address the shortcomings of existing technologies, the present invention aims to provide a method for dynamic analysis of alfalfa phenotypes and germplasm evaluation based on small-sample temporal segmentation and growth knowledge guidance.

[0007] To achieve the above objectives, the technical solution of the present invention is as follows:

[0008] A method for dynamic phenotypic analysis and germplasm evaluation of alfalfa based on few-sample temporal segmentation and growth knowledge guidance, comprising the following steps:

[0009] S1. Collect multi-temporal UAV remote sensing data of alfalfa in the field, construct the DINO-XMem few-sample instance segmentation network, and use the network to segment single alfalfa instances in the alfalfa images in the field.

[0010] S2. Based on the segmentation results of single alfalfa instances, high-throughput extraction of multi-dimensional phenotypic traits of alfalfa and inversion of key growth indicators are performed. Temporal phenotypic information at the single-plant scale and the plot scale is integrated to construct a multi-dimensional time-series high-throughput phenotypic dataset.

[0011] S3. Based on prior knowledge of alfalfa growth process, the knowledge-guided time-series growth mechanism of the multidimensional time-series high-throughput phenotypic dataset is decoupled to obtain multiple mutually orthogonal biological growth dimensions.

[0012] S4. Construct germplasm-specific characterization vectors by integrating the characteristics of various biological growth dimensions, and perform probabilistic clustering of the germplasm-specific characterization vectors using a Gaussian mixture model. Based on the clustering results, complete the functional classification of alfalfa germplasm resources, and realize dynamic analysis of alfalfa phenotypes and germplasm evaluation.

[0013] Further, step S1 specifically includes:

[0014] S11. Acquisition and Preprocessing of Multi-Temporal Remote Sensing Data

[0015] At multiple observation points throughout the alfalfa growth period, remote sensing images of the field were collected using drones. The original images were orthorectified, stitched together and pixel registered to generate orthophotos and canopy height models. A multi-temporal remote sensing dataset was constructed and divided into labeled support samples and unlabeled query samples.

[0016] Construction and training of S12 and DINO-XMem networks

[0017] The DINOv3 visual Transformer with frozen pre-trained parameters was used as the backbone network for feature extraction. A trainable LoRA adapter was embedded in the deep structure of the Transformer. Based on the support samples, a foreground feature memory bank of alfalfa and a complex background feature memory bank of the field were constructed to form a dual memory bank. The foreground feature enhancement of alfalfa plants and the suppression of background interference were carried out through the dual memory cross-attention mechanism to complete the construction and training of the DINO-XMem model.

[0018] S13, Dual-Memory Cross-Attentional Interaction and Conditional Feature Injection

[0019] The query sample is input into the backbone network to extract the query feature sequence. The query sample features are then interacted with the alfalfa foreground feature memory and the field complex background feature memory to generate foreground attention features and background attention features. The two attention features are then fused into a conditional feature increment using a differential strategy of foreground enhancement and background suppression. This increment is then injected into the query features of the backbone network.

[0020] S14. Multi-layer feature decoding and single-plant instance segmentation output

[0021] The conditionally enhanced features are subjected to multi-layer feature fusion, upsampling and boundary refinement to output a single alfalfa instance segmentation mask.

[0022] Further, step S13 specifically includes:

[0023] S131. Constructing an alfalfa-background dual-memory library

[0024] Based on support sample feature maps and binary mask Construct a high-confidence foreground clover texture memory library Interference background memory bank The specific formula is as follows:

[0025] (1)

[0026] (2)

[0027] in, Indicates support for sample number A feature token for each location; This represents the set of feature tokens that support the flattened feature map of the sample. , These are the mask values ​​corresponding to the i-th and j-th positions, respectively; Let the foreground purity function be . As the overall prototype of the foreground plant, Cosine similarity is used as a metric. The purity threshold, The threshold for selecting difficult negative samples;

[0028] S132, Differential Cross-Attention Injection

[0029] The following formula flattens the features of the query image into a sequence. The query features are used as query vectors, and then compared with the key / value pairs respectively. and Perform multi-head cross-attention interaction to generate foreground attention features. and background attention features :

[0030] (3)

[0031] (4)

[0032] Where N is the total number of observations; MHCA stands for Multi-head Cross-Attention Mechanism; The query feature matrix is ​​the target image, i.e., the query image. To support the sample, i.e., the key feature matrix of the example image; To support the value feature matrix of the sample; , and These are the learnable query, key, and value linear projection weight matrices, used to map features to a unified attention computation subspace;

[0033] A differential strategy of positive enhancement and negative suppression is adopted, and conditional feature increments are generated using the following formula. :

[0034] (5)

[0035] in, The background suppression coefficient is... Indicates feature level;

[0036] S133, Prior Guidance Space Gating

[0037] First calculate query features Prototype of the foreground plant Pixel-level cosine similarity is used to generate a prior similarity map. Then, the prior similarity map is used as a spatial gating signal to apply the feature increment of the differential attention output. Weighting is applied; finally, local write constraints for conditional features are achieved through Hadamard product and Sigmoid activation, specifically implemented using the following formula:

[0038] (6)

[0039] in, This represents the updated query feature sequence at layer (l+1). This represents a feedforward network used to perform nonlinear transformations on the weighted features; This indicates a normalization operation; This represents the updated query feature sequence at layer l. This represents the conditional feature increment generated by the difference across attention; It represents the Hadamardi (or Hadama) stack; This represents the Sigmoid activation function; This represents the learnable scaling factor.

[0040] Further, step S14 specifically includes:

[0041] S141. Multi-layer Feature Alignment and Fusion

[0042] Obtain the feature pyramid output by the feature extraction backbone network. ,in and through Convolutional layer Map features at different levels to a unified channel dimension. And using bilinear interpolation Spatial resolution of features at each level:

[0043] (7)

[0044] in, This indicates that the data has passed through a 1×1 convolutional layer (with uniform channel dimension) and bilinear interpolation has been applied. After spatial resolution alignment, the first Layer feature map; This represents the feature pyramid output by the feature extraction backbone network (i.e., the frozen DINOv3 encoder), where the [number]th [unit] is the [number]th [unit]. The original encoded feature map of the layer (the subscript "enc" represents the encoder). Representing the universal quantifier symbol in mathematics, the above feature channel mapping and resolution alignment operation is applicable to feature layer sets. This operation is performed on each specified level (i.e., levels 2, 5, 8, and 11).

[0045] The following formula is used to concatenate all aligned features across all levels, and then a convolutional reshaping module is applied. Processing generates fused features that combine global semantic features and local spatial features. :

[0046] (8)

[0047] in, [ ] indicates the feature channel concatenation operation, which stacks and merges multiple aligned feature maps within the brackets along the channel dimension, thereby fusing feature information at different depths; , , , These represent the aligned feature maps from layers 2, 5, 8, and 11 of the feature extraction backbone network DINOv3, respectively.

[0048] S142, Pixel-level boundary refinement and single-plant instance segmentation output

[0049] Based on multi-layer fusion features Generate coarse-grained segmentation prediction results and extract the original image I. img The high-frequency components in the sample are used to obtain high-frequency detail information; the high-frequency detail information is interacted with the fused features, and the boundary refinement residual is generated through the residual refinement function; the coarse-grained segmentation prediction result is added to the boundary refinement residual, and after processing by the Sigmoid activation function, the segmentation mask of a single alfalfa instance is output.

[0050] Furthermore, the segmentation mask for the single alfalfa instance As shown in the following formula:

[0051] (9)

[0052] in, This is a coarse-grained prediction function; This represents a fusion feature that integrates global semantic features and local spatial features. ( ) represents the Sigmoid activation function;

[0053] To utilize the original image The residual refinement function for high-frequency information is defined as:

[0054]

[0055] in, This represents a low-resolution feature map output by the backbone network or across attention modules; Represents the original input image Or high-resolution features in the shallow layer; Indicates a refinement decoder; This indicates a high-frequency feature encoder; Extracting the original image The high-frequency components in the image are the clear canopy outline and the gaps in the shadows between plants. A cascading symbol indicating feature splicing; This represents the feature projection and alignment operator.

[0056] Further, step S2 specifically includes:

[0057] S21. Generation of Region of Interest for a Single Plant: The segmentation mask of a single alfalfa instance output by the DINO-XMem network is vectorized to construct a region of interest for each alfalfa plant, and is assigned a variety identifier and spatial coordinate information to establish a variety-plant-spatiotemporal association index.

[0058] S22. High-throughput extraction of multi-dimensional phenotypic traits: Based on the region of interest of a single alfalfa plant, refined phenotypic traits of alfalfa plants are extracted from five dimensions: morphology, structure, color, texture, and size.

[0059] S23. Inversion and multi-scale supplementation of key growth indicators: Based on the canopy height model, invert the individual plant height and canopy coverage, and switch to the plot scale for population-level phenotypic statistics after alfalfa enters the canopy closure period.

[0060] S24. Construction of a multidimensional time-series high-throughput phenotypic dataset: The refined phenotypic traits at the individual plant scale and the population indicators at the micro-plot scale are spatiotemporally matched with the timestamps and spatial coordinates of multi-temporal observations to construct a multidimensional time-series high-throughput phenotypic dataset covering the individual plant to micro-plot scale and containing multidimensional traits and continuous time-series information.

[0061] Further, step S3 specifically includes:

[0062] S31. Calculation of production potential dimension: Based on the time series phenotypic data in the multidimensional time series high-throughput phenotypic dataset, the Logistic growth model is used to fit the plant height and canopy coverage growth curves, and the average plant height, cumulative canopy coverage, maximum growth rate and growth inflection point parameters of the whole cycle are extracted.

[0063] S32. Calculation of the regeneration resilience dimension: Based on the regeneration time trajectory of alfalfa after multiple harvests, the regeneration lag time, regeneration rate, regeneration peak baseline and seasonal stability index are extracted to construct a composite regeneration resilience index and a seasonal stability index.

[0064] S33. Calculation of the dimensions of the multidimensional phenotypic strategy: Principal component analysis is used to reduce the dimensionality of multiple alfalfa phenotypic traits and generate low-dimensional feature components.

[0065] S34. Decoupling of multidimensional biological mechanisms: The three types of features obtained from the calculation, namely production potential, regeneration resilience, and multidimensional phenotypic strategies, are orthogonally integrated to generate mutually independent biological dimensional features.

[0066] Further, step S31 specifically includes:

[0067] S311. Calculate the average plant height over the entire cycle using the following formula. :

[0068] (10)

[0069] in, For the i-th observation time point The plant height is N, where N is the total number of observations.

[0070] S312. Calculate the cumulative canopy coverage using the following formula. :

[0071] (11)

[0072] in, For time points Canopy coverage, For time points Canopy coverage, For inclusion The total number of time points throughout the entire cycle;

[0073] S313. Based on the first derivative of plant height, calculate the maximum growth rate using the following formula. :

[0074] (12)

[0075] in, This represents the function for finding the maximum value. and This represents two consecutive observation points within the rapid growth period. This represents the set of condition constraints for the analysis window. Indicates the observation time node Corresponding plant height, Indicates the observation time node Corresponding plant height, This indicates that the calculation of formula (12) is performed within the time window of the rapid growth phase.

[0076] Further, step S32 specifically includes:

[0077] S321. Calculate the composite regeneration index using the following formula. :

[0078] (13)

[0079] in, and These represent plant height and coverage on the first day of drone observation after mowing, respectively. Time value; and The value of the stubble reference point. For the number of days of regeneration;

[0080] S322. Based on the reciprocal of the coefficient of variation of the maximum plant height over three growth cycles, the seasonal stability index is calculated using the following formula. :

[0081] (14)

[0082] in, This represents the maximum plant height during the c-th growth cycle. and These represent the mean and standard deviation of the peak values ​​for the three crops, respectively. To prevent tiny quantities with a denominator of zero.

[0083] Further, step S4 specifically includes:

[0084] S41. Construction of germplasm-specific characterization vectors: The features of the three dimensions of production potential, regeneration resilience and multidimensional phenotypic strategy obtained by decoupling are standardized, and the three types of features after processing are spliced ​​and integrated to integrate the features of various biological growth dimensions and construct the germplasm-specific characterization vector for each alfalfa germplasm.

[0085] S42. Probability density clustering: The germplasm-specific representation vector of each alfalfa germplasm is input into the Gaussian mixture model. The optimal number of clusters is determined by the Bayesian information criterion and the silhouette coefficient. Unsupervised probability density clustering is performed, and the distribution characteristics of different cluster categories are represented by the covariance ellipse.

[0086] S43. Functional typing of alfalfa germplasm resources: Based on the probability density clustering results, alfalfa germplasm is functionally classified to complete the functional typing of alfalfa germplasm resources, and finally realizes dynamic analysis of alfalfa phenotypes and germplasm evaluation.

[0087] Compared with the prior art, the advantages of the present invention are:

[0088] (1) This invention aims to solve the problem of high-throughput phenotypic extraction and mechanism analysis of alfalfa plants under complex field conditions. It constructs the DINO-Pheno-Cluster framework, achieving a systematic breakthrough from microscopic individual segmentation to macroscopic germplasm screening. The DINO-XMem network constructed in this invention utilizes a frozen visual baseline model and a dual-memory mutual exclusion mechanism to effectively overcome field background interference and plant adhesion problems. Compared with traditional fully supervised models, it demonstrates excellent generalization ability with only a very small number of samples, ensuring high-fidelity extraction of individual plant morphology and structural traits before canopy closure, providing crucial data support for early breeding.

[0089] (2) This invention uses multidimensional trait analysis to confirm the energy trade-off of plant space occupation strategy with internal structure construction, and further discovers the unique mechanism of alfalfa in response to seasonal environmental stress through time-series dynamic monitoring. That is, the "dwarf dense planting" strategy of prioritizing horizontal cover expansion by inhibiting vertical growth under low autumn temperatures, which corrects the underestimation of production potential by single plant height monitoring.

[0090] (3) Based on the mechanism decomposition and clustering guided by knowledge of the growth process, this invention successfully constructed a functional evaluation system oriented towards breeding objectives. This system effectively corrects the bias of "only plant height" in traditional breeding, identifies high-density "dwarf and sturdy" germplasm that is easily overlooked, and locks in the ideal genotype that has both high yield potential and environmental adaptability, providing an interpretable decision-making basis for precision breeding. Attached Figure Description

[0091] Figure 1 This is a flowchart of the method for dynamic analysis of alfalfa phenotypic traits and germplasm evaluation based on few-sample temporal segmentation and growth knowledge in this invention.

[0092] Figure 2 This is a schematic diagram of the alfalfa phenotypic dynamic analysis and germplasm evaluation method based on few-sample temporal segmentation and growth knowledge guided by the present invention.

[0093] Figure 3 This is a diagram of the overall architecture of DINO-XMem in this invention;

[0094] Figure 4 This is a schematic diagram of the knowledge-guided mechanism-based decomposition-type time-series phenotypic feature extraction in this invention;

[0095] Figure 5 This is a schematic diagram of a high-throughput time-based analytics platform based on unmanned aerial vehicles (UAVs).

[0096] Figure 6 This is a schematic diagram illustrating the performance evaluation of DINO-XMem segmentation.

[0097] Figure 7 This diagram illustrates the comparison of segmentation effects of different models at different growth stages, with RG-seeding representing the regeneration seedling stage.

[0098] Figure 8 This is a schematic diagram of the mechanism-based decomposition clustering results guided by knowledge of the growth process. Detailed Implementation

[0099] To provide a better understanding of the structural features and effects achieved by the present invention, a detailed description is provided below, accompanied by preferred embodiments and accompanying drawings:

[0100] like Figure 1 and Figure 2 The method shown is a dynamic analysis of alfalfa phenotypic traits and germplasm evaluation based on few-sample temporal segmentation and growth knowledge. The method includes the following steps:

[0101] S1. Collect multi-temporal UAV remote sensing data of alfalfa in the field, construct the DINO-XMem few-sample instance segmentation network, and use the network to segment single alfalfa instances in the alfalfa images in the field.

[0102] S2. Based on the segmentation results of single alfalfa instances, high-throughput extraction of multi-dimensional phenotypic traits of alfalfa and inversion of key growth indicators are performed. Temporal phenotypic information at the single-plant scale and the plot scale is integrated to construct a multi-dimensional time-series high-throughput phenotypic dataset.

[0103] S3. Based on prior knowledge of alfalfa growth process, the knowledge-guided time-series growth mechanism of the multidimensional time-series high-throughput phenotypic dataset is decoupled to obtain multiple mutually orthogonal biological growth dimensions.

[0104] S4. Construct germplasm-specific characterization vectors by integrating the characteristics of various biological growth dimensions, and perform probabilistic clustering of the germplasm-specific characterization vectors using a Gaussian mixture model. Based on the clustering results, complete the functional classification of alfalfa germplasm resources, and realize dynamic analysis of alfalfa phenotypes and germplasm evaluation.

[0105] This invention constructs a such Figure 2 The integrated phenotyping framework (DINO-Pheno-Cluste) shown here, which combines "sensory denoising, mechanism analysis, and functional evaluation," aims to overcome the dual technical bottlenecks of traditional phenotyping, namely, coarse observation granularity and insufficient revelation of spatiotemporal mechanisms.

[0106] like Figure 2 As shown, the framework mainly consists of four progressive levels:

[0107] (1) At the visual perception level, a single-plant instance segmentation network (DINO-XMem) based on few-shot learning was constructed to address the challenges of complex field backgrounds and alfalfa plant adhesion. This network utilizes the powerful representation capabilities and dual-memory interaction mechanism of the visual basic model DINOv3 to accurately segment and extract pure alfalfa individual plants from field images, thus achieving perceptual denoising and target separation.

[0108] (2) At the level of phenotypic feature extraction, a multidimensional high-throughput phenotypic feature set is constructed based on the alfalfa single plant objects obtained by segmentation. This dataset not only includes core agronomic indicators such as plant height (PH) and canopy coverage (CC), but also deeply extracts phenotypic features in five dimensions: morphology, structure, texture, color, and size.

[0109] (3) At the level of mechanism analysis, a mechanism decomposition method guided by knowledge of growth process is proposed, which breaks through the limitations of traditional static indicator monitoring. Combining a priori knowledge of alfalfa growth, the high-frequency time-series trajectory of the whole growth season is deconstructed into three orthogonal biological dimensions: production potential, regeneration resilience and multidimensional phenotypic strategy, so as to realize the quantitative analysis of growth dynamic mechanism.

[0110] (4) At the functional evaluation level, based on the mechanism parameters extracted by deconstruction, the Gaussian mixture model (GMM) is used to perform probabilistic clustering and functional classification of alfalfa germplasm resources, thereby identifying ideal functional germplasm that are adapted to different breeding objectives (such as high yield, stable yield, and grazing tolerance), and providing decision support for alfalfa precision breeding.

[0111] Furthermore, such as Figure 2 As shown in the first module, step S1 specifically includes:

[0112] S11. Multi-temporal remote sensing data acquisition and preprocessing

[0113] Remote sensing images of alfalfa were collected at multiple observation points throughout its entire growth cycle (seedling stage, branching stage, and harvesting and regeneration stage) using drones equipped with RGB cameras. The original images were orthorectified, stitched, and registered to generate high-resolution orthophotos and a canopy height model (CHM), thus constructing a multi-temporal remote sensing dataset. The dataset was then divided into labeled support samples (including images and individual plant instance masks) and unlabeled query samples to form a few-sample segmented dataset.

[0114] Construction and training of S12 and DINO-XMem networks

[0115] A DINOv3 visual Transformer with frozen pre-trained parameters was used as the backbone network for feature extraction. A trainable LoRA adapter was embedded in the deep structure of the Transformer to enhance the feature discrimination ability of alfalfa canopy texture, plant structure, and field background differences while keeping the model parameter scale controllable. An alfalfa foreground feature memory (FG Memory, M) was constructed based on support samples. fg ) and field difficulty background feature memory bank (BG Memory, M bg A dual memory bank is formed, and the foreground features of alfalfa plants are enhanced and the background interference is suppressed through the dual memory cross-attention mechanism, thus completing the construction and training of the DINO-XMem model.

[0116] Specifically, addressing the technical challenges of scarce labeled samples, strong interference from field background textures, and severe adhesion of seedlings / branching plants in alfalfa phenotypic extraction, this invention proposes a small-sample dual-memory conditional segmentation network, DINO-XMem (DINOv3-based X-Attention Memory Network), for single-plant alfalfa phenotypic analysis. The overall architecture of this network is as follows: Figure 3 As shown. This network follows the classic support-query inference paradigm, using a very small number of labeled support images as conditions to guide the model to complete pixel-level instance segmentation of a single alfalfa plant in unlabeled query images. The workflow of this model consists of three stages: (1) support-side feature extraction and dual memory construction; (2) multi-level feature condition injection driven by cross-attention mechanism; and (3) enhanced multi-level feature decoding and instance segmentation boundary refinement.

[0117] To simultaneously leverage the general semantic features of large-scale pre-trained models and the specific texture features adapted to agricultural scenarios, this invention employs DINOv3 (Vision Transformer) as the backbone network for feature extraction, achieving efficient extraction of alfalfa phenotypic features. Unlike traditional supervised learning models, DINOv3 is based on a self-supervised learning paradigm, pre-trained on large-scale unlabeled data, and possesses excellent object-level semantic representation capabilities. Even without specific class labels, its attention map can spontaneously focus on salient plant objects in the image and effectively suppress interference from background noise such as soil and weeds. This "class-independent" general visual capability makes it the optimal feature extractor for few-shot segmentation tasks. However, the general DINOv3 features lack sensitivity to specific agronomical attributes of alfalfa (such as canopy texture and spiderweb structure). Therefore, this invention, based on freezing all parameters of the DINOv3 backbone network, introduces a low-rank adaptation (LoRA) fine-tuning strategy, embedding the LoRA module into the deep structure of the Transformer (layers 6-11). This design, while ensuring that the parameter scale is controllable, significantly improves the model's ability to distinguish subtle differences between alfalfa plants and soil background by fine-tuning the adaptation to the subtle differences in features between alfalfa and field soil background, thus achieving accurate adaptation of general pre-trained features to specific agricultural scenarios.

[0118] S13, Dual-Memory Cross-Attentional Interaction and Conditional Feature Injection

[0119] The query samples are input into the frozen DINOv3 backbone network and embedded with a LoRA adapter to extract query feature sequences (Query Tokens). These query sample features are then compared with the foreground feature memory (M). fg), Difficult Background Feature Memory (M) bg Multi-head cross-attention interaction is performed to obtain foreground attention (FG Attention) and background attention (BG Attention) features. Through a differential strategy of foreground enhancement and background suppression, the two attention features are fused into conditional feature increments and injected into the query features of the backbone network to strengthen plant features and suppress background interference.

[0120] Traditional few-shot semantic segmentation methods often compress support samples into a global prototype, which easily leads to the loss of local morphological and texture details and makes it difficult to capture the fragmented structure of alfalfa canopy edges. This results in adjacent plants appearing as if they are stuck together in the prediction results, making it impossible to distinguish individual plants. To address this, this invention designs an example-driven dual-memory cross-attention module, which replaces static prototype matching with retrieved patch-level memory interactions, improving the model's adaptability to complex field scenarios.

[0121] Step S13 specifically includes:

[0122] S131. Constructing an alfalfa-background dual-memory library

[0123] To explicitly decouple foreground and background features, this invention is based on support sample feature maps. and binary mask Construct a memory bank. Unlike traditional global mean pooling prototypes, this method preserves patch-level texture details to maintain sensitivity to canopy edge jaggedness and branching structures.

[0124] Based on support sample feature maps and binary mask Construct a high-confidence foreground clover texture memory library Interference background memory bank .

[0125] Foreground alfalfa texture memory library and interference background memory bank Defined as a set of feature tokens. It stores high-confidence alfalfa canopy feature tokens. They specifically collect difficult-to-interfere materials that are similar in color to alfalfa (such as green weeds and regrown stubble).

[0126] As shown in equations (1) and (2), alfalfa canopy features with high confidence are selected and stored using the foreground purity function. Simultaneously calculate the background token and the plant prototype. The similarity score was used to filter out difficult interfering elements (such as green weeds and regreening stubble) that were similar in color to alfalfa and stored their features. The aim is to enhance the model's ability to distinguish easily confused field backgrounds.

[0127] (1)

[0128] (2)

[0129] in, Indicates support for sample number A location-specific token. Let the foreground purity function be . As the overall prototype of the foreground plant, Cosine similarity is used as a metric. The purity threshold, The threshold for selecting negative samples.

[0130] Through this strategy, It stores high-confidence plant textures, and It specifically collects difficult background features that are easily confused, enhancing the model's ability to discriminate complex field backgrounds.

[0131] S132, Differential Cross-Attention Injection

[0132] In the inference phase, to effectively suppress false positives against the background, this invention employs a differential mutual exclusion mechanism. This mechanism explicitly excludes the background manifold in the feature space, thereby significantly reducing false positive responses against complex farmland backgrounds.

[0133] Flatten the features of the query image into a sequence. The query features are used as the Query, and then compared with the Key / Value pairs. and Perform multi-head cross-attention (MHCA) interaction to obtain foreground attention features. and background attention features :

[0134] (3)

[0135] (4)

[0136] Where N is the total number of observations; MHCA stands for Multi-head Cross-Attention Mechanism; The query feature matrix is ​​the target image, i.e., the query image. To support the sample, i.e., the key feature matrix of the example image; To support the value feature matrix of the sample; , and These are the learnable query, key, and value linear projection weight matrices, used to map features to a unified attention computation subspace.

[0137] A differential strategy of positive enhancement and negative suppression is adopted, utilizing the conditional feature increments generated by the following formula. :

[0138] (5)

[0139] in, The background suppression coefficient is... Indicates the feature level.

[0140] S133, Prior Guidance Space Gating

[0141] To further constrain the scope of conditional propagation and prevent background textures from being erroneously activated, this invention introduces a priori gating mechanism after the cross-attention mechanism.

[0142] First, calculate the query features. Prototype of the foreground plant Pixel-level cosine similarity is used to generate a prior similarity map. Then, the prior similarity map is used as a spatial gating signal to adjust the feature increment of the differential attention output. Weighting is applied, and local write constraints for conditional features are implemented through Hadamard product and Sigmoid activation. This ensures that support information is injected only into potential target regions with high prior confidence, preventing background textures from being erroneously activated.

[0143] (6)

[0144] in, Representing Hadamaji, It is the Sigmoid activation function. This is a learnable scaling factor.

[0145] This mechanism ensures that supporting information is injected only into potential target regions with high prior confidence, thus achieving a shift from global indiscriminate writing to controlled local writing.

[0146] S14. Multi-layer feature decoding and single-plant instance segmentation output

[0147] The conditionally enhanced features are subjected to multi-layer feature fusion, upsampling, and boundary refinement to output a clear, independent, and non-adhesive single-plant alfalfa instance segmentation mask, providing an accurate single-plant target carrier for subsequent phenotypic extraction. Specifically, such as... Figure 3As shown, the conditionally enhanced features are input into the Multi-Level Decoder. First, a coarse segmentation mask is generated by upsampling and feature fusion. Then, high-resolution image features are introduced to extract high-frequency detail features to refine the edges of the coarse segmentation mask. Finally, a clear and unadhesive single alfalfa instance segmentation mask is output, providing an accurate single-plant target carrier for subsequent phenotypic extraction.

[0148] Step S14 specifically includes:

[0149] S141. Multi-layer Feature Alignment and Fusion

[0150] To address the problem that a single feature level cannot simultaneously capture both the overall morphology (semantic consistency) and edge contour (spatial accuracy) of a single alfalfa plant, an enhanced multi-layer feature decoder is proposed. This decoder employs an "alignment-fusion-decoding" strategy, generating fused features that balance the recognition accuracy of both the overall morphology and edge contour of a single alfalfa plant.

[0151] First, obtain the feature pyramid output by the feature extraction backbone network. ,in .

[0152] Then, through Convolutional layer Map features at different levels to a unified channel dimension. And using bilinear interpolation Spatial resolution of features at each level:

[0153] (7)

[0154] in, This represents the result after passing through a 1×1 convolutional layer (with uniform channel dimension) and spatially aligned using bilinear interpolation (Up). Layer feature map; This represents the feature pyramid output by the feature extraction backbone network (i.e., the frozen DINOv3 encoder), where the [number]th [unit] is the [number]th [unit]. The original encoded feature map of the layer (the subscript "enc" represents the encoder). Representing the universal quantifier symbol in mathematics, the above feature channel mapping and resolution alignment operation is applicable to feature layer sets. This operation is performed on each specified level (i.e., levels 2, 5, 8, and 11).

[0155] Finally, all aligned features from each level are concatenated by channels and then processed through a convolutional reshaping module. Processing generates fused features that combine global semantic features and local spatial features. :

[0156] (8)

[0157] in, [ The symbol ] indicates the feature channel concatenation operation, which is used to stack and merge multiple aligned feature maps within the brackets along the channel dimension, thereby fusing feature information at different depths; , , , These represent the feature maps from layers 2, 5, 8, and 11 of the feature extraction backbone network (DINOv3), respectively, after alignment.

[0158] S142, Pixel-level boundary refinement and single-plant instance segmentation output

[0159] To achieve effective separation of individual alfalfa instances under intensive planting conditions, this invention introduces auxiliary boundary branches and uses coarse-grained prediction and high-frequency residual refinement to generate a single alfalfa instance segmentation mask.

[0160] First, based on multi-layer fusion features Generate coarse-grained segmentation prediction results.

[0161] Secondly, extract the original image I img The high-frequency components, namely the clear outline of the alfalfa canopy and the gaps in the shadows between plants, are used to obtain high-frequency detail information.

[0162] Next, high-frequency detail information is interacted with the fused features, and boundary refinement residuals are generated through the residual refinement function.

[0163] Finally, the coarse-grained segmentation prediction results are added to the boundary refinement residuals, and then processed by the Sigmoid activation function to output a segmentation mask of single alfalfa instances with clear edges, independent individuals, and no adhesion. This provides an accurate single-plant target carrier for the subsequent construction of a multi-dimensional time-series high-throughput phenotypic dataset.

[0164] The segmentation mask for a single alfalfa plant instance Defined as the sum of coarse-grained prediction and refined residual:

[0165] (9)

[0166] in, This is a coarse-grained prediction function. To utilize the original image The residual refinement function for high-frequency information is defined as:

[0167]

[0168] in, This represents a low-resolution feature map output by the backbone network or across attention modules; Represents the original input image or high-resolution features in a shallow layer; Indicates a refinement decoder; This indicates a high-frequency feature encoder; Extract high-frequency components from the original image, namely clear canopy outlines and inter-plant shadow gaps; A cascading symbol indicating feature splicing; This represents the feature projection and alignment operator.

[0169] This module utilizes these high-frequency details to trim the blurred boundaries in coarse-grained predictions, thereby physically cutting off the slight overlap between the canopies of adjacent plants, repairing jagged edges, and finally outputting a single alfalfa instance segmentation mask with clear edges and a complete and independent topological structure, providing an accurate single-plant target carrier for the subsequent construction of multi-dimensional time-series high-throughput datasets.

[0170] Furthermore, such as Figure 2 As shown in the second module, step S2 specifically includes:

[0171] S21. Generation of Region of Interest (ROI) for a single plant: The segmentation mask of the single alfalfa instance output by the DINO-XMem network is vectorized to construct an independent ROI for each alfalfa plant, and a unique variety identifier and spatial coordinate information are assigned to it to establish a variety-plant-spatiotemporal association index.

[0172] S22. High-throughput extraction of multi-dimensional phenotypic traits: Based on the region of interest of a single alfalfa plant, refined phenotypic traits of alfalfa plants are extracted from five dimensions: morphology, structure, color, texture, and size. The morphological dimension includes roundness, density, and symmetry; the structural dimension includes branching angle and canopy structure index; the color dimension includes RGB mean, HSV features, and color difference; the texture dimension includes GLCM contrast, homogeneity, and energy; and the size dimension includes canopy area, diameter, and perimeter. A total of 33 phenotypic traits were extracted.

[0173] S23. Inversion and Multi-Scale Supplementation of Key Growth Indicators: Based on the aforementioned canopy height model (CHM), key growth indicators such as individual plant height and canopy coverage are inverted. After alfalfa enters the canopy closure stage, the mutual shading of plant canopies makes it difficult to distinguish individual plant boundaries. At this time, the scale is switched to the plot level for population-level phenotypic statistics to supplement the time-series data of the entire growth period, ensuring the continuity and completeness of the data.

[0174] S24. Construction of a multidimensional time-series high-throughput phenotypic dataset: The refined phenotypic traits at the individual plant scale and the population indicators at the micro-plot scale are spatiotemporally matched with the timestamps and spatial coordinates of multi-temporal observations to unify the data benchmark and construct a multidimensional time-series high-throughput phenotypic dataset covering the individual plant to micro-plot scale and containing multidimensional traits and continuous time-series information.

[0175] Furthermore, such as Figure 2 As shown in the third module, step S3 specifically includes:

[0176] S31. Calculation of Production Potential Dimension: Based on the time-series phenotypic data in the multidimensional time-series high-throughput phenotypic dataset, the Logistic growth model is used to fit the plant height and canopy coverage growth curves. The average plant height, cumulative canopy coverage (AUC), maximum growth rate (Vmax), growth inflection point and other characteristic parameters are extracted to characterize the biomass accumulation capacity and canopy space occupation potential of alfalfa germplasm.

[0177] Specifically, based on multidimensional time-series high-throughput phenotypic data, the discontinuities in the time-series curve caused by mowing are corrected. In this embodiment, the second day after each mowing is set as the stubble retention reference point (…). Based on field observations, the plant height at that moment was normalized to 5 cm, and the canopy coverage was normalized to 5%. On this basis, three key parameters were extracted: average plant height over the entire growth cycle, cumulative canopy coverage, and maximum growth rate, to characterize the ability of germplasm resources to accumulate biomass and occupy ecological niches.

[0178] Step S31 specifically includes:

[0179] S311. Calculate the average plant height over the entire period using the following formula ( This parameter characterizes the average vertical structure advantage of a variety throughout the growing season.

[0180] (10)

[0181] in, Let represent the plant height at the i-th observation time point, and N represent the total number of observations.

[0182] S312. Calculate the cumulative canopy coverage using the following formula. ):

[0183] (11)

[0184] in, For time points Canopy coverage, For time points Canopy coverage, For inclusion The total number of time points throughout the entire growing season. The area under the canopy cover over time is calculated using the trapezoidal rule. This indicator comprehensively reflects the area-time integral of light energy intercepted by the canopy throughout the entire growing season and is highly correlated with dry matter accumulation.

[0185] S313. Based on the first derivative of plant height, calculate the maximum growth rate using the following formula. :

[0186] (12)

[0187] Among them, the maximum growth rate Used to capture the potential for explosive vertical growth in a variety during its rapid growth phase; This represents the function for finding the maximum value. and This represents two consecutive observation points within the rapid growth period. This represents the set of condition constraints for the analysis window. Indicates the observation time node Corresponding plant height, Indicates the observation time node Corresponding plant height, This indicates that the calculation of formula (12) is performed within the time window of the rapid growth phase.

[0188] Since the canopy is often close to closure at this stage, the rate of change in horizontal coverage is no longer sensitive, so this index is constructed based only on the first derivative of plant height.

[0189] S32. Calculation of Regeneration Resilience Dimension: Based on the regeneration time-series trajectory of alfalfa after multiple harvests, key parameters such as regeneration lag time, regeneration rate, regeneration peak baseline, and seasonal stability index are extracted to construct a composite regeneration resilience index and a seasonal stability index. These indices evaluate the recovery and regeneration capacity of alfalfa germplasm after harvesting and its cross-seasonal growth stability. Regeneration and resilience aim to evaluate the recovery speed of germplasm in response to harvesting disturbances and its stability in response to seasonal environmental fluctuations. This is the core basis for screening harvest-tolerant and durable varieties.

[0190] Step S32 specifically includes:

[0191] S321. Calculate the composite regeneration index using the following formula. :

[0192] (13)

[0193] in, and These are the plant height and coverage on the first day of drone observation after mowing (i.e., ...). (Time value) and The value of the stubble reference point. The number of days for regeneration.

[0194] Composite Regeneration Index The composite index, which includes vertical recovery and horizontal expansion, calculates the geometric mean of the rate of plant height growth and the rate of cover recovery in the early post-harvest (re-greening phase).

[0195] S322. Based on the reciprocal of the coefficient of variation (CV) of the maximum plant height over three growth cycles (Cycle 1-3), the seasonal stability index is calculated using the following formula. :

[0196] (14)

[0197] in, This represents the maximum plant height during the c-th growth cycle. and These represent the mean and standard deviation of the peak values ​​for the three crops, respectively. To prevent tiny quantities with a denominator of zero.

[0198] Because the coverage of superior varieties tends to reach saturation in different seasons, seasonal fluctuations in plant height are more sensitive in identifying stable-yielding varieties that are not sensitive to environmental changes. Seasonal stability index The higher the value, the more consistent the variety's performance is across different growing seasons.

[0199] S33. Calculation of multidimensional phenotypic strategy dimensions: Principal component analysis (PCA) was used to reduce the dimensionality of 33 alfalfa phenotypic traits, eliminate multicollinearity among phenotypic characteristics, and generate low-dimensional feature components to reflect the differentiated growth strategies of different alfalfa germplasms in terms of canopy morphology construction, resource allocation, and field environment adaptation.

[0200] To comprehensively characterize the adaptive differences of alfalfa germplasm in morphology, physiological metabolism, and structural texture, this invention uses principal component analysis (PCA) to reduce the dimensionality of the high-dimensional feature space based on an initial high-dimensional feature pool of five dimensions: morphology, structure, physiology, texture, and color. This eliminates multicollinearity among features and forms low-dimensional feature components. The selection of key features follows two criteria: (1) Statistical significance: Select the features with the highest factor loading values ​​on the top N principal components with the largest explained variance; (2) Biological orthogonality: Ensure that the selected features do not overlap in a biological sense to guarantee the completeness of the feature set in describing plant growth strategies.

[0201] S34. Decoupling of multidimensional biological mechanisms: The three types of features obtained from the calculation, namely production potential, regeneration resilience and multidimensional phenotypic strategies, are orthogonally integrated to construct three independent growth dimensions with clear biological significance, thereby realizing the decoupling transformation from high-dimensional dynamic phenotypic features to interpretable growth mechanisms.

[0202] To identify alfalfa germplasm with specific breeding value from high-dimensional, dynamic UAV phenotypic data, this invention proposes a mechanism decomposition method guided by knowledge of the growth process. This method, combined with alfalfa growth and development patterns, deconstructs complex time series data into three orthogonal biological dimensions: production potential, regeneration and resilience, and multidimensional phenotypic strategies. The extraction process of the production potential and regeneration resilience parameters is illustrated below. Figure 4 As shown. In Figure 4 In the middle, (a) dimension A: production potential. Based on single-period data, a Logistic growth curve (green solid line) is fitted, and the cumulative canopy coverage is extracted. (green shaded area), maximum growth rate ( (slope of the red tangent line) and average plant height over the entire period ( (b) Dimension B: Regeneration Resilience. The multi-crop time series trajectory spanning Cycle 1, Cycle 2, and Cycle 3 reflects the dynamic response of germplasm to mowing stress. Key parameters include the composite regeneration index (…). ), that is, relative to the stubble baseline ( The growth slope of ) and the seasonal stability index ( ), which is determined by the peak plant height of each crop ( It is derived from the reciprocal of the coefficient of variation.

[0203] Furthermore, such as Figure 2 As shown in the fourth module, step S4 specifically includes:

[0204] S41. Construction of germplasm-specific characterization vectors: The three dimensions of production potential, regeneration resilience, and multidimensional phenotypic strategy obtained from decoupling are standardized to effectively eliminate the dimensional differences between different features. The three types of features are then spliced ​​and integrated to construct a multidimensional characterization vector for each alfalfa germplasm.

[0205] S42. Clustering based on probability density of Gaussian mixture model: The multidimensional representation vector of each alfalfa germplasm is input into the Gaussian mixture model (GMM). The optimal number of clusters is determined by the Bayesian information criterion (BIC) and silhouette coefficient (Silhouette Score). Unsupervised probability density clustering is performed, and the distribution characteristics of different cluster categories are clearly represented by the covariance ellipse.

[0206] S43. Functional classification of alfalfa germplasm resources: Based on the clustering results of probability density, alfalfa germplasm is divided into functional categories such as high-yield and high-regeneration type, compact and sparse type, high-yield and stable type, and short, dense and grazing-tolerant type. The core advantages of each type of germplasm are clarified, providing accurate germplasm screening basis and scientific decision support for targeted breeding of alfalfa with high yield, grazing tolerance and wide adaptability.

[0207] The clustering process and typing results of the present invention will be described in detail below with reference to specific embodiments:

[0208] In this embodiment, k=4 was determined to be the optimal number of clusters. Although BIC analysis showed that the model fitting error was smallest (lowest BIC value) when k=6, the corresponding silhouette coefficient showed a significant decrease (e.g., Figure 8 (a) shows that the inter-class separation deteriorates, indicating over-segmentation, which mathematically forces the separation of groups with similar biological functions. In contrast, k=4 significantly reduces the BIC value while maintaining a high silhouette coefficient, achieving the best balance between model statistical fit and intra-cluster compactness. Furthermore, t-SNE dimensionality reduction visualization further confirms the rationality of this division. The four types of germplasm form clearly defined independent clusters in the feature space, and the confidence ellipsoids show extremely high intra-class aggregation, specifically as shown in... Figure 8 As shown in (b).

[0209] To verify the biological accuracy of the classification, this example incorporates ground biomass data measured at two harvesting periods for 92 representative varieties (covering 72.4% of the entire population). Combined with... Figure 8 (c) shows the multidimensional phenotypic strategy radar chart and Figure 8 The verification of the measured biomass shown in (d) is explained as follows:

[0210] (1) High-yield and high-regeneration integrated type (Type 4, N=75) is the core of the resource nursery, accounting for 59.05%. It has the highest canopy coverage (AUC). CC It has the best composite regeneration index and the dark green leaf color (lowest RGB_Mean_R).

[0211] (2) The upright and sparse type (Type 3, N=16) exhibits a significant vertical growth priority strategy. Its average plant height (Mean pH) is basically the same as that of Type 4, but its cumulative coverage is lower and its plant compactness is lower.

[0212] (3) High-yielding and durable type (Type 2, N=4) is a rare and specific germplasm. Its most prominent feature is that it has an extremely high seasonal stability index, which is about 3 times that of other groups.

[0213] (4) Dwarf and compact type (Type 1, N=32), which corresponds to the phenotype of short plant size and slow growth observed in the field. Its average plant height and regeneration rate are the lowest in the whole group.

[0214] Figure 8 This is a schematic diagram of the mechanism-based decomposition clustering results guided by knowledge of the growth process, where... Figure 8 (a) is a schematic diagram of the optimal cluster number k value based on the Bayesian Information Criterion (BIC) and the Silhouette Score; Figure 8 (b) is a schematic diagram of the distribution pattern and confidence ellipsoid aggregation features in the two-dimensional t-SNE feature space; Figure 8 (c) is a schematic diagram illustrating the differences in characteristics of different clusters in terms of production potential (Dim A), regenerative resilience (Dim B), and multidimensional phenotypic strategies (Dim C); Figure 8 (d) is a validation diagram of the clustering results based on ground-measured biomass and various subtypes; Figure 8 (e) is a spatial distribution map of each fractal. Figure 8 In (b), the ellipse represents the 95% confidence interval for each cluster. It's important to note that the cluster labels are determined by a Gaussian Mixture Model (GMM) in a high-dimensional feature space, not based on the two-dimensional projection positions in the graph. Therefore, the small amount of inter-cluster overlap or outliers appearing in the graph are attributed to the dimensionality reduction loss when projecting high-dimensional data onto a two-dimensional plane, which is normal in nonlinear manifold learning.

[0215] The four types of germplasm mentioned above correspond to distinctly different breeding potentials and production application values. The "high-yield, high-regeneration integrated type," with its excellent cover recovery ability and high biomass potential, is an ideal parent for breeding varieties tolerant of intensive cutting, suitable for establishing high-yield hay production bases, supporting shorter cutting cycles to increase annual cutting counts and improve average annual crude protein yield. The "upright, sparse type," due to its open canopy structure allowing more light to penetrate to the understory, is extremely valuable in legume-grass mixed planting systems, effectively mitigating the shading inhibition of grasses by legumes; simultaneously, its upright plant type improves livestock accessibility, making it suitable as a variety for establishing rotational grazing systems. Although the "high-yield, stable, and durable type" germplasm resources are scarce, its excellent seasonal stability makes it a key gene source for addressing climate change and breeding widely adaptable and durable varieties, effectively alleviating forage supply gaps caused by high summer temperatures. The "dwarf, sturdy, and dense" type is ideal for improving grazing grasslands to prevent soil erosion. Furthermore, by removing individuals with severely stunted growth, this group can be used to select truly grazing-tolerant and dense specific germplasm. This mechanism-based classification system allows breeders to move beyond single yield indicators and precisely target specific production objectives (such as mowing tolerance, stable yield, mixed sowing, or grazing tolerance) to identify suitable germplasm.

[0216] To more clearly illustrate the technical solution and actual performance of the present invention, the method described in the present invention will be described in detail below with reference to specific embodiments. This embodiment takes multiple alfalfa germplasm resources as the research object. Through time-series image acquisition by field drones, segmentation of single alfalfa instances, decoupling of growth mechanisms, and Gaussian mixture model clustering, the production potential, regeneration resilience, and multidimensional phenotypic strategy evaluation and functional typing of alfalfa germplasm resources are completed, verifying the feasibility and effectiveness of the method of the present invention.

[0217] This invention utilizes the DJI Mavic 3 Multispectral (M3M) drone (DJI Technology Co., Ltd., Shenzhen, China) as a platform for acquiring high-throughput phenotypic data. This platform integrates a visible light and multispectral imaging system, equipped with one 20-megapixel visible light camera and four 5-megapixel multispectral cameras, covering the Green, Red, Red Edge, and NIR bands respectively. The system integrates an RTK (Real-Time Kinematic) module, enabling the simultaneous acquisition of high spatial resolution structural information and spectral information related to crop physiology. Automatic flight paths are planned using DJI Pilot 2 software, with both forward and lateral overlap rates of 75%. Before each flight, ground correction images are taken using a radiometrically calibrated gray card.

[0218] This invention conducted full-cycle phenotypic monitoring and UAV remote sensing data acquisition of alfalfa across its three complete growth cycles in 2025, carrying out 12 flight missions. These missions covered key growth stages, including seedling, branching, budding, and initial flowering, as well as different periods before and after two harvests. Each flight mission encompassed five flight altitudes: 12m-45° oblique photography (GSD: 0.869 cm / pixel), 15m (GSD: 0.69 cm / pixel), 20m (GSD: 0.92 cm / pixel), 30m (GSD: 1.38 cm / pixel), and 60m (GSD: 2.77 cm / pixel). All flight missions were conducted daily between 10:00 and 14:00 under relatively stable lighting conditions and low wind speeds to minimize the impact of shadow variations, motion blur, and radiometric inconsistencies on image quality. The acquired raw aerial images were imported into PIX4Dmapper software (v4.5.6, Pix4D SA, Switzerland) for drone image stitching. For specific flight times and data processing procedures, please refer to [link to documentation]. Figure 5 Among them, July 7 and August 11, 2025 are the harvesting nodes.

[0219] Between May and September 2025, approximately 70,000 raw RGB and multispectral images were collected. The stitched UAV orthophotos were cropped to the study area size using a sliding window strategy, resulting in images of 1024 pixels. Image patches of 1024 pixels each. After filtering out blurry or invalid samples with no vegetation cover, an experimental dataset containing 849 high-quality image patches was finally constructed, totaling approximately 14,700 individual plant instances. It should be noted that the canopy closure period was not included in the construction of the individual plant segmentation dataset.

[0220] To reduce manual costs while maintaining annotation accuracy, this invention employs the ISAT (https: / / github.com / yatengLG / ISAT_with_segment_anything) semi-automatic annotation tool to construct ground truth for instance segmentation. During the annotation process, a high-precision pre-trained weight, sam_vit_h_4b8939.pth (ViT-H backbone, 2.6 GB), is loaded. Clicking on the region of interest automatically generates a high-precision mask that fits the plant edges. For complex samples with severe adhesion, minor manual adjustments are made. Each individual plant in each image is assigned a unique instance label, defined as plant_1 to plant_n, and distinguished by different colors.

[0221] To comprehensively evaluate the data efficiency advantages of the DINO-XMem algorithm proposed in this invention, two comparative settings were designed for the experiment:

[0222] (1) Full-supervision setting: The SOTA model used for comparison was trained using all 849 labeled images. The dataset was randomly divided into training, validation and test sets in a ratio of 7:2:1.

[0223] (2) Few-shot Setting: For inference of the DINO-XMem network and the classic few-shot segmentation (FSS) baseline model, only a very small number of samples are extracted from the fully supervised training set as the support set. A scale-temporal dual hierarchical sampling strategy is adopted to address the dual challenges brought about by resolution differences and phenological changes in UAV imagery. First level (scale grouping): The training data is first grouped by flight altitude. When constructing each N-way K-shot task, the support set and query set are strictly limited to the same scale group to eliminate interference caused by inconsistent ground resolution (GSD). Second level (temporal hierarchical): Within each scale group, to ensure that the very small support set can also cover the complete morphological evolution from seedling to mature plant, the data acquisition time axis is divided into three time windows: early regeneration / seedling stage (S1), rapid elongation stage (S2), and early canopy closure stage (S3). When constructing the K-shot (K= 5, 10) task, the support set samples are forced to be evenly distributed across the three time windows mentioned above (e.g., the 5-shot setting includes 1 sample from period S1, 2 samples from period S2, and 2 samples from period S3). All support set sampling uses a fixed random seed (Seed=42) to ensure reproducibility. Performance evaluation in the Few-shot setting was performed using an independent test set that was completely identical to that in the Full-supervision setting.

[0224] To verify the advancement and robustness of the proposed DINO-XMem network in complex field scenarios, it was systematically compared with current state-of-the-art (SOTA) instance segmentation models. The comparison baselines covered representative models with different technical architectures, specifically divided into two categories: (1) Classic fully supervised models, including single-stage efficient networks (YOLOv11-seg, YOLOv13-seg, YOLACT) and classic two-stage segmentation networks (Mask R-CNN). The training epochs were all 300. (2) Classic few-shot segmentation (FSS) models, including PANet, PFENet, and HSNet, were used to compare the generalization performance of the models under extremely low labeled samples. The training epochs for both the classic few-shot segmentation models and the DINO-XMem model were 20.

[0225] The model training and inference of this invention are both performed on a high-performance computing workstation, equipped with an NVIDIA GeForce RTX 5090D GPU and an AMD Ryzen Threadripper 7960X 24-core CPU. All input images are 1024. 1024 pixels. Average Intersection over Union (mIoU), Dice coefficient, and mAP were used. 50 and mAP 50-95 The segmentation model for each instance was evaluated using four metrics, including the coefficient of determination (COP). The three statistical indicators, root mean square error (RMSE) and mean absolute error (MAE), are used to quantitatively evaluate the consistency between the phenotypic information extracted by the UAV and the ground-based measured data.

[0226] This invention systematically evaluates DINO-XMem against mainstream fully supervised instance segmentation models and classic few-shot segmentation methods. Experimental results are as follows: Figure 6 As shown in Table 1, the results demonstrate that DINO-XMem exhibits significant advantages across almost all evaluation metrics. In the 10-shot setting, DINO-XMem achieves an mIoU of 89.54% and a Dice coefficient of 94.07%, significantly outperforming the strongest baseline model in fully supervised mode, YOLOv11-Seg (mIoU = 84.02%, Dice = 91.13%). In the more challenging 5-shot scenario, DINO-XMem's mIoU (87.29%) surpasses models trained with nearly a thousand fully labeled images, such as Mask R-CNN (81.11%) and YOLOv13-Seg (83.42%). In contrast, while the classic FSS model performs robustly on some metrics, its overall performance still lags significantly behind DINO-XMem, especially in terms of fine-grained instance segmentation at high IoU thresholds.

[0227] Table 1. Test results of different segmentation models

[0228]

[0229] This performance breakthrough has significant practical implications for breeding. Traditional fully supervised deep learning models heavily rely on large-scale manual annotation. According to actual measurements, annotating a single 1024×1024 alfalfa image takes approximately 7-20 minutes. For phenotypic identification of large-scale germplasm resources, the annotation cost of constructing a fully supervised dataset is extremely high. Experimental comparisons clearly demonstrate that the DINO-XMem method, using only a very small dataset, breaks the strong dependence of full-sample models on large-scale labeled data, achieving practical instance segmentation accuracy with higher data efficiency, and is more suitable for real-world applications where annotation resources are scarce.

[0230] To visually demonstrate the differences in segmentation between different models at the seedling stage, branching stage, regenerated seedling stage, and budding stage, representative models with superior comprehensive performance from each model system were selected for comparative visualization. The results are as follows: Figure 7 As shown, different models exhibit significantly different error types. While fully supervised models have advantages in macroscopic localization, they still have limitations in handling fine morphology and complex backgrounds. Although YOLOv11-seg has accurate localization, its mechanism of detecting before segmenting results in a noticeable "rectangular" characteristic at the mask edges (e.g., ...). Figure 7 As shown in the middle red circle), it did not perfectly match the natural serrated boundary of the plant; at the same time, it was missed in the regeneration seedling stage where the characteristics were not obvious (such as...). Figure 7 (As shown in the yellow box in the middle), this reflects the insufficient sensitivity of the detection-driven mechanism to weak feature targets. Mask R-CNN exhibits lower semantic discriminative power and is easily affected by weeds during the seedling stage, leading to false detections (such as...). Figure 7 (As shown in the blue box), and frequent missed cuttings occurred in the later stages of dense plant growth (such as...). Figure 7 (as shown in the yellow box) and incomplete boundary coverage (such as) Figure 7 The problems (circled in yellow) indicate that its generalization ability in complex farmland contexts still needs improvement. While PFENet (10-shot), serving as the FSS baseline, reduced data dependency, it exhibited severe boundary blurring and localized adhesion blurring phenomena during the budding stage at high canopy closure (e.g., Figure 7 As shown in the yellow circle, it cannot effectively handle physical isolation between instances.

[0231] In contrast, the DINO-XMem (10-shot) proposed in this invention demonstrates significant advantages with extremely low sample sizes. This model not only effectively suppresses weed interference during the seedling stage, solving the false detection problem faced by Mask R-CNN and PFENet, but also overcomes the rectangular artifacts of the YOLO series by leveraging the powerful self-supervised feature representation capabilities of DINOv3, achieving high-fidelity fitting of fine alfalfa edges. Crucially, during the heavily clustered budding stage, DINO-XMem can clearly define the boundaries between individual plants, avoiding the missed cuts and blurring issues of the baseline model. Experimental results demonstrate that DINO-XMem, by mining high-resolution general feature priors, not only corrects the structural defects of fully supervised models using only a very small number of samples, but also comprehensively surpasses existing mainstream algorithms in both semantic accuracy and spatial detail of segmentation.

[0232] In summary, achieving low-cost and high-precision single-plant segmentation is a prerequisite for analyzing crop growth dynamics in high-throughput phenotypic analysis in the field. The DINO-XMem model proposed in this invention, by integrating the general representation of a visual base model with a dual-memory interaction mechanism, demonstrates significant advantages over traditional methods in terms of data efficiency and robustness. Firstly, in terms of data efficiency, DINO-XMem overcomes the dependence of fully supervised models on massive amounts of labeled data. Classic fully supervised instance segmentation networks typically rely on inductive bias to learn features from scratch, requiring long training periods to fit the data distribution. In contrast, DINO-XMem utilizes the frozen DINOv3 backbone network to inherit robust object-level semantic priors, requiring only a very small number of parameters to be fine-tuned via LoRA. It can quickly adapt to field scenarios under extremely short training periods (20 epochs) and extremely low sample conditions. This "pre-training-fine-tuning" paradigm not only significantly reduces the time cost of phenotypic extraction but also endows the model with generalization resilience to cope with phenotypic differences at different growth stages. Secondly, addressing the detail loss caused by "prototype compression" in traditional few-shot segmentation methods, DINO-XMem's innovative "dual-memory differential interaction" mechanism plays a crucial role. By constructing dual memories for the foreground and background, the model not only explicitly suppresses weed interference similar to crop textures using negative background references, but also replaces coarse-grained prototype matching with fine-grained patch-level feature interactions. Combined with a boundary-first sampling strategy, it successfully reconstructs clear physical boundaries between semantically very similar adjacent plants. This mechanism improvement ensures the acquisition of high-fidelity single-plant masks before crop canopy closure, laying a solid data foundation for subsequent extraction of reliable plant height, coverage, and multi-dimensional morphological features.

[0233] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A method for dynamic phenotypic analysis and germplasm evaluation of alfalfa based on few-sample temporal segmentation and growth knowledge guidance, characterized in that, The method includes the following steps: S1. Collect multi-temporal UAV remote sensing data of alfalfa in the field, construct the DINO-XMem few-sample instance segmentation network, and use the network to segment single alfalfa instances in the alfalfa images in the field. S2. Based on the segmentation results of single alfalfa instances, high-throughput extraction of multi-dimensional phenotypic traits of alfalfa and inversion of key growth indicators are performed. Temporal phenotypic information at the single-plant scale and the plot scale is integrated to construct a multi-dimensional time-series high-throughput phenotypic dataset. S3. Based on prior knowledge of alfalfa growth process, the knowledge-guided time-series growth mechanism of the multidimensional time-series high-throughput phenotypic dataset is decoupled to obtain multiple mutually orthogonal biological growth dimensions. S4. Construct germplasm-specific characterization vectors by integrating the characteristics of various biological growth dimensions, and perform probabilistic clustering of the germplasm-specific characterization vectors using a Gaussian mixture model. Based on the clustering results, complete the functional classification of alfalfa germplasm resources, and realize dynamic analysis of alfalfa phenotypes and germplasm evaluation.

2. The method for dynamic analysis of alfalfa phenotypic traits and germplasm evaluation based on few-sample temporal segmentation and growth knowledge as described in claim 1, is characterized in that... Step S1 specifically includes: S11. Acquisition and Preprocessing of Multi-Temporal Remote Sensing Data At multiple observation points throughout the alfalfa growth period, remote sensing images of the field were collected using drones. The original images were orthorectified, stitched together and pixel registered to generate orthophotos and canopy height models. A multi-temporal remote sensing dataset was constructed and divided into labeled support samples and unlabeled query samples. Construction and training of S12 and DINO-XMem networks The DINOv3 visual Transformer with frozen pre-trained parameters was used as the backbone network for feature extraction. A trainable LoRA adapter was embedded in the deep structure of the Transformer. Based on the support samples, a foreground feature memory bank of alfalfa and a complex background feature memory bank of the field were constructed to form a dual memory bank. The foreground feature enhancement of alfalfa plants and the suppression of background interference were carried out through the dual memory cross-attention mechanism to complete the construction and training of the DINO-XMem model. S13, Dual-Memory Cross-Attentional Interaction and Conditional Feature Injection The query sample is input into the backbone network to extract the query feature sequence. The query sample features are then interacted with the alfalfa foreground feature memory and the field complex background feature memory to generate foreground attention features and background attention features. The two attention features are then fused into a conditional feature increment using a differential strategy of foreground enhancement and background suppression. This increment is then injected into the query features of the backbone network. S14. Multi-layer feature decoding and single-plant instance segmentation output The conditionally enhanced features are subjected to multi-layer feature fusion, upsampling and boundary refinement to output a single alfalfa instance segmentation mask.

3. The method for dynamic analysis of alfalfa phenotypic traits and germplasm evaluation based on few-sample temporal segmentation and growth knowledge as described in claim 2, is characterized in that... Step S13 specifically includes: S131. Constructing an alfalfa-background dual-memory library Based on support sample feature maps and binary mask Construct a high-confidence foreground clover texture memory library Interference background memory bank The specific formula is as follows: (1) (2) in, Indicates support for sample number A feature token for each location; This represents the set of feature tokens that support the flattened feature map of the sample. , These are the mask values ​​corresponding to the i-th and j-th positions, respectively; Let be the foreground purity function. As the overall prototype of the foreground plant, Cosine similarity is used as a metric. The purity threshold, The threshold for selecting negative samples; S132, Differential Cross-Attention Injection The following formula flattens the features of the query image into a sequence. The query features are used as query vectors, and then compared with the key / value pairs respectively. and Perform multi-head cross-attention interaction to generate foreground attention features. and background attention features : (3) (4) Where N is the total number of observations; MHCA stands for Multi-head Cross-Attention Mechanism; The query feature matrix is ​​the target image, i.e., the query image. To support the sample, i.e., the key feature matrix of the example image; To support the value feature matrix of the sample; , and These are the learnable query, key, and value linear projection weight matrices, used to map features to a unified attention computation subspace; A differential strategy of positive enhancement and negative suppression is adopted, and conditional feature increments are generated using the following formula. : (5) in, The background suppression coefficient is... Indicates feature level; S133, Prior Guidance Space Gating First calculate query features Prototype of the foreground plant Pixel-level cosine similarity is used to generate a prior similarity map. Then, the prior similarity map is used as a spatial gating signal to apply the feature increment of the differential attention output. Weighting is applied; finally, local write constraints for conditional features are achieved through Hadamard product and Sigmoid activation, specifically implemented using the following formula: (6) in, This represents the updated query feature sequence at layer (l+1). This represents a feedforward network used to perform nonlinear transformations on the weighted features. This indicates a normalization operation; This represents the updated query feature sequence at level l. This represents the conditional feature increment generated by the difference across attention; It represents the Hadamardi (or Hadama) stack; This represents the Sigmoid activation function; This represents the learnable scaling factor.

4. The method for dynamic analysis of alfalfa phenotypic traits and germplasm evaluation based on few-sample temporal segmentation and growth knowledge as described in claim 3, is characterized in that... Step S14 specifically includes: S141. Multi-layer Feature Alignment and Fusion Obtain the feature pyramid output by the feature extraction backbone network. ,in and through Convolutional layer Map features at different levels to a unified channel dimension. And using bilinear interpolation Spatial resolution of features at each level: (7) in, This indicates that the data has passed through a 1×1 convolutional layer and utilized bilinear interpolation. After spatial resolution alignment, the first Layer feature map; This represents the feature pyramid output by the feature extraction backbone network, where the first... The original encoded feature map of the layer; Symbols representing universal quantifiers in mathematics; The following formula is used to concatenate all aligned features across all levels, and then a convolutional reshaping module is applied. Processing generates fused features that combine global semantic features and local spatial features. : (8) in, [ ] indicates the feature channel concatenation operation, which stacks and merges multiple aligned feature maps within the brackets along the channel dimension, thereby fusing feature information at different depths; , , , These represent the aligned feature maps from layers 2, 5, 8, and 11 of the feature extraction backbone network DINOv3, respectively. S142, Pixel-level boundary refinement and single-plant instance segmentation output Based on multi-layer fusion features Generate coarse-grained segmentation prediction results and extract the original image I. img The high-frequency components in the sample are used to obtain high-frequency detail information; the high-frequency detail information is interacted with the fused features, and the boundary refinement residual is generated through the residual refinement function; the coarse-grained segmentation prediction result is added to the boundary refinement residual, and after processing by the Sigmoid activation function, the segmentation mask of a single alfalfa instance is output.

5. The method for dynamic analysis of alfalfa phenotypic traits and germplasm evaluation based on few-sample temporal segmentation and growth knowledge as described in claim 4, characterized in that, The segmentation mask for the single alfalfa plant instance As shown in the following formula: (9) in, This is a coarse-grained prediction function; This represents a fusion feature that integrates global semantic features and local spatial features. ( ) represents the Sigmoid activation function; To utilize the original image The residual refinement function for high-frequency information is defined as: ； in, This represents a low-resolution feature map output by the backbone network or across attention modules; Represents the original input image Or high-resolution features in the shallow layer; Indicates a refinement decoder; This indicates a high-frequency feature encoder; Extracting the original image The high-frequency components in the image are the clear canopy outline and the gaps in the shadows between plants. A cascading symbol indicating feature splicing; This represents the feature projection and alignment operator.

6. The method for dynamic analysis of alfalfa phenotypic traits and germplasm evaluation based on few-sample temporal segmentation and growth knowledge as described in claim 5, is characterized in that... Step S2 specifically includes: S21. Generation of Region of Interest for a Single Plant: The segmentation mask of a single alfalfa instance output by the DINO-XMem network is vectorized to construct a region of interest for each alfalfa plant, and is assigned a variety identifier and spatial coordinate information to establish a variety-plant-spatiotemporal association index. S22. High-throughput extraction of multi-dimensional phenotypic traits: Based on the region of interest of a single alfalfa plant, refined phenotypic traits of alfalfa plants are extracted from five dimensions: morphology, structure, color, texture, and size. S23. Inversion and multi-scale supplementation of key growth indicators: Based on the canopy height model, invert the individual plant height and canopy coverage, and switch to the plot scale for population-level phenotypic statistics after alfalfa enters the canopy closure period. S24. Construction of a multidimensional time-series high-throughput phenotypic dataset: The refined phenotypic traits at the individual plant scale and the population indicators at the micro-plot scale are spatiotemporally matched with the timestamps and spatial coordinates of multi-temporal observations to construct a multidimensional time-series high-throughput phenotypic dataset covering the individual plant to micro-plot scale and containing multidimensional traits and continuous time-series information.

7. The method for dynamic analysis of alfalfa phenotypic traits and germplasm evaluation based on few-sample temporal segmentation and growth knowledge as described in claim 6, is characterized in that... Step S3 specifically includes: S31. Calculation of production potential dimension: Based on the time series phenotypic data in the multidimensional time series high-throughput phenotypic dataset, the Logistic growth model is used to fit the plant height and canopy coverage growth curves, and the average plant height, cumulative canopy coverage, maximum growth rate and growth inflection point parameters of the whole cycle are extracted. S32. Calculation of the regeneration resilience dimension: Based on the regeneration time trajectory of alfalfa after multiple harvests, the regeneration lag time, regeneration rate, regeneration peak baseline and seasonal stability index are extracted to construct a composite regeneration resilience index and a seasonal stability index. S33. Calculation of the dimensions of the multidimensional phenotypic strategy: Principal component analysis is used to reduce the dimensionality of multiple alfalfa phenotypic traits and generate low-dimensional feature components. S34. Decoupling of multidimensional biological mechanisms: The three types of features obtained from the calculation, namely production potential, regeneration resilience, and multidimensional phenotypic strategies, are orthogonally integrated to generate mutually independent biological dimensional features.

8. The method for dynamic analysis of alfalfa phenotypic traits and germplasm evaluation based on few-sample temporal segmentation and growth knowledge as described in claim 7, is characterized in that... Step S31 specifically includes: S311. Calculate the average plant height over the entire period using the following formula. : (10) in, For the i-th observation time point Plant height, N is the total number of observations; S312. Calculate the cumulative canopy coverage using the following formula. : (11) in, For time points Canopy coverage, For time points Canopy coverage, For inclusion The total number of time points throughout the entire cycle; S313. Based on the first derivative of plant height, calculate the maximum growth rate using the following formula. : (12) in, This represents the function for finding the maximum value. and This represents two consecutive observation points within the rapid growth period. This represents the set of condition constraints for the analysis window. Indicates the observation time node Corresponding plant height, Indicates the observation time node Corresponding plant height, This indicates that the calculation of formula (12) is performed within the time window of the rapid growth phase.

9. The method for dynamic analysis of alfalfa phenotypic traits and germplasm evaluation based on few-sample temporal segmentation and growth knowledge as described in claim 8, is characterized in that... Step S32 specifically includes: S321. Calculate the composite regeneration index using the following formula. : (13) in, and These represent plant height and coverage on the first day of drone observation after harvesting, i.e. Time value; and The value of the stubble reference point. For the number of days of regeneration; S322. Based on the reciprocal of the coefficient of variation of the maximum plant height over three growth cycles, the seasonal stability index is calculated using the following formula. : (14) in, This represents the maximum plant height during the c-th growth cycle. and These represent the mean and standard deviation of the peak values ​​for the three harvests, respectively. To prevent tiny quantities with a denominator of zero.

10. The method for dynamic analysis of alfalfa phenotypic traits and germplasm evaluation based on few-sample temporal segmentation and growth knowledge as described in claim 9, characterized in that, Step S4 specifically includes: S41. Construction of germplasm-specific characterization vectors: The features of the three dimensions of production potential, regeneration resilience and multidimensional phenotypic strategy obtained by decoupling are standardized, and the three types of features after processing are spliced ​​and integrated to integrate the features of various biological growth dimensions and construct the germplasm-specific characterization vector for each alfalfa germplasm. S42. Probability density clustering: The germplasm-specific representation vector of each alfalfa germplasm is input into the Gaussian mixture model. The optimal number of clusters is determined by the Bayesian information criterion and the silhouette coefficient. Unsupervised probability density clustering is performed, and the distribution characteristics of different cluster categories are represented by the covariance ellipse. S43. Functional typing of alfalfa germplasm resources: Based on the probability density clustering results, alfalfa germplasm is functionally classified to complete the functional typing of alfalfa germplasm resources, and finally realizes dynamic analysis of alfalfa phenotypes and germplasm evaluation.

Images