A transferable plant growth coordinate construction method based on agronomic supervision shaping and significant subspace alignment and application thereof

Abstract

The application discloses a kind of based on agricultural supervision shaping and significant subspace alignment transferable plant growth coordinate construction method and its application, it is related to computer vision technical field.The present application trains potential space projector using source domain single trait or multi-trait supervised training, discards trait prediction output after training convergence, only retains intermediate latent feature;Carry out principal component analysis to the latent feature, define the projection score of sample in the first principal direction as plant growth coordinate;After freezing visual encoder, set low-rank residual adapter before freezing source domain projector, realize no label target domain significant subspace light alignment.The present application can output unified continuous plant growth coordinate, support cross-species, cross-scene no label migration and development deviation monitoring;Core process relies on mature algorithm module to realize, can be packaged as server interface, edge end application, agricultural management platform module or control system component, with good industrial transformation prospect.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and in particular to a method for constructing transferable plant growth coordinates based on agronomically supervised shaping and salient subspace alignment, and its application. Background Technology

[0002] Plant phenotyping monitoring, environmental control in protected agriculture, and plant growth prediction all require obtaining numerical values ​​characterizing the current developmental state of plants from continuously acquired images. Existing image phenotyping methods typically train prediction models separately for fresh weight, dry weight, plant height, crown width, leaf area, number of organs, or stress category. The output of such models corresponds to a pre-selected single task. Once the crop type, variety, cultivation method, camera perspective, or lighting conditions are changed, labels often need to be re-acquired and retrained, making it difficult to form a unified state variable that can be reused by time-series prediction, anomaly detection, and control algorithms.

[0003] Obtaining agricultural labels is costly. Indicators such as fresh weight and dry weight typically require destructive sampling, while leaf area, plant height, and canopy width require manual measurement or specialized equipment, making it difficult to achieve the acquisition frequency of ordinary RGB images. Therefore, data structures with "many images, few labels" are common in production scenarios. While direct regression models can predict specific traits using a small number of labels, their final output is still limited to that trait and cannot naturally describe the overall developmental process of plants under the combined changes of multiple traits.

[0004] General-purpose visual models can extract high-dimensional features from RGB images, and unsupervised methods such as Principal Component Analysis (PCA) can also extract the main directions of change from high-dimensional features. However, if PCA is directly applied to visual features that have not been task-shaped, the first principal component may mainly reflect background, lighting, viewpoint, container, variety differences, or plant identity, rather than necessarily representing growth and development. Manually constructed indicators such as green area, outline, perimeter, roundness, and edge density are also easily affected by occlusion, canopy overlap, imaging scale, and color drift, making it difficult to stably form developmental coordinates across scenes.

[0005] On the other hand, there are significant domain shifts among different crops, varieties, and collection systems. Target scenes often lack agronomic labels consistent with source scenes, making supervised fine-tuning unsuitable. Existing domain adaptation methods are mostly geared towards classification tasks, relying on classification entropy, pseudo-labels, or statistical matching across the entire feature space. For continuous trait regression and its latent space, information truly relevant to plant development is often concentrated in only a few directions, and blindly aligning all features may weaken the original developmental structure.

[0006] Therefore, those skilled in the art are dedicated to developing a method for constructing transferable plant growth coordinates based on agronomically supervised shaping and significant subspace alignment, and its applications. Summary of the Invention

[0007] In view of the above-mentioned deficiencies of the prior art, the technical problem to be solved by the present invention is to develop a method for constructing plant growth coordinates that can migrate across scenes without target domain labels.

[0008] To achieve the above objectives, this invention provides a method for constructing transferable plant growth coordinates based on agronomically supervised shaping and significant subspace alignment, and its application.

[0009] Furthermore, a method for constructing transferable plant growth coordinates based on agronomically supervised shaping and salient subspace alignment includes the following steps: S1. Collect source domain plant RGB images with agronomic trait tags and extract image spatial markers using a frozen visual encoder; S2. For the visual encoder that outputs spatial markers, remove the CLS marker, register marker, or other non-spatial markers contained in the spatial markers, and perform normalization and average pooling on the remaining spatial markers respectively; for the visual encoder that outputs feature maps or image-level vectors, use global pooling, flattening, or corresponding image-level aggregation methods to obtain image-level features. S3. Input the image-level features into the source domain nonlinear projector and output compact latent features; S4. Use a single-task or multi-task agronomic trait prediction head to supervise the training of the source domain nonlinear projector. S5. After the source domain nonlinear projector training converges, fix all parameters, discard the agronomic trait prediction output, and retain only the intermediate potential features after agronomic supervision shaping. S6. Fit the PCA model using the intermediate latent features of the source domain training set, and define the projection score of the sample on the first principal direction of PCA as the plant growth coordinate.

[0010] Furthermore, in step S1, the RGB image needs to be scaled and cropped to a fixed resolution before pixel normalization is completed; The frozen visual base model is selected from one of DINOv2, CLIP, MAE, ResNet, or a visual autoencoder.

[0011] Furthermore, in step S3, the nonlinear projector is a multilayer perceptron with GELU activation.

[0012] Furthermore, in step S4, supervised training is carried out using mean squared error loss; The agronomic traits are selected from one or more of the following: fresh weight, dry weight, plant height, crown width, leaf area, number of leaves, water content, and nutritional quality. The prediction head for the agronomic traits is a linear layer or a lightweight regression network.

[0013] Furthermore, in step S6, after obtaining the projection score in the first principal direction of PCA, the positive and negative directions of the coordinates are corrected according to the acquisition time information or non-destructive anchor points, and the coordinates are subjected to z-score normalization or 0 to 1 interval normalization as needed.

[0014] Furthermore, it also includes lightweight alignment of the salient subspace of the target domain, with the following specific steps: Step 1: Perform PCA on intermediate latent features after agronomically supervised shaping, retaining the previous... Identify principal directions and record the variance of each principal direction; Step 2: Calculate the significant weights of each main direction: If it is a single-task agronomic trait prediction branch, take the weight vector of the corresponding agronomic trait prediction head; if it is a multi-task branch, take the absolute value of each element of the weight of the trait prediction head and then calculate the mean; based on the single-task prediction head weight vector or the mean weight vector of the multi-task prediction head, calculate the unnormalized weights of each source domain main direction, and then normalize the weights of all main directions. Step 3: After freezing the visual encoder and before the source domain nonlinear projector, a low-rank residual adapter is set. The adapter performs residual transformation on the target domain image-level features according to the following formula:

[0015] in, For target domain image-level features, To adapt the target domain image-level features, For parameters Low-rank residual adapter, For dimensionality reduction mapping, For dimensional mapping, For image-level feature dimensions, It is a low-rank dimension and , It is a learnable residual metric.

[0016] Step 4: Acquire RGB images of target domain plants without agronomic trait labels. Extract spatial labels from the images using a frozen visual encoder. Obtain latent features of the target domain through a low-rank residual adapter and a source domain nonlinear projector that has been trained, converged, and has fixed parameters. Map the latent features of the target domain to the salient subspace of the source domain and then calculate the mean and variance of each principal direction in batches. Calculate the bidirectional KL divergence of the feature distributions of the source and target domains for each PCA principal direction. Use the aforementioned normalized salient weights to weight the KL divergence corresponding to each principal direction to obtain the unlabeled alignment loss function and complete the model training and optimization. During the training phase, only the target domain adaptation parameters are updated. Step 5: After training is completed, the target domain PCA is independently fitted to the adapted target latent features and the plant growth coordinates are extracted. The agronomic trait labels of the target domain do not participate in the alignment training and coordinate fitting.

[0017] Furthermore, in step four, the target domain adaptation parameters include the parameters of the low-rank residual adapter and the scaling and bias parameters of the LayerNorm layer in the source domain nonlinear projector. The remaining parameters of the visual encoder, the agronomic trait prediction head, and the source domain nonlinear projector do not participate in the parameter update.

[0018] Furthermore, it also includes time-progressive monitoring steps for crops in target domains without similar agronomic trait labels: Label-free salient subspace alignment and adaptation are performed on continuous crop growth images to extract plant growth coordinates. Crop images of the same plant are sorted according to the acquisition time. Based on the correlation between coordinate difference and acquisition time, the consistency of coordinate temporal sorting, and the smoothness of coordinate trajectory, the relative development progress of crops is quantitatively characterized, realizing cross-species crop growth time-series monitoring under the condition of no similar agronomic trait labels.

[0019] Furthermore, a method for analyzing and regulating crop growth includes the following steps: A. Collect crop time-series RGB images using multiple fixed cameras, process the images at each time point using the method described in any one of claims 1 to 8, and complete coordinate orientation to obtain plant growth coordinates; B. Divide all time series coordinates into pre-disturbance and post-disturbance sequences; using the number of days elapsed as the independent variable, fit a growth baseline using the coordinates of the pre-disturbance sequence, and extrapolate this baseline to the post-disturbance period; the growth baseline adopts one of the following: a quadratic trend model, a linear regression model, a spline interpolation model, an exponential model, or a state-space model. C. Calculate the normalized backoff for each observation point after the disturbance; D. Calculate the average value of the normalized backoff amount at all time points after the disturbance of a single camera to obtain the average backoff amplitude of a single camera; then calculate the mean and standard deviation of the average backoff amplitude of a single camera corresponding to multiple cameras; alarm thresholds can be set based on historical normal batches, or the normalized backoff amplitude can be directly input into the environmental controller to complete the growth analysis and control.

[0020] Furthermore, a transferable plant growth coordinate construction system based on agronomically supervised shaping and salient subspace alignment includes an image acquisition module, a visual feature extraction module, a source domain supervised shaping module, a target domain adaptation module, a coordinate extraction module, a temporal analysis module, and an application interface module.

[0021] In a preferred embodiment 1 of the present invention, the process of constructing the source domain plant growth coordinates is described in detail.

[0022] In another preferred embodiment 2 of the present invention, the significant subspace alignment and cross-species migration process of the unlabeled target domain is described in detail.

[0023] In another preferred embodiment 3 of the present invention, the time-progression monitoring process for target domains without similar agronomic labels is described in detail.

[0024] In another preferred embodiment 4 of the present invention, the irrigation interruption deviation detection process based on plant growth coordinates is described in detail.

[0025] In another preferred embodiment 5 of the present invention, the composition and production deployment process of the plant growth coordinate construction system are described in detail.

[0026] Technical effects: 1. The method of this invention can generate one-dimensional continuous plant growth coordinates, which serve as a unified state variable to support various higher-level tasks such as growth curve plotting, treatment group comparison, maturity ranking, time series prediction, and intelligent environmental control. Validation results in the lettuce source domain show that, under the single-task scheme, the coordinates exhibit absolute subclass weighted differential correlation with both fresh and dry weights. The values ​​reached 0.958 and 0.964 respectively; the multi-task scheme affected plant height, crown width, and leaf area. The results achieved R values ​​of 0.934, 0.898, and 0.966, significantly outperforming traditional manual visual indicators. After fusing the coordinates of this invention with manual features such as green area, perimeter, roundness, Hu moment, and edge density, the verification of five agronomic traits (R values) was further enhanced. 2 It can still be improved by 0.464 to 0.601, proving that the coordinate contains developmental information beyond the manual phenotypic indicators.

[0027] 2. The method of this invention can achieve cross-species and cross-scenario migration under label-free conditions. In the cross-domain migration experiment from lettuce to tomato, the multi-task combination with a significant subspace alignment scheme showed a weighted differential correlation of absolute subclasses on fresh weight, plant height, and leaf area. The scores were 0.768, 0.810, and 0.816 respectively, significantly outperforming the multi-task representation scheme without adaptation. On the Komatsuna dataset lacking homologous agronomic labels, the temporal difference correlation of the multi-task + SSA scheme was [not specified]. With a score of 0.759, a sorting accuracy of 0.930, and a trajectory smoothness of 0.533, it can stably maintain the cross-species time-progression structure.

[0028] 3. The method of this invention can quantify continuous developmental deviations and can be used for developmental deviation detection or abnormal early warning. In a lettuce irrigation interruption experiment with 24 fixed cameras, all camera samples output positive regression; the average normalized regression amplitude of the multi-task + SSA scheme was 0.579±0.086, which has higher sensitivity for developmental deviation identification compared with the original DINOv2 feature.

[0029] 4. This invention has a low barrier to entry for engineering implementation and effectively reduces annotation costs. Source domain training only requires a batch of reference samples with agronomic labels, and the trait prediction head can be discarded during online deployment. Target domain adaptation requires no manual annotation throughout the entire process, significantly reducing the annotation workload for new crops, new varieties, new growth cycles, and new camera systems. The rank of the low-rank adapter is much smaller than the dimension of the input features, and the number of parameters involved in the update and the training computational cost are far lower than full-model fine-tuning, supporting rapid cold-start adaptation in offline batch mode. This solution can run using only ordinary RGB cameras, existing image databases, and general-purpose GPUs or servers, without requiring 3D scanning equipment, professional spectroscopic instruments, or continuous destructive sampling operations.

[0030] 5. This invention does not mandate that the absolute values ​​of plant growth coordinates for different crops remain consistent. Priority is given to ensuring that coordinates within a single domain change smoothly and orderly with development, and that they correlate with multiple agronomic traits. When a unified scale is required across systems, linear calibration can be achieved using source domain PCA basis vectors, a small number of anchor point samples, sowing time, and non-destructive dimensional measurements. The entire process is developed based on mature modules such as image acquisition, neural network inference, matrix factorization, and statistical alignment. It can be implemented by executing programs within a computer-readable medium via a processor. It can be packaged as a cloud service, edge inference device program, plant factory controller component, or agricultural management platform software module, possessing industrial transformation value and suitable for large-scale phenotypic monitoring scenarios in greenhouses, plant factories, seedling bases, and breeding platforms.

[0031] The following will further explain the concept, specific structure and technical effects of the present invention in conjunction with the embodiments and accompanying drawings, so as to fully understand the purpose, features and effects of the present invention. Attached Figure Description

[0032] Figure 1This is a flowchart of the overall process for source domain supervised shaping, target domain salient subspace alignment, and coordinate extraction of plant growth coordinates in this invention. Detailed Implementation

[0033] The following description, with reference to the accompanying drawings, illustrates several preferred embodiments of the present invention to make its technical content clearer and easier to understand. The present invention can be embodied in many different forms, and the scope of protection of the present invention is not limited to the embodiments mentioned herein.

[0034] Example 1: Construction of source area plant growth coordinates

[0035] This embodiment uses RGB images of lettuce with agronomic trait labels as the source domain data. The source domain data contains 388 images, each corresponding to one or more measurements of fresh weight, dry weight, plant height, crown width, and leaf area. To compare the effects of different characterization schemes, this embodiment simultaneously tests the original DINOv2, Single-task shaping, Multi-task shaping, and manual visual feature schemes. The construction steps are as follows: 1. Image preprocessing. The input image is uniformly converted to RGB format, scaled and cropped to a fixed resolution according to the default method of the visual encoder, and then normalized.

[0036] 2. Visual Encoder Freeze-out Visual Feature Extraction. The image spatial label matrix is ​​extracted using a visual encoder with fixed parameters (hereinafter referred to as "frozen visual encoder"). ,in, For image spatial labeling matrix, The number of image block markers retained after removing the CLS marker and register marker. The feature dimensions are labeled for each image patch. This embodiment uses DINOv2 with register labels, and the feature dimensions are... The frozen visual encoder does not participate in parameter updates during subsequent training.

[0037] 3. Image-level pooling. LayerNorm normalization is performed on each spatial label, and then... The average of the labels is used to obtain the image-level vector. The order can be written as: .in, Index the image patches. For the first Image block markers, This represents the LayerNorm operation. These are image-level features.

[0038] 4. Supervise the shaping process. Input Projector Output m-dimensional compact latent features , m is the potential dimension, in this embodiment The projector employs a multilayer perceptron with GELU activation. One or more linear prediction heads are then connected to the projector. Single-task branches predict only standardized fresh weight; multi-task branches predict standardized fresh weight, dry weight, plant height, canopy width, and leaf area separately.

[0039] 5. Loss Calculation. Mean squared error is used for single-task calculations. ,in, The number of training samples. For the first Standardized fresh weight prediction for each sample. This corresponds to the standardized fresh weight true value. For multi-task processing, the training set mean is subtracted from each trait and divided by the standard deviation, then the average of the five mean squared errors is used. To prevent training dominated by traits with large dimensions, among which, In this embodiment, the number of agronomic traits is [not specified]. , For the first The first sample Standardized trait prediction values.

[0040] 6. Preserve latent features. After training, fix the projector parameters, discard or bypass the prediction head, and do not use predicted values ​​such as fresh weight as the final plant state. Output only intermediate latent features for each image. .

[0041] 7. Extract the growth coordinates of dominant source domain plants. Fit PCA using only the latent features of the training set to obtain the mean. First main direction and other main directions. For potential features as The plant growth coordinates of the sample are defined as follows: in, The projection score of the sample onto the first principal direction of PCA, i.e., the source domain plant growth coordinates. This represents the transpose of a vector.

[0042] 8. Orientation and Scale Processing. Since the sign of the PCA direction is arbitrary, the positive direction can be determined based on fresh weight, sowing time, or other developmental references, so that later or larger plants correspond to higher source domain plant growth coordinates. Depending on the application requirements, c can be further normalized to a z-score or between 0 and 1; this processing does not change the sample order or latent space geometry.

[0043] In this embodiment, the absolute subclass weighted differential correlation of the source domain validation set As shown in Table 1, the Single group showed the best results in fresh weight and dry weight, while the Multi group showed the best results in plant height, crown width, and leaf area. This indicates that the growth coordinates extracted after supervised shaping were optimal. It can represent the common changes of multiple developmental traits in a concentrated manner.

[0044] Table 1

[0045] Further validation of the five traits was achieved by incorporating hand-manufactured proxy variables consisting of green area, average green index, perimeter, compactness, roundness, Hu moment, and edge density into the source domain plant growth coordinates. The increments of 0.464 to 0.601 indicate that the coordinates of this invention do not simply repeat canopy color or area information.

[0046] Example 2: Significant subspace alignment of unlabeled target domains and cross-species migration

[0047] This embodiment transfers the lettuce projector trained in Embodiment 1 to the tomato target domain. The tomato target domain contains 476 valid RGB images. Fresh weight, plant height, and leaf area are used only for final effect evaluation and are not involved in adaptation training or coordinate extraction. To compare the effects of different representation and adaptation schemes, this embodiment simultaneously tests the original DINOv2, Single, Multi, Single+SSA, and Multi+SSA schemes. The specific steps are as follows: 1. Construct source domain statistics. This involves analyzing the latent features of the source domain. Calculate the mean And perform PCA, retain the previous Composed of main directions Record the variance in each direction This embodiment .

[0048] 2. Calculate the weights for significant directions. For a single-task branch, use the weight vector of the fixed fresh weight prediction head. For multi-task branches, the absolute values ​​of the prediction head weights for each trait are taken element by element and then averaged to obtain the result. For the first Main direction of each source domain Calculate unnormalized weights Then normalize to get The more sensitive the prediction head is to the source domain direction, the higher its proportion in the alignment.

[0049] 3. Set up a low-rank residual adapter.

[0050] A low-rank residual adapter is set after the frozen visual encoder and before the source domain nonlinear projector. This adapter performs a residual transformation on the target domain image-level features according to the following formula:

[0051] in, For target domain image-level features, To adapt the target domain image-level features, For parameters Low-rank residual adapter, For dimensionality reduction mapping, For dimensional mapping, For image-level feature dimensions, It is a low-rank dimension and , It is a learnable residual metric.

[0052] 4. Obtain the salient subspace representation of the target. The target image is obtained through an adapter and a frozen source domain projector. , then calculate .in, For latent features of the target domain in the source domain The projection along each principal direction. For each target mini-batch, the mean is calculated along the k-th direction. and variance .

[0053] 5. Optimize the unlabeled alignment loss. The source domain corresponds to... Target batch corresponding Calculate the two-way KL divergence of the two Gaussian distributions for each direction, and then... Weighted summation: .in, Indicates a Gaussian distribution. Denotes KL divergence, This represents the significant subspace alignment loss.

[0054] 6. Limit trainable parameters. During training, keep the visual encoder, trait prediction head, and source domain projector body frozen, and only update the low-rank residual adapter; in an alternative, simultaneously update the scaling and bias of the LayerNorm within the projector. Optimize with Adam for 20 epochs, with a learning rate of... .

[0055] 7. Extract plant growth coordinates from the target domain. Independently fit the target domain PCA to the adapted latent features and extract the coordinates. The agronomic labels in the target domain do not participate in adaptation, PCA fitting, or coordinate extraction. When coaxial display of the source domain is required, the target features can also be projected onto the source domain PCA basis vectors.

[0056] The cross-species migration results of lettuce to tomato in each group are shown in Table 2. The multi-task + SSA approach achieved the highest values ​​in all three traits. This demonstrates that significant subspace alignment can improve growth coordinates under the condition of no target label.

[0057] Table 2

[0058] In the parameterized comparison, mean-variance matching and mismatch results were essentially the same, while CORAL full covariance alignment actually reduced performance; when only the low-rank adapter was trained, the average of the Single and Multi tomato traits was significantly improved. The values ​​were 0.763 and 0.797 respectively, and after jointly updating LayerNorm, they were 0.765 and 0.798 respectively, indicating that the performance improvement mainly comes from the low-rank residual adapter located before the source domain projector.

[0059] Example 3: Time-progress monitoring of target domains without similar agronomic tags

[0060] This embodiment uses 300 continuous growth images of Komatsuna (Japanese mustard greens). This data does not have the same agronomic trait labels as the source domain, but only has timestamps and plant identification information. The timestamps and identification are not used in coordinate construction, but only for post-evaluation. To compare the effects of different characterization and adaptation schemes, this embodiment simultaneously tests the original DINOv2, Single, Multi, Single+SSA, and Multi+SSA schemes. After completing the label-free adaptation and extracting the target domain plant growth coordinates according to Embodiment 2, the images of each plant are sorted by time.

[0061] The evaluation includes three aspects: First, the correlation between pairwise coordinate differences and time differences among samples of the same plant. Secondly, the accuracy of sorting is determined by whether the order of coordinates between any two time points is consistent with the time. Thirdly, after normalizing the coordinates of individual plants, the trajectory smoothness is obtained by removing the net positions of the start and end points and gradually summing the absolute displacements; the closer the value is to 1, the less the backswing. The results for each group are shown in Table 3. Table 3

[0062] Multi+SSA achieved the highest temporal difference correlation and trajectory smoothness, with a sorting accuracy second only to the original DINOv2. These results demonstrate that the coordinates of this invention can represent relative developmental progression in cross-species scenarios where no similar agronomic labels exist.

[0063] Example 4: Irrigation Interruption Deviation Detection Based on Plant Growth Coordinates

[0064] This embodiment uses a full-view time series of lettuce images acquired by 24 fixed cameras. Irrigation was interrupted from December 8, 2025, and observations before this point are defined as the pre-perturbation sequence, while those after are defined as the post-perturbation sequence. To compare the deviation recognition performance of different feature representation schemes, this embodiment simultaneously tests five schemes: the original DINOv2, Single, Multi, Single+SSA, and Multi+SSA. The following steps are performed for each camera and each representation: 1. Extract and orient plant growth coordinates from images at each time point. .in, This is an index for the observation time point.

[0065] 2. Based on the number of days elapsed As the independent variable, only the coordinates before perturbation are used to fit the quadratic trend. This trend is then extrapolated to the perturbation. The quadratic trend is only used in this embodiment; in practice, it can be replaced by linear regression, splines, exponents, state-space models, or other normal growth baselines.

[0066] 3. Calculate the normalized backoff amount for each time point after the disturbance. . Indicates reality

[0067] When the actual coordinates are lower than the normal extrapolated trajectory, it can correspond to an absolute decline, stagnant growth, or a slowdown in progress. Among these, For time point The normalized backoff quantity, The set of observation points before the disturbance. This is the index at the time point before the disturbance. When... When the actual coordinates are lower than the normal extrapolated trajectory, it indicates that the actual coordinates are lower than the normal extrapolated trajectory.

[0068] 4. Calculate the average over all time points after the disturbance to obtain the single-camera backoff amplitude. , Then calculate the mean and standard deviation for multiple cameras. Alarm thresholds can be set based on historical normal batches, or... Directly input to the environmental controller. Among them, This is the set of observation points after the disturbance. This represents the number of observation points after the disturbance. Subsequently, multiple cameras were... Calculate the mean and standard deviation, and set alarm thresholds based on historical normal batches, or... or Input environment controller.

[0069] The baseline fitting correlation coefficients and multi-camera average normalized backoff magnitudes for each scheme before perturbation are shown in Table 4.

[0070] Table 4

[0071] All 24 cameras showed positive regression, with Multi+SSA exhibiting the largest average regression. This embodiment demonstrates that the present invention can identify adverse environmental disturbances through continuous state deviations without requiring pre-training of a "water shortage category".

[0072] Example 5: System Composition and Production Deployment Method

[0073] A plant growth coordinate construction system for implementing the present invention includes at least: an image acquisition module for receiving plant RGB images from a fixed camera, mobile terminal, robot, or historical database; a visual feature extraction module for outputting image-level vectors by freezing the visual encoder; a source domain supervised shaping module for training a projector and a trait prediction head; a target domain adaptation module for training a low-rank residual adapter based on the salient subspace of the source domain; a coordinate extraction module for fitting PCA and outputting plant growth coordinates; a time series analysis module for performing coordinate orientation, normalization, trend fitting, and deviation calculation; and an application interface module for outputting coordinates and anomaly indicators to growth monitoring, yield prediction, environmental control, robot operation, or agricultural management platforms.

[0074] During production deployment, the source domain model can be trained first on a reference batch with destructive measurements. Then, for each new crop or imaging system, a batch of unlabeled images can be collected to complete SSA adaptation. During the online phase, a plant growth coordinate is output after each image acquisition and written to the database along with environmental sensor data, historical coordinates, and control actions. The upper-level system can adjust environmental parameters based on the crop development status represented by the plant growth coordinates, or use the coordinates as input to time-series prediction and intelligent control algorithms.

[0075] This method can be implemented by a processor executing a program stored on a computer-readable medium, and can also be deployed as a cloud service, edge inference device, plant factory controller, or built-in software module for agricultural robots.

[0076] The principle of the method for constructing transferable plant growth coordinates in this invention is as follows: 1. This invention utilizes source domain single-trait or multi-trait supervised training of a latent space projector. After training convergence, the predicted trait output is discarded, retaining only intermediate latent features. Principal component analysis (PCA) is performed on these latent features, and the projection score of the sample in the first principal direction is defined as the plant growth coordinate. Agronomic trait loss does not directly define or optimize the first principal direction of PCA, but rather forces the projector to reorganize frozen visual features, making shared developmental changes such as fresh weight, dry weight, plant height, crown width, and leaf area more concentrated in a compact latent space. The PCA retains one or more principal components; when multiple principal components are retained, the first principal component is the principal growth coordinate.

[0077] 2. For visual encoders that output spatial labels, this invention removes CLS labels, register labels, or other non-spatial labels, performs LayerNorm normalization and average pooling on a label-by-label basis; for visual encoders that output feature maps or image-level vectors, the corresponding image-level aggregation method is used to obtain image-level vectors. Single-task or multi-task supervised shaping is performed using a compact nonlinear projector; before multi-task training, the agronomic trait data are standardized to avoid single-dimensional traits dominating training loss. This process reduces the scheme's dependence on image segmentation, manual contour extraction, and single color thresholding.

[0078] 3. This invention sets up a low-rank residual adapter between the frozen visual encoder and the source domain projector; constructs a salient subspace based on the latent features of the source domain, and assigns corresponding weights to the sensitivity of each principal direction of the source domain according to the trained trait prediction head; the target image features are input into the frozen source domain projector after low-rank residual transformation, and further projected to several principal directions of the source domain; bidirectional KL divergence is used to match the Gaussian distribution statistics of the source domain samples and the target batch in each principal direction; the higher the sensitivity of the trait prediction head to a certain principal direction, the greater the weight of that principal direction in the alignment loss. During the training phase, only the low-rank adapter is updated, and the scaling and bias parameters of the LayerNorm inside the projector can also be updated synchronously to avoid drift of the entire model parameters.

[0079] 4. After label-free alignment of the target domain, this invention independently fits PCA and extracts plant growth coordinates using only the latent features of the target domain. It unifies the positive direction of the coordinates by combining the acquisition time, known developmental references, or a small number of non-destructive anchor points, and performs z-score normalization or interval normalization as needed. The sign of the PCA principal direction does not change the feature geometry. Independent fitting allows each target domain to extract its most significant developmental changes, without requiring the target domain to have the same agronomic labels as the source domain. When a unified numerical scale is needed across scenarios, linear calibration can be completed using the source domain PCA basis vectors or anchor point samples.

[0080] 5. This invention uses plant growth coordinates within the time series before the disturbance to fit the expected normal growth trend. The difference between the observed coordinates after the disturbance and the extrapolated trend is normalized based on the range of coordinate values ​​before the disturbance, resulting in a normalized backtracking amount that can be compared across cameras. This amount is then used as an input for alarm, prediction, or control. As long as the observed trajectory is lower than the normal extrapolated trajectory, it can be quantified as a positive backtracking amount, which can reflect an absolute decline, a plateau period, or a slowdown in the rate of advancement.

[0081] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for constructing transferable plant growth coordinates based on agronomically supervised shaping and salient subspace alignment, characterized in that, Includes the following steps: S1. Collect source domain plant RGB images with agronomic trait tags and extract image spatial markers using a frozen visual encoder; S2. For the visual encoder that outputs spatial markers, remove the CLS marker, register marker, or other non-spatial markers contained in the spatial markers, and perform normalization and average pooling on the remaining spatial markers respectively; for the visual encoder that outputs feature maps or image-level vectors, use global pooling, flattening, or corresponding image-level aggregation methods to obtain image-level features. S3. Input the image-level features into the source domain nonlinear projector and output compact latent features; S4. Use a single-task or multi-task agronomic trait prediction head to supervise the training of the source domain nonlinear projector. S5. After the source domain nonlinear projector training converges, fix all parameters, discard the agronomic trait prediction output, and retain only the intermediate potential features after agronomic supervision shaping. S6. Fit the PCA model using the intermediate latent features of the source domain training set, and define the projection score of the sample on the first principal direction of PCA as the plant growth coordinate.

2. The method for constructing transferable plant growth coordinates as described in claim 1, wherein in step S1, the RGB image needs to be scaled and centered to a fixed resolution before pixel normalization is completed; The frozen visual base model is selected from one of DINOv2, CLIP, MAE, ResNet, or a visual autoencoder.

3. The method for constructing transferable plant growth coordinates as described in claim 1, wherein in step S3, the nonlinear projector is a multilayer perceptron with GELU activation.

4. The method for constructing transferable plant growth coordinates as described in claim 1, wherein in step S4, mean square error loss is used for supervised training; The agronomic traits are selected from one or more of the following: fresh weight, dry weight, plant height, crown width, leaf area, number of leaves, water content, and nutritional quality. The prediction head for the agronomic traits is a linear layer or a lightweight regression network.

5. The method for constructing movable plant growth coordinates as described in claim 1, wherein in step S6, after obtaining the projection score in the first principal direction of PCA, the positive and negative directions of the coordinates are corrected according to the acquisition time information or non-destructive anchor points, and the coordinates are subjected to z-score normalization or 0 to 1 interval normalization processing as needed.

6. The method for constructing transferable plant growth coordinates as described in claim 1, further comprising lightweight alignment of the salient subspace of the target domain, the specific steps of which are as follows: Step 1: Perform PCA on intermediate latent features after agronomically supervised shaping, retaining the previous... Identify principal directions and record the variance of each principal direction; Step 2: Calculate the significant weights of each main direction: If it is a single-task agronomic trait prediction branch, take the weight vector of the corresponding agronomic trait prediction head; if it is a multi-task branch, take the absolute value of each element of the weight of the trait prediction head and then calculate the average. Based on the single-task prediction head weight vector or the mean weight vector of the multi-task prediction head, calculate the unnormalized weights of each source domain principal direction, and then normalize the weights of all principal directions. Step 3: After freezing the visual encoder and before the source domain nonlinear projector, a low-rank residual adapter is set. The adapter performs residual transformation on the target domain image-level features according to the following formula: in, For target domain image-level features, To adapt the target domain image-level features, For parameters Low-rank residual adapter, For dimensionality reduction mapping, For dimensional mapping, For image-level feature dimensions, It is a low-rank dimension and , A learnable residual metric; Step 4: Acquire RGB images of target domain plants without agronomic trait labels. Extract spatial labels from the images using a frozen visual encoder. Obtain latent features of the target domain through a low-rank residual adapter and a source domain nonlinear projector that has been trained, converged, and has fixed parameters. Map the latent features of the target domain to the salient subspace of the source domain and then calculate the mean and variance of each principal direction in batches. Calculate the bidirectional KL divergence of the feature distributions of the source and target domains for each PCA principal direction. Use the aforementioned normalized salient weights to weight the KL divergence corresponding to each principal direction to obtain the unlabeled alignment loss function and complete the model training and optimization. During the training phase, only the target domain adaptation parameters are updated. Step 5: After training is completed, the target domain PCA is independently fitted to the adapted target latent features and the plant growth coordinates are extracted. The agronomic trait labels of the target domain do not participate in the alignment training and coordinate fitting.

7. The method for constructing transferable plant growth coordinates as described in claim 6, wherein in step four, the target domain adaptation parameters include the parameters of the low-rank residual adapter and the scaling and bias parameters of the LayerNorm layer in the source domain nonlinear projector, and the remaining parameters of the visual encoder, the agronomic trait prediction head, and the source domain nonlinear projector do not participate in parameter updates.

8. The method for constructing transferable plant growth coordinates as described in claim 6, further comprising a crop time-progression monitoring step in a target domain without similar agronomic trait labels: Label-free salient subspace alignment and adaptation are performed on continuous crop growth images to extract plant growth coordinates. Crop images of the same plant are sorted according to the acquisition time. Based on the correlation between coordinate difference and acquisition time, the consistency of coordinate temporal sorting, and the smoothness of coordinate trajectory, the relative development progress of crops is quantitatively characterized, realizing cross-species crop growth time-series monitoring under the condition of no similar agronomic trait labels.

9. A method for analyzing and regulating crop growth, characterized in that, Includes the following steps: A. Acquire crop time-series RGB images using multiple fixed cameras, process the images at each time point using the method described in any one of claims 1 to 7, and complete coordinate orientation to obtain plant growth coordinates; B. Divide all time series coordinates into pre-perturbation and post-perturbation sequences; use the number of days elapsed as the independent variable, fit the growth baseline using the coordinates of the pre-perturbation sequence, and extrapolate the baseline to the post-perturbation period; The growth baseline is adopted using one of the following models: a quadratic trend model, a linear regression model, a spline interpolation model, an exponential model, or a state-space model. C. Calculate the normalized backoff for each observation point after the disturbance; D. Calculate the average value of the normalized backoff amount at all time points after the disturbance of a single camera to obtain the average backoff amplitude of a single camera; then calculate the mean and standard deviation of the average backoff amplitude of a single camera corresponding to multiple cameras; alarm thresholds can be set based on historical normal batches, or the normalized backoff amplitude can be directly input into the environmental controller to complete the growth analysis and control.

10. A transferable plant growth coordinate construction system based on agronomically supervised shaping and significant subspace alignment, characterized in that, It includes an image acquisition module, a visual feature extraction module, a source domain supervised shaping module, a target domain adaptation module, a coordinate extraction module, a time series analysis module, and an application interface module.

Images