Intelligent decision method for load capacity of fruit trees based on pollen tube behavior and carbon accumulation balance

Abstract

The application discloses a fruit tree load intelligent decision-making method based on pollen tube behavior and carbon accumulation balance, and belongs to the field of fruit tree growth management. The method collects physiological data such as fruit tree pollen tube growth and leaf photosynthesis in real time through multiple sensors, combines environmental factors and carbon balance parameters, constructs a multi-source data fusion model, and realizes fine perception and dynamic optimization and regulation and control of fruit load under different growth states of fruit trees. The system introduces an abnormal data detection and self-adaptive evolution mechanism to improve the adaptability and robustness of the model to complex field environments and various varieties. The application can effectively improve the yield and quality of fruit trees, optimize the utilization efficiency of carbon resources, and promote the intelligent and fine management of orchards.

Description

Technical Field

[0001] This invention belongs to the field of fruit tree growth management, and more specifically relates to an intelligent decision-making method for fruit tree load based on the balance between pollen tube behavior and carbon accumulation. Background Technology

[0002] Fruit load, a key agronomic parameter affecting fruit yield, fruit quality, and resource utilization efficiency, has always been a core challenge in modern fruit tree management and high-efficiency production. Traditional fruit load decisions rely primarily on experience or methods based on single physiological indicators, often failing to fully reflect the actual needs of fruit trees in complex environments and variable growth conditions. This can easily lead to excessive or insufficient fruit load, resulting in physiological fruit drop, carbon nutrient imbalance, and quality decline. In recent years, with the development of agricultural informatization and intelligent sensing technologies, some research has begun to focus on load management methods based on fruit tree physiological, environmental, and imaging data. However, these technologies still suffer from significant problems, including insufficient dynamic reflection of key physiological processes, inadequate handling of data anomalies, and limited model adaptability.

[0003] Especially in the context of refined fruit tree management, pollen tubes serve as a crucial window into the reproductive physiology of fruit trees. Their growth dynamics can directly reflect floral activity and pollination / fertilization status. However, current methods have not yet achieved high-precision real-time capture and effective utilization of pollen tube growth dynamics. Meanwhile, fruit tree carbon balance, as an important physiological basis for measuring fruit tree growth and yield potential, has a complex coupling relationship with flower and fruit load, influenced by various environmental factors and growth processes. A complete technical solution that can simultaneously integrate pollen tube dynamics and carbon balance information for intelligent optimization decision-making is still lacking. Therefore, there is an urgent need to construct an intelligent decision-making method for flower and fruit load that can integrate pollen tube growth dynamics and fruit tree carbon balance information, possessing abnormal data processing and adaptive evolution capabilities, to achieve precise regulation of fruit tree load and improve production efficiency, thereby promoting the intelligent and refined development of orchard management. Summary of the Invention

[0004] The technical problem this invention aims to solve is how to achieve intelligent, precise, and dynamic optimization and control of the flower and fruit load of fruit trees, overcome the bottlenecks of existing load decision-making methods such as insufficient utilization of pollen tube growth and carbon balance dynamic information, lack of ability to process abnormal data, and weak model adaptability, and achieve accurate decision-making on the flower and fruit load of fruit trees under different growth states and environmental conditions, so as to improve fruit tree yield, fruit quality and carbon resource utilization efficiency, and promote the refined and intelligent management of orchards.

[0005] To achieve the above objectives, the present invention employs the following technical solution: the method comprises:

[0006] Step 1: Using multispectral high-resolution imaging and deep learning segmentation network, obtain pollen tube growth time-series image data after fruit tree pollination. The spatial boundary and length features of pollen tubes are automatically extracted by deep learning segmentation network. Combined with wavelet packet clustering regression algorithm, abnormal samples are identified and corrected to form stable and accurate pollen tube growth dynamic data.

[0007] Step 2: Collect multidimensional data on carbon balance of fruit trees, perform standardization, multidimensional feature fusion and modeling, and complete feature mining and typical pattern recognition of carbon balance multidimensional data.

[0008] Step 3: The pollen tube growth dynamic data and carbon balance multidimensional data are fused together, and a collaborative expression model is constructed using a dual adaptive interactive network.

[0009] Step 4: Based on the multi-objective global and local optimization algorithm and combined with the multi-objective fitness aggregation principle, the flower and fruit load is optimized.

[0010] Step 5: Collect actual feedback data such as flower and fruit load and carbon balance after implementation, and use memory-enhanced incremental learning algorithm to self-learn and correct the model.

[0011] In one approach, the multispectral high-resolution imaging is used to continuously scan samples of fruit trees after pollination to obtain pollen tube growth time-series images. Each frame of the image is processed to automatically complete accurate segmentation and feature extraction of the pollen tubes.

[0012] In one approach, the identification of outliers in pollen tube growth dynamics data employs a combination of wavelet packet decomposition and spectral clustering.

[0013] First, the original time series data is decomposed into multiple scales. Then, a similarity matrix between samples is constructed in each scale space. The main group and outliers are distinguished by spectral clustering. Finally, a weighted regression prediction algorithm is used to adaptively correct outliers to ensure the stability and continuity of the time series data.

[0014] In one approach, the modeling process of the carbon balance multidimensional data includes the synchronous acquisition and standardization of carbon dioxide exchange, photosynthetic rate, respiration rate, temperature, light intensity and humidity. Embedded feature fusion and subspace clustering methods are used to achieve the compression and fusion modeling of multi-source heterogeneous features, and typical dynamic patterns of carbon balance are automatically obtained through model selection.

[0015] In one approach, the pollen tube growth dynamics data and carbon balance feature data are collaboratively modeled using a dual adaptive interactive network. This network includes a bidirectional cross-attention mechanism and an adaptive gating structure, which are used to dynamically capture the coupling relationship between the two types of data and automatically adjust the spatiotemporal importance of the input features.

[0016] In one approach, the multi-objective decision optimization employs a parallel and complementary approach using genetic algorithms and artificial bee colony algorithms. First, the genetic algorithm performs global search and optimization, and then the artificial bee colony algorithm performs local depth mining of the neighborhood solution space of excellent individuals. Pareto non-dominated sorting is used to achieve multi-objective trade-offs, and the final load scheme is selected in conjunction with actual production objectives.

[0017] In one approach, the self-learning correction step is implemented through a memory-enhanced incremental learning algorithm. After collecting new feedback on flower and fruit load and carbon budget, it is not necessary to retrain with all historical data. Instead, it is achieved through incremental updates and retention of key historical knowledge.

[0018] In one scheme, the flower and fruit load decision method supports fine-grained customization of the flower and fruit load of a single fruit tree or a partition unit, and can calculate the optimal load scheme for different growth stages, different environmental conditions and different cultivars.

[0019] Beneficial effects of this invention:

[0020] This invention achieves scientific decision-making and dynamic regulation of fruit tree flower and fruit load by integrating precise perception and intelligent analysis of pollen tube growth dynamics and multidimensional characteristics of fruit tree carbon balance. It not only effectively overcomes the limitations of traditional methods in acquiring and utilizing key physiological information but also significantly improves the model's adaptability to abnormal data and changes in different varieties and environments. Applying this invention can enhance the scientific and personalized level of fruit tree flower and fruit load regulation, significantly improve fruit quality and yield, optimize carbon resource utilization efficiency, reduce management costs, reduce resource waste, and promote the digitalization, intelligentization, and sustainable development of orchard production, demonstrating broad prospects for widespread application. Attached Figure Description

[0021] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0022] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Typical embodiments of the invention are shown in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete.

[0023] Unless otherwise defined, all technical and scientific terms used in this invention have the same meaning as understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. To facilitate understanding, the invention will now be described more fully with reference to the accompanying drawings. Typical embodiments of the invention are shown in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to make the disclosure of the invention more thorough and complete.

[0024] like Figure 1 As shown, the intelligent decision-making method for fruit tree load based on the balance between pollen tube behavior and carbon accumulation has the following specific implementation steps:

[0025] Step 1: Fine-grained acquisition of pollen tube growth dynamic data and adaptive correction of outliers

[0026] Multispectral high-resolution imaging and a deep learning segmentation network were used to acquire time-series image data of pollen tube growth. In the process of fine acquisition and adaptive outlier correction of pollen tube growth dynamic data, samples from pollinated fruit trees were first continuously scanned using a multispectral high-resolution imaging system to obtain pollen tube growth image sequences at different time points. Each image frame was processed by a deep learning segmentation network (such as an improved U-Net or a custom residual attention network) to automatically segment and extract the spatial boundary and length features of the pollen tubes, forming a time-series growth dataset. Where x t Let be the growth length or structural feature vector of the pollen tube at time t.

[0027] A dynamic anomaly detection algorithm based on wavelet packet clustering regression (different from the commonly used LOF / IsolationForest) is adopted to identify and correct abnormal sample points to ensure data stability and accuracy.

[0028] To achieve adaptive correction of outliers, an improved wavelet packet clustering regression (WPCR) algorithm is introduced. First, for the original time-series data x... t Wavelet packet decomposition is performed to obtain the multi-scale decomposition coefficients {C}. j,k}, where j represents the decomposition level and k is the frequency band index, i.e. Where ψ j,k (t) is the wavelet packet basis function.

[0029] Subsequently, spectral clustering was used to cluster samples for each scale coefficient space. Specifically, a similarity matrix between samples was first constructed. Where C i Let σ be the feature representation of sample i in the wavelet packet coefficient space, and σ be the scaling factor. Then, calculate the Laplacian matrix L = DW, where D is the degree matrix. Then, perform eigenvalue decomposition on L, select the first k eigenvectors for K-means clustering, and thus divide the time series points into the main group and the outlier group.

[0030] Points whose clustering results are more than a preset threshold θ away from the main cluster center are identified as outliers, and their values ​​are corrected using intra-cluster weighted regression prediction. The weighted regression prediction function is as follows: Where N is the set of temporal neighborhoods in the same cluster, w t =exp(-α|tt′|) is the weighting coefficient that decays with time and distance, and α is a parameter for adjusting the decay magnitude.

[0031] The above process ensures that the pollen tube time-series data, accurately collected through multispectral and deep networks, undergoes improved wavelet packet decomposition and clustering to identify anomalies, and is adaptively corrected using weighted regression, thereby guaranteeing the high quality and stability of the data used for subsequent analysis. For example, x t C represents the length of the pollen tube at time t. j,k W represents the multiscale coefficients under wavelet packet decomposition at that moment. i,i′ Let represent the similarity between samples i and i' in the feature space, and N represent the neighborhood set of normal points. w represents the predicted value after outlier correction. t This is the weighting factor.

[0032] Step 2: Multidimensional Data Modeling and Feature Fusion of Fruit Tree Carbon Balance

[0033] Construct a multi-timescale hybrid dataset of CO2 exchange, photosynthesis, and respiration in fruit trees; in the process of multidimensional data modeling and feature fusion of carbon balance in fruit trees, it is first necessary to construct a dataset covering CO2 exchange rates. A multi-timescale hybrid dataset of core carbon flux parameters such as photosynthetic rate (P) and respiration rate (R). Specifically, high-precision instruments such as infrared gas analyzers (IRGA) are used to simultaneously monitor fruit tree leaves or canopies, recording data at different time intervals (e.g., minutes, hours, days). P (s) and R (s) This also incorporates key environmental factors such as temperature (T), light intensity (I), and humidity (H). The entire dataset is organized as follows:

[0034]

[0035] , where t (s) is the time label, and S is the total number of samples.

[0036] The "Composite Embedded Subspace Clustering (CESC)" algorithm is applied to aggregate multidimensional carbon flow data with multiple features such as time and environmental factors to obtain the dynamic feature distribution of carbon balance.

[0037] To eliminate scale differences between different data dimensions, the data matrix X needs to be standardized so that each feature follows a normal distribution with a mean of 0 and a variance of 1. Specifically: Where x ij Let μ be the original value of the j-th feature of the i-th sample. j and σ j These are the mean and standard deviation of the feature, respectively.

[0038] In the feature fusion modeling stage, the Composite Embedded Subspace Clustering (CESC) algorithm is employed. This algorithm maps the joint representation X′ of high-dimensional carbon flow and environmental features to φ through feature embedding. Projecting onto a low-dimensional subspace Z captures the correlation between heterogeneous features. The mapping process can be expressed as Z = φ(X′) = X′W, where W is the feature embedding weight matrix. For low-dimensional fusion features, d' is much smaller than the original dimension d. To optimize W, CESC introduces subspace sparsity constraints, minimizing the objective function.

[0039]

[0040] Where C is the reconstruction coefficient matrix, ||·|| F Let W be the Frobenius norm, λ be the sparse regularization coefficient, and ||W|| be the Frobenius norm. 2,1 Promote column sparsity to select the most relevant subset of features.

[0041] After completing the low-dimensional embedding, adaptive clustering is performed in Z-space, and the number of clusters K is automatically determined using a Gaussian Mixture Model (GMM) combined with the Bayesian Information Criterion (BIC). The center of each cluster is μ. k Covariance Σ k Describe a typical carbon balance characteristic distribution, with clustering probability derived from...

[0042]

[0043] Given, where p ik Let N(·|μ) be the prior probability of the k-th class. k ,Σ kLet z be the Gaussian distribution density function. i Let be the low-dimensional vector of the i-th sample.

[0044] Ultimately, the CESC algorithm not only achieves feature fusion and dimensionality reduction representation of multidimensional carbon flow and environmental factors, but also uncovers the dynamic feature distribution pattern of fruit tree carbon balance, providing a solid data foundation and criteria for subsequent intelligent decision-making on fruit tree carbon allocation and flower and fruit load across time, space, and samples. In the above formula, X represents the original carbon balance feature data, X′ represents the standardized data, φ represents the embedding mapping, W represents the feature weight, Z represents the embedded feature, C represents the reconstruction coefficient, and μ... k , Σ k For cluster centers and covariance, π k p is the prior probability. ik This represents the cluster probability distribution.

[0045] Step 3: Co-expression of pollen tube growth dynamics and carbon balance model

[0046] Based on the "Dual Adaptive Interaction Network (DAIN)," a relationship model between pollen tube growth rate and carbon balance index is established to achieve dynamic causal mapping, rather than a single linear regression or traditional deep network.

[0047] In the synergistic expression of pollen tube growth dynamics and the carbon balance model, the fusion of data from both is crucial. Firstly, the time series of pollen tube growth characteristics obtained in step 1 above is used... Where x t This represents the pollen tube growth length or morphological index at time t after outlier correction; combined with the multidimensional carbon balance characteristics obtained in step 2. The carbon balance features are embedded in the subspace for time period s. To achieve collaborative modeling of these two features, it is necessary not only to consider the linear correspondence between individual variables but also to describe the temporal interaction and nonlinear coupling between them. Therefore, a Dual Adaptive Interaction Network (DAIN) is introduced. Its core idea is to construct a dynamic bidirectional cross-attention structure between the two sets of feature streams, supplemented by an adaptive feature weighting mechanism, to form a high-resolution causal relationship mapping.

[0048] The DAIN model adjusts feature weights in real time to improve the accuracy and robustness of flower and fruit load prediction.

[0049] The input to DAIN is the time-aligned pollen tube features. and carbon balance characteristics First, each representation is mapped to a higher-order representation by its respective encoder. Where f p (·), f c(·) represents a nonlinear neural network encoder. Subsequently, the DAIN core's bidirectional cross-attention mechanism calculates the interaction attention matrix between pollen tube features and carbon balance features at any time t:

[0050]

[0051] Where h p,t and h c,t Let W be the eigenvector at time t. q W k Let A be a learnable weight matrix. The attention distribution A is... ptc Used to weightedly fuse features from each other to generate interactive representations:

[0052] Furthermore, DAIN designs an adaptive feature gating system to adjust the impact of each feature on the model output in real time. This is illustrated by the predicted flower and fruit load value y. t For example, the final prediction function is

[0053] Where g(·) is a nonlinear decision mapping, and ⊙ denotes element-wise multiplication. The adaptively learned feature weight gating vector dynamically reflects the influence weights of different physiological and environmental factors. The weight gating mechanism automatically adjusts the weights for the current features through an independent network, achieving a robust response to sudden anomalies and external disturbances.

[0054] Through the aforementioned collaborative interaction and adaptive weight adjustment, DAIN can accurately capture the deep coupling relationship between pollen tube growth rate and carbon balance, achieving high-precision dynamic prediction of flower and fruit load under complex conditions. In the above formula, and These are the aligned pollen tube and carbon balance feature inputs, H, respectively. p H c For its higher-order representation, W q W k For attention weights, A ptc For interactive attention matrix, and For attention-weighted features, w p,t w c,t For adaptive weight gating, g(·) is the output layer decision mapping, and y t To predict the flower and fruit load at the predicted time. Step 4: Multi-objective adaptive flower and fruit load decision optimization.

[0055] We introduce a multi-objective parallel genetic bee colony optimization algorithm (MO-FPBGO) for flowering. This algorithm combines the global search capability of the genetic algorithm with the local optimum capability of the artificial bee colony algorithm, and simultaneously optimizes multiple objectives (flower and fruit load, carbon utilization rate, and high-quality fruit rate), which is significantly different from single optimization algorithms.

[0056] In the multi-objective adaptive flower and fruit load decision optimization stage, the flower and fruit load predictions output by the aforementioned pollen tube dynamics and carbon balance interaction models, along with physiological data, are used as the basis to further introduce a multi-objective parallel genetic bee colony optimization algorithm (MO-FPBGO) for flowering to achieve comprehensive decision-making. This algorithm integrates the global population search capability of the Genetic Algorithm (GA) with the efficient local space search capability of the Artificial Bee Colony (ABC) algorithm, and simultaneously optimizes flower and fruit load (F) and carbon utilization efficiency (η) through a multi-objective evolutionary mechanism. c Key indicators such as the rate of high-quality fruit (q) are used to overcome the limitations of traditional single-objective or single-optimization strategies. In specific implementation, the decision variable F = {f1, f2, ..., f...} is first used. N} represents the flower and fruit load of each fruit tree or unit in the system, where N is the number of fruit trees, and the individual is encoded as chromosome X = [f1,...,f N The model constraints are reflected in the carbon balance C. balance (F), Pollen tube activity A p (F) and fruit development potential P f (F) Three aspects, specifically:

[0057]

[0058] A p (F)≥θ p

[0059] P f (F)≥θ f

[0060] Where P i and R i Carbon input and output, E fruit,i θ represents the consumption per unit load of fruit. c ,θ p ,θ f This represents the corresponding physiological threshold.

[0061] The MO-FPBGO population initialization is based on historical observation data and candidate load schemes output by the DAIN model, forming an initial individual set P. (0)The genetic evolution process includes selection, crossover, and mutation: the selection stage employs non-dominated sorting based on crowding distance for M objective functions, namely maximizing the rate of high-quality fruit.

[0062]

[0063] Maximize carbon utilization

[0064]

[0065] Furthermore, it is necessary to ensure a positive carbon budget while minimizing the risk of total fruit overload.

[0066]

[0067] Where q i (f i Let r be a function of fruit quality as a function of fruit load. i (f i The overload risk function is denoted by . Crossover and mutation operations dynamically adjust parameters through adaptive probability distributions, increasing global exploration capabilities.

[0068] After each genetic evolution cycle, the algorithm enters the local depth mining phase of the artificial bee colony. Here, each "worker bee" is represented by the currently superior individual X. k Based on this, a new solution X is generated through neighborhood perturbation. k′ =X k +φ(X k -X j ), where φ is a random number in the interval [-1, 1], and X j For other individuals in the current population, through the objective function set F(X) k′ The system performs selection and replacement of inferior products. The local search step size and worker bee allocation number are adaptively adjusted based on population crowding and iteration convergence to ensure sufficient exploration in different regions of the multi-objective Pareto front.

[0069] Multi-objective fitness aggregation employs a hierarchical selection method based on Pareto dominance relations. For any two solutions X... a ,X b If for all m targets f m There is f m (X a )≥f m (X b And there exists f m′ (X a )>f m′ (X b ), then X a DominateX bFinally, by analyzing the non-dominated front solution set P... * Based on the analysis, and according to the actual production objectives (such as maximizing the weight of high-quality fruit rate or maximizing carbon utilization), the final recommended optimal load scheme X is selected. * That is, the optimal flower and fruit load configuration for each fruit tree / unit.

[0070] Among them, f i For the flower and fruit load of a single plant or unit, q i (f i E represents the percentage of high-quality fruit as a function of the fruit load. fruit,i Carbon consumption per unit load, P i For carbon input, R i For carbon loss due to respiration, C balance (F) represents the global carbon balance, P * For a multi-objective, non-dominated Pareto front, X * The final optimal decision solution is obtained through MO-FPBGO, which uses a collaborative global and local, multi-objective parallel evolution approach to ensure that the final output flower and fruit load scheme achieves adaptive optimality under multiple objectives such as carbon balance, quality, and yield.

[0071] Step 5: Decision Feedback and Self-Learning Correction

[0072] Collect actual flower and fruit load data and subsequent carbon budget changes; apply the "Memory Enhanced Incremental Learning Algorithm (MEIL)" to dynamically learn and correct the model, adapt to long-term changes in environment, year, variety, etc., and achieve continuously evolving fine load management.

[0073] After implementing the multi-objective flower and fruit load decision-making scheme, it is necessary to continuously collect the flower and fruit load of each unit of the fruit tree during the actual production process. and subsequent carbon balance change data This includes dynamic information such as fruit development, carbon exchange, and quality indicators, forming a time-series observation set. in For the characteristics of the i-th new sample (such as environment, variety, management measures, etc.), The decision-making results or carbon balance feedback are based on actual observations. To achieve dynamic correction and self-learning of the model, a memory-enhanced incremental learning algorithm (MEIL) is used to continuously update the decision-making model to adapt to long-term changes in environment, year, variety, etc.

[0074] The core idea of ​​MEIL lies in the combined use of recently observed data. new With historical high-value sample memory bank The model parameters θ are dynamically adjusted. First, for each new observation, a sample selection mechanism S(·) is used to add data with strong representativeness, significant errors, or those from special environments to the memory bank, enabling M to reflect the breadth of features from multiple years, environments, and varieties. During incremental learning, the model loss function is composed of a weighted average of the new and old data.

[0075]

[0076] Where f(·θ) is the profit and loss function related to the target decision prediction model (such as DAIN or MO-FPBGO output), l(·) is the loss term (such as mean squared error or multi-target loss), and α∈[0,1] is the balancing weight between the new data and the memory data. After each round of incremental training, θ is updated using optimization methods such as gradient descent.

[0077] Where η is the dynamic learning rate, which can be adaptively adjusted according to the rate of change of observed data or prediction error.

[0078] Memory enhancement mechanisms also include sample forgetting and reactivation strategies. If certain historical samples are not triggered by new environments for a long period of time, their weights gradually decay, satisfying ω. j ←γω j 0 < γ < 1, and a confidence threshold is used to determine whether to eliminate a sample. When encountering new environments, extreme climates, or new varieties, clustering (CESC subspace clustering) is used to determine whether the importance of a certain type of sample needs to be increased, thereby adjusting the weight distribution of the memory bank to achieve proactive adaptation to new trends.

[0079] Through MEIL dynamic self-learning correction, the entire fruit and flower load management system can achieve long-term knowledge accumulation and continuous evolution, maintaining sensitivity and robustness to long-term changes in environment, year, variety, and management strategies, and realizing precise and highly adaptable dynamic load optimization decisions for fruit trees. In the above formula, D... new Let M be the newly collected observation sample set, θ be the historical memory bank, l be the model parameters, α be the loss function, α be the weighting coefficient of the new and old samples, η be the learning rate, and ω be the model parameter. j γ represents the weight of historical samples, and γ is the forgetting decay coefficient.

[0080] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0081] It should be understood that the above detailed description of the technical solutions of the present invention with reference to preferred embodiments is illustrative and not restrictive. Those skilled in the art can modify the technical solutions described in the embodiments or make equivalent substitutions for some of the technical features based on reading this specification; however, these modifications or substitutions do not cause the essence of the corresponding technical solutions to depart from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A smart decision-making method for fruit tree load based on pollen tube behavior and carbon accumulation balance, characterized in that: The method includes: Step 1: Using multispectral high-resolution imaging and deep learning segmentation network, obtain pollen tube growth time-series image data after fruit tree pollination. The spatial boundary and length features of pollen tubes are automatically extracted by deep learning segmentation network. Combined with wavelet packet clustering regression algorithm, abnormal samples are identified and corrected to form stable and accurate pollen tube growth dynamic data. Step 2: Collect multidimensional data on carbon balance of fruit trees, perform standardization, multidimensional feature fusion and modeling, and complete feature mining and typical pattern recognition of carbon balance multidimensional data. Step 3: The pollen tube growth dynamic data and carbon balance multidimensional data are fused together, and a collaborative expression model is constructed using a dual adaptive interactive network. Step 4: Based on the multi-objective global and local optimization algorithm and combined with the multi-objective fitness aggregation principle, the flower and fruit load is optimized. Step 5: Collect actual feedback data on flower and fruit load and carbon balance after implementation, and use a memory-enhanced incremental learning algorithm to self-learn and correct the model.

2. The intelligent decision-making method for fruit tree load based on pollen tube behavior and carbon accumulation balance according to claim 1, characterized in that: The multispectral high-resolution imaging is used to continuously scan samples after fruit tree pollination to obtain pollen tube growth time sequence images. Each frame of the image is processed to automatically complete accurate segmentation and feature extraction of the pollen tubes.

3. The intelligent decision-making method for fruit tree load based on pollen tube behavior and carbon accumulation balance according to claim 1, characterized in that: The outlier identification in the pollen tube growth dynamic data was achieved by wavelet packet decomposition combined with spectral clustering. First, the original time series data is decomposed into multiple scales. Then, a similarity matrix between samples is constructed in each scale space. The main group and outliers are distinguished by spectral clustering. Finally, a weighted regression prediction algorithm is used to adaptively correct outliers to ensure the stability and continuity of the time series data.

4. The intelligent decision-making method for fruit tree load based on pollen tube behavior and carbon accumulation balance according to claim 1, characterized in that: The modeling process of the carbon balance multidimensional data includes the synchronous acquisition and standardization of carbon dioxide exchange, photosynthetic rate, respiration rate, temperature, light and humidity. Embedded feature fusion and subspace clustering methods are used to realize the compression and fusion modeling of multi-source heterogeneous features, and typical dynamic patterns of carbon balance are automatically obtained through model selection.

5. The intelligent decision-making method for fruit tree load based on pollen tube behavior and carbon accumulation balance according to claim 1, characterized in that: The pollen tube growth dynamics data and carbon balance feature data are collaboratively modeled using a dual adaptive interactive network. This network includes a bidirectional cross-attention mechanism and an adaptive gating structure, which are used to dynamically capture the coupling relationship between the two types of data and automatically adjust the spatiotemporal importance of the input features.

6. The intelligent decision-making method for fruit tree load based on pollen tube behavior and carbon accumulation balance according to claim 1, characterized in that: The multi-objective decision optimization adopts a parallel and complementary approach of genetic algorithm and artificial bee colony algorithm. First, the genetic algorithm completes the global search and selection, and then the artificial bee colony algorithm performs local depth mining of the neighborhood solution space of excellent individuals. The Pareto non-dominated sorting is used to realize the multi-objective trade-off, and the final load scheme is screened in combination with the actual production target.

7. The intelligent decision-making method for fruit tree load based on pollen tube behavior and carbon accumulation balance according to claim 1, characterized in that: The self-learning correction steps are implemented through a memory-enhanced incremental learning algorithm. After collecting new actual feedback on flower and fruit load and carbon balance, it is not necessary to retrain all historical data. Instead, incremental updates and retention of key historical knowledge are used.

8. The intelligent decision-making method for fruit tree load based on pollen tube behavior and carbon accumulation balance according to claim 1, characterized in that: The flower and fruit load decision method supports fine customization of the flower and fruit load of a single fruit tree or a zone unit, and can calculate the optimal load scheme for different growth stages, different environmental conditions and different cultivars.

Images