Orchard management and protection intelligent auxiliary decision-making method and system
By collecting multi-source data and constructing decision trees and pruning reinforcement learning models, the problems of adaptability and intelligence in fruit tree pruning decisions were solved, achieving high-precision branch identification and pruning action optimization, and providing visual feedback.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CETC BIGDATA RES INST CO LTD
- Filing Date
- 2026-03-30
- Publication Date
- 2026-05-26
AI Technical Summary
Existing fruit tree pruning decision-making schemes lack adaptability, have low intelligence, rigid decision-making rules, are unable to make effective decisions in complex environments, and rely on a single data source, resulting in insufficient global judgment.
Collect multi-source data (environment, multispectral images, tree structure point clouds), process them through normalization and time synchronization, use a decision tree model to determine the suitability of pruning, combine support vector machines to identify unhealthy and redundant branches, and build a pruning reinforcement learning model to make intelligent decisions and output pruning actions.
It enables intelligent and adaptive optimization management of fruit tree pruning, improves the accuracy of branch identification and pruning results, provides visual feedback, and reduces human error.
Smart Images
Figure CN121935568B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of orchard management, specifically to an intelligent auxiliary decision-making method and system for orchard management. Background Technology
[0002] Pruning is a crucial aspect of orchard production management, directly impacting the trees' access to sunlight and ventilation, as well as subsequent yield and quality. Traditional pruning relies heavily on manual experience, with operators pruning based on the tree's appearance and growth habits. This approach suffers from strong subjectivity, inconsistent standards, and low precision, easily leading to over- or under-pruning, thus affecting the tree's structural balance and fruit yield.
[0003] There are two main existing fruit tree pruning decision-making schemes:
[0004] (1) Fruit tree pruning based on point cloud skeleton analysis: The fruit tree skeleton is extracted using point cloud data and topological analysis is performed for automatic identification and geometric feature analysis of the trunk and branches. However, it can only identify which branches need to be pruned, but cannot make decisions on where and how to prune.
[0005] (2) Fruit tree pruning method based on neural network and fuzzy decision: The convolutional neural network identifies which branches need to be pruned and the fuzzy decision provides the pruning points. However, the fuzzy decision algorithm is a pruning scheme based on fixed rules, which is too rigid and difficult to make effective decisions in complex situations.
[0006] The existing scheme has the following main drawbacks: (1) The decision layer lacks adaptability and has a low level of intelligence; (2) The decision rules are rigid and have poor generalization ability. When the external meteorological conditions or physiological state deviate from the training samples, it is easy to produce misjudgment or inefficient pruning results; (3) Based on a single data source (such as point cloud or RGB image), it is impossible to make a global judgment. Summary of the Invention
[0007] This invention provides an intelligent auxiliary decision-making method and system for orchard management to solve the problems of strong reliance on human experience, insufficient consideration of environmental factors, low accuracy of branch identification, and lack of feedback optimization of pruning results in existing orchard pruning management.
[0008] On the one hand, the present invention provides an intelligent auxiliary decision-making method for orchard management, the method comprising:
[0009] Collect environmental data, multispectral images, and tree structure point cloud data from the orchard to obtain different types of multi-source data;
[0010] By normalizing and time-synchronizing the data, the multi-source data is uniformly mapped into a standardized input format to obtain a multi-source input dataset.
[0011] Determine whether the current pruning requirements are met based on the environmental data in the multi-source input dataset;
[0012] If so, then based on the multispectral images in the multi-source input dataset, unhealthy branches are identified, and a priority pruning target set is generated;
[0013] Based on the tree structure point cloud data in the multi-source input dataset, redundant branches are identified, and a secondary priority pruning target set is generated.
[0014] The data from the multi-source input dataset, the priority pruning target set, and the second-priority pruning target set are input into the pruning reinforcement learning model, and the pruning action is determined based on the output of the pruning reinforcement learning model.
[0015] Optionally, the method further includes performing any one or more of the following preprocessing steps on the multi-source data:
[0016] The environmental data is subjected to noise filtering and missing data repair.
[0017] Illumination compensation and band correction are performed on the multispectral images;
[0018] The point cloud data of the tree structure is subjected to three-dimensional coordinate transformation and spatial registration.
[0019] Optionally, determining whether the current pruning operation requirements are met based on the environmental data in the multi-source input dataset includes:
[0020] The environmental data is input into the decision tree model to determine whether the current conditions meet the requirements for pruning operations. The decision tree model includes multiple leaf nodes, each corresponding to a parameter combination and pruning suitability level composed of multiple key indicators. The multiple key indicators include at least: temperature, humidity, soil moisture content, and light intensity.
[0021] Optionally, the step of identifying unhealthy branches based on the multispectral image includes:
[0022] Spectral features, color index, and texture features of the branch region are extracted from the multispectral image to construct a multidimensional optical feature vector.
[0023] The multidimensional optical feature vectors are input into the first support vector machine to determine the health status of branches, so as to distinguish between healthy branches and unhealthy branches;
[0024] Output the spatial location and identification results of unhealthy branches, and generate the corresponding priority pruning target set.
[0025] Optionally, the step of identifying redundant branches based on the tree structure point cloud data and generating a secondary priority pruning target set includes:
[0026] Geometric features of branches are extracted from the tree structure point cloud data, and the geometric features include at least: length, diameter, growth direction, and angle with the trunk and adjacent branches;
[0027] The geometric features are input into a second support vector machine to determine the state of redundant branches, thereby identifying redundant branches.
[0028] Based on the results of redundant branch identification, a second-priority pruning target set is generated, and the spatial location of the branches is output.
[0029] Optionally, the pruning action includes: pruning position, pruning angle, and pruning length; the method further includes:
[0030] The pruning action is mapped to the 3D model corresponding to the tree structure point cloud data; the 3D model is rendered, the branches to be pruned are marked, and unhealthy branches and redundant branches are highlighted; the rendering result is output.
[0031] Optionally, the method further includes constructing the pruning reinforcement learning model in the following manner:
[0032] Collect point cloud data of fruit tree structure and environmental data, and label unhealthy branches and redundant branches;
[0033] Define a set of executable pruning actions, which includes: pruning position, pruning angle, and pruning length;
[0034] Define a reward function to evaluate the effect of each pruning action;
[0035] A pruning reinforcement learning model is iteratively trained using DQN to obtain the optimal pruning action in the current state. The input to the pruning reinforcement learning model includes: tree structure point cloud data, environmental data, and annotation information. The output of the pruning reinforcement learning model includes: Optionally, the method further includes:
[0036] The pruning results were evaluated, and the evaluation results were standardized and quantified to obtain the quantified evaluation results.
[0037] The pruning reinforcement learning model is adaptively optimized and trained based on the quantified evaluation results.
[0038] On the other hand, the present invention also provides an intelligent auxiliary decision-making system for orchard management, the system comprising:
[0039] The data acquisition module is used to collect environmental data, multispectral images, and tree structure point cloud data from the orchard, obtaining different types of multi-source data.
[0040] The preprocessing module is used to normalize and synchronize the data with time to map the multi-source data into a standardized input format to obtain a multi-source input dataset.
[0041] The suitability assessment module is used to input environmental data from the multi-source input dataset into the decision tree to determine whether the current pruning operation requirements are met.
[0042] The unhealthy branch identification module is used to identify unhealthy branches based on multispectral images in the multi-source input dataset, and generate a priority pruning target set, when the suitability judgment module determines that the current pruning operation requirements are met.
[0043] The redundant branch identification module is used to identify redundant branches based on the tree structure point cloud data in the multi-source input dataset and generate a secondary priority pruning target set.
[0044] The decision module is used to input the data from the multi-source input dataset, the priority pruning target set, and the secondary priority pruning target set into the pruning reinforcement learning model, and determine the pruning action based on the output of the pruning reinforcement learning model.
[0045] Optionally, the unhealthy branch identification module includes:
[0046] An optical feature extraction unit is used to extract spectral features, color index, and texture features of the branch region from the multispectral image and construct a multidimensional optical feature vector.
[0047] The first classification unit is used to input the multidimensional optical feature vector into the first support vector machine to determine the health status of branches, so as to distinguish between healthy branches and unhealthy branches.
[0048] The first output unit is used to output the spatial location and identification results of unhealthy branches, and generate the corresponding priority pruning target set.
[0049] Optionally, the redundant branch identification module includes:
[0050] A geometric feature extraction unit is used to extract the geometric features of branches from the tree structure point cloud data. The geometric features include at least: length, diameter, growth direction, and angle with the trunk and adjacent branches.
[0051] The second classification unit is used to input the geometric features into the second support vector machine to determine the state of redundant branches, so as to identify redundant branches.
[0052] The second output unit is used to generate a set of secondary priority pruning targets based on the redundant branch identification results, and output the spatial location of the branches.
[0053] Optionally, the pruning action includes: pruning position, pruning angle, and pruning length; the system further includes:
[0054] The rendering module is used to map the pruning action to the 3D model corresponding to the tree structure point cloud data; render the 3D model, mark the branches to be pruned, highlight unhealthy branches and redundant branches; and output the rendering results.
[0055] Optionally, the system further includes:
[0056] The model building module is used to build the pruned reinforcement learning model;
[0057] The model building module includes:
[0058] The data acquisition and processing unit is used to collect point cloud data of the tree structure and environmental data of fruit trees, and to label unhealthy branches and redundant branches.
[0059] The setting unit is used to define a set of executable pruning actions, which includes: pruning position, pruning angle, and pruning length; and to define a reward function to evaluate the effect of each pruning action.
[0060] The training unit is used to iteratively train the pruning reinforcement learning model through DQN to obtain the optimal pruning action in the current state; the input of the pruning reinforcement learning model includes: tree structure point cloud data, environmental data, and annotation information; the output of the pruning reinforcement learning model includes: a set of executable pruning actions.
[0061] Optionally, the system further includes:
[0062] The evaluation module is used to evaluate the pruning results and standardize and quantify the evaluation results to obtain the quantified evaluation results.
[0063] The model building module is also used to adaptively optimize and train the pruning reinforcement learning model based on the quantized evaluation results.
[0064] On the other hand, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to perform the steps of the intelligent auxiliary decision-making method for orchard management.
[0065] The intelligent auxiliary decision-making method and system for orchard management provided by this invention comprehensively collects multi-source information such as meteorological, soil, tree structure and multispectral images, and integrates machine learning and reinforcement learning technologies to realize the judgment of pruning suitability, the identification of branch health and structure, the intelligent decision-making of pruning actions and the visualization feedback of results, thereby realizing intelligent and adaptive optimization management of fruit tree pruning. Attached Figure Description
[0066] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly described below. Obviously, the drawings described below are merely some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings without creative effort.
[0067] Figure 1 This is a flowchart of an intelligent auxiliary decision-making method for orchard management provided in an embodiment of the present invention;
[0068] Figure 2 This is a flowchart illustrating the construction of the pruning reinforcement learning model in an embodiment of the present invention;
[0069] Figure 3 This is a schematic diagram of a structure of the intelligent auxiliary decision-making system for orchard management provided in an embodiment of the present invention;
[0070] Figure 4 This is a schematic diagram of a model building module in an embodiment of the present invention;
[0071] Figure 5 This is another structural schematic diagram of the intelligent auxiliary decision-making system for orchard management provided in an embodiment of the present invention. Detailed Implementation
[0072] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.
[0073] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.
[0074] This invention provides an intelligent auxiliary decision-making method and system for orchard management, aiming to overcome the problems of strong reliance on manual experience, insufficient consideration of environmental factors, low branch identification accuracy, and lack of feedback optimization in existing orchard pruning management. This solution can comprehensively collect multi-source information such as meteorological, soil, tree structure, and multispectral images, and integrate machine learning and reinforcement learning techniques to achieve pruning suitability judgment, branch health and structure identification, intelligent decision-making on pruning actions, and visualized feedback of results, thereby realizing intelligent and adaptive optimization management of fruit tree pruning.
[0075] like Figure 1 The diagram shown is a flowchart of an intelligent auxiliary decision-making method for orchard management provided by an embodiment of the present invention, which includes the following steps:
[0076] In step 101, environmental data, multispectral images, and tree structure point cloud data of the orchard are collected to obtain different types of multi-source data.
[0077] In step 102, through normalization and time synchronization processing, the multi-source data is uniformly mapped to a standardized input format to obtain a multi-source input dataset.
[0078] To ensure data quality, before normalizing and synchronizing the collected multi-source data, further preprocessing can be performed on these data, including but not limited to one or more of the following:
[0079] (1) Noise filtering and missing data repair of environmental data.
[0080] The environmental data may include, but is not limited to, data such as temperature, humidity, wind speed, soil moisture, and light intensity.
[0081] Specifically, for the collected environmental data, a moving average filter or a Gaussian filter can be used to remove short-term fluctuations and sensor noise; for missing data points, interpolation methods or average values based on nearby time periods can be used to repair them to ensure the continuity of the time series; for outliers, they can be removed and corrected by setting upper and lower limit thresholds or by anomaly detection based on Z-score (standard score).
[0082] (2) Perform illumination compensation and band correction on multispectral images.
[0083] Specifically, it may include any one or more of the following processes:
[0084] 1) Perform illumination compensation on the acquired multispectral images and reduce the impact of illumination changes on the reflectivity of branches by using a dark channel prior algorithm;
[0085] 2) Perform radiometric correction on images of each band, converting the raw DN (Digital Number) values acquired by the sensor into reflectance values to eliminate differences caused by different sensors or different shooting conditions and the influence of shooting conditions;
[0086] 3) Perform illumination equalization processing on multispectral images, such as using adaptive histogram equalization or the Retinex algorithm, to reduce the impact of uneven illumination on branch feature extraction;
[0087] 4) Perform geometric correction on the multispectral images to ensure spatial alignment of each band.
[0088] (3) Perform three-dimensional coordinate transformation and spatial registration on the point cloud data of the tree structure.
[0089] Specifically, the collected tree structure point cloud data can be unified into a coordinate system, transforming point cloud data collected at different scanning positions or at different times into the same reference coordinate system; the iterative nearest point algorithm or feature matching registration method can be used to spatially align multiple frames of point clouds; the point cloud can be downsampled or denoised, such as by voxel grid filtering, to reduce the amount of data and remove outliers.
[0090] Normalization of multi-source data refers to normalizing or standardizing various types of data so that they are mapped to a uniform numerical range, which facilitates subsequent processing by reinforcement learning models.
[0091] Time synchronization of multi-source data refers to synchronizing data from different sources in time, aligning multispectral images, point cloud data, and environmental data according to a unified timestamp to form multi-source data tuples at the same point in time.
[0092] By performing noise filtering, missing data repair, illumination compensation, band correction, 3D registration, and normalization on environmental data, multispectral images, and tree structure point cloud data, the standardization and synchronization of multi-source heterogeneous data can be achieved, providing high-quality, reliable, and consistent input data for pruning decisions.
[0093] In step 103, the environmental data in the multi-source input dataset is used to determine whether the current conditions meet the requirements for pruning. If yes, proceed to step 104; otherwise, wait for suitable conditions.
[0094] Specifically, the environmental data can be input into a decision tree model to determine whether the current conditions meet the requirements for pruning operations. The decision tree model is used to judge whether meteorological conditions and soil conditions meet the requirements for pruning operations based on environmental data, and outputs a pruning suitability assessment result.
[0095] In this embodiment of the invention, the decision tree model includes multiple leaf nodes, each leaf node corresponding to a parameter combination and pruning suitability level composed of multiple key indicators; the multiple key indicators include at least: temperature, humidity, soil moisture content and light intensity.
[0096] The nodes in the decision tree model can be constructed based on key indicators. In a non-limiting embodiment, the process of constructing the decision tree model is as follows: Based on historical orchard environmental data and expert experience, key environmental indicators (including temperature, humidity, wind speed, soil moisture content, and light intensity) are used as input variables. Each environmental indicator is divided into value ranges to form candidate branch nodes, and the optimal splitting indicator is selected using information gain or the Gini index. Tree branches are recursively constructed, with each internal node representing a conditional judgment of the environmental indicator, and each leaf node corresponding to the pruning suitability level. Furthermore, cross-validation can be performed on the training data to ensure the generalization ability of the decision tree model and the accuracy of the pruning suitability judgment.
[0097] When using a decision tree model to determine whether the current conditions for pruning are met, the current environmental data can be input into the trained decision tree model. Conditional judgments are made along the tree branches. When the environmental data meets the conditions of a certain branch, the search continues downwards along that branch until a leaf node is reached. The pruning suitability level corresponding to the leaf node is output, such as "Suitable," "Generally Suitable," or "Unsuitable."
[0098] Furthermore, the pruning fitness level output by the decision tree model can be mapped to a numerical score as needed, serving as a reference reward value or constraint for subsequent pruning reinforcement learning models.
[0099] By using decision tree models to assess the suitability of pruning, we can quickly and accurately determine the feasibility of pruning fruit trees under different environmental conditions, avoid the risks of pruning due to unsuitable weather or soil conditions, and improve the efficiency of fruit tree health management.
[0100] In step 104, if the requirements for pruning are met, unhealthy branches are identified based on multispectral images in the multi-source input dataset, and a priority pruning target set is generated.
[0101] Specifically, this may include the following steps:
[0102] (1) Extract the spectral features, color index and texture features of the branch region from the multispectral image and construct a multidimensional optical feature vector.
[0103] Specifically, the branch regions in the image can be segmented. For example, segmentation algorithms based on thresholding, edge detection, or deep learning can be used to obtain the branch pixel regions. Within the branch regions, spectral features (reflectance values of each band), color indices (such as NDVI (Normalized Difference Vegetation Index) and GNDVI (Green NDVI)) and texture features (such as GLCM (Gray-level Co-occurrence Matrix) and LBP (Local Binary Pattern)) can be extracted. The extracted features are then combined into a multi-dimensional optical feature vector, which serves as input data for determining the health status of the branches.
[0104] (2) Input the multidimensional optical feature vector into the first support vector machine to determine the health status of branches, so as to distinguish between healthy branches and unhealthy branches.
[0105] The first support vector machine classifies branches by learning from labeled healthy and unhealthy branch samples; it outputs the health status of each branch, distinguishing it as a healthy branch or an unhealthy branch.
[0106] Based on the classification results of the first support vector machine, the spatial location of branches judged to be unhealthy can also be marked in the three-dimensional tree model.
[0107] (3) Output the spatial location and identification results of unhealthy branches, and generate the corresponding priority pruning target set.
[0108] The above method can accurately identify the health status of fruit tree branches, effectively distinguishing between diseased, dead, and healthy branches. Compared with traditional manual inspection, this method has the advantages of high efficiency, automation, and quantifiability, reducing human error.
[0109] In step 105, redundant branches are identified based on the tree structure point cloud data in the multi-source input dataset, and a secondary priority pruning target set is generated.
[0110] Specifically, this may include the following steps:
[0111] (1) Extract the geometric features of branches from the tree structure point cloud data, wherein the geometric features include at least: length, diameter, growth direction, and angle with the trunk and adjacent branches.
[0112] The methods for extracting each geometric feature bead are as follows:
[0113] 1) Length: Calculate the principal direction length of the centroid sequence of the branch point cloud, that is, the spatial Euclidean distance from the start point to the end point of the branch;
[0114] 2) Diameter: Select a set of cross-sectional points on the branch axis and use the least squares fitting circle algorithm to obtain the average cross-sectional diameter;
[0115] 3) Growth direction vector: Principal component analysis is used to extract the first principal component direction of the branch point cloud, which is taken as the main growth direction of the branch;
[0116] 4) Angle with the main trunk: Calculate the angle between the branch direction vector and the main trunk direction vector;
[0117] 5) Angle with adjacent branches: Calculate the angle between the direction vectors of adjacent branches;
[0118] 6) Height coordinate: The Z-coordinate value of the centroid of the branch, used to distinguish between upper vigorous branches and lower old and weak branches.
[0119] (2) Input the geometric features into the second support vector machine to determine the state of redundant branches in order to identify redundant branches.
[0120] In this embodiment of the invention, the second support vector machine may employ a radial basis function kernel to enhance nonlinear classification capabilities. Training samples consist of expert-annotated data, and categories may include, for example, "normal branches," "crossing branches," "overgrown branches," and "weak old branches." The second support vector machine optimizes the penalty coefficient C and kernel parameter γ through cross-validation to achieve optimal classification performance. After training, the model can automatically output the category label based on the geometric features of the input branches.
[0121] (3) Generate a set of secondary priority pruning targets based on the redundant branch identification results, and output the spatial location of the branches.
[0122] Specifically, for each branch input during the inference phase, the second support vector machine outputs its category result and generates a second-priority pruning target set based on the recognition result.
[0123] Furthermore, branches identified as "crossing branches," "overgrown branches," or "weak old branches" can be spatially labeled in the 3D tree model, with different colors or symbols used to distinguish different types of branches. The labeled branches also record their spatial coordinates, connecting nodes, and relative position parameters to the trunk.
[0124] The above methods can accurately identify redundant branches of fruit trees, such as crossing branches, vigorous shoots, and old and weak branches, enabling optimized management of branch structure and providing a reliable reference for subsequent pruning reinforcement learning decisions.
[0125] In step 106, the data from the multi-source input dataset, the priority pruning target set, and the second-priority pruning target set are input into the pruning reinforcement learning model, and the pruning action is determined based on the output of the pruning reinforcement learning model.
[0126] In this embodiment of the invention, the pruning action may include, but is not limited to, pruning position, pruning angle, and pruning length.
[0127] Furthermore, in some embodiments, the pruning action can be mapped to a 3D model corresponding to the tree structure point cloud data; the 3D model can be rendered, the branches to be pruned can be marked, and unhealthy and redundant branches can be highlighted; the rendering result can be output. For example, the rendering result can be output to a display terminal or remote management interface for orchard managers to view and interact with, in order to assist orchard managers in completing fruit tree pruning.
[0128] Specifically, the optimal pruning action parameters output by the pruning reinforcement learning model, including pruning position coordinates, pruning angle, and pruning length, can be mapped to the three-dimensional structural data of the fruit tree. During the mapping process, a KD-Tree (k-dimensional tree) is used to establish the correspondence between pruning points and the three-dimensional branch model, ensuring precise alignment between the pruning position and the actual branch model. The system then calls a 3D rendering engine (such as Open3D, Three.js, or Unity) to visualize the 3D tree model. Different colors and transparency are used for layered rendering based on the branch type and pruning status; for example, branches to be pruned are highlighted in red; unhealthy branches are marked in orange; and redundant branches are outlined in blue.
[0129] After rendering, the system generates an interactive 3D visualization interface, allowing users to view the tree structure and pruning details through rotation, zoom, and click. To enhance the visual experience, the system also supports before-and-after pruning comparison views, displaying changes in branch distribution and areas of improved lighting through dual windows or sliding comparisons. Rendering results can be output in real-time to a local display terminal or transmitted over a network to a remote orchard management platform. Managers can view the pruning effects on their terminals or fine-tune the pruning plan in the interactive interface (e.g., modifying pruning point positions or adjusting angles), and the system will update the rendered model in real-time based on these modifications.
[0130] Through the above visualization rendering process, the present invention can present the pruning decision results output by the pruning reinforcement learning model in a three-dimensional interactive form, enabling orchard managers to clearly understand the spatial distribution and structural optimization effect of the pruning scheme.
[0131] In this embodiment of the invention, a pruning reinforcement learning model can be constructed using tree structure data and environmental features as the state space, and pruning position, pruning angle, and pruning length as the action space, in order to obtain the optimal pruning action.
[0132] like Figure 2The diagram shown is a flowchart of constructing the pruned reinforcement learning model in an embodiment of the present invention, including the following steps:
[0133] In step 201, point cloud data of the tree structure and environmental data of the fruit tree are collected, and unhealthy branches and redundant branches are marked.
[0134] The annotation information can be added to the tree structure point cloud data based on the recognition results of the first and second support vector machines.
[0135] In step 202, a set of executable pruning actions is defined, which includes: pruning position, pruning angle, and pruning length; a reward function is defined to evaluate the effect of each pruning action.
[0136] The reward function can be determined by comprehensively considering parameters such as light uniformity, ventilation, flower bud retention rate, and pruning penalty coefficient.
[0137] In step 203, the pruning reinforcement learning model is iteratively trained using DQN (Deep Q-Network) to obtain the optimal pruning action in the current state.
[0138] The input to the pruning reinforcement learning model includes: tree structure point cloud data, environmental data, and annotation information; the output of the pruning reinforcement learning model includes: a set of executable pruning actions.
[0139] The steps for iteratively training a pruning reinforcement learning model include:
[0140] (1) Store the experience consisting of the obtained state, action, reward and next state into the experience replay pool, and initialize the experience set with random or predefined actions before training begins.
[0141] (2) Randomly extract small batches of experience from the experience replay pool to train the Q network, and iteratively update the Q network parameters by minimizing the error between the current Q network output and the target Q value.
[0142] (3) Every predetermined number of training steps, copy the current Q network parameters to the target network to stabilize the training process.
[0143] (4) Continue iterative training until the reinforcement learning model can output the optimal pruning action under a given state.
[0144] Specifically, the structural features of the fruit tree, environmental parameters, and the branch health and structural labels output by the preceding module are collectively used to form the state space S of the reinforcement learning. The state vector includes the following:
[0145] Branch structure parameters: branch length, diameter, growth angle, spatial coordinates, and angle with adjacent branches;
[0146] Environmental parameters such as light intensity, wind speed, humidity, and temperature reflect the characteristics of the current working conditions.
[0147] Branch label information: The identification results of unhealthy branches and redundant branches are embedded in binary or category encoding form;
[0148] Local growth density features: Canopy density is reflected by calculating the number of branches per unit space, the degree of overlap, or the leaf area index. These features are normalized and then input into a reinforcement learning model to describe the overall state before pruning.
[0149] Discretize the pruning operation parameters to form a finite set of actions A, which may include the following information:
[0150] Pruning locations: Divide the branch length into several pruning points;
[0151] Pruning angle: Using the main axis of the branch as a reference, set several discrete angles, such as 15°, 30°, 45°, etc.
[0152] Pruning length: Set the shortening ratio, such as 10%, 20%, 30%, etc.
[0153] Each action can be represented as a triplet. This refers to the pruning point, pruning angle, and pruning length. The model learns from the state... To action The mapping is used to find the optimal pruning scheme under different conditions.
[0154] To guide the model towards optimizing fruit tree growth, a comprehensive reward function R was designed, the specific structure of which can be expressed as follows:
[0155] ;
[0156] in, This indicates the uniformity of light exposure on the tree after pruning. This represents ventilation based on canopy airflow simulation. This indicates the flower bud retention rate, which is the ratio of the number of flower buds retained after pruning to the number of flower buds before pruning. This indicates the penalty for pruning, which is the proportion of total branches removed. These represent the weights of light, ventilation, flower bud retention, and penalty coefficient, respectively, and were obtained through experimental adjustments.
[0157] The pruning reinforcement learning model uses a DQN structure. During training, the following steps are performed:
[0158] Experience collection and storage: During simulation or actual pruning, record the experience quadruple consisting of state, action, reward value, and next state. And store it in the experience replay pool;
[0159] Experience replay: Randomly select a small batch of samples from the experience pool for training to break data correlation;
[0160] Q-network update: The parameters are updated by minimizing the mean squared error between the current Q-network prediction and the target Q-value, which is defined as:
[0161] ;
[0162] in, Indicates the discount factor. Indicates the next state One of the possible actions in the set of all possible actions. This represents the largest Q-value among all possible actions output by the target network in the next state;
[0163] Target network synchronization: Every fixed number of steps, the current Q network parameters are copied to the target network to improve training stability;
[0164] An ε-greedy strategy is adopted to balance exploration and utilization. In the early stage of training, actions are randomly selected with a high probability, and then gradually converge to the optimal action strategy in the later stage.
[0165] After training, the DQN model can output the optimal pruning action in any state, realizing an adaptive pruning strategy for different tree structures and environmental conditions.
[0166] Through the aforementioned reinforcement learning training process, the pruning reinforcement learning model can automatically learn the correlation between pruning actions and changes in fruit tree structure over multiple iterations, achieving dynamic optimization of pruning strategies. This part, in conjunction with the aforementioned identification of unhealthy branches and redundant branches, forms a closed-loop intelligent orchard management system from data perception to autonomous pruning decision-making.
[0167] In some embodiments, the pruned reinforcement learning model can also be adaptively optimized, as follows:
[0168] (1) Evaluate the pruning results and standardize and quantify the evaluation results to obtain the quantified evaluation results.
[0169] The evaluation of pruning results can be conducted by experts to determine whether the pruning actions achieved the expected goals. The expert evaluation results are then standardized and quantified. Specifically, the pruning actions can be presented to fruit tree experts or managers in a three-dimensional visualization for manual evaluation. Experts, based on their experience with fruit tree growth and cultivation standards, score the pruning effect from multiple dimensions, including the rationality of pruning, the integrity of retained branches, the balance of the canopy structure, and the degree of improvement in tree health. The scoring uses a five- or ten-level system, and the subjective evaluation results are converted into standardized quantitative scores through weighted averaging or fuzzy membership functions.
[0170] (2) Adaptive optimization training of the pruning reinforcement learning model based on the quantified evaluation results.
[0171] Specifically, the quantified evaluation results are input into the pruning reinforcement learning model to adjust the weight parameters in the reward function, thereby optimizing the model's decision-making performance in subsequent pruning action selection. If the expert score is high, the weights related to positive indicators such as light uniformity and flower bud retention rate are appropriately increased through gradient ascent; if the score is low, the weights of the pruning penalty term are strengthened through gradient descent, thus favoring a conservative pruning strategy in subsequent training.
[0172] Based on the updated policy of the reinforcement learning model, the optimal action is selected in the next round of pruning decisions to achieve adaptive optimization of the model. Specifically, during the optimization process, the pruned reinforcement learning model updates the reward function, recalculates the value estimate of each state-action pair, and re-executes the training or sampling process according to the new policy function.
[0173] After several iterations, the model can gradually learn pruning behavior patterns consistent with expert evaluations, achieving a dynamic fusion of expert knowledge and the pruning reinforcement learning model.
[0174] By introducing an adaptive optimization mechanism based on expert evaluation feedback, the present invention can dynamically adjust the reward function parameters during the training of the pruning reinforcement learning model, making the model learning results more consistent with human expert experience and the growth patterns of fruit trees.
[0175] Accordingly, embodiments of the present invention also provide an intelligent auxiliary decision-making system for orchard management, such as... Figure 3 The diagram shown is a structural schematic of an intelligent auxiliary decision-making system for orchard management provided in an embodiment of the present invention.
[0176] In this embodiment, the intelligent auxiliary decision-making system 300 for orchard management includes the following modules:
[0177] Data acquisition module 301 is used to acquire environmental data, multispectral images and tree structure point cloud data of the orchard to obtain different types of multi-source data;
[0178] Preprocessing module 302 is used to normalize and synchronize the processing time to map the multi-source data into a standardized input format to obtain a multi-source input dataset.
[0179] The suitability judgment module 303 is used to input environmental data from the multi-source input dataset into the decision tree to determine whether the current pruning operation requirements are met.
[0180] The unhealthy branch identification module 304 is used to identify unhealthy branches based on the multispectral images in the multi-source input dataset and generate a priority pruning target set when the suitability judgment module determines that the current pruning operation requirements are met.
[0181] The redundant branch identification module 305 is used to identify redundant branches based on the tree structure point cloud data in the multi-source input dataset and generate a secondary priority pruning target set.
[0182] The decision module 306 is used to input the data from the multi-source input dataset, the priority pruning target set, and the secondary priority pruning target set into the pruning reinforcement learning model, and determine the pruning action based on the output of the pruning reinforcement learning model. The pruning action may include, for example, pruning position, pruning angle, and pruning length.
[0183] One non-limiting embodiment of the unhealthy branch identification module 304 may include the following units:
[0184] An optical feature extraction unit is used to extract spectral features, color index, and texture features of the branch region from the multispectral image and construct a multidimensional optical feature vector.
[0185] The first classification unit is used to input the multidimensional optical feature vector into the first support vector machine to determine the health status of branches, so as to distinguish between healthy branches and unhealthy branches.
[0186] The first output unit is used to output the spatial location and identification results of unhealthy branches, and generate the corresponding priority pruning target set.
[0187] One non-limiting embodiment of the redundant branch identification module 305 may include:
[0188] A geometric feature extraction unit is used to extract the geometric features of branches from the tree structure point cloud data. The geometric features include at least: length, diameter, growth direction, and angle with the trunk and adjacent branches.
[0189] The second classification unit is used to input the geometric features into the second support vector machine to determine the state of redundant branches, so as to identify redundant branches.
[0190] The second output unit is used to generate a set of secondary priority pruning targets based on the redundant branch identification results, and output the spatial location of the branches.
[0191] In this embodiment of the invention, the pruning reinforcement learning model can be trained by a corresponding model building module. For example... Figure 4 The diagram shown is a structural schematic of the model building module in an embodiment of the present invention.
[0192] A non-limiting embodiment of the model building module 400 may include:
[0193] The data acquisition and processing unit 401 is used to acquire point cloud data of the tree structure and environmental data of the fruit tree, and to mark unhealthy branches and redundant branches.
[0194] Setting unit 402 is used to define a set of executable pruning actions, the set of pruning actions including: pruning position, pruning angle and pruning length; and to define a reward function to evaluate the effect of each pruning action;
[0195] Training unit 403 is used to iteratively train a pruning reinforcement learning model through DQN to obtain the optimal pruning action in the current state; the input of the pruning reinforcement learning model includes: tree structure point cloud data, environmental data, and annotation information; the output of the pruning reinforcement learning model includes: a set of executable pruning actions.
[0196] In some embodiments, the intelligent auxiliary decision-making system 300 for orchard management provided in this invention may further include: an evaluation module ( Figure 3 (Not shown in the image) is used to evaluate the pruning results and to standardize and quantify the evaluation results to obtain the quantified evaluation results.
[0197] Accordingly, the model building module 400 is also used to adaptively optimize and train the pruning reinforcement learning model based on the quantized evaluation results.
[0198] like Figure 5 The diagram shown is another structural schematic of the intelligent auxiliary decision-making system for orchard management provided in an embodiment of the present invention.
[0199] and Figure 3 The difference in the illustrated embodiment is that, in this embodiment, the orchard management intelligent auxiliary decision-making system 300 further includes a rendering module 307, which is used to map the pruning action to the three-dimensional model corresponding to the tree structure point cloud data; render the three-dimensional model, mark the branches to be pruned, highlight unhealthy branches and redundant branches; and output the rendering results.
[0200] The specific implementation methods of the above modules and units can be referred to the relevant descriptions in the previous embodiments of the present invention, and will not be repeated here.
[0201] The intelligent auxiliary decision-making method and system for orchard management provided by this invention comprehensively collects multi-source information such as meteorological, soil, tree structure and multispectral images, and integrates machine learning and reinforcement learning technologies to realize the judgment of pruning suitability, the identification of branch health and structure, the intelligent decision-making of pruning actions and the visualization feedback of results, thereby realizing intelligent and adaptive optimization management of fruit tree pruning.
[0202] Compared to existing technologies, it has the following advantages:
[0203] 1. The present invention comprehensively evaluates the growth environment and structural status of fruit trees based on fused multi-source data, which can provide an objective basis for decision-making on pruning timing and intensity, avoid misjudgment problems caused by a single data source, and thus significantly improve the accuracy and timeliness of pruning suitability judgment.
[0204] 2. This invention utilizes machine learning and multispectral recognition technology to extract spectral reflectance characteristics, color indices, and texture information from branch regions. Combined with geometric structural features, this allows for comprehensive discrimination, accurately identifying unhealthy or redundant branches such as diseased, dead, and overly vigorous shoots. The identification results serve as input labels for a reinforcement learning model, providing high-quality data support for subsequent pruning optimization. This significantly improves the accuracy and reliability of pruning target selection, reduces unnecessary removal of healthy branches, and enhances the scientific rigor of pruning operations.
[0205] 3. This invention introduces a reinforcement learning decision-making mechanism. By constructing a pruning reinforcement learning model with the fruit tree structure and environmental characteristics as the state space, it achieves intelligent optimization decisions for pruning actions. Moreover, the policy function can be continuously updated in multiple rounds of training, autonomously learning the optimal pruning scheme under different tree vigor and environmental conditions. This effectively overcomes the randomness and uncertainty of human experience, forming an adaptive pruning strategy for individual fruit trees, and achieving efficient, accurate, and sustainably optimized intelligent orchard management.
[0206] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the present invention is not limited to the described order of actions, because according to the present invention, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to the present invention.
[0207] The present invention also provides a storage medium, which is a computer-readable storage medium storing a computer program thereon, the computer program being executable when it runs. Figure 1 or Figure 2The method shown may include some or all of the steps. The storage medium may include read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk, etc. The storage medium may also include non-volatile memory or non-transitory memory, etc.
[0208] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer program are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data provider to another website, computer, server, or data provider via wired or wireless means.
[0209] The embodiments of the present invention have been described in detail above. Specific implementation methods have been used to illustrate the present invention. The descriptions of the embodiments above are merely for the purpose of helping to understand the method and apparatus of the present invention, and are only a part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention, and the content of this specification should not be construed as a limitation of the present invention. Therefore, any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A smart auxiliary decision-making method for orchard management, characterized in that, The method includes: Collect environmental data, multispectral images, and tree structure point cloud data from the orchard to obtain different types of multi-source data; By normalizing and time-synchronizing the data, the multi-source data is uniformly mapped into a standardized input format to obtain a multi-source input dataset. Determine whether the current pruning requirements are met based on the environmental data in the multi-source input dataset; If so, then based on the multispectral images in the multi-source input dataset, unhealthy branches are identified, and a priority pruning target set is generated; Based on the tree structure point cloud data in the multi-source input dataset, redundant branches are identified, and a secondary priority pruning target set is generated. The data from the multi-source input dataset, the priority pruning target set, and the second-priority pruning target set are input into the pruning reinforcement learning model, and the pruning action is determined based on the output of the pruning reinforcement learning model. The method further includes constructing the pruning reinforcement learning model in the following manner: Collect point cloud data of fruit tree structure and environmental data, and label unhealthy branches and redundant branches; Define a set of executable pruning actions, which includes: pruning position, pruning angle, and pruning length; Define a reward function to evaluate the effect of each pruning action; The pruning reinforcement learning model is trained iteratively using DQN to obtain the optimal pruning action in the current state. The input of the pruning reinforcement learning model includes: tree structure point cloud data, environmental data, and annotation information. The output of the pruning reinforcement learning model includes: a set of executable pruning actions.
2. The intelligent auxiliary decision-making method for orchard management according to claim 1, characterized in that, The method further includes performing one or more of the following preprocessing steps on the multi-source data: The environmental data is subjected to noise filtering and missing data repair. Illumination compensation and band correction are performed on the multispectral images; The point cloud data of the tree structure is subjected to three-dimensional coordinate transformation and spatial registration.
3. The intelligent auxiliary decision-making method for orchard management according to claim 1, characterized in that, The step of determining whether the current pruning operation requirements are met based on the environmental data in the multi-source input dataset includes: The environmental data is input into the decision tree model to determine whether the current conditions meet the requirements for pruning operations. The decision tree model includes multiple leaf nodes, each corresponding to a parameter combination and pruning suitability level composed of multiple key indicators. The multiple key indicators include at least: temperature, humidity, soil moisture content, and light intensity.
4. The intelligent auxiliary decision-making method for orchard management according to claim 1, characterized in that, The method of identifying unhealthy branches based on the multispectral image includes: Spectral features, color index, and texture features of the branch region are extracted from the multispectral image to construct a multidimensional optical feature vector. The multidimensional optical feature vectors are input into the first support vector machine to determine the health status of branches, so as to distinguish between healthy branches and unhealthy branches; Output the spatial location and identification results of unhealthy branches, and generate the corresponding priority pruning target set.
5. The intelligent auxiliary decision-making method for orchard management according to claim 1, characterized in that, The step of identifying redundant branches based on the tree structure point cloud data and generating a secondary priority pruning target set includes: Geometric features of branches are extracted from the tree structure point cloud data, and the geometric features include at least: length, diameter, growth direction, and angle with the trunk and adjacent branches; The geometric features are input into a second support vector machine to determine the state of redundant branches, thereby identifying redundant branches. Based on the results of redundant branch identification, a second-priority pruning target set is generated, and the spatial location of the branches is output.
6. The intelligent auxiliary decision-making method for orchard management according to claim 1, characterized in that, The pruning action includes: pruning position, pruning angle, and pruning length; the method further includes: The pruning action is mapped to the 3D model corresponding to the tree structure point cloud data; the 3D model is rendered, the branches to be pruned are marked, and unhealthy branches and redundant branches are highlighted; the rendering result is output.
7. The intelligent auxiliary decision-making method for orchard management according to claim 1, characterized in that, The method further includes: The pruning results were evaluated, and the evaluation results were standardized and quantified to obtain the quantified evaluation results. The pruning reinforcement learning model is adaptively optimized and trained based on the quantified evaluation results.
8. An intelligent auxiliary decision-making system for orchard management, characterized in that, The system includes: The data acquisition module is used to collect environmental data, multispectral images, and tree structure point cloud data from the orchard, obtaining different types of multi-source data. The preprocessing module is used to normalize and synchronize the data with time to map the multi-source data into a standardized input format to obtain a multi-source input dataset. The suitability assessment module is used to input environmental data from the multi-source input dataset into the decision tree to determine whether the current pruning operation requirements are met. The unhealthy branch identification module is used to identify unhealthy branches based on multispectral images in the multi-source input dataset, and generate a priority pruning target set, when the suitability judgment module determines that the current pruning operation requirements are met. The redundant branch identification module is used to identify redundant branches based on the tree structure point cloud data in the multi-source input dataset and generate a secondary priority pruning target set. The decision module is used to input the data in the multi-source input dataset, the priority pruning target set, and the secondary priority pruning target set into the pruning reinforcement learning model, and determine the pruning action based on the output of the pruning reinforcement learning model. The system also includes: The model building module is used to build the pruned reinforcement learning model; The model building module includes: The data acquisition and processing unit is used to collect point cloud data of the tree structure and environmental data of fruit trees, and to label unhealthy branches and redundant branches. The setting unit is used to define a set of executable pruning actions, which includes: pruning position, pruning angle, and pruning length; and to define a reward function to evaluate the effect of each pruning action. The training unit is used to iteratively train the pruning reinforcement learning model through DQN to obtain the optimal pruning action in the current state; the input of the pruning reinforcement learning model includes: tree structure point cloud data, environmental data, and annotation information; the output of the pruning reinforcement learning model includes: a set of executable pruning actions.
9. The intelligent auxiliary decision-making system for orchard management according to claim 8, characterized in that, The unhealthy branch identification module includes: An optical feature extraction unit is used to extract spectral features, color index, and texture features of the branch region from the multispectral image and construct a multidimensional optical feature vector. The first classification unit is used to input the multidimensional optical feature vector into the first support vector machine to determine the health status of branches, so as to distinguish between healthy branches and unhealthy branches. The first output unit is used to output the spatial location and identification results of unhealthy branches, and generate the corresponding priority pruning target set.
10. The intelligent auxiliary decision-making system for orchard management according to claim 8, characterized in that, The redundant branch identification module includes: A geometric feature extraction unit is used to extract the geometric features of branches from the tree structure point cloud data. The geometric features include at least: length, diameter, growth direction, and angle with the trunk and adjacent branches. The second classification unit is used to input the geometric features into the second support vector machine to determine the state of redundant branches, so as to identify redundant branches. The second output unit is used to generate a set of secondary priority pruning targets based on the redundant branch identification results, and output the spatial location of the branches.
11. The intelligent auxiliary decision-making system for orchard management according to claim 8, characterized in that, The pruning action includes: pruning position, pruning angle, and pruning length; the system also includes: The rendering module is used to map the pruning action to the 3D model corresponding to the tree structure point cloud data; render the 3D model, mark the branches to be pruned, highlight unhealthy branches and redundant branches; and output the rendering results.
12. The intelligent auxiliary decision-making system for orchard management according to claim 8, characterized in that, The system also includes: The evaluation module is used to evaluate the pruning results and standardize and quantify the evaluation results to obtain the quantified evaluation results. The model building module is also used to adaptively optimize and train the pruning reinforcement learning model based on the quantized evaluation results.
13. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is run by the processor, it executes the steps of the intelligent auxiliary decision-making method for orchard management as described in any one of claims 1 to 7.