Search

CN-122022055-A - Tree potential control method based on young shoot detection

CN122022055ACN 122022055 ACN122022055 ACN 122022055ACN-122022055-A

Abstract

The invention relates to a tree vigor control method based on young shoot detection, and belongs to the technical field of agricultural intelligent monitoring and tree vigor control. The method utilizes improved YOLO11-OG to detect new shoots and old shoots in an image, introduces a mixed MetaFormer architecture combined with CGLU in a backbone network to enhance local and global feature extraction, adopts BiFPN to perform multi-scale feature fusion in a neck network, adopts CGLU functions in a detection head to solve the problem of category imbalance, calculates the rate of the new shoots of the citrus trees based on the detection result, quantitatively evaluates the growth vigor of the tree body according to the rate of the new shoots, and generates a corresponding pruning or fertilizing control strategy. The invention can effectively solve the problems of difficult identification of the young shoots, easy omission of the slender target, unbalanced sample and the like under the complex background, and realizes the closed-loop intelligent control from the young shoots to the agronomic execution while keeping the light weight of the model so as to facilitate the deployment of the mobile terminal.

Inventors

  • QU HONGCHUN
  • WU BAILIN
  • NIE QIANQIAN

Assignees

  • 重庆邮电大学

Dates

Publication Date
20260512
Application Date
20260205

Claims (10)

  1. 1. A tree potential control method based on young shoot detection is characterized by comprising the following steps: S1, acquiring an image of a citrus tree in a natural orchard environment to obtain an initial data set; S2, preprocessing and fine labeling are carried out on the initial data set, new pin and old pin categories are distinguished, a target data set is obtained, and the target data set is divided into a training set, a verification set and a test set; S3, constructing an improved YOLO11-OG model serving as a new pin detection network model by taking a YOLO11n network model as a deep learning reference model, wherein the improved YOLO11-OG model comprises the steps of replacing a backbone network of the reference model with a hybrid CAFormerCGLU structure to enhance local-global feature representation, replacing a neck network of the reference model with a bidirectional feature pyramid network BiFPN to realize efficient multi-scale feature fusion, adopting a class perception sliding loss function CASL as a loss function of the model, and dynamically distributing differentiated training weights for samples with different difficulties according to the average value of the cross-correlation ratio of each detection class prediction frame and the real frame; S4, training and evaluating the new pin detection network model by using a target data set to acquire an optimal model weight; s5, acquiring a real-time orchard image of a region to be detected, inputting the real-time orchard image into a young shoot detection network model for acquiring optimal weight, and outputting category, position coordinates and quantity statistics of the young shoots and the old shoots in the image; s6, calculating a new shoot rate based on a detection result, wherein the new shoot rate is the proportion of the number of new shoots to the total number of the branches; And S7, judging the growth potential state of the tree body according to the comparison result of the shoot rate and the preset tree potential threshold value, and generating a structured agricultural control instruction based on the growth potential state.
  2. 2. The treeing control method based on the treeing detection of claim 1 wherein in step S3, the mixing CAFormerCGLU structure is a mixing MetaFormer structure combined with a convolution gating linear unit introduced into a backbone network, and specifically includes replacing a C3K2 module in the YOLO11 backbone network with a c3k2_mcc module and a c3k2_msc module, wherein the c3k2_mcc module adopts a MFConvCGLU structure, uses convolution operation as a token mixer to extract local spatial features, and the c3k2_msc module adopts a MFSE-CGLU structure, uses a self-attention mechanism as a token mixer to capture global semantic information, wherein CGLU represents a convolution gating linear unit for dynamically modeling channel features through a depth separable convolution and gating mechanism.
  3. 3. The tree potential control method based on shoot detection according to claim 2, wherein in step S3, the convolution operation in the MFConvCGLU structure is defined as: Wherein, the And (3) with Respectively a point-by-point convolution layer, Is a deep convolution layer; activating a function for StarReLU, defined as: Wherein, the And Is a learnable channel-by-channel scaling and offset parameter.
  4. 4. The tree potential control method based on the treetop detection according to claim 1, wherein in step S3, the feature fusion is performed by adopting a bidirectional feature pyramid network BiFPN, which comprises constructing a bidirectional feature flow path from top to bottom and from bottom to top, and introducing a learnable dynamic weighting mechanism at a feature fusion node, wherein a fusion formula of the dynamic weighting mechanism is as follows: Wherein, the In order to output the feature map, Represent the first The characteristic map is entered in the form of a plurality of input features, For the corresponding weight to be able to learn, To prevent a small constant with zero denominator.
  5. 5. The tree vignetting control method based on the shoot detection according to claim 1, wherein in step S3, the class-aware sliding loss function CASL focuses on difficult samples and minority samples by assigning different weights to samples in different cross-over intervals IoU, and the total loss is calculated by: Wherein, the Is a category-aware sliding loss; Is a normalization factor, i.e. the total number of samples; as a total number of categories, Is the first BCE loss of the individual samples, Is a category of Corresponding based on the IoU mean value Is a function of the dynamic weights of the (c), For model pair number The result of the prediction of the individual samples, Is the corresponding true value.
  6. 6. The treeing control method according to claim 5, wherein the dynamic weight function The calculation formula of (2) is as follows: Wherein, the The IoU value of the sample is represented, Is a category of IoU mean value of (c).
  7. 7. The method of claim 1, wherein in step S7, the step of generating a structured agricultural control command comprises if the shoot rate is higher than a first predetermined threshold Judging that the tree vigor is overgrown, generating a control instruction for increasing pruning strength or reducing nitrogen fertilizer application, if the shoot rate is lower than a second preset threshold value Judging that the tree vigor is insufficient in growth, and generating a control instruction for increasing the fertilization amount or reducing the branch thinning proportion.
  8. 8. A system for implementing the treeing control method according to any one of claims 1 to 7, characterized in that the system comprises: the image acquisition module is used for acquiring an image of a citrus tree canopy in an orchard environment; The model reasoning module is used for detecting the network model by the new pin and outputting detection results of the new pin and the old pin; the tree potential analysis module is used for calculating the shoot rate according to the detection result and evaluating the tree potential; And the decision execution module is used for generating and outputting an agronomic management suggestion or a control instruction according to the tree potential evaluation result.
  9. 9. The system of claim 8, wherein the decision execution module is further configured to encode the generated control instructions into structured data packets and send the structured data packets to an orchard automated execution device via a wireless communication module, the execution device comprising a programmable fertilizer controller, a robotic trimming device, or an agricultural unmanned aerial vehicle flight control system.
  10. 10. The system of claim 8, wherein the decision execution module further receives a job feedback signal from an execution device and dynamically calibrates a shoot rate threshold or a control strategy parameter for a subsequent cycle based on the feedback signal.

Description

Tree potential control method based on young shoot detection Technical Field The invention belongs to the technical field of agricultural intelligent monitoring and tree potential control, and relates to a tree potential control method based on young sprout detection, which is applied to citrus orchard management based on a target detection technology of computer vision, the method is particularly suitable for real-time detection and tree vigor assessment of citrus shoots, and provides scientific basis for accurate pruning, fertilization and yield prediction of citrus orchards. Background Citrus is an important cash crop, and the yield and quality of citrus are directly related to the economic benefit of growers and the development of related industrial chains. In the growth cycle of citrus, the growth status of young shoots is a key indicator reflecting the nutritional level, physiological activity and future yield potential of the tree. Excessive growth of young shoots can cause unbalance of nutrition growth and reproductive growth of the tree body, cause flower and fruit drop, and influence fruit quality, while insufficient growth or too little young shoots can limit accumulation of photosynthetic products, inhibit normal development of the tree body, and cause yield reduction. Therefore, the method can accurately and timely monitor the growth dynamics of the citrus young shoots, and has important significance for making scientific pruning, fertilization and use strategies of plant growth regulators. In the traditional orchard management mode, the identification and statistics of young shoots mainly depend on manual observation. The agricultural personnel or fruit growers need to go deep into the orchard personally and observe the tree vigor by experience and naked eyes. The method has the advantages that firstly, the efficiency is extremely low, the real-time monitoring requirement of a modern large-scale intensive orchard is difficult to meet, secondly, the artificial observation has extremely strong subjectivity, different personnel have different judgment standards for tree vigor, quantitative data records are difficult to form, long-term data analysis and accurate management are not facilitated, and finally, the labor intensity of artificial inspection is high and the artificial inspection is easily limited by factors such as weather, topography and the like. With the rise of accurate agricultural concepts, an automatic inspection mode based on an unmanned aerial vehicle and an intelligent vision system gradually becomes a research hotspot. The fruit tree canopy information is automatically identified by utilizing a computer vision technology, and remarkable results are obtained in the fields of fruit counting, pest and disease detection and the like. However, for the automated detection of citrus shoots, there are still great technical challenges facing today, mainly in the following aspects: 1) The characteristic similarity is high, the distinguishing difficulty is high, the citrus is a evergreen fruit tree, and the color and the texture of the new shoots, the old shoots (mature branches) and the background leaves are highly similar. Particularly, in the middle and later stages of young shoot growth, the color of the young shoot is gradually changed from light green to dark, and the difference between the young shoot and old leaves is reduced, so that the traditional algorithm based on the color threshold value or simple texture features is difficult to distinguish effectively. 2) The target form is slender, the multi-scale characteristics are complex, the young shoots are usually in a slender form, and the size difference among the sprouting period, the stretching period and the leaf spreading period is huge. In images, elongated targets tend to occupy fewer pixels and feature information is easily lost during downsampling of a multi-layer convolutional network, resulting in small and elongated targets being missed. 3) The natural environment is complex, the interference factors are multiple, the illumination condition in the natural orchard environment is changed severely (such as backlight, shadow and highlight), and serious shielding and overlapping phenomena exist between branches and leaves, the background is extremely messy, and extremely high requirements are provided for the feature extraction capability of the model. 4) The category imbalance is serious, namely, the number of old shoots and background leaves is far more than that of new shoots in the image shot in the actual orchard. This extreme imbalance in class distribution can result in deep learning models that tend to learn the features of most classes (old shoots) while ignoring few classes (young shoots) during the training process, thereby reducing the accuracy of detection of the young shoots. 5) The light weight and real-time requirements of the model are that the orchard inspection generally depends on mobile equipment with