Apple tree intelligent pruning system based on cross-modal fusion and double-engine decision

Abstract

The application discloses an apple tree intelligent pruning system based on cross-modal fusion and double-engine decision, and belongs to the technical field of intelligent agriculture. The system comprises a multi-source data acquisition module, a cross-modal phenotype parameter extraction module, a "rule-data" double-engine pruning decision module, a lightweight edge inference module and a closed-loop execution feedback module. By fusing laser radar point cloud and high-resolution images, accurate extraction of branch geometry and texture features is realized; in combination with a pruning standard rule base and a data-driven MLP model, a decision mechanism with rigid error-proofing and flexible adaptation is constructed; pruning, quantization and knowledge distillation techniques are adopted to compress the model, and edge-end single-tree pruning schemes are supported. The system is suitable for high-spindle apple tree winter pruning operations, especially in non-standard orchards, can significantly reduce labor costs and improve high-quality fruit yield, and has the advantages of high precision, strong adaptability and low-cost promotion.

Description

Technical Field

[0001] This invention relates to the field of smart agriculture technology, specifically to an intelligent apple tree pruning system based on cross-modal fusion and dual-engine decision-making. Background Technology

[0002] my country is the world's largest apple producer, with an output of 51.2851 million tons in 2024, accounting for more than 55% of the global total. With the mainstream adoption of dwarfing, high-density planting and tall spindle-shaped trees, the demand for refined winter pruning in orchards is increasingly urgent. However, current pruning still relies heavily on manual labor, and the aging of skilled pruners and labor shortages have become a key bottleneck restricting the high-quality development of the industry. Existing intelligent pruning technologies are still insufficient to meet actual needs.

[0003] Current research both domestically and internationally exhibits significant shortcomings: While foreign countries have achieved small-scale intelligent pruning on regular tree shapes such as grapes and UFO cherries, the high cost of their equipment and the algorithms designed for standardized orchards make them unsuitable for the complex structures of tall spindle-shaped apple trees and the high degree of non-standardization in Chinese orchards. Domestic research, on the other hand, largely remains in the laboratory stage, lacking field validation. At the perception level, traditional 2D vision methods are susceptible to interference from strong light, backlight, and leaf occlusion, resulting in a high rate of branch misjudgment. While 3D point cloud processing can acquire spatial information, existing segmentation algorithms have weak capabilities in analyzing sparse and interlaced branches, failing to meet the demands of real-time operations. At the decision-making level, most systems rely on a single AI model or a simple rule base, failing to integrate pruning standards, expert experience, and individual tree vigor differences. This leads to pruning recommendations lacking scientific rigor and safety, making them difficult for fruit growers to trust. More importantly, existing models are generally bulky and computationally expensive, hindering efficient operation on edge devices and severely restricting the large-scale deployment of the technology.

[0004] Therefore, there is an urgent need for an intelligent pruning system that combines high-precision perception, highly reliable decision-making, and low-cost deployment capabilities. This system should be able to accurately identify the branch structure under complex tree shapes, scientifically generate pruning plans that conform to national standards and are adapted to individual tree vigor, and operate in real time in resource-constrained field environments. Summary of the Invention

[0005] The present invention aims to solve the aforementioned problems by providing an intelligent apple tree pruning system based on cross-modal fusion and dual-engine decision-making.

[0006] To solve the above-mentioned technical problems, the technical solution provided by the present invention is as follows: the system includes the following functional modules: The multi-source data acquisition module is used to simultaneously acquire the target apple tree's 3D point cloud data, high-resolution 2D image data, and high-precision spatial positioning information; The cross-modal phenotypic parameter extraction module is used to preprocess, segment, and fuse the point cloud data and image data, and extract the geometric parameters and spatial topological relationships of the trunk, first-level branches, and second-level branches; The "rules-data" dual-engine pruning decision module is used to combine pruning standards, expert experience, and individual orchard scenarios to generate structured pruning suggestions; A lightweight edge inference module for efficiently running perception and decision-making models on resource-constrained edge devices; The closed-loop execution feedback module is used to convert pruning decisions into mechanical control commands and verify the execution effect through visual feedback; The system is deployed in a three-tier architecture of “cloud training – edge inference – terminal execution”, realizing full-process automation from perception to execution.

[0007] As an improvement, the multi-source data acquisition module includes: Industrial-grade RGB camera with a resolution of 4096×2160, a frame rate of ≥25fps, and a focal length of 24mm; 16-line lidar with a ranging range of 0.5–10m and a point cloud density of ≥200 points / cm²; The high-precision RTK positioning device has a planar positioning accuracy of ±1cm and an elevation accuracy of ±2cm. The camera, lidar, and RTK device are installed on an adjustable-height mobile platform and synchronized in time via the PTP protocol. The acquisition frequencies are 25fps for images and 10Hz for point clouds.

[0008] As an improvement, the cross-modal phenotypic parameter extraction module further includes: The point cloud processing submodule uses voxel grid downsampling, statistical filtering and radius filtering to remove noise, and uses the RANSAC plane segmentation algorithm to remove ground points. Then, the improved PointNet++ model is used to perform semantic segmentation on the canopy point cloud. The sphere query radius of PointNet++ is set to 2cm, 4cm and 6cm, corresponding to feature dimensions of 128, 256 and 512, and an attention mechanism is introduced in the feature fusion layer. The image processing submodule uses CLAHE adaptive histogram equalization and Gaussian bilateral filtering for preprocessing, and inputs it into the YOLO26n-seg lightweight segmentation model. After being pre-trained by COCO-Seg, the model is fine-tuned on a fruit tree-specific dataset, embeds the EPGO attention mechanism, and uses a weighted fusion loss function of Dice Loss, CIoU Loss and Focal Loss. The cross-modal fusion submodule extracts image feature points using the SIFT algorithm, extracts point cloud geometric features using the FPFH algorithm, performs bidirectional matching using the FLANN matcher, and achieves spatial alignment by solving the homography matrix using RANSAC, with a calibration error of <1cm. After fusion, a 256-dimensional phenotypic feature vector is constructed for subsequent decision-making.

[0009] As an improvement, the "rule-data" dual-engine pruning decision module includes: The rules engine submodule adopts a production rule representation method, encoding the apple pruning standard "DB13 / T 2904-2018" and expert experience into 120 executable logical rules. These rules are assigned weights according to branch type, constructing a "rule base-inference engine-fact base" architecture. The inference engine uses forward chain reasoning, and a weighted voting mechanism is used for conflicting rules to ensure that the probability of mistakenly pruning high-yield branches is <1%. The data engine submodule is built on a multilayer perceptron (MLP). The input is a 256-dimensional cross-modal fusion feature vector, the hidden layer structure is 512→256→128, the activation functions are ReLU and Sigmoid, and the output is the pruning action category and the three-dimensional pruning position coordinates. The training data includes 10,000+ expert-annotated samples, which are augmented to 500,000+ samples, and a scene labeling system is introduced to support adaptation to non-standard orchards. The collaborative optimization submodule executes a closed-loop process of "rule pre-screening - data fine-tuning - expert review - feedback update": the rule engine pre-filters 30% of clear samples, and the data engine processes fuzzy samples; when the two engines make conflicting decisions, they are submitted to experts for review, and the results are used to update the rule weights and incrementally train the data engine to ensure that the overall decision-making consistency rate with experts is ≥92%.

[0010] As an improvement, the lightweight edge inference module operates as follows: S1. Perform structured pruning on the YOLO26n-seg model, retaining 60% of the core C2f-N modules and removing redundant neurons; S2. TensorRT is used to perform INT8 quantization on YOLO26n-seg, and 1000 orchard images are used for calibration to ensure that the segmentation accuracy loss is ≤2%; S3. The overall decision-making model is quantized using FP16 mixed precision quantization, and the lightweight student model is trained by knowledge distillation with a temperature coefficient of T=2 using a high-precision teacher model. S4. Finally, the original 1.8GB model was compressed to 12MB, with a compression rate of 99.35%, and CPU / GPU hybrid inference was achieved on edge devices through ONNX Runtime, with a single tree pruning scheme generation time of ≤45 seconds.

[0011] As an improvement, the closed-loop execution feedback module includes: The decision output unit generates a standardized pruning report, which includes a 3D visualized pruning plan, confidence score, and target coordinates of the robotic arm. The execution control unit transmits trimming commands to the end effector of the trimming robot via the CAN bus; The visual feedback unit acquires local images after the pruning action is completed, compares the state of the branches before and after pruning, and triggers a replanning mechanism if the deviation exceeds a threshold, forming a closed-loop control of "perception-decision-execution-feedback".

[0012] This invention proposes an intelligent apple tree pruning system based on cross-modal fusion and dual-engine decision-making, which significantly improves the accuracy, efficiency, and applicability of intelligent orchard management. By integrating 3D point cloud and 2D image perception, the system achieves branch segmentation mIoU ≥ 90%, phenotypic parameter extraction accuracy ≥ 92%, and spatial alignment error less than 1cm in non-standard orchard scenarios such as complex lighting and branch cross-shading. It innovatively constructs a "rule-data" dual-engine decision-making mechanism, embedding pruning standards and expert experience to ensure pruning safety (probability of mistakenly pruning high-yielding branches < 1%), and achieving personalized adaptation of "one policy per tree" through data-driven models, with an overall decision-making consistency rate with experts exceeding 92%. Simultaneously, it adopts a lightweight strategy integrating pruning, quantification, and knowledge distillation, compressing the model size to 12MB and reducing the single-tree solution generation time to within 45 seconds. This allows for deployment on edge devices costing only a few thousand yuan, significantly lowering the application threshold and providing a practical technical path to solve the "pruning worker shortage" and promote the quality and efficiency improvement of the apple industry. Attached Figure Description

[0013] Figure 1 This is a schematic diagram of the intelligent apple tree pruning system based on cross-modal fusion and dual-engine decision-making according to the present invention. Figure 2 This is a field image of an embodiment of the intelligent apple tree pruning system based on cross-modal fusion and dual-engine decision-making of the present invention; Figure 3 This is an analysis diagram of an embodiment of the intelligent apple tree pruning system based on cross-modal fusion and dual-engine decision-making of the present invention; Figure 4 This is data from an embodiment of the intelligent apple tree pruning system based on cross-modal fusion and dual-engine decision-making, as described in this invention. Figure 1 ; Figure 5 This is data from an embodiment of the intelligent apple tree pruning system based on cross-modal fusion and dual-engine decision-making, as described in this invention. Figure 2 ; Figure 6 This is data from an embodiment of the intelligent apple tree pruning system based on cross-modal fusion and dual-engine decision-making, as described in this invention. Figure 3 . Detailed Implementation

[0014] The present invention will now be described in further detail with reference to the accompanying drawings.

[0015] Combined with appendix Figure 1 An intelligent apple tree pruning system based on cross-modal fusion and dual-engine decision-making, the system includes the following functional modules: A multi-source data acquisition module is used to simultaneously acquire 3D point cloud data, high-resolution 2D image data, and high-precision spatial positioning information of the target apple tree. The multi-source data acquisition module includes an industrial-grade RGB camera with a resolution of 4096×2160, a frame rate of ≥25fps, and a focal length of 24mm; and a 16-line lidar with a ranging range of 0.5–10m and a point cloud density of ≥200 points / cm². The high-precision RTK positioning device has a planar positioning accuracy of ±1cm and an elevation accuracy of ±2cm. The camera, lidar, and RTK device are mounted on an adjustable-height mobile platform and synchronized in time via the PTP protocol. The acquisition frequencies are 25fps for images and 10Hz for point clouds. A cross-modal phenotypic parameter extraction module is used to preprocess, segment, and fuse the point cloud data and image data, extracting the geometric parameters and spatial topological relationships of the trunk, first-level branches, and second-level branches. This module further includes: a point cloud processing submodule, which uses voxel grid downsampling, statistical filtering, and radius filtering to remove noise, and uses the RANSAC plane segmentation algorithm to remove ground points; subsequently, an improved PointNet++ model is used to perform semantic segmentation on the canopy point cloud, where the sphere query radius of PointNet++ is set to 2cm, 4cm, and 6cm, corresponding to feature dimensions of 128, 256, and 512, and an attention mechanism is introduced in the feature fusion layer; and an image processing submodule, which uses CLAHE adaptive histogram equalization and Gaussian bilateral filtering for preprocessing, and inputs the data into the YOLO26n-seg lightweight segmentation model; this model is pre-trained on COCO-Seg and fine-tuned on a fruit tree-specific dataset, embedding the EPGO attention mechanism, and employing Dice Loss, CIoU Loss, and Focal... The loss function for weighted fusion is as follows: A cross-modal fusion submodule extracts image feature points using the SIFT algorithm and point cloud geometric features using the FPFH algorithm. Bidirectional matching is performed using a FLANN matcher, and spatial alignment is achieved by solving the homography matrix using RANSAC, with a calibration error <1cm. A 256-dimensional phenotypic feature vector is constructed after fusion for subsequent decision-making. The "rule-data" dual-engine pruning decision module combines pruning standards, expert experience, and individual orchard scenarios to generate structured pruning recommendations. This module includes: a rule engine submodule, employing a production rule representation to encode the apple pruning standard (DB13 / T 2904-2018) and expert experience into 120 executable logical rules, assigning weights according to branch type, and constructing a "rule base-inference engine-fact base" architecture. The inference engine uses forward chain reasoning, and a weighted voting mechanism is used for conflicting rules to ensure that the probability of mistakenly pruning high-yielding branches is <1%; and a data engine submodule, built on a multilayer perceptron (MLP) architecture, with input being a 256-dimensional cross-modal fusion feature vector, a hidden layer structure of 512→256→128, activation functions of ReLU and Sigmoid, and output being pruning dynamics. The system includes classification and 3D pruning location coordinates; the training data contains 10,000+ expert-annotated samples, expanded to 500,000+ samples through data augmentation, and introduces a scene labeling system to support adaptation to non-standardized orchards; the collaborative optimization submodule executes a closed-loop process of "rule pre-screening—data fine-tuning—expert review—feedback update": the rule engine pre-filters 30% of clear samples, and the data engine processes fuzzy samples; when the two engines' decisions conflict, they are submitted for expert review, and the results are used to update the rule weights and incrementally train the data engine, ensuring that the overall decision-making consistency rate with experts is ≥92%. A lightweight edge inference module is used to efficiently run perception and decision-making models on resource-constrained edge devices; the operation steps of the lightweight edge inference module are as follows: S1. Perform structured pruning on the YOLO26n-seg model, retaining 60% of the core C2f-N modules and removing redundant neurons; S2. TensorRT is used to perform INT8 quantization on YOLO26n-seg, and 1000 orchard images are used for calibration to ensure that the segmentation accuracy loss is ≤2%; S3. The overall decision-making model is quantized using FP16 mixed precision quantization, and the lightweight student model is trained by knowledge distillation with a temperature coefficient of T=2 using a high-precision teacher model. S4. Finally, the original 1.8GB model was compressed to 12MB, achieving a compression rate of 99.35%. On edge devices, CPU / GPU hybrid inference was implemented using ONNX Runtime, with a single tree pruning scheme generation time of ≤45 seconds. The closed-loop execution feedback module is used to convert pruning decisions into mechanical control commands and verify the execution effect through visual feedback. The closed-loop execution feedback module includes: a decision output unit, which generates a standardized pruning report, including a 3D visualized pruning plan, confidence score, and target coordinates of the robotic arm; an execution control unit, which transmits pruning commands to the end effector of the pruning robot via a CAN bus; and a visual feedback unit, which acquires local images after the pruning action is completed, compares the state of branches before and after pruning, and triggers a replanning mechanism if the deviation exceeds a threshold, forming a closed-loop control of "perception-decision-execution-feedback".

[0016] The following are specific embodiments of the present invention: Example 1: System Hardware Platform Setup Mobile data collection platform An electric tricycle was selected as the carrier, and a height-adjustable support frame was installed (height adjustment range 1.5–2.5m). Equipped with: Industrial camera: Sony IMX series, 4K resolution (4096×2160), 25fps, 24mm fixed focal length lens; LiDAR: RoboSense RS-LiDAR-16, 16 lines, 10m range, 320,000 points / second frequency; RTK module: Qianxun Location FindCM, positioning accuracy ±1cm (plane) / ±2cm (elevation); Synchronization triggering: The camera and radar are synchronized at the frame level through PTP (Precision Time Protocol), with a time error of <1ms.

[0017] Edge inference devices It uses NVIDIA Jetson AGX Orin (32GB RAM, 200 TOPS INT8 computing power) as the edge computing unit; Install the lightweight model (12MB ONNX format), CAN communication driver, and ROS2 middleware; It is connected to a six-DOF pruning robot arm (equipped with a high-precision torque sensor at the end) via a CAN bus.

[0018] Cloud training platform Alibaba Cloud ECS instance: 8-core CPU, 16GB memory, Tesla V100 GPU; Storage: MySQL (stores structured parameters such as branch angles and diameters), MongoDB (stores .pcd point clouds, .png images, and .json annotations); Training framework: PyTorch 2.0 + CUDA 12.1 + Open3D + Ultralytics YOLO.

[0019] Example 2: Cross-modal sensing process execution Data collection Data was collected from 100 fruit trees in a Red Fuji apple orchard (non-standardized, 15-year-old trees) in Yantai, Shandong. Data was collected in different time periods: strong light (12:00), weak light (7:00), and backlight (17:00). 10 frames of images + 10 frames of point cloud data were collected for each tree. Simultaneously record wind speed (<3 level) and light intensity (500–10000 lux).

[0020] Preprocessing Image: CLAHE enhancement → Gaussian bilateral filter → scaling to 640×640 → RGB → BGR conversion; Point cloud: Voxel downsampling (0.5cm³) → Statistical filtering (K=20) → Radius filtering (r=0.8cm) → RANSAC ground segmentation.

[0021] Cross-modal alignment and fusion Extract SIFT feature points from the image (approximately 200 points / frame) and FPFH features from the point cloud (neighborhood 5cm); After FLANN matching, RANSAC is used to remove mismatches, and the homography matrix H is solved. The branch mask output by YOLO26n-seg is projected onto the point cloud space to correct the misjudgment of leaves in the point cloud segmentation. The depth information of the point cloud is used in reverse to complete the outline of the broken branches in the image due to occlusion.

[0022] Parameter extraction Output the following for each branch: diameter (mm), length (cm), angle with the trunk (°), and spatial coordinates (x, y, z). Calculate population phenotypes such as canopy volume (m³), branch density (branches / m³), and average spacing between first-order branches (cm).

[0023] Example 3: Dual-engine decision-making and execution Rule engine running Input: A branch of a certain order has a diameter of 2.8cm, an angle of 55°, and a distance of 18cm from its neighboring branch; Rule matching: "Weak branches with a spacing of less than 20cm should be removed" + "Branches with an angle of less than 60° should be removed first"; The inference engine determines that this branch is a weak branch and recommends its removal.

[0024] Data engine running Input a 256-dimensional fused feature vector (including texture roughness, growth potential index, surrounding branch density, etc.); MLP output: Thinning (confidence 94.7%), pruning location coordinates (x=1.23, y=0.87, z=1.05).

[0025] Dual-engine synergy If both parties agree, the data is output directly; if there is a conflict (such as a rule suggesting retention or a data suggestion suggesting truncation), it is submitted for expert review. The review results are stored in the feedback database for the next round of model fine-tuning and rule weight updates.

[0026] Mechanical execution and feedback The edge-side generates a JSON-formatted instruction: {"action": "remove", "position": [1.23, 0.87,1.05], "confidence": 0.947}; The robotic arm moves to the target point and completes the cutting; After execution, the camera takes a partial image and compares it to see if the branch has been removed. If not, a retry is triggered.

[0027] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structural methods to the embodiments without departing from the spirit of the present invention, such designs should fall within the protection scope of the present invention.

Claims

1. An intelligent apple tree pruning system based on cross-modal fusion and dual-engine decision-making, characterized in that, The system includes the following functional modules: The multi-source data acquisition module is used to simultaneously acquire the target apple tree's 3D point cloud data, high-resolution 2D image data, and high-precision spatial positioning information; The cross-modal phenotypic parameter extraction module is used to preprocess, segment, and fuse the point cloud data and image data, and extract the geometric parameters and spatial topological relationships of the trunk, first-level branches, and second-level branches; The "rules-data" dual-engine pruning decision module is used to combine pruning standards, expert experience, and individual orchard scenarios to generate structured pruning recommendations; A lightweight edge inference module for efficiently running perception and decision-making models on resource-constrained edge devices; The closed-loop execution feedback module is used to convert pruning decisions into mechanical control commands and verify the execution effect through visual feedback; The system is deployed in a three-tier architecture of "cloud training - edge inference - terminal execution" to achieve full automation from perception to execution.

2. The intelligent apple tree pruning system based on cross-modal fusion and dual-engine decision-making according to claim 1, characterized in that, The multi-source data acquisition module includes: Industrial-grade RGB camera with a resolution of 4096×2160, a frame rate of ≥25fps, and a focal length of 24mm; 16-line lidar with a ranging range of 0.5–10m and a point cloud density of ≥200 points / cm²; High-precision RTK positioning device, with a horizontal positioning accuracy of ±1cm and an elevation accuracy of ±2cm; The camera, lidar, and RTK device are mounted on an adjustable-height mobile platform and synchronized in time via the PTP protocol. The acquisition frequencies are 25fps for images and 10Hz for point clouds.

3. The intelligent apple tree pruning system based on cross-modal fusion and dual-engine decision-making as described in claim 1, characterized in that, The cross-modal phenotypic parameter extraction module further includes: The point cloud processing submodule uses voxel grid downsampling, statistical filtering and radius filtering to remove noise, and uses the RANSAC plane segmentation algorithm to remove ground points. Then, the improved PointNet++ model is used to perform semantic segmentation on the canopy point cloud. The sphere query radius of PointNet++ is set to 2cm, 4cm and 6cm, corresponding to feature dimensions of 128, 256 and 512, and an attention mechanism is introduced in the feature fusion layer. The image processing submodule uses CLAHE adaptive histogram equalization and Gaussian bilateral filtering for preprocessing, and inputs it into the YOLO26n-seg lightweight segmentation model. After being pre-trained by COCO-Seg, the model is fine-tuned on a fruit tree-specific dataset, embeds the EPGO attention mechanism, and uses a weighted fusion loss function of Dice Loss, CIoU Loss and Focal Loss. The cross-modal fusion submodule extracts image feature points using the SIFT algorithm, extracts point cloud geometric features using the FPFH algorithm, performs bidirectional matching using the FLANN matcher, and achieves spatial alignment by solving the homography matrix using RANSAC, with a calibration error of <1cm. After fusion, a 256-dimensional phenotypic feature vector is constructed for subsequent decision-making.

4. The intelligent apple tree pruning system based on cross-modal fusion and dual-engine decision-making according to claim 1, characterized in that, The "rule-data" dual-engine pruning decision module includes: The rules engine submodule adopts a production rule representation method, encoding the apple pruning standard "DB13 / T 2904-2018" and expert experience into 120 executable logical rules. These rules are assigned weights according to branch type, constructing a "rule base-inference engine-fact base" architecture. The inference engine uses forward chain reasoning, and a weighted voting mechanism is used for conflicting rules to ensure that the probability of mistakenly pruning high-yield branches is <1%. The data engine submodule is built on a multilayer perceptron (MLP). The input is a 256-dimensional cross-modal fusion feature vector, the hidden layer structure is 512→256→128, the activation functions are ReLU and Sigmoid, and the output is the pruning action category and the three-dimensional pruning position coordinates. The training data includes 10,000+ expert-annotated samples, which are augmented to 500,000+ samples, and a scene labeling system is introduced to support adaptation to non-standard orchards. The collaborative optimization submodule executes a closed-loop process of "rule pre-screening - data fine-tuning - expert review - feedback update": the rule engine pre-filters 30% of clear samples, and the data engine processes fuzzy samples; when the two engines make conflicting decisions, they are submitted to experts for review, and the results are used to update the rule weights and incrementally train the data engine to ensure that the overall decision-making consistency rate with experts is ≥92%.

5. The intelligent apple tree pruning system based on cross-modal fusion and dual-engine decision-making according to claim 1, characterized in that, The lightweight edge inference module operates as follows: S1. Perform structured pruning on the YOLO26n-seg model, retaining 60% of the core C2f-N modules and removing redundant neurons; S2. TensorRT is used to perform INT8 quantization on YOLO26n-seg, and 1000 orchard images are used for calibration to ensure that the segmentation accuracy loss is ≤2%; S3. The overall decision-making model is quantized using FP16 mixed precision quantization, and the lightweight student model is trained by knowledge distillation with a temperature coefficient of T=2 using a high-precision teacher model. S4. Finally, the original 1.8GB model was compressed to 12MB, with a compression rate of 99.35%, and CPU / GPU hybrid inference was achieved on edge devices through ONNXRuntime, with a single tree pruning scheme generation time of ≤45 seconds.

6. The intelligent apple tree pruning system based on cross-modal fusion and dual-engine decision-making according to claim 1, characterized in that, The closed-loop execution feedback module includes: The decision output unit generates a standardized pruning report, which includes a 3D visualized pruning plan, confidence score, and target coordinates of the robotic arm. The execution control unit transmits trimming commands to the end effector of the trimming robot via the CAN bus; The visual feedback unit acquires local images after the pruning action is completed, compares the state of the branches before and after pruning, and triggers a replanning mechanism if the deviation exceeds a threshold, forming a closed-loop control of "perception-decision-execution-feedback".

Images