Autonomous Drone-Based Pruning

• Autonomous drone-based pruning is the integration of UAVs with stereo vision, deep learning, and real-time control to automate hazardous branch removal.
• Systems employ semantic skeletonization and stereo depth estimation to accurately detect and localize branches, addressing challenges like occlusion.
• Performance evaluations show high detection and precise cut success rates, though further work is needed to optimize sensor robustness and processing speed.

Autonomous drone-based pruning is the integration of aerial robotics, computer vision, and real-time control to enable the detection, localization, and removal of branches from trees in agricultural and forestry settings. Such systems aim to automate hazardous and labor-intensive pruning tasks through advances in 3D perception, semantic modeling, trajectory planning, and end-effector actuation. State-of-the-art pipelines employ hardware and algorithms for stereo or depth sensing, deep learning–based branch segmentation, semantic skeletonization of tree structure, and onboard trajectory planning to execute precision pruning in complex outdoor environments (You et al., 2021, Lin et al., 26 Sep 2024, Lin et al., 1 Oct 2024, Lin et al., 5 Dec 2025).

1. System Architectures and Hardware Components

Autonomous pruning systems utilize lightweight multi-rotor UAVs with stereo-vision sensors, robotic pruning implements (scissor or blade end-effectors), and embedded compute units. Standard hardware choices include:

• Sensing: ZED Mini stereo cameras (baseline 60–120 mm, 1080p resolution, 30 Hz) to capture left/right images for disparity estimation (Lin et al., 26 Sep 2024, Lin et al., 5 Dec 2025).
• Flight Platform: DJI Matrice or custom quadcopters (payload 1–2 kg), with 2-DOF gimbals for precise end-effector positioning (Lin et al., 1 Oct 2024, Lin et al., 5 Dec 2025).
• Computation: NVIDIA Jetson Xavier NX (6-core GPU, 16 GB RAM) for on-device real-time inference, SGBM, and motion control.
• Pruning End-Effector: Lightweight electric scissor mechanisms (stall torque ≥2.5 kg·cm, cut time <0.1 s), mounted under the fuselage or on articulated gimbals (Lin et al., 26 Sep 2024, Lin et al., 1 Oct 2024, Lin et al., 5 Dec 2025).

Calibration ensures accurate stereo extrinsics and vision-based pose estimation via checkerboard-based Zhang-Hartley procedures, rectification, and bundle adjustment to sub-pixel reprojection errors (Lin et al., 26 Sep 2024, Lin et al., 1 Oct 2024).

2. 3D Perception and Semantic Skeletonization

The key challenge is to reconstruct an actionable, semantically labeled model of the tree to identify valid pruning targets. Two principal paradigms are employed:

A. Semantic Skeletonization

The tree is represented as a rooted, directed acyclic graph $S = (N, E, \ell)$:

• $N = \{n_i\}$: Superpoint nodes in $\mathbb{R}^3$, each a centroid of a point cluster.
• $E \subset N \times N$: Directed edges $e = (n_{parent}\rightarrow n_{child})$.
• $$\ell: E \rightarrow \{\text{Trunk}, \text{Support}, \text{Leader}, \text{SideBranch}\}$$ assigns semantic labels reflecting orchard biology (You et al., 2021).

Key geometric/topological constraints (hard/soft) are enforced:

• Label progression: If $e_p \rightarrow e_s$, then $$\text{Order}(\ell(e_p)) \le \text{Order}(\ell(e_s))$$
• Turn-angle penalties: Large angular deviation between consecutive edges is discouraged.
• Growth-direction penalties: Edge orientation enforced according to biological priors (e.g., supports horizontal, leaders vertical).

Population-based constrained search is applied, iteratively expanding skeletons from the base node $n_B$, scoring candidate edge-label extensions by global reward functions incorporating both classification confidence (from CNN branch classifiers) and the geometric/topological penalty terms. Branch orderings (trunk=0 to side branch=3) guide allowable cuts.

B. End-to-End Stereo/Segmentation Pipeline

Systems focused on radiata pine pruning typically use direct visual segmentation and stereo depth pipelines:

• Segmentation: Performed via YOLOv8s-seg (anchor-free, segmentation head) on RGB stereo frames resized to 640×640. Output is a per-pixel branch mask and box/objectness score (Lin et al., 26 Sep 2024, Lin et al., 1 Oct 2024, Lin et al., 5 Dec 2025).
• Disparity: Semi-Global Block Matching (SGBM) computes disparity maps using $\sim$5×5 block sizes, filtered with Weighted Least Squares (WLS).
• Depth recovery: $z = \frac{f_x\cdot B}{d}$, with masked pixels in the segmentation supplying robust 3D branch centroids. Outlier filtering (median, IQR/MAD) and morphological closing are applied for refinement (Lin et al., 26 Sep 2024, Lin et al., 1 Oct 2024, Lin et al., 5 Dec 2025).

Deep stereo learning models (PSMNet, NeRF-Stereo) have demonstrated further improvements in edge precision and localization, at the expense of increased computational latency (NeRF-Stereo: 0.17 Hz) compared to SGBM (30 Hz) (Lin et al., 1 Oct 2024).

3. Branch Detection and Depth Estimation

Branch semantic segmentation and depth estimation are central to precision pruning:

Model	mAP_box50–95 (%)	mAP_mask50–95 (%)	Depth RMSE @1 m
Mask R-CNN X101	85.5	11.6	—
YOLOv8n-seg	98.9	77.4	0.018 m*
YOLOv8s-seg	99.5	82.0	0.021–0.035 m
SGBM	—	—	0.018–0.05 m
NeRF-Stereo	—	—	0.018 m

*Performance varies by pipeline and dataset; NeRF-Stereo achieves lowest RMSE but at high latency (Lin et al., 26 Sep 2024, Lin et al., 1 Oct 2024, Lin et al., 5 Dec 2025).

Branch centroid 3D localization uses projection formulas: $$X = (u - o_x)\cdot z / f_x,\quad Y = (v - o_y)\cdot z / f_y$$

Selected branch centroids feed into the motion controller. Precision is typically ±3–5 mm in low-occlusion, close-range (<2 m) scenarios (Lin et al., 1 Oct 2024).

4. Control Pipeline and Pruning Execution

End-to-end pruning consists of target selection, trajectory planning, and actuation:

1. Flight planning: Pre-planned orbits at multiple heights, with onboard or ground-ROS trajectory generation. Collision avoidance is maintained via bounding spheres and obstacle maps (You et al., 2021, Lin et al., 5 Dec 2025).
2. Targeting: 3D cut-point (e.g., $p_\text{cut}$) computed from the skeleton (at the side branch just beyond the cambial ring) or from branch centroid via YOLO/stereo pipeline (You et al., 2021, Lin et al., 26 Sep 2024).
3. Motion execution: MAVROS or custom controller moves the UAV to a cut offset (e.g., 0.1–0.5 m), aligns end-effector by yaw/pitch, and approaches with velocity capped for stability (e.g., 0.2 m/s).
4. Pruning: Visual feedback loop ensures the branch centroid is centered and within distance/angular tolerance (±5 mm, ±3°). The cutter (electric scissor, micro-servo) is triggered (Lin et al., 1 Oct 2024, Lin et al., 5 Dec 2025).

Low-latency operation is realized by limiting superpoint/edge search population (K ≈ 100), pre-emptive branch selection, and parallelizing flight and perception tasks (You et al., 2021).

5. Performance Evaluation and Experimental Results

• Detection rate: 87–94% across radiata pine field trials (Lin et al., 26 Sep 2024, Lin et al., 5 Dec 2025).
• Successful cut rate: 85–90% on detected branches, with failures attributed primarily to occlusion or backlighting.
• False positive approach: 3%.
• Depth accuracy: RMSE of 15–42 mm for ranges 1–2 m (YOLOv8/SGBM or NeRF-stereo) (Lin et al., 1 Oct 2024).
• Latency: YOLOv8n-seg + SGBM, 28 Hz end-to-end (indoor); YOLOv8s-seg + SGBM, ~0.8 s per stereo pair; full field UAV system, ~2 Hz (Lin et al., 26 Sep 2024, Lin et al., 1 Oct 2024, Lin et al., 5 Dec 2025).
• Skeletonization accuracy: 70% global skeleton correctness, with label-specific edit rates: Trunk 0.00, Support 0.42, Leader 0.21, SideBranch 0.50 (You et al., 2021).
• Runtime (skeletonization): ≈120 s for 150–300 superpoints, with main cost in the population-based search (You et al., 2021).

6. Limitations and Future Directions

Significant challenges and development fronts remain:

• Occlusion: Dense foliage and complex lighting degrade segmentation and depth; recall drops 12% in occluded cases (Lin et al., 1 Oct 2024).
• Domain adaptation: Most training sets are small and lab-based; larger, field-diverse datasets with variable illumination are required for robust deployment (Lin et al., 1 Oct 2024).
• Range: All surveyed systems have 2 m effective operational range; depth noise grows with distance due to stereo baseline limitations (Lin et al., 26 Sep 2024, Lin et al., 5 Dec 2025).
• Sensor robustness: Uniform/low-texture bark causes disparity holes; specular highlights induce failures. Event-based or hybrid LiDAR-stereo sensing may alleviate these.
• Processing speed: SGBM enables near-real-time pruning (0.5–2 Hz); deep-learning stereo models (e.g., NeRF-supervised) achieve highest accuracy but are not yet real-time for deployed UAVs (Lin et al., 1 Oct 2024, Lin et al., 5 Dec 2025).
• End-effector stability: Wind gusts and tool slip require compliant end-effectors and force sensing (Lin et al., 26 Sep 2024).
• Automation pipeline: Joint optimization of segmentation/depth, real-time neural stereo networks quantized for embedded hardware, and closed-loop force feedback are ongoing research priorities (Lin et al., 1 Oct 2024).

7. Comparative Summary and System Benchmarks

Characteristic	Semantics-guided Skeletonization (You et al., 2021)	YOLO/SGBM Integration (Lin et al., 5 Dec 2025)
Tree Model	Directed acyclic semantic skeleton	Pixel-level mask (YOLOv8s-seg)
Sensing	Depth (stereo/LiDAR/RGB-D)	Stereo RGB (ZED Mini)
Detection Accuracy	70% skeleton edit ratio	mAP_mask50–95: 82%
Depth Accuracy @1 m	—	RMSE: 15–21 mm
Runtime/throughput	120 s (skeleton, offline); onboard: 2–3 s/iter	480 ms/frame; 2 Hz
Field cut success	—	88–90% (radiata pine, <2 m)
Limiting factors	Search runtime, occlusion, superpoint errors	Depth at >2 m, lighting, occlusion

These systems conclusively demonstrate the practicality of drone-based pruning using stereo vision and semantic or direct branch detection, achieving high cut accuracy and moderate throughput under real-world constraints. Future advances in field dataset scale, sensor fusion, and embedded neural inference are likely pathways to operational deployment in commercial forestry and orchard management (You et al., 2021, Lin et al., 26 Sep 2024, Lin et al., 1 Oct 2024, Lin et al., 5 Dec 2025).