AppleQSM: Geometry-Based 3D Characterization of Apple Tree Architecture in Orchards

The architecture of apple trees plays a pivotal role in shaping their growth and fruit-bearing potential, forming the foundation for precision apple management. Traditionally, 2D imaging technologies were employed to delineate the architectural traits of apple trees, but their accuracy was hampered by occlusion and perspective ambiguities. This study aimed to surmount these constraints by devising a 3D geometry-based processing pipeline for apple tree structure segmentation and architectural trait characterization, utilizing point clouds collected by a terrestrial laser scanner (TLS). The pipeline consisted of four modules: (a) data preprocessing module, (b) tree instance segmentation module, (c) tree structure segmentation module, and (d) architectural trait extraction module. The developed pipeline was used to analyze 84 trees of two representative apple cultivars, characterizing architectural traits such as tree height, trunk diameter, branch count, branch diameter, and branch angle. Experimental results indicated that the established pipeline attained an R2 of 0.92 and 0.83, and a mean absolute error (MAE) of 6.1 cm and 4.71 mm for tree height and trunk diameter at the tree level, respectively. Additionally, at the branch level, it achieved an R2 of 0.77 and 0.69, and a MAE of 6.86 mm and 7.48° for branch diameter and angle, respectively. The accurate measurement of these architectural traits can enable precision management in high-density apple orchards and bolster phenotyping endeavors in breeding programs. Moreover, bottlenecks of 3D tree characterization in general were comprehensively analyzed to reveal future development.

1 Case Study: Characterization for Breeding Programs Eight three-year-old Gala-W apple trees grown on their own roots (derived tissue culture) in a Cornell Agritech Orchard (latitude: 42.880 N, longitude: 77.006 W) in Geneva, NY, USA, were used as an extra dataset to prove the versatility of the developed processing pipeline.These trees were planted at a spacing of 4.27 m (14 feet) by 0.91 m (36 inches) in 2020.Notably, these trees grow downward branches (Figure S1).
The developed apple tree characterization pipeline has also demonstrated effectiveness when applied to apple trees with altered shoot growth from upward to downward.Despite the unique and unconventional branch pattern, the pipeline has consistently yielded favorable results, showcasing a positive correlation between the traits under examination (R 2 = 0.59, MAE of 10.98°, and MAPE of 12.80%) (Figure S11).This achievement underscores the robustness and adaptability of the pipeline, promising a useful tool for high-throughput phenotyping of diverse tree growth habits that are relevant in apple genetic studies and breeding programs.It is worth noting that the error originated from the measurement protocol difference became more pronounced when trees exhibited an unconventional branch pattern.The skeleton points after tree skeletonization

Tables
The biology-aware weighted graph The optimal minimum spanning tree in G W (i.e., the coarse skeleton points) The branch origin clusters The spectral clusters after the spectral clustering algorithm The most critical parameters to make the pipeline achieve the optimal performance are: 1) downsampling -k.This "Iteration" parameter determines the resolution of the downsampled tree.A higher iteration level divides the original point cloud into smaller grids, resulting in a higher-resolution downsampled tree, 2) skeleton extraction -ϵ which determines the number of skeleton points and the average distance between skeleton points.An increase in this value will lead to fewer skeleton       Figure S8: Branch root clustering results using improper sphere pruning radius.Using a large sphere pruning radius leads to over-segmentation issues, whereas using a small sphere pruning radius leads to under-segmentation issues.Table S4: The number of segmented branches for eachh order from AdQSM.

Figure S2 :
Figure S2: Illustration of Hilbert curve based downsampling method.Each individual branch was labeled by different colors to better demonstrate the process.The color coding was to intuitively illustrate the detail-preserving properties of the HC-based downsampling by visualizing how points belonging to different branches were sampled.

Figure S3 :
FigureS3: Biology aware graph and MST algorithm.S1 is the skeleton points with the raw connectivity (cycles and discontinuity exist in S1).A weighted graph was developed by taking each skeleton point as vertex and each connection as edges.The weight of edges is computed by a weighted sum of the total thickness of two vertices and Euclidean distance between two vertices.The MST algorithm was used to generate the optimal subgraph without cycles.S2 is the same skeleton points with the improved connectivity.The more red the vertex is the larger the weight is.

Figure S4 :
Figure S4: Illustration of the maximum direction matching (MDM) method for branch unique matching.

Figure S5 :
FigureS5: Illustration of trunk and branch refinement process.A cylinder pruning operation takes each coarse trunk skeleton point as center and points inside the cylinder were selected for the trunk refinement.The points were divided into cross-sections and CPC optimization was used to generate improved center points for each cross-section.Improved center points were post-processed by a local and global RANSAC to remove outliers produced because of irregular cross-sections.S1 and S2 are the improved trunk skeleton points output.Trunk skeleton points in S2 were used to fit a cubic spline and N points were uniformly sampled from the fitted spline as the final refined trunk skeleton points.

Figure S6 :
Figure S6: Illustration of scanning setup for row 13, row 15, and row 16.The distance between scans was shown on top of the dotted line (Zoom-in).The average distance of scans for row 13 is around 3 m, which presented a high overlapping between scans and led to low registration error.The scanning setup for row 15 and row 16 is a more standard setup to balance the scanning efficiency and data quality.

Figure S7 :
Figure S7: Illustration of branch level architecture trait field measurement and branch angle sensor measurement.In-field branch diameter was measured by an equilifruit disc (Valent LLC, USA) and in-field branch inclination angle was measured using a digital angle finder (CamRad 82305, Guilin GemRed Sensor Technology Co., Ltd, China) by placing it in the crotch.Sensor branch inclination angle was measured by computing the angle between two vectors in the 3D space.

Figure S9 :
FigureS9: Showcase of the trunk estimation outliers.A: Red and blue points represent the trunk and branch skeleton points, respectively.The trunk skeleton starts from a wrong point because the side branch extends to a lower height than the trunk.B: Two clusters were produced by the 3D DBSCAN algorithm, and the blue cluster was the trunk cluster.However, a bunch of branch points were also included in the blue cluster, leading to a significantly larger trunk estimation.C: Only half of the trunk points exist in the original point cloud.D: The field measurement turns out to be inaccurate by using CloudCompare to manually measure the trunk diameter (i.e., 58.5 mm).

Figure S10 :
Figure S10: Showcase of failure and inconsistency in branch inclination angle estimation.A: The branch origin points were incorrectly identified because of noise so the estimation has a significant deviation.C and D: Different measurements in the inconsistent measurement approach.

Figure S11 :
Figure S11: Branch angle estimation results of trees in the Cornell Agritech Orchard in Geneva.

Figure S12 :
Figure S12: Branch angle estimation results using an artificial tree.

Figure S13 :
Figure S13: The distribution of the number of primary branches characterized by TreeQSM in 50 independent runs.The ground truth was indicated by the red number

Figure S15 :
Figure S15: Branch diameters and inclination angles extracted using the AdQSM method.

Table S1 :
Summary of apple trees and point cloud data from the Cornell Agritech Orchard in Geneva.

Table S2 :
Summary of the meaning of symbols used in this study.

Table S3 :
Critical pipeline parameters used in this study.
Trunk and Branch Segmentation -N min DBSCAN minimum number for clusters (#points) 3 Trunk and Branch Segmentation -d ext Projection distance threshold (m) 0.06 Trunk and Branch Segmentation -R