SDrAwberry: Scale-referenced Depth-Anything-3 long-sequence reconstruction for Strawberry canopy volume estimation

Authors: HUANG, ZIJING - University Of Florida; LEE, WON SUK - University Of Florida; QIN, RUOYAO - University Of Florida; MEDEIROS, HENRY - University Of Florida; Jeon, Hongyoung; Zhu, Heping

Submitted to: Computers and Electronics in Agriculture
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 6/18/2026
Publication Date: 6/24/2026
Citation: Huang, Z., Lee, W., Qin, R., Medeiros, H., Jeon, H., Zhu, H. 2026. SDrAwberry: Scale-referenced Depth-Anything-3 long-sequence reconstruction for Strawberry canopy volume estimation. Computers and Electronics in Agriculture. 252. Article 112088. https://doi.org/10.1016/j.compag.2026.112088.
DOI: https://doi.org/10.1016/j.compag.2026.112088

Interpretive Summary: Measuring the volume of strawberry bushes in commercial fields is challenging due to the long, repetitive layout of strawberry beds. However, accurate volume estimation is important for predicting strawberry yield, and accurate yield prediction is beneficial for U.S. strawberry growers to better manage their strawberry farm business. This work introduces SDrAwberry, a scale referenced 3D reconstruction and phenotyping pipeline, that processes a series of synchronized RGB images captured by a ground robot. This method uses advanced image processing algorithms to measure the volume of each strawberry bush from three camera views. Experiments across eight strawberry beds show strong agreement with manual plant counts (Mean absolute percentage error (MAPE) of 5.36%) and reliable volume measurement (MAPE of 3.95%). The results demonstrate that the development from this work could enable field scale, metrically meaningful phenotyping from field RGB images through batching, fusion, and scale referencing. For U.S. strawberry farmers, this work could reduce labor requirements for monitoring of crop vigor, early detection of underperforming plants, and data-driven decisions for irrigation and fertilization.

Technical Abstract: Canopy volume is a key structural trait for crop monitoring and phenotyping platforms, but reliable plant-level estimation in commercial strawberry fields remains challenging due to repetitive canopy structure, limited texture on bed surfaces, and the need to process long, nearly linear image sequences without loop closure. This paper presents SDrAwberry, a scale-referenced long-sequence 3D reconstruction and plant-level canopy volume estimation, pipeline built on synchronized tri-view RGB image data. A ground robot captured long sequences of RGB image sets from eight strawberry plant beds over a growing season. Then the data were partitioned into overlapping temporal batches to enable memory-efficient inference. Each batch was exported in a standard representation, and submaps were fused into a shared global frame by extracting fixed cross-batch 3D–3D correspondences from overlapping feature tracks and estimating motion-constrained stitching consistent with bed-aligned traversal. Redundant points in overlapping regions were removed incrementally, and voxel-hash fusion produced a compact global point cloud. Plant instances were obtained from the fused reconstruction using Hue, Saturation, and Value-guided plant point selection, robust support-plane estimation, and DBSCAN clustering with ground-like cluster rejection. Instance point clouds were converted to watertight alpha-shape concave hull meshes, and physical volumes were recovered using reference-box scale calibration. Canopy volume estimations of SDrAwberry showed strong agreement between clustering and manual plant counts with a mean absolute percentage error (MAPE) of 5.36% and consistent metric scale recovery with accurate reference-box volume self-validation (overall MAPE 3.95%). These results indicate that feed-forward multi-view geometry transformers can be extended to field-scale, metrically meaningful phenotyping via batching-and-fusion and scale-referenced volumetric reconstruction.