ASABE Technical Library - Abstract

Click on “Download PDF” for the PDF version or on the title for the HTML version.
If you are not an ASABE member or if your employer has not arranged for access to the full-text, view access and subscription options

Automated 3D Phenotyping of Strawberry Plants: A Fast and Memory-Efficient Deep Learning Approach

Published by the American Society of Agricultural and Biological Engineers, St. Joseph, Michigan www.asabe.org

Citation:  2026 ASABE Annual International Meeting  2600884.(doi:10.13031/aim.202600884)
Authors:   Partho Ghose, Al Bashir, Yibin Wang, Azlan Zahid
Keywords:   3D reconstruction, Deep Learning, PointNet, Plant Phenotyping, Strawberry

Abstract. Accurate extraction of strawberry phenotypic traits is essential for assessing genetic variability, supporting breeding programs, and optimizing yield estimation. However, high-throughput, nondestructive 3D organ-level segmentation remains a significant challenge due to the complex morphology of strawberry plants. In this study, we propose Berry-PointNet++, an optimized deep learning framework for the automated segmentation of leaves, stems, and fruits from 3D point cloud data. Our pipeline employs a smartphone camera to capture multi-angle video, followed by a Structure from Motion (SfM) and Multi-View Stereo (MVS) workflow implemented using COLMAP to reconstruct high-fidelity 3D models. The proposed approach builds on the PointNet++ architecture for segmentation, which suffers from limited feature representation and high memory consumption. To overcome these limitations, we introduce two key architectural enhancements: integrating the Convolutional Block Attention Module (CBAM) into the multilayer perceptron layers to improve feature discrimination, and replacing standard convolutions with Partial Convolutions (PConv) to significantly reduce memory usage during training. The proposed Berry-PointNet++ framework achieved an Overall Accuracy (OA) of 97.45% and an mIoU of 92.71% while reducing training memory consumption to 2.6 GB, demonstrating its effectiveness for accurate, efficient, and memory-efficient organ-level segmentation of strawberry plants. Yet, the limited dataset size may limit the generalizability of the proposed framework across different strawberry plant structures. Future work will focus on collecting larger and more diverse point cloud datasets to improve model robustness and scalability.

(Download PDF)    (Export to EndNotes)