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

AgriDINO-seg: A DINOv3-Based Foundation Model for Multi-Crop Semantic Segmentation: A Case Study on Blueberry and Peanut

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

Citation:  2026 ASABE Annual International Meeting  2601108.(doi:10.13031/aim.202601108)
Authors:   Rui-Feng Wang, Daniel J. Petti, Changying Li
Keywords:   Blueberry, DINOv3, Peanut, Self-Supervised Learning, Visual Foundation Models.

Abstract. Agricultural visual perception is challenged by strong cross-crop variability, heterogeneous acquisition conditions, and limited annotated data, which reduce the robustness of conventional supervised segmentation methods. To address these issues, this study proposes AgriDINO-seg, a Self-Distillation with No Labels v3 (DINOv3)-based vision foundation model for multi-crop segmentation. AgriDINO-seg integrates annotation standardization, frozen DINOv3 feature extraction, and a lightweight patch-level segmentation decoder within a unified pipeline. This design translates pretrained semantics into stable agricultural foreground masks through efficient downstream adaptation. The framework was validated on three semantic segmentation tasks with distinct visual characteristics: blueberry bruising segmentation, blueberry fruit segmentation, and peanut pod segmentation. Experimental results showed consistent performance gains as backbone scale increased from ViT-S to ViT-L, with improvements observed in mIoU and Dice across both validation and test sets (e.g., +5.8 % mIoU and +5.0 % Dice on the peanut pod test set). Qualitative analysis further confirmed that AgriDINO-seg produced more coherent masks, improved region continuity, and more reliable foreground coverage under heterogeneous imaging conditions. These findings demonstrate that AgriDINO-seg provides an effective and lightweight solution for multi-crop agricultural semantic segmentation. In addition, the stable fruit-region segmentation results suggest a promising future direction for blueberry robotic harvesting, in which AgriDINO-seg may serve as a front-end for blueberry cluster detection through aggregation-aware reasoning.

(Download PDF)    (Export to EndNotes)