Zero-shot learning for phenotyping of maize cob geometric traits in diverse environments

Maize cob geometric traits are critical for breeding programs and yield estimation, yet traditional phenotyping methods remain labor-intensive, costly, and lack scalability. Existing deep learning approaches typically require model retraining or parameter adjustments for new varieties or environments, creating barriers for non-experts, users with limited resources, and field applications. To overcome limitations in data dependency, high costs, and model generalization, this study explores zero-shot learning (ZSL) framework that integrates Grounding DINO and a lightweight shape-corrected segment anything model (SAM) for maize cob segmentation and geometric traits estimation. To evaluate the performance of the ZSL method, extensive experiments were conducted on a public maize cob dataset from Peru and four self-collect datasets covering both laboratory and in-field conditions from China. The framework achieves 100% detection accuracy in labs and 98% in fields, with lightweight MobileSAM delivering 99.6% accuracy for maize cob segmentation at optimized efficiency of 12.90 FPS. Morphological corrections remove distracting shapes to enhance the accuracy of segmentation results in the field scenarios. Estimated traits exhibit strong correlations with manual measurements (r = 0.958–0.987). The framework achieved a remarkable correlation (r = 0.95) between cob area and actual yield, validating its practical utility. This study demonstrates the feasibility of ZSL for maize cob phenotyping in laboratory and field environments, enabling the first fully zero-shot pipeline for accurate geometric trait extraction and yield estimation. This work provides a computationally efficient, generalizable solution that bridges laboratory precision with field applicability in maize phenomics.