Wheat seedling vigor classification under field conditions using low-cost ground-level RGB imagery and a lightweight attention model

作者：Runzhi Xu, Lixin Qiu, Fei Yuan, Dashuai Wang, Kang Yu, Wei Li, Jun Lu, Xiaojun Liu, Yongchao Tian, Yan Zhu, Weixing Cao, Qiang Cao, Qiang Cao · 发表于：European Journal of Agronomy · 年份：2026 · DOI：10.1016/j.eja.2026.128266 · 研究领域：Smart Agriculture and AI、Remote Sensing in Agriculture、Spectroscopy and Chemometric Analyses

Rapid and reliable assessment of wheat seedling vigor is essential for guiding early-season field management and ensuring stable crop establishment. However, traditional field assessment methods are labor-intensive, subjective, and difficult to scale under heterogeneous production environments. There is therefore a need for practical, scalable approaches that enable consistent and efficient seedling vigor monitoring in field conditions. This study developed a practical, scalable smartphone image-based approach for wheat seedling vigor classification under field conditions. RGB canopy images were collected from seven winter wheat production sites under diverse environmental and management conditions. A lightweight model combining a mobile-friendly backbone with multi-level attention mechanisms was developed to enhance canopy feature representation. Model performance was evaluated using both a fixed train–validation–test split and a location-based cross-validation strategy. Under the fixed split, the model achieved 89.60% accuracy, 89.83% macro-F1, and a Cohen’s kappa of 0.86. Location-based cross-validation yielded 85.90 ± 7.58% accuracy, 84.26 ± 5.98% macro-F1, and a Cohen’s kappa of 0.81 ± 0.17, indicating robust cross-site generalization. Grad-CAM analysis further showed that the model primarily focused on canopy-related regions, supporting its interpretability. The proposed approach enables reliable classification of wheat seedling vigor from smartphone RGB images and demo...