An integrated YOLOv11-based framework for wheat spike phenotyping segmentation and grain yield estimation

Le résumé fourni par la source

Spike count and spatial arrangement are among the strongest determinants of grain yield in wheat, yet reliable spike-level phenotyping under field conditions remains challenging. Field-acquired images are often affected by occlusion, heterogeneous illumination, and dense canopy overlap. In addition to these visual complexities, the effort required to generate large-scale instance-level annotations makes it difficult to build the extensive labeled datasets typically needed for robust segmentation models. This paper addresses both problems. The core methodological contribution is a semi-automated iterative annotation workflow: a YOLOv11x-seg model trained on a small manually annotated set is applied to unlabeled field images, and its predictions, after post-processing to remove duplicated, fragmented, and merged detections, are incorporated back into training. To prevent augmentation from distorting the training distribution, we introduce a distribution-aware augmentation strategy guided by Fréchet Inception Distance (FID), which retains only those augmented samples that remain within an acceptable distance from the original data distribution. Together, these components allowed us to build an effective training set from 3,899 high-resolution RGB images (4000 × 3000 pixels) of durum wheat collected at the CREA Research Centre for Cereal and Industrial Crops, with substantially reduced manual annotation effort. On the independent test set(242 images), the final YOLOv11x-seg model (M5 model) achieved a mask-level precision of 86.73%, recall of 83.02%, F1-score of 84.83%, mAP@50 of 89.42%, and mAP@50:95 of 60.51%. Spike masks were used to derive image-based traits including spike count, spike density, canopy coverage, spike area, spatial distribution, and vegetation indices. Their relationships with measured grain yield were explored through statistical analysis and machine-learning-based yield estimation.Both statistical and machine-learning analyses demonstrated that image-derived spike traits provided meaningful information for grain yield estimation, with canopy coverage showing the strongest positive association with yield. Using repeated nested cross-validation with out-of-fold (OOF) predictions, XGBoost achieved the highest yield estimation performance ( R OOF 2 = 0.312 , RMSE = 106.61 g/plot), supporting the potential of near-image phenotyping for late-stage yield estimation in wheat. These results show that semi-automated iterative annotation can enable practical wheat spike segmentation and image-based phenotyping under realistic open-field conditions. While grain yield estimation should be interpreted within the context of the experimental setting, the proposed framework highlights the value of image-derived spike traits for late-stage phenotyping and yield assessment in breeding experiments rather than for early-season yield forecasting.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.