Counting wheat heads using a simulation model

Abstract

Numerous studies have reported a significant positive correlation between wheat yield and the quantity of wheat heads. However, collecting data on wheat heads in the field poses a challenge for several reasons, including the uncontrollable nature of the environment, inconsistent data quality, and ambiguous data truth. To address these challenges, we developed a simulation strategy to replicate the conditions of a real wheat field, which enabled the data collection process to be conducted indoors over a short period. After applying grayscale image processing to process the simulated wheat images, we trained and tested nine deep learning models: Faster-RCNN, YOLOv7, YOLOv8, CenterNet, SSD, RetinaNet, EfficientDet, Deformable-DETR and DINO. Our results indicated that YOLOv7 performed the best (R² = 0.963, RMSE = 2.463). We then compared our model trained on simulated wheat data to a model trained on real wheat data (R² = 0.963 vs 0.972, RMSE = 2.463 vs 2.692). We also achieved good model performance on five test sets: GWHD, SDAU2021-SDAU2024. The results demonstrated the efficacy of our simulation, which provides an efficient and convenient strategy for the precision agriculture community.