Tiller number, an important growth parameter for rice cultivation, is still being assessed manually. This work investigated the influence of dataset composition on performance of deep learning models for tiller number estimation in rice. Four datasets were constructed for early tillering, active tillering, and maximum tillering by applying the concepts of mixed varieties, class balance, and data augmentation. YOLOv4 models were trained to estimate tiller numbers using each constructed dataset. Then their performance was evaluated. Results demonstrated that the models trained with datasets created using a combination of mixed variety, class balance, and augmentation showed the best performance for estimating the tiller number at the three tillering stages with a mAP range of 68.8–86.4 %.
View full abstract