2021 Volume 83 Issue 5 Pages 391-406
The monitoring of rice tiller number is one of the most tedious and the time consuming task of rice cultivation. In this work, we use deep learning as an alternative method to estimate tiller number from images captured by a field robot at three tillering stages: early stage, active stage, and maximum stage, for the two Japanese rice varieties of Fukuhibiki and Haenuki. Three types of YOLOv4 models were trained to estimate the tiller number: models aimed at estimating actual tiller numbers, models trained on classes of grouped tiller numbers, and models trained with classes based on a tiller number histogram. In the experiments, the tiller number histogram based models achieved the highest scores of mAP at the three tillering stages of early, active, and maximum: 62.3, 67.5, and 73.5 for Fukuhibiki variety, 61.3, 63.5, and 49.8 for Haenuki variety.
