Click on “Download PDF” for the PDF version or on the title for the HTML version.
If you are not an ASABE member or if your employer has not arranged for access to the full-text, Click here for options.
Applying Deep Learning for Tiller Detection by Field Robot in Rice Cultivation
Published by the American Society of Agricultural and Biological Engineers, St. Joseph, Michigan www.asabe.org
Citation: 2021 ASABE Annual International Virtual Meeting 2100311.(doi:10.13031/aim.202100311)
Authors: Dhirendranath Singh, Shigeru Ichiura, Thanh Tung Nguyen, Mitsuhiko Katahira
Keywords: class range, deep learning, field robot, rice, tiller detection, YOLOv4,
Abstract. Deep learning is increasingly being applied to some of the more challenging tasks in agriculture. The major advantage it offers is the reduction of labour and the ability to analyze large quantities of data in relatively short time. In rice cultivation, the monitoring of tiller number at the vegetative growth stage remains one of the most tedious and time consuming task at the field level, as it is still being done manually. Alternative methods, focus on systems that detect tillers in potted rice plants which are not suitable for in-field application, as such, a new approach is needed. In this study we explore the use of deep learning for detecting tiller number of rice plants in field conditions at three stages; early, active and maximum tiller. Images for training models were collected using a field robot, from an experimental plot at the Field Science Center, Faculty of Agriculture, Yamagata University, Japan during June-July 2020. The number of tiller per plant was manually counted for each plant in the captured images. Three types of YOLOv4 models were trained to estimate tiller numbers; 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. Evaluation of models showed the trained models could not accurately detect actual tiller number, but good results (mAP; 62.3,61.3, 67.5) could be obtained with the tiller number histogram based class models.
(Download PDF) (Export to EndNotes)