Articles
Smartphone-based strawberry plant growth monitoring using YOLO
Article number
1377_8
Pages
69 – 76
Language
English
Abstract
In recent years, as agricultural production has become more commercialized and large-scaled, there has been a marked increase in the need to qualitatively and quantitatively assess plant condition and the know-how of cultivation management.
For example, weekly measurements of plant growth using various instruments are becoming more common in greenhouse horticultural production.
However, to evaluate strawberry plant growth, the number of flowers and fruits are typically counted which is both labor and cost intensive.
We therefore developed a cost-effective phenotyping technique that can be used by ordinary farmers to evaluate plant growth.
The technique employs a deep learning based object detection model to count the number of flowers and fruits of the strawberry canopy.
We used a smartphone to capture color images and the YOLOv3 object detection algorithm to construct a flower and fruit detection model.
To consider the changes in the shape and color that flowers undergo as they mature into fruit, 400 images were annotated with four labels, i.e., flower, pre-immature fruit, post-immature fruit, and mature fruit, and these images were then used for training the model.
The developed model was tested using 20 images that were not included in the training data set.
The developed model successfully distinguished among the four labels with F-measures ranging from 0.69 to 0.83. Based on these initial results, measurements were conducted once every 1 to 2 weeks over a three-month period.
The number of mature fruit detected was compared with the number of harvested fruits.
The correlation coefficient was 0.95, indicating that the developed deep learning model was sufficiently accurate to evaluate strawberry growth.
For example, weekly measurements of plant growth using various instruments are becoming more common in greenhouse horticultural production.
However, to evaluate strawberry plant growth, the number of flowers and fruits are typically counted which is both labor and cost intensive.
We therefore developed a cost-effective phenotyping technique that can be used by ordinary farmers to evaluate plant growth.
The technique employs a deep learning based object detection model to count the number of flowers and fruits of the strawberry canopy.
We used a smartphone to capture color images and the YOLOv3 object detection algorithm to construct a flower and fruit detection model.
To consider the changes in the shape and color that flowers undergo as they mature into fruit, 400 images were annotated with four labels, i.e., flower, pre-immature fruit, post-immature fruit, and mature fruit, and these images were then used for training the model.
The developed model was tested using 20 images that were not included in the training data set.
The developed model successfully distinguished among the four labels with F-measures ranging from 0.69 to 0.83. Based on these initial results, measurements were conducted once every 1 to 2 weeks over a three-month period.
The number of mature fruit detected was compared with the number of harvested fruits.
The correlation coefficient was 0.95, indicating that the developed deep learning model was sufficiently accurate to evaluate strawberry growth.
Authors
S. Toda, T. Sakamoto, Y. Imai, R. Maruko, T. Kanoh, N. Fujiuchi, K. Takayama
Keywords
counting flowers, counting fruits, deep learning, greenhouse, plant diagnosis
Groups involved
Online Articles (117)