Relationship between the tomato plants’ population size for imaging and the calculated fruit density by AI-based object detection

The speaking plant approach (SPA) is a paradigm that the crop cultivation conditions should be optimized based on the physiological states of the crop.
This has been regarded as an indispensable key concept for the environmental control of current greenhouses.
A hanging-type imaging robot developed in our previous study enabled us to automatically capture color images of dozens of meter-wide tomato plants in a commercial greenhouse.
In this study, we applied robot imaging to cherry tomato plants of 50-m width, comprising more than hundred plants, on a single cultivation gutter.
The obtained color images were analyzed using the deep learning-based object detection algorithm YOLOv3 to count the number of tomato fruits.
The average tomato fruit density (total fruit number detected with YOLOv3 divided by the width of the target population) of a 50 m-width tomato plants’ population (AFD-50 m) was 37.8±18.8% (C.V.). On the other hand, the average tomato fruit density of 30 m wide (AFD-30 m) varied from 37.0 to 40.3 fruits m^‑1; AFD-20 m varied from 35.2 to 41.4 fruits m^‑1; AFD-10 m varied from 31.1 to 41.6 fruits m^‑1. This result indicates that smaller tomato plants’ population size for imaging provides greater variations of the average tomato fruit density.
This was caused by the heterogeneous distribution of tomato fruits along the tomato plants on a single cultivation gutter.
In this case, a measurement of 30 m wide was considered sufficient for large-scale monitoring of the average tomato fruit density.

International Symposium on New Technologies for Sustainable Greenhouse Systems: GreenSys2023

K. Tauchi, N. Fujiuchi, T. Kanoh, S. Toda, K. Takayama

chlorophyll fluorescence imaging, greenhouse, plant diagnosis, robot, speaking plant approach, stem elongation

https://doi.org/10.17660/ActaHortic.2025.1426.50