ASPEN study case: real time in situ tomato detection and localization for yield estimation

Abstract

As human population continue to increase, our food production system is challenged. With tomatoes as the main indoor produced fruit, the selection of adapter varieties to each specific condition and higher yields is an imperative task if we wish to supply the growing demand of coming years. To help farmers and researchers in the task of phenotyping, we here present a study case of the A gro s cope ph en otyping tool (ASPEN) in tomato under indoor conditions. We prove that using the ASPEN pipeline it is possible to obtain real time in situ yield estimation not only in a commercial-like greenhouse level but also within growing line. To discuss our results, we analyse the two main steps of the pipeline in a desktop computer: object detection and tracking, and yield prediction. Thanks to the use of YOLOv5, we reach a mean average precision for all categories of 0.85 at interception over union 0.5 with an inference time of 8 ms, who together with the best multiple object tracking (MOT) tested allows to reach a 0.97 correlation value compared with the real harvest number of tomatoes and a 0.91 correlation when considering yield thanks to the usage of a SLAM algorithm. Moreover, the ASPEN pipeline demonstrated to predict also the sub following harvests. Confidently, our results demonstrate in situ size and quality estimation per fruit, which could be beneficial for multiple users. To increase accessibility and usage of new technologies, we make publicly available the required hardware material and software to reproduce this pipeline, which include a dataset of more than 850 relabelled images for the task of tomato object detection and the trained YOLOv5 model[1] [1]https://github.com/camilochiang/aspen

Plant phenotyping relevance

ASPENはトマト果実の検出・追跡から収量、果実サイズ、品質を推定する画像ベースの表現型解析パイプラインであり、手法の評価と再現可能なソフトウェア・データセット提供が中心です。

Code and data availability

The authors explicitly make publicly available the ASPEN pipeline software/hardware materials, a dataset of 850+ relabelled tomato images, and the trained YOLOv5 model via their GitHub repository.