Forecasting growth dynamics of hydroponic kale in controlled environment agriculture through vision-based phenotyping and time-series modeling

Abstract

In controlled environment agriculture (CEA), accurate yield forecasting remains challenging due to reliance on environmental sensor data, which fails to capture plants’ dynamic morphological responses to growth conditions. This study bridges the gap by establishing a vision-based framework to forecast plant growth dynamics over prediction windows of 2, 4, and 8 days using automated phenotyping and time-series modelling. A plant phenotype monitoring framework was implemented using commercially available cameras and off-the-shelf deep learning-based models (YOLO). The robustness of the YOLO and time-series models was rigorously evaluated under a range of treatment conditions, including a control, salt stress levels at 3, 6, and 9 ppt, and different root architectures (single-root and split-root) in hydroponic greenhouse trials conducted over two growing seasons. Top-view images of the plants were collected using GoPro and Raspberry Pi cameras, and different YOLOv8 instance segmentation model variants were trained on four image datasets to extraction of morphological traits such as area, major, and minor axes. Results indicated that YOLOv8 generalized well, achieving mAP50 for bounding boxes and masks in the range of 0.897 – 0.952 and 0.896 – 0.947, respectively. Model-derived morphological parameters effectively captured growth differences across salt levels and root architectures, with split-root plants showed resiliency under salt stress compared to single-root. Comparisons between physical measurements and image-derived parameters such as major and minor axes yielded high R² values of 0.85 and 0.92 for single-root systems, and 0.90 and 0.84 for split root systems. Additionally, the area parameter obtained from images showed an R² of 0.882 when compared with plant fresh weight. ARIMA model used to forecast the plant area parameters over 2-, 4-, and 8-days windows and evaluated using MAPE. Notably, the 2-day forecasts for single-root plants under 9 ppt salt stress yielded the lowest MAPE values (3.99 in the fall and 1.70 in the spring), although 8-day forecasts at higher salt concentrations exhibited generally larger errors. For split-root plants, the 4 days forecast under 3 ppt salt stress produced a MAPE of 7.13 in the fall, while in the spring, the 8 days forecast at 9 ppt achieved a MAPE of 2.08. The forecasted area values demonstrated R² values of 0.623, 0.671, and 0.75 for the 2-, 4-, and 8-day forecast windows respectively when compared with fresh weight, indicating that the area parameter is a reliable predictor of yield. These findings confirm that morphological changes capture environmental influences and can be reliably forecasted, introducing a scalable, data-driven method to predict yield in CEA while helping growers optimize resource usage and reduce productivity risks.

Plant phenotyping relevance

画像ベースの植物表現型取得と時系列予測を中心に、形態形質の抽出、モデル性能評価、物理測定との検証を行っているため、方法論文として適格。

Code and data availability

公開論文であることは確認できましたが、現在の公式API・許可済み取得経路では本文を自動取得できませんでした。

No evidence-backed public reproduction asset is currently recorded.