Infrared Thermography in Plant Factories: Solving Spatiotemporal Variations Via Machine Learning

ABSTRACT Infrared thermography (IRT) for real‐time stress detection in plant factories (PFs) remains largely unexplored. Hence, this study investigates the feasibility of implementing IRT in PFs, using machine learning (ML) to address the challenges in information processing. Herein, purple basil plantlets were subjected to root dehydration within a pilot‐scale PF, and canopy temperature was monitored at regular intervals using a thermal camera. Subsequently, eight ML models using the ‘support vector machines’ algorithm were tested for stress detection. Our findings revealed that differences in canopy temperature due to microenvironmental variations led to inaccurate representation of stress. Nonetheless, binary classification models trained using plants at medial and high stress overcame this issue by identifying stressed samples with 81%–94% accuracy. However, although models trained with medially stressed samples performed well for all stress levels, models trained using highly stressed samples failed to identify medial stress reliably. Additionally, ternary and quaternary classification models were able to identify unstressed samples but could not distinguish between different levels of stress. Hence, binary classification models trained using medially stressed samples overcame spatiotemporal variations in canopy thermal profile most effectively and provided probabilistic estimates of plant stress within the PF most consistently.