Infrared thermography for plant stress detection in vertical farms: Investigating spatiotemporal variations and exploring solutions via machine learning

Abstract

Application of infrared thermography (IRT) for real-time plant stress detection has grown rapidly in recent years. Although the technology has been well established for crops grown in fields and glasshouses, its feasibility for vertical farms has not been tested extensively. In this study, temporal monitoring of stress induced by root dehydration in purple basil plantlets inside a vertical farm was performed to identify bottlenecks in real-time stress detection via IRT. Subsequently, potential solutions were investigated via machine learning by implementing support vector machines for supervised classification. Edge effects as well as proximity to air vents were identified as the major causes of positional variation in plant temperature that could lead to misprediction of stress. Binary, ternary, and quaternary classification models were trained using thermal images from two, three, and four levels of stress, respectively, to assess model performance. Binary classification models trained with plants experiencing medial and high levels of stress were able to identify stressed plants with high accuracy (81–94%). Further, binary models trained using plants under medial levels of stress generated a continuous probability distribution for stress prediction when plotted against plant temperature. In contrast, models trained using samples experiencing high stress generated distinct probabilistic clusters for the unstressed and highly stressed plants, but were unable to classify medial stress samples reliably. Similarly, ternary and quaternary models were able to better predict very high and very low levels of stress than intermediate stress levels. Hence, our findings suggest that binary classification models trained using samples under medial levels of stress would be helpful in overcoming spatiotemporal variations in canopy thermal profile by providing reliable probabilistic estimates of plant stress within a vertical farming system. Key points Plant stress detection in vertical farms via thermal imaging may be challenging because perceptible plant temperature can be strongly influenced by its microenvironment. Thermal image analysis via supervised machine learning allows the development of robust prediction models that can overcome such factors to identify stressed plants. Binary classification machine learning models can reliably identify stressed plants as well as provide probabilistic estimates for the degree of stress.

Plant phenotyping relevance

赤外線熱画像による植物ストレス状態の検出と、機械学習による分類・確率推定が研究の中心であり、垂直農場での技術的課題と性能を評価している。

Code and data availability

The paper describes thermal imaging of purple basil and an SVM-based ML pipeline, but no public deposit of the thermal images, phenotype data, or author analysis code is mentioned. The only URLs given (flirextractor on PyPI, scipy.stats docs) are generic third-party libraries, not paper-specific assets.

No evidence-backed public reproduction asset is currently recorded.