Automated hyperspectral data collection system for monitoring greenhouse plant health

Authors: SI, GAOSHOUTONG - The Ohio State University; LING, PETER - The Ohio State University; Testen, Anna; Zhu, Heping; Jeon, Hongyoung

Submitted to: Smart Agricultural Technology
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 6/17/2026
Publication Date: 6/17/2026
Citation: Si, G., Ling, P., Testen, A.L., Zhu, H., Jeon, H. 2026. Automated hyperspectral data collection system for monitoring greenhouse plant health. Smart Agricultural Technology. 14. Article 102331. https://doi.org/10.1016/j.atech.2026.102331.
DOI: https://doi.org/10.1016/j.atech.2026.102331

Interpretive Summary: Hyperspectral sensing (HS) could provide unique capabilities for detecting subtle changes in plant health. But applying this technology inside commercial greenhouses has been challenging due to constantly changing light conditions, irregular plant canopies, and labor-intensive measurements. This research developed a mobile, canopy scale HS platform designed specifically to overcome these limitations in production greenhouse environments. A watering boom in a greenhouse equipped with a position system carried the system which performed automatic measurement by adjusting sensing distance, monitoring real time sunlight illumination, calibrating the hyperspectral sensor, and collecting spectral reflectance measurements for each plant sample. Five experiments were conducted in a commercial-style greenhouse using tomato plants inoculated with and without bacterial spot as a case study to evaluate the HS system performance under realistic operating conditions. The study identified several environmental factors that influence HS data quality such as seasonal spectral distribution variations, high-humidity-driven noise in the near infrared range, and plant stress due to short term environmental changes. A targeted preprocessing workflow was introduced to overcome the environmentally induced variations in environmental factors. This improved the quality of reflectance measurements, reduced variability within healthy and diseased groups, and enhanced their separability. The data quality improvement increased classification accuracy with linear support vector machine and logistic regression by 5.7% and 7.7%, respectively. This work could help U.S. greenhouse farmers detect disease early and reliably, reducing chemical use, improving crop quality, and increasing labor efficiency. In addition, this work could be a cornerstone of precision integrated pest management for greenhouse production.

Technical Abstract: Hyperspectral sensing (HS) offers substantial potential for monitoring plant health. However, its deployment in greenhouses remains challenging due to fluctuating illumination, varied canopy structure, and the lack of automated data-collection systems. In this research, a mobile canopy-scale HS platform was developed that integrated a mobile boom-based positioning unit, depth-guided sensing distance control, real-time sunlight monitoring unit, and automated calibration for individual samples. The platform was tested in a production-style greenhouse through five independent experiments using tomato bacterial spot as a case study, ensuring reliable and repeatable canopy-level spectral acquisition under normal greenhouse operation conditions. The case study revealed several factors that need to be addressed in assuring HS data quality collected in a greenhouse environment. Seasonal climate variation across experiments resulted in measurable differences in spectra distributions while no significant diurnal differences were observed between morning and afternoon measurements in the visible range. Increased air moisture during rainy conditions introduced higher spectral noise in the near-infrared region, and short-term transportation transient environment influenced spectral characteristics of plants. To mitigate the environmentally induced variabilities, a preprocessing workflow was implemented to remove spectra influenced by transient lighting changes or inconsistent sensor-to-canopy distance, resulting in high-quality reflectance measurements. After the preprocessing, both healthy and diseased groups showed narrower standard deviations across wavelengths, and separation between groups improved in several cases. t-test results also confirmed that these changes were statistically significant rather than random. Additionally, classification performance improved despite fewer training samples (n = 1235 vs. 1773). The accuracy of linear support vector machine (SVM) and logistic regression increased by 5.7% and 7.7% with preprocessed data respectively. Overall, the automated data collection combined with the environmental-aware preprocess enhanced HS data quality and provided practical guidance for reliable plant health monitoring assistance in dynamic greenhouse environments.