Plant Image Science (Plant Image Sci)
Si-Hong Kim1,†
, Yun-Hee Shin1,†, Soo Hyun Park1, Taek Sung Lee2, Hyoung Seok Kim1,*
1Smart Farm Research Center, Korea Institute of Science and Technology (KIST), Gangneung 25451, Republic of Korea
2Research and Development, PhytoWorks Inc., Gangneung 25451, Republic of Korea
†These authors contributed equally to this work.
Correspondence to Hyoung Seok Kim, Email: hkim58@kist.re.kr
Plant Image Sci. 1:3. https://doi.org/10.65971/PIS.2025.1.3
Received on December 08, 2025, Revised on December 16, 2025, Accepted on December 30, 2025, Published on December 31, 2025.
© Author(s). This is an Open Access article distributed under the terms of the Creative Commons CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Digital phenotyping has evolved from simple imaging-based trait measurement to a core technology enabling large-scale analysis of genotype × environment × management (G × E × M) interactions. This review highlights the manner in which digital phenotyping simultaneously supports two major research trajectories in modern agro-ecosystems. In engineering-driven approaches, imaging and sensor data are increasingly used to enhance the physical control of climate, irrigation, and crop protection through model-based and AI-assisted decision systems. In life science-centered approaches, high-throughput phenotyping and multi-omics integration provide mechanistic insights into plant stress responses, developmental plasticity, and complex trait regulation. Despite significant progress, both pathways face limitations in addressing the biological complexity and climatic unpredictability of crop systems. We discuss the emerging opportunities to integrate these domains through a two-layer AI framework that combines real-time sensing and actuation (“Physical AI”) with ontology- and LLM-based reasoning systems capable of synthesizing biological knowledge and generating strategic policies. Together, these developments position digital phenotyping as a bridging technology and a foundation for adaptive, resilient, and knowledge-driven agro-ecosystem management.
digital phenotyping, physical artificial intelligence, plant image science, ontology-based reasoning
Digital phenotyping has emerged as an imaging-based approach for quantifying plant traits and monitoring plant growth in breeding programs (Li et al. 2014). Advances in sensors, automation, and computation have expanded these methods to modern phenomics, supporting the large-scale analysis of genotype × environment × management (G × E × M) interactions (Fiorani and Schurr 2013; Yang et al. 2020).
Currently, digital phenotyping supports two major research trajectories. Engineering-driven approaches use imaging and sensing technologies to enhance precision crop protection, environmental control, and simulation-based cultivation strategies (Mahlein 2016). In parallel, life science-centered approaches use phenomics and multi-omics integration to understand stress responses and identify the genetic factors underlying complex traits (Araus and Cairns 2014).
Although both domains have advanced significantly, increasing climatic variability and narrowing crop genetic diversity have highlighted their limitations when applied independently (Yang et al. 2020). Importantly, both rely on digital phenotyping as a shared source of quantitative and scalable measurements for monitoring plant responses and validating biological hypotheses (Pieruschka and Schurr 2019).
This review summarizes global developments in engineering- and life science-oriented approaches, emphasizes digital phenotyping as a bridging technology, and discusses the pathways for integrating phenomic data and computational methods to manage complex agro-ecosystems.
Digital phenotyping is currently central to engineering-driven systems because it provides quantitative indicators of the canopy structure, stress responses, and water-use dynamics that inform climate and irrigation management (Li et al. 2014). As imaging quality improved, phenotyping enabled the earlier detection of stress and more plant-centered control strategies (Mahlein 2016; Pieruschka and Schurr 2019).
Recent developments have integrated phenotyping with environmental and soil data to support artificial intelligence (AI)-assisted forecasting of yield, evapotranspiration, irrigation needs, and microclimatic conditions (van Klompenburg et al. 2020). Imaging-based disease and nutrient diagnosis further enhance precision interventions. Digital phenotyping has also contributed to the development of simulation-based control and early digital-twin applications (Table 1).
Table 1. Representative digital phenotyping technologies used in engineering-driven physical control
| Digital phenotyping technology | Target crops | Physical control objective | References |
|---|---|---|---|
| RGB imaging | Tomato, cucumber, lettuce | Growth estimation, early stress detection, climate/irrigation adjustment | Li et al. (2014); Pieruschka and Schurr (2019) |
| Hyperspectral imaging | Wheat, barley, maize | Disease/nutrient diagnosis, precision spraying | Mahlein (2016) |
| Thermal imaging | Greenhouse tomato, grapevine | Water use monitoring, irrigation scheduling | Khan et al. (2019) |
| 3D / LiDAR | Tomato, pepper, fruit crops | Canopy modeling, robotics, forecasting | Paulus (2019) |
RGB, red-green-blue; 3D, three dimensions; LiDAR, light detection and ranging.
However, the physical control frameworks face persistent constraints. High installation and operational costs limit their adoption beyond high-tech systems (Wang et al. 2025). AI models require large, high-quality datasets and often struggle during rare or extreme weather events (Materia et al. 2024). Most engineering approaches remain environment-centric and do not explicitly account for plant physiological mechanisms or G × E × M interactions, thereby reducing their robustness under multi-stress scenarios (Yang et al. 2020). Growers also maintain manual overrides because of concerns regarding transparency and reliability.
Moreover, current physical-control architectures are optimized for short-term environmental responses and rarely incorporate biological objectives such as resilience or long-term trait expression. Digital phenotyping can help expand these frameworks by providing plant-level signals that link control actions to physiological outcomes; however, this potential remains underutilized. These limitations underscore the need to couple engineering systems with biologically informed reasoning-capable frameworks.
Life science-centered research aims to explain the manner in which genetic variations interact with the environment and management to produce phenotypes. Traditional multi-environment regional trials quantified G × E interactions for major traits such as yield and stress tolerance but provided limited temporal resolution (Hoyos-Villegas et al. 2025).
High-throughput phenotyping (HTP) has transformed this process by enabling dynamic, nondestructive measurement of morphological and physiological traits through RGB, hyperspectral, thermal, and 3D imaging (Li et al. 2021; Zhao et al. 2019). Fieldbased HTP using UAVs and ground systems further supports large-scale G × E assessments (Xu and Li 2022; Adak et al, 2024).
Multi-omics integration—transcriptomics, metabolomics, and proteomics—has added mechanistic insight into the regulatory networks underlying complex traits (Younas 2025). Recent efforts have incorporated environmental and management metadata to extend analyses to G × E × M frameworks (Großkinsky et al. 2018) (Table 2).
Table 2. Summary of phenomics approaches in life science-centered research
| Phenomics approach | Typical crops | Main objective | References |
|---|---|---|---|
| Controlled-environment HTP | Rice, wheat, barley, tomato | Growth dynamics, stress responses, genotype screening | Yang et al. (2020) |
| Field-based HTP (UAVs, ground robots, tractor-mounted sensors) | Cereals, legumes, maize | G × E trials, stress tolerance, yield prediction | Xu and Li (2022); Wang et al. (2025) |
| Multi-omics + phenomics integration | Mapping populations, breeding panels | Linking genes to traits, stress pathway mapping | Younas (2025) |
| Dynamic phenotyping (time-series phenotyping) | Model plants, cereals | Developmental dynamics, stress progression | Chang et al. (2021); Dhondt et al. (2014) |
HTP, high throughput system; UAV, unmanned aerial vehicle; G × E, genotype × environment.
However, despite these advances, several challenges remain unresolved. HTP platforms require significant infrastructure and generate datasets that require intensive preprocessing (Sheikh et al. 2024). Multi-omics data provide mechanistic depth but limited predictive power for polygenic and environmentally sensitive traits. Field phenotyping remains affected by environmental noise, complicating its integration with controlled-environment observations. Translating phenomics and multi-omics outputs into breeding or management decisions remains slow.
However, digital phenotyping, offers a common quantitative layer that can support cross-environment harmonization, enabling AI models to connect molecular-level signals with field-relevant phenotypes. As G × E × M structures become more complex under climate variability, integrating life science insights with sensing, modeling, and AI systems will be essential to achieve scalable and context-aware biological decision-making.
Physical AI systems—combining sensing, prediction, and actuation—enable the real-time monitoring and automated adjustment of climate, irrigation, and crop protection. However, they remain environment-centric and struggle to incorporate biological knowledge and long-term reasoning.
Ontology engineering and knowledge graphs offer structured representations of genes, traits, physiological processes, soils, and environments, thereby supporting a mechanistic understanding (Pech-May and Rios-Toledo 2023). Plant trait ontologies and phenotype knowledge bases facilitate standardized trait descriptions and improve the cross-study interpretation of genotypephenotype relationships (Dumschott et al. 2023).
Large language models (LLMs) further enhance reasoning by synthesizing heterogeneous biological and environmental information. In medicine and biology, they have been used to generate hypotheses and interpret multimodal evidences (Singhal et al. 2023). Coupled with retrieval-augmented generation, LLMs can translate biological objectives into decision rules or scenario sets. In robotics, LLMs have supported translating abstract goals into procedural control instructions (Zeng et al. 2023).
In agriculture, ontology–LLM frameworks can define long-term utility functions—resilience, yield stability, and resource efficiency—and translate them into rules that physical AI can operationalize. Digital phenotyping serves as the connective layer between these two systems, offering quantitative and multimodal signals that can ground LLM outputs and contextualize ontology-based reasoning. As these datasets expand, they may reduce hallucinations, strengthen ontology alignments, and improve biological validities. This two-way interaction enables closed-loop systems in which physical AI updates real-time states, and the ontologyLLM layer refines strategies and interprets emerging phenotypic patterns. Despite challenges—including incomplete ontologies and risks of hallucination—this integrated architecture represents a promising path toward biologically informed adaptive agricultural management.
Digital phenotyping has become a key technology linking engineering-driven physical control with life science-based genome-tophenome research. Its ability to quantify plant responses with high spatial and temporal resolutions strengthens real-time decision systems and mechanistic trait interpretations (Mahlein 2016; Pieruschka and Schurr 2019).
The integration of physical AI with ontology–LLM frameworks offers a promising path toward adaptive, biologically grounded decision-making. Physical AI provides real-time state estimation, whereas ontology–LLM systems can interpret G × E × M interactions and generate long-term strategies (Singhal et al. 2023; Zeng et al. 2023).
Future progress depends on advances in multimodal data integration, transparent AI models, and high-quality agricultural ontologies. Digital phenotyping will continue to serve as a measurement tool and as a bridging layer unifying empirical control and biological reasoning for resilient and sustainable agro-ecosystem management.
Conceptualization: Kim SH, Shin YH. Methodology and supervision: Kim HS; Investigation: Kim SH, Shin YH; Resources: Park SH, Lee TS; Writing – original draft: Kim SH, Shin YH, Park SH, Lee TS, Kim HS; Writing – review and editing: Kim HS; Project administration and funding acquisition: Kim HS.
The authors declare no conflicts of interest.
This research was funded by the Smart Farm Innovation Technology Development Program, funded by the Ministry of Agriculture, Food and Rural Affairs (MAFRA) and carried out with the support of the Korea Institute of Planning and Evaluation for Technology in Food, Agriculture and Forestry (iPET) (grant number RS-2025-02306502 to Kim HS).
The data are available from the corresponding author upon reasonable request.
Adak A, DeSalvio AJ, Arik MA, Murray SC. 2024. Field-based high-throughput phenotyping enhances phenomic and genomic predictions for grain yield and plant height across years in maize. G3 (Bethesda). 14(7):jkae092. https://doi.org/10.1093/g3journal/jkae092
Araus JL, Cairns JE. 2014. Field high-throughput phenotyping: the new crop breeding frontier. Trends Plant Sci. 19(1):52–61. https://doi.org/10.1016/j.tplants.2013.09.008
Chang S, Lee U, Hong MJ, Jo YD, Kim JB. 2021. Time-series growth prediction model based on U-net and machine learning in Arabidopsis. Front Plant Sci. 12:721512. https://doi.org/10.3389/fpls.2021.721512
Dhondt S et al. 2014. High‐resolution time‐resolved imaging of in vitro Arabidopsis rosette growth. Plant J. 80(1):172–184. https://doi.org/10.1111/tpj.12610
Dumschott K et al. 2023. Ontologies for increasing the FAIRness of plant research data. Front Plant Sci. 14:1279694. https://doi.org/10.3389/fpls.2023.1279694
Fiorani F, Schurr U. 2013. Future scenarios for plant phenotyping. Annu Rev Plant Biol. 64:267–291. https://doi.org/10.1146/annurev-arplant-050312-120137
Großkinsky DK, Syaifullah SJ, Roitsch T. 2018. Integration of multi-omics techniques and physiological phenotyping within a holistic phenomics approach to study senescence in model and crop plants. J Exp Bot. 69(4):825–844. https://doi.org/10.1093/jxb/erx333
Hoyos‐Villegas V et al. 2025. Affordable phenomics: expanding access to enhancing genetic gain in plant breeding. Plant Phenome J. 8(1):e70034. https://doi.org/10.1002/ppj2.70034
Khan S, Narvekar M, Hasan M, Charolia A, Khan A. 2019. Image processing based application of thermal imaging for monitoring stress detection in tomato plants. In: 2019 International Conference on Smart Systems and Inventive Technology (ICSSIT). IEEE. p 1111–1116. https://doi.org/10.1109/ICSSIT46314.2019.8987900
Li D et al. 2021. High-throughput plant phenotyping platforms (HT3P) as a novel tool for estimating agronomic traits from the lab to the field. Front Bioeng Biotechnol. 8:623705. https://doi.org/10.3389/fbioe.2020.623705
Li L, Zhang Q, Huang D. 2014. A review of imaging techniques for plant phenotyping. Sensors. 14(11):20078–20111. https://doi.org/10.3390/s141120078
Mahlein AK. 2016. Plant disease detection by imaging sensors – parallels and specific demands for precision agriculture and plant phenotyping. Plant Dis. 100(2):241–251. https://doi.org/10.1094/pdis-03-15-0340-fe
Materia S et al. 2024. Artificial intelligence for climate prediction of extremes: state of the art, challenges, and future perspectives. Wiley Interdiscip Rev Clim Chang. 15(6):e914. https://doi.org/10.1002/wcc.914
Paulus S. 2019. Measuring crops in 3D: using various sensing techniques for plant phenotyping. Plant Methods. 15:103. https://doi.org/10.1186/s13007-019-0490-0
Pech-May F, Rios-Toledo G. 2023. Development of agricultural knowledge graphs through the reuse of ontologies. Comput Sist. 27(4):1191–1201. https://doi.org/10.13053/cys-27-4-4789
Pieruschka R, Schurr U. 2019. Plant phenotyping: past, present, and future. Plant Phenomics. 2019:7507131. https://doi.org/10.34133/2019/7507131
Sheikh M et al. 2024. Integrating artificial intelligence and high-throughput phenotyping for crop improvement. J Integr Agric. 23(6):1787–1802. https://doi.org/10.1016/j.jia.2023.10.019
Singhal K et al. 2023. Large language models encode clinical knowledge. Nature. 620:172–180. https://doi.org/10.1038/s41586-023-06291-2
van Klompenburg T, Kassahun A, Catal C. 2020. Crop yield prediction using machine learning: a systematic literature review. Comput Electron Agric. 177:105709. https://doi.org/10.1016/j.compag.2020.105709
Wang RF, Qu HR, Su WH. 2025. From sensors to insights: Technological trends in image-based high-throughput plant phenotyping. Smart Agric Technol. 12:101257. https://doi.org/10.1016/j.atech.2025.101257
Wang Y et al. 2025. A comprehensive review of proximal spectral sensing devices and diagnostic equipment for field crop growth monitoring. Precision Agric. 26:54. https://doi.org/10.1007/s11119-025-10251-3
Xu R, Li C. 2022. A review of high-throughput field phenotyping systems: focusing on ground robots. Plant Phenomics. 2022:9760269. https://doi.org/10.34133/2022/9760269
Yang W et al. 2020. Crop phenomics and high-throughput phenotyping: past decades, current challenges, and future perspectives. Mol Plant. 13(2):187–214. https://doi.org/10.1016/j.molp.2020.01.008
Younas MU et al. 2025. Multi-omics approaches in plant biology: decoding agronomic traits for sustainable agriculture. Plant Stress. 18:101118. https://doi.org/10.1016/j.stress.2025.101118
Zeng F et al. 2023. Large language models for robotics: a survey [preprint]. arXiv:2311.07226. https://doi.org/10.48550/arXiv.2311.07226
Zhao C et al. 2019. Crop phenomics: current status and perspectives. Front Plant Sci. 10:714. https://doi.org/10.3389/fpls.2019.00714