European Journal of Sustainable Agroecosystems

Background: Drought stress represents one of the most critical threats to global cereal production and food security under climate change. While genomic selection and molecular breeding have advanced exponentially, the ability to accurately, non-destructively, and dynamically characterize target crop phenotypes under field conditions remains the primary limiting factor—the 'phenotypic bottleneck.' High-throughput phenotyping (HTP) has emerged as a disruptive technological paradigm to alleviate this constraint.

Objective: This critical review provides a comprehensive literature synthesis evaluating remote and proximal sensing technologies, deployment platforms, and computer vision pipelines engineered to decode drought-adaptive traits in major cereal crops (Triticum aestivum, Oryza sativa, Zea mays, Sorghum bicolor, and Hordeum vulgare).

Methodology: A systematic literature search was executed using Scopus, Web of Science, PubMed, ScienceDirect, and Google Scholar for studies published between 2016 and 2025. Following a rigorous PRISMA framework, inclusion criteria isolated peer-reviewed journal articles optimizing multi-sensor approaches, field/greenhouse robotics, and machine learning methods targeting morphological, physiological, and architectural adaptations to water deficits.

Major Findings: Evidence synthesis indicates that while multi-spectral and thermal sensors successfully capture canopy temperature depression and water index dynamics as proxies for stomatal conductance, they are prone to significant confounding effects caused by microclimate variations and soil background interference. High-resolution LiDAR and three-dimensional (3D) point-cloud voxelization offer unparalleled structural precision for detecting rolling leaves, canopy height reductions, and tillering dynamics, yet they suffer from processing bottlenecks and high computational costs. Furthermore, non-destructive root phenotyping via X-ray Computed Tomography (CT) and mini-rhizotrons reveals profound architectural plasticities, but translating these controlled-environment discoveries into open-field performance metrics remains highly problematic due to soil heterogeneity.

Conclusion: Critical analysis highlights three main knowledge gaps: (1) a distinct lack of standardized, cross-species phenotypic ontologies, (2) the 'black box' nature of deep learning pipelines that lack biological interpretability, and (3) limited multi-environment validation of digital traits across variable environmental conditions. Bridging these gaps requires a paradigm shift toward physics-informed machine learning and edge-computing sensor networks to accelerate the selection of climate-resilient ideotypes.