Deep Learning for Automated Posture Annotation in Tall Grass Genotypes Using Multi-View UAV Imagery

Abstract

Accurate measurement of posture traits, such as curvature and droop, can provide critical insight into the structural phenotype and stress response of tall grasses, including key perennial bioenergy crops. These posture traits vary across genotypes and growth stages and can influence both productivity and harvestability. This study builds on previous efforts to quantify whole-plant posture, extending the analysis to incorporate deep learning and multi-view UAV imagery. High-resolution oblique, sideways, and nadir images were collected over experimental plots comprising multiple grass genotypes with diverse morphologies. A deep learning framework, incorporating instance segmentation models, was developed to classify plants based on posture metrics extracted from multiple viewing angles. By training on oblique and nadir views together, this work explores the degree to which plant posture can be inferred from each perspective, helping evaluate the contribution of each data acquisition approach and exploring the possibility of combined data acquisitions. Validation against manually annotated datasets evaluates the ability of the approach to differentiate posture traits at high resolution, enabling non-destructive, plant-specific phenotyping at scale. The work supports longitudinal monitoring of canopy structure and facilitates the integration of posture descriptors into breeding programs. It also provides a framework for future fusion with LiDAR and Structure-from-Motion (SfM) data to enhance 3D modeling of genotype-specific canopy architectures in switchgrass or other tall grasses.

Code and data availability

The supplied block contains only the title page and abstract; no public dataset, image, code, or model availability statement with an authors' URL is present.

No evidence-backed public reproduction asset is currently recorded.