Autonomous Phenotyping for Vineyard Disease Monitoring: Agreement and Bias Relative to Human Ratings

Cole Regnier¹, Yiyuan Lin², David Combs¹, Yu Jiang¹, Kaitlin M Gold¹, Rye Weber¹

¹ School of Integrative Plant Science, Cornell University, Geneva, NY 14456, USA
² School of Electrical and Computer Engineering, College of Engineering, Cornell University, Ithaca, NY 14850, USA

Abstract

Vineyard disease severity assessments are often labor intensive and prone to rater bias. The PhytoPatholoBotv2 (PPBv2) is an autonomous, high-throughput phenotyping robot equipped with dual high-resolution stereo cameras for objective disease monitoring at scale. This study evaluated systematic bias and agreement between PPBv2 measurements and human ratings for grapevine powdery mildew (GPM) and downy mildew (GDM). Georeferenced PPBv2 imagery was collected at four timepoints on a panel basis (3–4 vines/panel) in a Vitis vinifera (cv. Chardonnay) vineyard with two randomized complete block design trials evaluating 36 (GPM) and 24 (GDM) treatments (4 panels/treatment). Human raters visually assessed disease severity, incidence and calculated percent damage (incidence ✕ severity) on 20 leaves per panel. PPBv2 imagery was processed using a custom multimodal segmentation model to estimate percent infected canopy area. AUDPC was calculated for both methods, and treatment efficacy rankings were compared using Spearman and Kendall correlations (n = 36, n = 24). For GPM, Treatment efficacy rankings were strongly correlated (Spearman = 0.915; Kendall τ = 0.752). Agreement at the panel-level observations was assessed using Bland-Altman analysis (n = 570). Heteroscedasticity and proportional bias were observed, with divergence increasing at higher disease severity. Mean bias was 10.99% with limits of agreement of −12.81 and 34.79%. GDM analyses are ongoing. These results indicate PPBv2 can approximate treatment efficacy rankings while enabling rapid, large-scale disease monitoring.

Citation

@misc{Regnier2026AutonomousPhenotyping,
  author       = {Regnier, Cole and Lin, Yiyuan and Combs, David and Jiang, Yu and Gold, Kaitlin M. and Weber, Rye},
  title        = {Autonomous Phenotyping for Vineyard Disease Monitoring: Agreement and Bias Relative to Human Ratings},
  year         = {2026},
  month        = aug,
  howpublished = {Poster abstract P-254, Plant Health 2026},
  note         = {American Phytopathological Society, Rhode Island Convention Center, Providence, RI, USA. Meeting held August 1--4, 2026; poster presented August 4, 2026},
  url          = {https://apsnet.confex.com/apsnet/2026/meetingapp.cgi/Paper/29336}
}