What’s Healthy Now and What’s Next? Integrating Drone-Based Canopy Health Mapping and Machine Learning for Risk Prediction in Strawberries

Monday, August 4, 2025

10:15 AM - 10:30 AM HST

Location: Room 312

Presenting Author(s)

JD

Jon R. Detka, PhD

California State University Monterey Bay
Marina, CA, USA

Author(s)

FM

Forrest S. Melton

Senior Research Scientist
California State University, Monterey Bay, CA; NASA Ames Research Center, Moffett Field, CA
Monterey, California, United States
AP

AJ Purdy

Senior Research Scientist
California State University, Monterey Bay, CA ; NASA Ames Research Center, Moffett Field, CA
Monterey, California, United States
OD

Oleg Daugovish

Strawberry and Vegetable Crop Advisor
University of California Cooperative Extension
Ventura, California, United States
CG

Chris Greer

Area Cooperative Extension Advisor - Integrated Pest Management
UC Cooperative Extension San Luis Obispo County
San Luis Obispo, California, United States
CC

Clarence Codod

Postdoctoral Scholar
Department of Plant Sciences, University of California, Davis
Davis, California, United States
FM

Frank M. Martin, Research Plant Pathologist

Research Plant Pathologist
U.S. Department of Agriculture, Agricultural Research Services
Salinas, California, United States

Abstract Text: Timely detection of plant health decline is critical for strawberry production. Early pest management interventions can minimize crop losses, yet traditional scouting can miss emerging spatiotemporal patterns. This study integrates drone-based canopy health mapping with machine learning risk prediction to support scouting efforts and inform management approaches. Using weekly multispectral drone imagery in a commercial field in Oxnard, California we identified locations of canopy stress via spatial clustering analysis of vegetation indices and canopy area changes. Hotspot analysis (Getis-Ord Gi*) provided grower-friendly risk maps, enabling targeted scouting. To enhance detection, we trained XGBoost Cox models to predict plant health decline and map high-risk zones based on the top 25% of risk scores. Model performance was evaluated with expanding window cross-validation. Models performed best during peak season (Concordance Index (CI) = 0.92, in April; 0.87 in early June), accurately ranking plant risk. However, CI declined by late June (0.52) due to rapid widespread canopy loss reducing predictive variation. Discrepancies between training and test CI (e.g., June: Train CI = 0.81, Test CI = 0.65) suggest increased overfitting as disease prevalence intensified. Earlier season models successfully forecasted high-risk zones that later experienced severe canopy loss due to soilborne diseases. Combining hotspot mapping for scouting with survival modeling for long-term risk prediction ensures practical, interpretable, and actionable precision agriculture solutions.