Machine learning and High-Resolution Satellite Imagery for Red Crown Rot Detection in Illinois Soybean

Monday, August 4, 2025

9:45 AM - 10:00 AM HST

Location: Room 312

Presenting Author(s)

BP

Bruno D. Pugliese

University of Illinois Urbana-Champaign
Urbana, Illinois, United States

Author(s)

JP

Juan A. Paredes, PhD

Researcher
University of Illinois Urbana-Champaign
Urbana, Illinois, United States
NM

Nicolas Martin

Associate Professor
University of Illinois Urbana-Champaign
Urbana, Illinois, United States
DB

Dennis Bowman

University of Illinois Urbana-Champaign
Urbana, Illinois, United States
BC

Boris X. Camiletti

Assistant Professor
University of Illinois Urbana-Champaign
Urbana, Illinois, United States

Abstract Text: Soybean plants severely affected by Red Crown Rot (RCR) experience premature death, creating non-uniform patches. This symptom expression may facilitate disease detection using remote sensing. The aim of this study was to develop a random forest model to detect RCR-affected areas in soybean fields using high-resolution satellite imagery. Ground truth data (GTD) were collected in 2024 from three soybean fields in Illinois at the R6 stage. Fifty random points per field were classified as diseased (≥5% symptomatic plants within a 3m² area) or healthy. PlanetScope satellite imagery (3m resolution) was acquired for each field, temporally aligned with GTD assessments. Using Google Earth Engine, a random forest classifier was trained with spectral features, including red, green, blue, near-infrared (NIR), and red-edge reflectance, as well as the Normalized Difference Vegetation Index, Green Normalized Difference Vegetation Index, and Normalized Difference Red Edge Index. The model was trained using 70% of the GTD, and its performance was evaluated using a confusion matrix with the remaining 30% and visual comparisons with drone imagery. The model achieved an overall accuracy of 75%, and visual analysis revealed that the classified affected areas matched those observed in drone imagery. This study confirms the potential of using satellite imagery to detect RCR-affected areas across fields, enabling the identification of 'hot spots' for targeted management practices. However, further research involving more fields, seasons, and regions is necessary to enhance model consistency.