Machine learning pipeline for 3D maize plant architecture analysis using unmanned aerial systems

Overview

This solution delivers a machine learning pipeline designed to transform 2D imagery captured by unmanned aerial systems (UAS) into detailed 3D point clouds of maize plant architecture. By analyzing plant structure in three dimensions, the technology supports precision agriculture decision-making related to crop productivity, fertilizer application, and pest control.

The pipeline addresses six critical maize phenotyping tasks: individual maize extraction, classification of plant organs into stem, leaf, and kernel, leaf segmentation and labeling, early tassel detection, analysis of physiological growth dynamics, and computation of canopy volume. These capabilities enable growers and agronomists to move beyond flat, 2D assessments and leverage full structural information for crop management.

Technical specifications

Core approach:

• Machine learning algorithms process 3D point clouds reconstructed from multi-viewpoint UAS imagery
• RGB imagery serves as the primary data source, making the approach cost-effective and scalable
• Regression analysis correlates imagery features with field-measured traits such as above-ground biomass and yield
• Cross-validation at the plot level ensures robust model performance

Validation metrics used:

• Accuracy
• Coefficient of determination (R2)
• p-value
• Root mean square error (RMSE)
• Average systematic error
• Mean percent standard error
• Nash and Sutcliffe index

Key phenotyping outputs:

• Individual maize plant extraction from field imagery
• Organ-level classification (stem, leaf, kernel)
• Leaf segmentation with individual labeling
• Early-stage tassel detection
• Growth dynamics analysis over the season
• Quantitative canopy volume measurements

Prior performance benchmarks:

• Over 90% accuracy achieved in soybean yield prediction using multi-sensor UAS imagery fusion
• Canopy roughness trait derived from UAS-based 3D data linked to biomass with R2 greater than 0.5

Technology readiness level

The underlying methodology has been validated in soybean trials conducted across 5 hectares at the Agronomy Center for Research and Education (ACRE) in West Lafayette, Indiana, during the 2018 growing season. Ground measurements including above-ground biomass and yield were collected alongside biweekly UAS flights at multiple growth stages.

The maize-specific pipeline is currently at an earlier stage of development. Future validation will involve a low-population trial planted at ACRE during the 2021 season, with digital outcomes tested against field measurements. The study is expected to take approximately one year, including new dataset generation and codification of the automatic pipelines. Industry partnership funding is being sought to support key personnel and material costs during this validation phase.