Detecting Sorghum Plant and Head Features from Multispectral UAV Imagery

Area of Science:

• Plant breeding
• Agricultural science
• Remote sensing

Background:

• Unmanned aerial vehicles (UAVs) with multispectral cameras are valuable for high-throughput phenotyping (HTP) in plant breeding.
• Stitching multiple images from UAV missions reduces resolution and quality for large-area analysis.
• High-quality reflectance data from single nadir images can overcome these limitations.

Purpose of the Study:

• To develop and validate a pipeline for deriving high-resolution reflectance data from raw multispectral UAV imagery.
• To apply this data for accurate phenotyping, specifically estimating plant and head counts in sorghum breeding plots.
• To assess the discriminative ability of spectral bands for plant traits.

Main Methods:

• A pipeline involving imagery calibration, spectral band alignment, backward calculation, and plot segmentation was developed.
• The pipeline was optimized to estimate the number of plants and count sorghum heads per plot.
• Reflectance data from single nadir images was utilized for phenotyping.

Main Results:

• High coefficients of determination (0.90 for plants, 0.86 for heads) were achieved for estimates using derived nadir images.
• Spectral bands demonstrated significant discriminative ability for sorghum head colors (red and white).
• Accurate segmentation of crop organs at the canopy level was achieved across diverse plots with minimal machine learning training.

Conclusions:

• The developed pipeline effectively derives high-resolution reflectance data from multispectral UAV imagery.
• This approach enhances accuracy in estimating plant and head counts for sorghum breeding.
• The method offers a robust solution for crop phenotyping, applicable across various field plots.