Applications of uncrewed aerial systems for rapid phenotyping using recurrent high-resolution imagery with currant (Ribes spp.) as a model berry shrub crop

Wolske, Eric Thomas

Permalink

https://hdl.handle.net/2142/116226 Copy

Description

Title

Applications of uncrewed aerial systems for rapid phenotyping using recurrent high-resolution imagery with currant (Ribes spp.) as a model berry shrub crop

Author(s)

Wolske, Eric Thomas

Issue Date

2022-07-15

Director of Research (if dissertation) or Advisor (if thesis)

Branham, Bruce

Doctoral Committee Chair(s)

• Juvik, John
• Leakey, Andrew
• Pritts, Marvin

Department of Study

Crop Sciences

Discipline

Crop Sciences

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• Ribes
• Currants
• Uncrewed Aerial Systems
• Remote Sensing
• Machine Learning
• Random Forest

Abstract

The development of uncrewed aerial systems (UAS) for remote sensing (RS) has advanced the phenotyping throughput of plants and enabled the rapid monitoring of crop traits. We explored UAS-RS for the measurement and monitoring of berry shrub crops, focusing on currants (Ribes spp.) as a model fruit crop for aerial surveys. Initially, we determined the optimal flight parameters to measure shrub crop morphology. We correlated destructively harvested black currant (Ribes nigrum) and aronia (Aronia melanocarpa) biomass during crop dormancy and active growth with UAS-derived parameters produced from red, green, blue (RGB) (up to r2=0.89) and multispectral (MS) cameras (up to r2=0.94). Growth of 24 cultivars of currants was monitored with UAS-RS for the 2020 and 2021 growing seasons using recurrent, low-altitude missions to collect RGB datasets. Random forest regression, a machine learning algorithm, predicted currant volume for each mission date (up to r2=0.83), while random forest classification provided accurate growth stage predictions throughout the growing season (misclassification = 7.5%). The importance of random forest regression for UAS-derived measurements of currants was shown during our final experiment to predict 2021 yield using the same currant cultivar trial. Random forest regression was compared with generalized linear regression with lasso regularization, gradient boosted trees, and k-nearest neighbors for yield prediction using RGB and/or MS datasets (up to r2=0.53). Hundreds of UAS-derived structural and spectral statistics for each plant were surveyed for importance in yield estimation and prediction. A detailed pipeline is presented for the acquisition and processing of UAS photographs into a workable format, the extraction of key features into functioning datasets, and the statistical and machine learning methods capable of accurate currant phenotyping.

Graduation Semester

2022-08

Type of Resource

Thesis

Copyright and License Information

Copyright 2022 Eric Wolske

Owning Collections