High-throughput phenotyping using aerial images for the prediction of agronomic traits in soybean

Previous studies on soybean (Glycine max L. Merr.) that apply high-throughput phenotyping (HTP) to predict agronomic traits such as date of physiological maturity, pubescence color, lodging, plant biomass, and grain yield have been carried out using artificial intelligence and aerial images taken from drones. Hand-held devices recording individual indices (e.g., NDVI, normalized difference vegetation index), multispectral radiometers, and spectroradiometers recording the continuous range of the electromagnetic light spectrum have also been used for predictive studies. More recently, hyperspectral cameras mounted in drones have also been developed, which take advantage of quickly recording a huge amount of spectral data. Nonetheless, several of these studies have used only hundreds of data and validated the models in data subsets of the same environments from where they were trained. In turn, beyond research studies, breeding programs have only recently started using HTP platforms for predicting agronomic traits on a large scale using spectral data and artificial intelligence methods. Soybean breeding programs need to record annually phenotypic traits for thousands of experimental lines grown in plant rows to select those that should be evaluated in preliminary yield tests. This is a time-consuming task considering that, in the end, much of the data collected is not used because the vast majority of the plant rows will be discarded due to their low grain yield. Using the Random forest machine learning algorithm and time series of RGB and multispectral images taken from a drone, this work aimed to study in three breeding experiments of plant rows how generalized prediction models can be trained to 1) predict the date to reach the physiological maturity (R8 stage), and 2) classify pubescence color (gray, light tawny, and tawny). This was done by studying how the image features, the number and time between flights, and germplasm relationships, impact the predictions and classifications, respectively. For predicting the R8 stage, it was found that compared to the full set of 8–10 flights (R2= 0.91–0.94; RMSE= 1.8–1.3 days), using data from three to five fights before harvest had almost no effect on the prediction error (RMSE increase ~0.1 days). While for classifying pubescence color, training the model with a time series of images (four drone flights before or at maturity), a higher overall accuracy was achieved than with a single flight at maturity, when the leaves fall to the ground and pubescence color is better observed. The overall accuracy was 86.55% using the four drone flights (Kappa= 0.7976), and the precision for gray, light tawny, and tawny pubescence were 0.893, 0.788, and 0.915, respectively. For predicting the R8 stage, similar prediction accuracy was achieved using either multispectral or an affordable RGB camera, and the excess green index (ExG) was found to be the important feature in making predictions. While for classifying pubescence color, even though the red/blue index was the most important image feature according to the scores of the variable importance analysis, the use of the multispectral indices blue NDVI and green NDVI was also helpful, mainly to increase the precision of classifying gray and light tawny pubescence. Using a model trained with data from two previous years and using fielding notes from check cultivars planted in the test season, the R8 stage was predicted in 2020 with an error of 2.1 days. In the case of pubescence color, the model tested in this same independent environment (2020) had a meaningful decrease in the overall accuracy, 65.86%, compared to the validation results shown above. Still, the tested model obtained a fair to good model reliability (Kappa= 0.4874), which could be improved by phenotyping pod color since it interferes with the pubescence color in the background. Applying a HTP pipeline like the one used in this study would help save time or allow evaluating a higher number of plots using similar resources. In both cases, this means that the efficiency of the soybean breeding programs could be improved when characterizing thousands of plant rows each season.

Name thesis:degree_name

Ph.D.

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Crop Sciences

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2024

Author dc:creator

• Perez Gonzalez, Osvaldo M.

Contributors dc:contributor

• Diers, Brian W
• Bohn, Martin O
• Martin, Nicolas F
• Arbelaez, Juan D

• High-throughput Phenotyping
• Machine Learning
• Random Forest
• Principal Component Analysis
• Canonical Discriminant Analysis
• Physiological Maturity
• Pubescence Color

Statement dc:rights

• Copyright 2024 Osvaldo Perez

Language dc:language

en, eng

Handle dc:identifier

https://hdl.handle.net/2142/125674

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F125674