In Silico Root Trait Phenotyping for Maize in C-K Deficit and Toxic Conditions using Convolutional Neural Networks

The study of root system architecture enables botanists and cultivators to understand the growth patterns of roots, which are dynamically affected by factors such as soil nutrient content. The analysis and optimization of root networks is a topic of interest to improve crop health and yields. The usage of machine learning techniques for the automation of root system analysis is currently not a mainstream practice. The presence of organic carbon and potassium in soil contributes to the growth, vitality, and water retention of plants. Sufficient amounts of these nutrients would then aid the productivity of maize crops. This study presents convolutional neural network models (CNN) that are developed as a means of extracting traits and information from synthetic root images. The trained CNN models were able to predict morphological characteristics of length, surface area, volume, and weight; metabolic traits of root respiration and exudate cost; and classify the amount of organic carbon and potassium present in soil. The six regression models achieved R
²
values of 0.922 to 0.953, while the two classification models achieved 100% classification accuracy. By compiling the results of the models together, it can be identified that potassium content has a more profound effect on maize root system architectures. In addition, these models were able to predict batches of images within 2.8 to 4.5 seconds. With these results, it can be concluded that properly trained CNN models can provide major benefits to the field of plant root studies.