Hyperparameter Optimization of Plant Root Segmentation
Board Location: #102
Discipline: Computer Sciences and Information Management
Subcategory: Computer Science & Information Systems
Session: 2
Camille Catolos - Skyline College
Co-Author(s): Daniela Ushizima - Lawrence Berkeley National Laboratory, 1 Cyclotron Rd, Berkeley, CA 94720
Bioenergy collected from plants’ biomass is a renewable source that can reduce greenhouse gas emissions. Due to the root’s vital part in growth and sustainability, scientists study root biomass to optimize crop cultivation. Traditional manual root research has been laborious, yet automated root image analysis has the potential to streamline the process. Segmentation of root images is needed to estimate biomass, but it is challenging due to inherent thin structures and issues with image acquisition. Researchers often use image enhancement filters and machine learning for better results. However, hyperparameter tuning in machine learning models can affect segmentation performance. This study develops a pipeline for exploring the parameter space and grid search’s impact on different supervised classification models. Three models: Random Forest, Multiple-Layer Perceptron, and Histogram Gradient Boosting Classifier are trained and tested on millions of samples from the 1667 images of Triticum aestivum L roots. The experiment finds that grid search may not be the best technique to hyper-tune a large data set due to the technique’s brute force approach, resulting in long computational time unless the experiment downsamples the parameter space or size of training data via feature extraction or pre-processing. Grid search optimization is compared to random search. The study finds that random search can search a wider parameter space and produces competing results with grid search.
Funder Acknowledgement(s): This work was supported in part by the U.S. Department of Energy, Office of Science, Office of Workforce Development for Teachers and Scientists (WDTS) under the Community College Internship (CCI) program.
Faculty Advisor: Daniela Ushizima, dushizima@lbl.gov
Role: My responsibilities in the project was improving the plant root segmentation pipeline by tuning the parameters of machine learning classification models: random forest, histogram gradient boosting classification, and multi-layer perceptron neural network. Prior to utilizing the models, I had to pre-process and extract features from a Triticum aestivum L root data set of 1667 images. I then developed a pipeline which compared the performance of each classifier under different sets of parameters. The best parameters were found via two different methods: grid and random search. Finally, I compiled my results into a research paper and poster which were presented at Lawrence Berkeley's summer 2023 internship poster session.