U-Net segmentation of 3D micro-CT images of rooted soils using label data from multi-physics simulators
The goal of this project was to develop a machine learning model to segment micro-CT images of vegetated soil. This tool aimed to automate the segmentation process, namely to assign soil, air, water, and plant root label to individual pixels of 3D images. By providing a more efficient means of analyzing soil-vegetation interactions, this project sought to make a significant impact in the field by enhancing the use of x-ray tomography images to quantify the effects of vegetation on soil properties.
The impact of this project lies in its ability to introduce a machine learning model based on an existing image analysis pipeline developed by the LA. The anticipated tool should have the potential to change the way researchers study the microscale mechanics of reinforced soil by providing an automated, data-driven method for segmenting complex datasets. For us, this project was particularly important because it builds on our previous work with x-ray images of vegetated soil. Having created a not-automated pipeline that identifies each label within an image, we understood the challenges researchers face when dealing with such complex data. A machine learning tool that automates this process can be a significant aid, especially for researchers who lack extensive image processing knowledge handling complex multiphase images, containing root system, water content variation, and soil matrix.
Moving forward, our plan is to continue developing the tool to reduce the segmentation error and enhance its accuracy. We aim to reach a point where the tool can consistently and reliably produce segmented 3D images that meet the needs of soil mechanics research. As with any scientific endeavor, we are committed to making the final tool open-source and accessible to everyone. We invite feedback and collaboration from the scientific community to help refine and improve the tool, ultimately contributing to better research outcomes. Further information about the tool will be provided upon its release, and we will encourage interested researchers to explore its potential applications in their work.
An artificial brain for interpreting and accelerating physics-based simulations of granular materials
Soil health is essential for infrastructure and plant growth. The UNSAT project aims to improve 3D soil image processing using machine learning and high-fidelity simulation data to study water transport and root-soil interactions during maize growth.
This repository provides a workflow to transform X-rays data originally stored in .TIFF format, into .h5. Overall, this makes the data more compressed and easier to process in machine learning pipelines, such as unsat.