Presenters-0555
Abstract
Segmentation, or the classification of pixels (grid cells) in imagery, is ubiquitously applied in the natural sciences. An example close to the CSDMS community might be translating images of earth surface into arrays of land cover to be used as model initial conditions, or to test model output. Manual segmentation is often prohibitively time-consuming, especially when images have significant spatial heterogeneity of colors or textures.
This Clinic is focused on demonstrating a machine learning method for image segmentation using two software tools: The first is “Doodler”, a fast, semi-automated, method for interactive segmentation of N-dimensional (x,y,N) images into two-dimensional (x,y) label images. It uses human-in-the-loop ML to achieve consensus between the labeler and a model in an iterative workflow. Second, we will demonstrate Segmentation Zoo, a python toolbox to segment imagery with a variety of deep learning models that uses output from Doodler with existing models, or train entirely new models.
Ideally the clinic will be divided into two separate days. Day 1 would be a short introductory lecture, a live code demo, and then homework — participants will doodle imagery to gain familiarity with the software and create training data for a segmentation model. Day 2 would be a short introductory lecture on machine learning, and a live code demo for how to use doodled images in Segmentation Zoo (i.e., the images that participants doodled).
There are two concrete goals for the clinic: 1) demonstrate how participants can use these two tools, and; 2) a group authored dataset of doodled images that will be placed in a Zenodo repository with all participants who contribute as coauthors.
Doodler preprint: https://doi.org/10.31223/X59K83
Doodler repository: https://github.com/dbuscombe-usgs/dash_doodler
Doodler Website: https://dbuscombe-usgs.github.io/dash_doodler/
Segmentation Zoo repository: https://github.com/dbuscombe-usgs/segmentation_zooPlease acknowledge the original contributors when you are using this material. If there are any copyright issues, please let us know (CSDMSweb@colorado.edu) and we will respond as soon as possible.
Of interest for:
