2023 CSDMS meeting-069

Edit registration

Pay registration

(if you haven't already)

"CA" is not in the list (NO STATE, Alabama, Alaska, Arizona, Arkansas, California, Colorado, Connecticut, D.C., Delaware, ...) of allowed values for the "CSDMS meeting state" property.
Log in (or create account for non-CSDMS members)
Forgot username? Search or email:CSDMSweb@colorado.edu

Browse  abstracts

<<

>>

Back to meeting

The Doodleverse: A python ecosystem for geoscientific image segmentation

Daniel Buscombe, (he/him),Marda Science LLC / USGS Santa Cruz , United States. dbuscombe@gmail.com

Evan Goldstein, UNCG Greensboro North Carolina, United States. ebgoldst@uncg.edu

Sharon Fitzpatrick, USGS Santa Cruz California, United States.

Venus Ku, USGS Santa Cruz California, United States.

Cameron Bodine, Northern Arizona University Flagstaff Arizona, United States.

The Doodleverse (https://github.com/Doodleverse) is an ecosystem of Python software, data, and Machine Learning (ML) models for the application of image segmentation. Image segmentation is pixelwise classification, and is ubiquitously applied across Earth sciences. Imagery is any type of gridded data, including numerical model inputs and outputs. As such, image segmentation is a potentially useful generic tool in numerical modeling exercises, which will be demonstrated using a case study in this poster and epub.

Doodleverse workflows are fully reproducible, such that it is possible to entirely reconstruct a labeled dataset and model from scratch by anyone on any computing platform. There are 3 main software; 1) “Doodler”, a human-in-the-loop ML tool for interactive image segmentation, 2) “Segmentation Gym”, for training image segmentation models, facilitating model experimentation, and 3) “Segmentation Zoo”, a repository of trained models that each do specific tasks, along with code implementation examples. Deep learning models are based on Keras/Tensorflow. Currently, the UNet, Residual UNet, and Segformer model architectures are available.

The focus now is building downstream and demonstrative applications that use Segmentation Zoo models for specific data retrieval, extraction and mapping tasks. They include 1) “CoastSeg”, for mapping coastal shoreline dynamics using satellite imagery; 2) “Seg2Map”, for generic landuse/cover and landform mapping from publicly available high-resolution imagery; and 3) “PingMapper”, for mapping river and lake substrates from sidescan sonar imagery. Watch out for more Doodleverse applications in the future!