A Reproducible and Reusable Pipeline for Segmentation of Geoscientific Imagery. Issue 9 (23rd September 2022)

• Record Type: Journal Article
• Authors: Buscombe, D.; Goldstein, E. B.
• Is Part Of: Earth and space science. Volume 9:Issue 9(2022)
• Page Start: n/a
• Page End: n/a
• Publication Date: 2022-09-23
• Subjects: machine learning -- deep learning -- gridded data -- image segmentation -- data science -- image classification; Space sciences -- Periodicals; Geophysics -- Periodicals; 500.5
• DOI: 10.1029/2022EA002332
• Languages: English
• ISSNs: 2333-5084

Abstract

Abstract: Segmentation of Earth science imagery is an increasingly common task. Among modern techniques that use Deep Learning, the UNet architecture has been shown to be a reliable for segmenting a range of imagery. We developed software–Segmentation Gym–to implement a data‐model pipeline for segmentation of scientific imagery using a family of UNet models. With an existing set of imagery and labels, the software uses a single configuration file that handles data set creation, as well as model setup and model training. Key benefits of this software are (a) the focus on reproducible data set creation and modeling, and (b) the ability for quick model experimentation through changes to a configuration file. Quick experimentation permits researchers to prototype different model architectures, sizes, and adjust common hyperparameters to find a suitable model. We demonstrate the use of the software using a data set of 419 labeled Landsat‐8 scenes of coastal environments and compare results across two model architectures, five model sizes, and three loss functions. This demonstration highlights that our software enables rapid, reproducible experimentation to determine optimal hyperparameters for specific data sets and research questions. Plain Language Summary: A common task for Earth scientists is to divide a satellite or aerial image into specific classes. For example, an image of the coastline might be assigned certain pixels as being water, beach, and land. In the Deep Learning world, this is called segmentation. We wrote a piece of software that helps researchers train Deep Learning models to do segmentation on all types of imagery. A major problem with making Deep Learning models is dealing with all the choices on which model to use and quickly testing many options. We have designed our code in such a way that it can easily be adjusted, and will work in many applications and for many common types of Earth science image data sets. Key Points: We develop software for segmentation of geoscientific imagery with fully convolutional deep neural network models The software presents options for users, but relies on a reusable template that allows for rapid experimentation We demonstrate an example workflow with Landsat 8 imagery, and compare loss functions, model size, and model architectures