Making ML models accessible with TorchGeo and the STAC Machine Learning Model extension

[rbavery]

Birds of a Feather Session Application

Making ML models accessible with TorchGeo and the STAC Machine Learning Model extension:

Organizer(s):
Ryan Avery, Wherobots
Isaac Corley, Wherobots, TorchGeo

Space Requirements:
What sort of space would you ideally have to meet with your group?
Ideally enough room for 50 people, multiple tables or other sit down sections for breakout sessions. A screen to show slides and graphics. Outlets for charging nearby.

Would it be an issue if other groups are meeting in the same room at the same time?.

No though ideally some separation and we don't conflict with needing a projector/screen at the same time.

Session Description:
Discovery geospatial models that are relevant for a particular use case, trained on a geography of interest, and documented enough to run it is difficult. To address this problem, the STAC Machine Learning Model (MLM) Extension standardizes descriptions of machine learning models trained on overhead imagery. The STAC MLM extension supports describing "foundational" models that generate image embeddings as well as computer vision models that produce classification, detection, or segmentation results.

TorchGeo, a framework for training deep remote sensing models, also supports these same kinds of models, allowing for an integration where TorchGeo can export models that can be easily catalogued and run in different execution environments on the correct STAC data.

In this session we will present a short overview of the STAC MLM Extension, present some use cases, and discuss how we can make it better so that it is easier to find machine learning models and run them on the correct datasets. We'll also talk about progress on integrating STAC MLM into TorchGeo, so that any model trained in TorchGeo is more portable and can be run with minimal dependencies.

This will mostly be a discussion based format, so bring any questions or feedback you have about TorchGeo or the STAC MLM extension!

Expected Attendance:
50