[ESIP-all] Upcoming Webinar on GeoAI Foundation Model on Nov. 13, 12-1 EST

[WenWen Li]

Dear colleagues,

We cordially invite you to join us for the upcoming Webinar sponsored by
the NSF CyberTraining project <https://cyber2a.github.io/>. Given the
interest in deepening the use of AI in Earth and Environmental Sciences, we
have had the pleasure to invite Dr. Paolo Fraccaro and Dr. Daniela
Szwarcman from IBM research to introduce the newly released geospatial
foundation model Prithvi jointly developed by IBM and NASA. Detailed
webinar information can be found below. We also attached the flyer for the
Webinar and would really appreciate it if you could circulate around your
network.

*Title*: A Framework for Building and Finetuning Geospatial Foundation
Models
*Speakers*: Dr. Paolo Fraccaro
<https://research.ibm.com/people/paolo-fraccaro> and Dr. Daniela Szwarcman
<https://research.ibm.com/people/daniela-szwarcman>, IBM research
*Time*: November 13, 2023, 12-1pm EST
*To register*: https://cyber2a.github.io/webinar-registration
*Abstract*:
Foundation models are artificial intelligence (AI) models that are
pre-trained on large unlabeled datasets through self-supervision and then
fine-tuned for different downstream tasks. There is increasing interest in
the scientific community to investigate whether this approach can be
successfully applied to domains beyond natural language processing and
computer vision to effectively build generalist AI models that make use of
different types of data. Here, IBM and NASA present the first end-to-end
framework for pre-training and fine-tuning foundation models efficiently
from a large source of geospatial data. We have implemented and applied
this framework to produce Prithvi, a geospatial foundation model
pre-trained on multispectral satellite imagery from the NASA Harmonized
Landsat-Sentinel 2 (HLS) dataset. The framework supports automated
statistical smart sampling strategies based on whether, land cover and
other datasets to maximize impact and minimize waste of resources (e.g.
avoiding areas and time ranges that would not bring any new information).
Prithvi is a Temporal Vision Transformer that includes positional and
temporal embeddings, which was trained on IBM Cloud Vela cluster (NVIDIA
A100 GPUs) using a Masked Auto Encoder approach and Mean Squared Error loss
function for a total of 10k GPUs hours. We demonstrated using the
fine-tuning workflows built in our framework that Prithvi could be
successfully fine-tuned to produce state-of-the-art AI models for Earth
observation tasks: flood mapping, burn scar identification and
multi-temporal crop classification. We carefully studied the impact of
Prithvi's pre-trained weights on the downstream tasks by comparing learning
curves for 1) fine-tuning the whole model, 2) fine-tuning only the
downstream task decoder, 3) training the model without taking advantage of
Prithvi's pre-trained weights. Furthermore, given the scarcity of labeled
data for Earth observation tasks, we progressively decreased the amount of
labeled data available for fine-tuning the model to assess data efficiency.
This analysis showed that using Prithvi we could achieve peak performance
on test data quicker and with less training data (i.e. up to 50% less).
Finally, in order to increase the impact of this work, the pre-trained
model and fine-tuning workflows have been made publicly available through
Hugging Face (https://huggingface.co/ibm-nasa-geospatial
<https://urldefense.com/v3/__https://huggingface.co/ibm-nasa-geospatial__;!!IKRxdwAv5BmarQ!dJpbD9Smzq372WC9oDu6IM7mya1lvQQpa4_NkgdSELtAtRT23lLkgIdhSjaqz0ozpqhR8rxdTHUDu9HY$>
).

Cheers,
Wenwen
-- 
Wenwen Li, PhD, Professor
School of Geographical Sciences and Urban Planning
Director, Cyberinfrastructure and Computational Intelligence Lab (CICI
<http://cici.lab.asu.edu>)
Research Director, Spatial Analysis Research Center (SPARC)
Graduate Faculty, Computer Science Program
Honors Faculty, Barrett Honors College at Arizona State University
Phone: (480)-727-5987.  Web: http://www.public.asu.edu/~wenwenl1
Office: COOR 5644, Arizona State University, P.O. Box 875302 Tempe, AZ
85287-5302
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.esipfed.org/pipermail/esip-all/attachments/20231025/7d80c165/attachment-0001.htm>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: Cyber2A Szwarcman Fraccaro Nov 13 Flier.pdf
Type: application/pdf
Size: 528841 bytes
Desc: not available
URL: <https://lists.esipfed.org/pipermail/esip-all/attachments/20231025/7d80c165/attachment-0001.pdf>