Foundation Models for Earth Observation with Hamed Alemohamad from Clark University
There are now a few different AI foundation models available for Earth Observation (EO) data. These vast neural networks can be rapidly fine-tuned for many downstream tasks, making them a highly versatile and appealing tool.
Today on Impact AI, I am joined by Hamed Alemohammad, Associate Professor in the Department of Geography at Clark University, Director of the Clark Center for Geospatial Analytics, and former Chief Data Scientist of the Radiant Earth Foundation, to discuss the applications of foundation models for remote sensing. Hamed’s research interests lie at the intersection of geographic information science and geography, using observations and analytical methods like machine learning to better understand the changing systems of our planet.
In this episode, he shares his perspective on the myriad purposes that foundation models serve and offers insight into training and fine-tuning them for different downstream applications. We also discuss how to choose the right one for a given project, ethical considerations for using them responsibly, and more. For a glimpse at the future of foundation models for remote sensing, tune in today!
Key Points:
- A look at Hamed’s professional journey and the research topics he focuses on today.
- Defining foundation models and the purposes they serve.
- The vast amount of data and resources required to train and fine-tune a foundation model.
- Ways to determine whether or not a foundation model will be beneficial.
- How foundation models improve generalizability for downstream tasks.
- Factors to consider when selecting a foundation model for a given downstream task.
- Insight into the future of foundation models for remote sensing.
- Hamed’s advice for machine learning teams looking to give foundation models a try.
- His take on the impact of foundation models in the next three to five years.
- Ethical considerations for the responsible use of AI that apply to foundation models too.
Quotes:
“[Foundation models] are pre-trained on a large amount of unlabeled data. Secondly, they use self-supervised learning techniques – The third property is that you can fine-tune this model with a very small set of labeled data for multiple downstream tasks.” — Hamed Alemohammad
“It takes a lot to train a model, but you would not [do it] as frequently as you would [fine-tune] the model. You can use shared resources from different teams to do that - share it as an open-source model, and then anybody can fine-tune it for their downstream application.” — Hamed Alemohammad
“The promising future [for foundation models] will be combining different modes of data as input.” — Hamed Alemohammad
“There is a lot to do and the community is eager to learn, so if people are looking for challenging problems, I would encourage them to explore [the foundation model domain] and work with domain experts.” — Hamed Alemohammad
Links:
Hamed Alemohammad, Clark University
Foundation Models for Generalist Geospatial Artificial Intelligence
HLS Multi-Temporal Crop Classification Model on Hugging Face
Resources for Computer Vision Teams:
LinkedIn – Connect with Heather.
Computer Vision Insights Newsletter – A biweekly newsletter to help bring the latest machine learning and computer vision research to applications in people and planetary health.
Computer Vision Strategy Session – Not sure how to advance your computer vision project? Get unstuck with a clear set of next steps. Schedule a 1 hour strategy session now to advance your project.
Custom Vision Model Assessment – Foundation models are popping up everywhere – do you need one for your proprietary image dataset? Get a clear perspective on whether you can benefit from a domain-specific foundation model.
You Might Like
