Are plant species encoded in the model currently?

[legel]

Great work overall to the whole team behind this.

On the research collaboration side of things, next week I'm leading a NSF program on DeepEarth in Boulder, Colorado for AI-supported species-specific landscape-scale fire resilience prediction, and we've been wanting to build on top of LFMC 2.0.

It seems this encoding of LFMC 2.0 does not (yet) incorporate the plant species? I'm curious if the Galileo model has any means to potentially integrate that as another modality, and if this has been explored at all? As a hint, you might consider taking an LLM and extracting embeddings from a pre-trained model, and then combining that encoding during the multimodal sensor fusion.

Happy to discuss further, feel free to ping me at lance@ecodash.ai for collaboration on this. Our work is integrated with the needs of landscape designers and a key requirement is that they be able to evaluate fire risk of individual and new plant species across landscapes. This is something that in some form I expect to move forward with next week, so please reach out if this sounds interesting to help bring everything together with the outstanding work your team has already done here.

[APatrickJ]

Thanks for the interest! We'd definitely like to discuss collaboration opportunities. I'll reach out over email.

Your DeepEarth work looks really impressive and meaningful.

It seems this encoding of LFMC 2.0 does not (yet) incorporate the plant species? I'm curious if the Galileo model has any means to potentially integrate that as another modality, and if this has been explored at all? As a hint, you might consider taking an LLM and extracting embeddings from a pre-trained model, and then combining that encoding during the multimodal sensor fusion.

The original Globe-LFMC 2.0 dataset does have detailed "Species collected" and coarser "Species functional type" data. In our evaluation we didn't incorporate these fields, however I think future work could potentially benefit from some of these data. The former may be too granular for model fine-tuning, but the latter contains broader categorizations like tree, shrub, grass, and forb, which could be very helpful.

[legel]

Thanks @APatrickJ for the wonderful reply!

Thanks as well for the very kind comment on DeepEarth. This morning I mentioned to my group your team's welcome for potential collaboration, and there was some excitement here as well.

Firstly I encourage leaning into the species-level granularity. Mainly because the physical chemistry of these species varies so dramatically, that I think too much information is lost in the categorizations. But also because we can encode that detail.

In the spirit of "open science", I'm very happy to share a few updates from here in Boulder, Colorado, typing from the NSF/NCAR/UCAR campus.

• AlphaEarth embeddings: we've extracted regions around LFMC 2.0, can add/compare with Galileo LFMC model.
• NAIP -> V JEPA 2 embeddings: same, will be working with 0.6m spatial resolution RGB vision features

Other cool projects too, please do drop an email, and I'll be happy to help organize collaboration! As another cool perspective, I think we could share insights / collaborate on DeepEarth + Galileo for the greater good of planetary foundation modeling.

PS speaking on behalf of the outstanding university leadership that organized this NSF I-GUIDE program, I can say that we're excited to put forward a proposal by February 2026 for this NSF FIRE program, and we could also join forces on that.

"Projects developed by a wide array of groups including, for example, academics, ... scientists, ... industry partners, ... working together to generate new knowledge of the interactions among biological, social, geoscientific, and engineering processes encompassing multiple fields, scales, and perspectives on wildland fire are encouraged."