AI Lab Seminar: Foundation Model Approach for Global Terrestrial Carbon Stock Mapping

Monday 7 Jul 2025, 09:30 → 10:30 Europe/Berlin

Seminar Room Ground Floor (Oncoray)

In this event, we showcase AI-based research conducted at HZDR and bridge the gap to the scientific community in Dresden. Everyone is invited to join! 

Abstract

Accurate and frequently updated global terrestrial carbon stock maps are crucial for understanding climate change dynamics, informing land management practices, and verifying carbon sequestration efforts. Traditional mapping approaches often rely on localized models that struggle with scalability, data heterogeneity, and generalization across diverse ecosystems. In this talk, we introduce the 3D-ABC foundation model approach for global terrestrial carbon stock mapping. This methodology leverages large-scale, self-supervised learning on diverse remote sensing Earth observation data, including optical, multispectral, Synthetic Aperture Radar (SAR), and LiDAR. We will focus on how masked image modeling is applied on multimodal data for this particular Earth Observation application.

Our Host

We are grateful that Oncoray is hosting this event! We hope that this caters to a large audience from all HZDR sites and provides apt opportunity to attend in-person.

Address

Building 130
Händelallee 26
01309 Dresden

Map location

Want to Join?

Please register so we can plan for extra chairs if required.

09:30 → 09:33 Welcome

Speaker: Peter Steinbach (HZDR)

09:33 → 10:13 Foundation Model Approach for Global Terrestrial Carbon Stock Mapping.

Accurate and frequently updated global terrestrial carbon stock maps are crucial for understanding climate change dynamics, informing land management practices, and verifying carbon sequestration efforts. Traditional mapping approaches often rely on localized models that struggle with scalability, data heterogeneity, and generalization across diverse ecosystems. In this talk, we introduce the 3DABC foundation model approach for global terrestrial carbon stock mapping. This methodology leverages large-scale, self-supervised learning on diverse remote sensing Earth observation data, including optical, SAR, and LiDAR. We will focus on how the Massively Multimodal Masked Modeling technique is being utilized for this Earth Observation application.

Speakers: Aldino Rizaldy (Helmholtz Institute Freiberg for Resource Technology), Weikang Yu

10:13 → 10:30 Q&A

Speaker: Peter Steinbach (HZDR)