We believe that the best and most innovative products, like our API and newly released Studio product, are built on a foundation of rigorous science and engineering excellence. 

LGND scientists are working at the forefront of the science of Earth embeddings - and active participants in academic exchanges. 

Today we are highlighting the publications co-authored by LGND scientists Sam Barrett and Konstantin Klemmer over the last year - ranging from work on machine learning methods to broader community perspectives. We've grouped them into three themes that mirror how this research feeds back into our products:

• Understanding what Earth embeddings actually encode
• Building better architectures on top of that understanding
• Turning the results into real-world tools and impact for our customers.

These studies are conducted with academic collaborators from around the world and, in parts, published in leading machine learning and remote sensing venues like the International Conference on Machine Learning (ICML) or IEEE Geoscience and Remote Sensing Magazine (GRSM): 

Understanding Earth embeddings encodings

Before we can build on Earth embeddings, we need to know what they actually capture — their structure, geometry, and limits.

Earth Embeddings: Towards AI-centric Representations of our Planet — A perspectives paper defining the "Earth embeddings" field and its research priorities. Preprint (EarthArXiv), Dec 2025. Klemmer, Rolf, Russwurm, Camps-Valls, et al.

Measuring the Intrinsic Dimension of Earth Representations — Quantifies how much information Earth embeddings actually encode, exposing their structure and redundancy. ICLR 2026. Rao, Rußwurm, Klemmer, Rolf.

Characterizing AlphaEarth Embedding Geometry for Agentic Environmental Reasoning — Testing AlphaEarth embeddings for agentic tasks in environmental modeling. Preprint (arXiv), Apr 2026. Rahman, Barrett, Last.

OT on the Map: Quantifying Domain Shifts in Geographic Space — Uses optimal transport to measure geographic distribution shift and predict cross-region model generalization. Preprint (arXiv), Apr 2026. Zhang, Betti, Klemmer, Rolf, Alvarez-Melis.

Building better architectures

These findings feed directly into how we design and train the next generation of geospatial foundation models.

Localized, High-resolution Geographic Representations with Slepian Functions — Proposes Slepian-function location encoders that capture fine-grained regional detail. ICML 2026. Rao, Crasto, Ooms, Rolnick, Klemmer, Rußwurm.

SITS-DECO: A Generative Decoder Is All You Need For Multitask Satellite Image Time Series Modelling — A GPT-style decoder that treats EO as next-token prediction and beats larger foundation models on crop-type. Preprint (arXiv), Oct 2025. Barrett, Sow. classification.

Spatial Representation Learning Beyond Pixels — Argues that geospatial foundation models should fuse raster imagery with vector semantics for richer, human-centric place representations. Preprint (arXiv), Jun 2026. Knoblauch, Li, Mai, Klemmer, Gao, Li.

From research to real-world impact

Closing the loop: turning the science into tools, standards, and outcomes our customers can use.

A Blueprint for Integrated Climate Intelligence — A community roadmap for combining Earth observation, climate modeling, and ML into actionable climate intelligence. Preprint (EarthArXiv), May 2026. Rodriguez-Pardo et al. (incl. Klemmer).

EarthEmbeddingExplorer — A web app for cross-modal (text↔image) retrieval over global satellite imagery using Earth embeddings. ML4RS Workshop, ICLR 2026. Zheng, Wu, Wu, Zhao, Li, Czerkawski, Klemmer.

Additionally, we publish shorter, more experimental insights on the LGND Devlog, be sure to check them out as well.