StefaLand: An Efficient Geoscience Foundation Model That Improves Dynamic Land–Surface Predictions

ICLR 2026 Conference Submission22153 Authors

Keywords: Foundation models, Spatiotemporal modeling, Geoscience AI, Land-surface prediction, Climate impacts, Transfer learning

TL;DR: A compute-efficient geoscience foundation model that learns landscape interactions and improves spatial generalization for streamflow, soil composition, landslides, and soil moisture.

Abstract: Managing natural resources, meeting growing societal needs, and reducing risks from floods, droughts, wildfires, and landslides require models that can accurately predict climate-driven land–surface responses. Traditional models of environmental impact, whether process-based or task-specific machine learning, often struggle with spatial generalization because they are trained on limited observations and can degrade under concept drift. Recently proposed vision foundation models trained on satellite imagery demand massive compute, and they are often ill-suited for dynamic land surface prediction tasks. We introduce StefaLand, a generative spatiotemporal Earth foundation model centered on landscape interactions. StefaLand is demonstrated to improve predictions on four important tasks across five datasets: streamflow, soil moisture, soil composition and landslides, compared to previous state-of-the-art methods, showing especially strong ability to generalize across diverse landscapes, including data-scarce regions. The model builds on a masked autoencoder architecture that learns deep joint representations of landscape attributes, and its design reflects a deliberate integration of ideas adapted to geoscience. These include a location-aware architecture that fuses static and time-series inputs, an attribute-based rather than image-based representation that drastically reduces compute demands, and residual fine-tuning adapters that strengthen knowledge transfer across tasks. Their alignment with domain knowledge enables StefaLand to deliver robust performance on various dynamic land–surface tasks. StefaLand can be pretrained and finetuned on commonly-available academic compute resources compared with commercial foundation models, yet consistently outperforms state-of-the-art supervised learning baselines and fine-tuned vision foundation models. To our knowledge, this is the first geoscientific land-surface foundation model that demonstrably improves dynamic land surface interaction prediction tasks and supports a wide range of downstream applications.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 22153