v292: TerraBytes at ICML 2025

2026-06-04 · Source: Proceedings of Machine Learning Research · Field: Science & Research — Environmental Science & Earth Systems, Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, quick

Summary

Volume 292 of the TerraBytes ICML Workshop Proceedings, held on July 19, 2025, in Vancouver, BC, presents eleven papers focused on advancing global datasets and models for Earth Observation. Key contributions include a human-in-the-loop method for whale detection in high-resolution satellite imagery and lessons from deploying geospatial foundation models like WorldCereal. Other research explores resampling augmentation for time series contrastive learning in remote sensing and improving air pollution forecasting via multi-variable data alignment. The volume also addresses challenges in cloud-based geospatial benchmarking, evaluates LLMs, and investigates the effects of label availability and temporal resolution on fine-tuning geospatial foundation models. Further papers detail optimizing cloud-to-GPU throughput for deep learning with Earth Observation data, generating high-resolution LFMC maps for wildfire risk, and label-efficient hyperspectral image classification.

Key takeaway

For Machine Learning Engineers developing Earth Observation solutions, this volume offers critical insights into current challenges and methods. You should explore human-in-the-loop detection for specific object identification and consider lessons from real-world geospatial foundation model deployments like WorldCereal. Evaluate techniques for optimizing cloud-to-GPU throughput and leveraging multi-modal data to enhance data-efficiency and out-of-distribution generalization in your satellite imagery applications.

Key insights

The TerraBytes workshop showcases diverse Earth Observation advancements, from specific applications to foundational model deployment and data optimization.

Principles

• Geospatial foundation models require real-world deployment lessons.
• Human-in-the-loop methods enhance satellite imagery analysis.
• Multi-modal data improves ML data-efficiency and generalization.

In practice

• Deploy geospatial foundation models like WorldCereal.
• Use resampling augmentation for remote sensing time series.
• Optimize cloud-to-GPU throughput for EO deep learning.

Topics

• Earth Observation
• Geospatial Foundation Models
• Satellite Imagery Analysis
• Deep Learning Optimization
• Time Series Contrastive Learning
• Air Pollution Forecasting

Best for: Computer Vision Engineer, AI Scientist, Research Scientist, Machine Learning Engineer

Related on AIssential

• Contrasting Local and Global Modeling with Machine Learning and Satellite Data: A Case Study Estimating Tree Canopy Height in African Savannas — Local SatML models can outperform global or fine-tuned global models for specific regional tasks.
• Paper Digest: ICCV 2025 Papers & Highlights — Multimodal AI, 3D reconstruction, and video understanding are advancing rapidly with novel models and benchmarks.
• Simplifying Image Classification with Azure AutoML for Images: A Practical Guide — Azure AutoML for Images automates complex computer vision tasks, enabling rapid deployment of production-ready models.
• Google Maps is about to get a big dose of AI — Google's new generative AI features enhance enterprise geospatial analysis and visualization across its mapping platforms.
• LatticeVision: Image to Image Networks for Modeling Non-Stationary Spatial Data — Image-to-image networks globally estimate non-stationary spatial parameters faster and more accurately than local methods.
• MapSR: Prompt-Driven Land Cover Map Super-Resolution via Vision Foundation Models — MapSR enables high-resolution land-cover mapping with minimal annotation and compute by decoupling supervision from training.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Proceedings of Machine Learning Research.