Title: Leveraging Geospatial Foundation Models to Advance the Sustainable Development Goals

Foundation Models (FMs)—large-scale, pre-trained artificial intelligence systems that have transformed natural language processing and computer vision—are now driving significant advances in geospatial analysis and Earth Observation (EO). These models offer enhanced generalization across various tasks, superior scalability, and the ability to adapt efficiently even with limited labeled data. Yet, despite the swift expansion of geospatial FMs, their practical application and alignment with global sustainability objectives have received insufficient attention.

To address this gap, we present SustainFM, a robust benchmarking framework anchored in the 17 Sustainable Development Goals. This framework encompasses a wide array of tasks, spanning from predicting asset wealth to detecting environmental hazards. Through this study, we deliver a thorough, interdisciplinary evaluation of geospatial FMs, shedding light on their potential to help achieve sustainability targets.

Our analysis yields several key insights:

1. Although they are not universally better than existing methods, FMs frequently surpass traditional approaches in performance across a variety of tasks and datasets.
2. Assessing these models requires looking beyond simple accuracy. Responsible deployment must prioritize transferability, generalization capabilities, and energy efficiency.
3. FMs provide scalable, SDG-aligned solutions that hold broad utility for addressing intricate sustainability issues.

Crucially, we call for a fundamental shift in approach: moving away from model-centric development toward impact-driven deployment. This transition must be supported by rigorous metrics focusing on energy efficiency, resilience to domain shifts, and ethical considerations.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC