Geospatial Foundation Models to Enable Progress on Sustainable Development Goals

TL;DR

SustainFM introduces a SDG-aligned benchmarking framework for geospatial foundation models, addressing a gap in evaluating real-world impact beyond task accuracy. It systematically compares a diverse set of pretraining paradigms (contrastive, masked-image modeling, supervised) across 16 SDG-grounded tasks using multi-sensor EO data, while incorporating energy efficiency and deployment impact into the assessment. The study finds that foundation models often outperform traditional approaches but emphasizes that gains are task- and data-dependent, with significant variability across datasets and datasets. By highlighting energy costs, transferability, and ethical considerations, SustainFM argues for impact-driven deployment and responsible AI practices to leverage geospatial FMs for sustainable development.

Abstract

Foundation Models (FMs) are large-scale, pre-trained artificial intelligence (AI) systems that have revolutionized natural language processing and computer vision, and are now advancing geospatial analysis and Earth Observation (EO). They promise improved generalization across tasks, scalability, and efficient adaptation with minimal labeled data. However, despite the rapid proliferation of geospatial FMs, their real-world utility and alignment with global sustainability goals remain underexplored. We introduce SustainFM, a comprehensive benchmarking framework grounded in the 17 Sustainable Development Goals with extremely diverse tasks ranging from asset wealth prediction to environmental hazard detection. This study provides a rigorous, interdisciplinary assessment of geospatial FMs and offers critical insights into their role in attaining sustainability goals. Our findings show: (1) While not universally superior, FMs often outperform traditional approaches across diverse tasks and datasets. (2) Evaluating FMs should go beyond accuracy to include transferability, generalization, and energy efficiency as key criteria for their responsible use. (3) FMs enable scalable, SDG-grounded solutions, offering broad utility for tackling complex sustainability challenges. Critically, we advocate for a paradigm shift from model-centric development to impact-driven deployment, and emphasize metrics such as energy efficiency, robustness to domain shifts, and ethical considerations.

Figures (6)

• Figure 1: Summary of the SustainFM benchmark. By bringing together domain experts and AI specialists (SDG 17: Partnerships for the Goals), and providing a collection of 16 datasets aligned with the first 16 SDGs, from six continents, spatial resolutions ranging from 0.5 m to 30 m, and multiple tasks grounded in real-world applications, SustainFM aims to provide a testbed to analyze the applicability and real impact of foundation models.
• Figure 2: Geographic coverage of the SustainFM benchmark, which spans over 200 regions across 6 continents to support a wide range of EO tasks aligned with the SDGs.
• Figure 3: SDGs tasks and data samples in the SustainFM benchmark.
• Figure 4: Model-wise comparison of mean F1-score across five datasets associated with different SDGs. This visualization highlights the performance differences among evaluated models, and provides insights into their generalization capabilities and robustness.
• Figure 5: Comprehensive evaluation of energy efficiency, carbon footprint, and predictive performance across diverse models and datasets. (a) Training energy consumption in kWh, (b) inference energy consumption in kWh, and (c) associated CO2 emissions in kg across all evaluated models. (d–h) Joint comparison of mF1 and CO2 emissions for five representative datasets (GAZAdeepDAV, GloSoForID, Ombria, HLS Burns, and CLCD), illustrating the trade-off between model accuracy and environmental sustainability.