GEO-Bench-2: A Capability-Aware Benchmark
for Geospatial Foundation Models

Naomi Simumba1†, Nils Lehmann2,9†, Paolo Fraccaro1†‡, Hamed Alemohammad3, Geeth De Mel1, Salman Khan5, Manil Maskey6, Nicolas Longepe7, Xiao Xiang Zhu2,9, Hannah Kerner4, Juan Bernabe-Moreno1, Alexandre Lacoste8‡
1 IBM Research Europe 2 Technical University Munich 3 Clark University 4 Arizona State University
5 MBZUAI 6 NASA IMPACT 7 ESA Φ-lab 8 ServiceNow AI Research 9 Munich Center for Machine Learning (MCML)
† Equal contribution ‡ Corresponding author
📄 Paper 💻 Code 📘 Docs 🏆 Leaderboard ⬇️ Download 🤗 Hugging Face

Related projects

💻 GEO-Bench-VLM 💻 GEO-Bench-1
GEO-Bench-2 Overview

GEO-Bench-2 is a large-scale, capability-aware benchmark for evaluating Geospatial Foundation Models through fine-tuning-based evaluation across diverse sensing modalities, temporal contexts, and downstream applications. It emphasizes open licensing, reproducibility, and capability-specific evaluation, enabling the community to measure progress in perception, reasoning, and generalization within the geospatial domain.

For zero-shot evaluation of Geospatial Vision-Language Models, please refer to our complementary benchmark GEO-Bench-VLM.