Now We Know? A Systematic Comparison of TerraMind and THOR

2026-07-20

AI Summary

This study addresses the limitations of existing geospatial foundation model (GFM) benchmarks, which rely heavily on aggregate scores and fail to elucidate the root causes of performance differences. Through controlled experiments, it systematically compares two architecturally distinct GFMs—TerraMind, emphasizing large-scale pretraining, and THOR, prioritizing flexible tokenization at inference—across ten remote sensing tasks. The analysis investigates the impact of patch size, decoder complexity, fine-tuning strategies, input modalities, and model scale. Introducing a diagnostic ablation framework, the work reveals that architectural choices often outweigh model identity in determining performance. Both models leverage multimodal generative pretraining and compute-adaptive architectures to uniformly process Sentinel-1/2/3 data, enabling cross-modal generation and variable-resolution inputs. Results demonstrate that model superiority is highly task- and data-dependent, with no universally optimal solution, yet yield a set of generalizable design hypotheses for future GFMs.

Abstract

Benchmarks for Geospatial Foundation Models (GFMs) increasingly rank models by aggregate score, but such rankings obscure why models differ: how much of the gap is architecture, how much is decoder capacity, and how much is a use-case-specific artefact? This study addresses that gap through a controlled comparison of two GFMs developed under European Space Agency's $Φ$-lab with contrasting design philosophies: THOR, which introduces a compute-adaptive architecture supporting variable patch sizes and unifies Sentinel-1, -2, and -3 data at their native resolutions; and TerraMind, a multimodal generative GFM pretrained with a dual-scale token/pixel objective that enables any-to-any cross-modal generation (Thinking-in-Modalities) to infer missing sensors at inference time. Rather than reporting a single leaderboard, we investigate the axes along which the two architectures actually differ - patch size, decoder complexity, finetuning regime, input modality, and model scale - across ten use cases spanning segmentation and regression in diverse domains, including climate disaster response, methane leak detection, snow monitoring, or sea ice mapping. We find that architectural design choices - patch size and decoder type in particular - explain more performance variance than model identity itself, that the two models embody complementary investment strategies (pretraining-time scale for TerraMind versus inference-time tokenisation for THOR), and that correctly interpreting results requires dataset-level characterisation. The resulting picture is not a single winner but a set of hypotheses and a diagnostic ablation methodology that we expect to generalise to future GFMs beyond THOR and TerraMind.

Problem

• Geospatial Foundation Models
• benchmarking
• model comparison
• performance variance
• architecture design

Innovation

• Geospatial Foundation Models
• controlled comparison
• compute-adaptive architecture
• cross-modal generation
• diagnostic ablation methodology