Scaling Laws for Geospatial Foundation Models: A case study on PhilEO Bench

TL;DR

This work analyzes scaling laws for Geospatial Foundation Models by systematically varying architecture (CNNs, Transformers, Mamba), model size (44M–300M), and dataset scale (0.5TB, 2TB, 23TB) using PhilEO Bench tasks. It compares Geo-Aware U-Net, ViT-UPerNet, and Mamba across regression (building/road density) and segmentation (land cover) tasks, benchmarking against TerraMind and Prithvi-EO-2.0. Key findings show CNNs excel in low-shot regression, ViTs dominate semantic segmentation with large pretraining, and Mamba offers efficiency with competitive accuracy, though gains depend on data scale and domain. The work provides public code and pretrained models to study scaling laws in GFMs and emphasizes the important role of dataset design, temporal sampling, and domain coverage in EO pretraining.

Abstract

Foundation Models (FMs) have achieved state-of-the-art performance across domains by leveraging large-scale pretraining. In Earth Observation (EO), the availability of petabyte-scale satellite archives has recently enabled the development of GeoSpatial Foundation Models (GFMs). Yet, fundamental questions remain regarding how dataset size, model architecture, and size interact to determine downstream performance. In this work, we systematically explore this design space by pretraining and fine-tuning models on three dataset scales: PhilEO Globe (0.5TB), FastTOM (2TB, introduced here), and MajorTOM (23TB). We evaluate three architectural families: Geo-Aware U-Net (CNN), ViT-UPerNet (Transformer), and Mamba (State-Space Model); across model sizes ranging from 44M to 300M parameters. All models are benchmarked on the PhilEO Bench, covering: road density and building density regression, and land cover segmentation, and are compared against existing GFMs such as TerraMind and Prithvi-EO-2.0. Our results show that CNN-based models remain highly competitive in low-shot settings, with a 200M-parameter Geo-Aware U-Net outperforming larger architectures on regression tasks. However, when scaling to multi-terabyte datasets, ViT-UPerNet achieves the best performance, particularly for semantic segmentation on MajorTOM (23TB). Finally, we provide the first extensive evaluation of Mamba models in EO, highlighting their potential efficiency advantages, though further large-scale pretraining is required to fully match CNNs and ViTs. All code, pretrained models, and the FastTOM dataset are released publicly, enabling reproducibility and further exploration of scaling laws for GFMs.

Figures (4)

• Figure 1: Performance on PhilEO benchmark tasks for models pretrained on GlobeEO 0.5TB: Building density regression (top), Road density estimation (middle), and Land cover mapping (bottom), across different $n$-shot settings, where the key takeaways are that pretraining helps, i.e. Geo-Aware U-Net 44M versus U-Net 44M (Scratch), and the UPerNet decoder helps, i.e. ViT UPerNet compared to ViT CNN, where the latter means that a CNN decoder is used rather than multi-scale features in the UPerNet decoder.
• Figure 2: Evaluation over PhilEO Bench downstream tasks for various n-shots and for different model pretraining strategies on FastTOM 2TB: Building density regression (top), Road density estimation (middle), and Land cover mapping (bottom), where the key takeaways are that for pixel-wise regression dowsntream tasks, U-Net 200M for most n-shots outperforms the other models, and for semantic segmentation land cover mapping, the RS3Mamba is effective.
• Figure 3: PhilEO benchmark results for models pretrained on MajorTOM 23TB: Building density regression (top), Road density estimation (middle), and Land cover mapping (bottom), across $n$-shot settings, where the key takeaways are that for pixel-wise regression tasks, U-Net 44M for most $n$-shots has good performance, and for large $n$-shots, ViT has competitive performance. We also note that for semantic segmentation downstream tasks, ViT UPerNet outperforms U-Net.
• Figure 4: Further evaluation and experiments of Mamba SSM models over PhilEO Bench downstream tasks for various $n$-shots for various models pretrained on ImageNet: Building density regression (Top), Road density estimation (Middle), and Land cover mapping (Bottom).