AnySat: An Earth Observation Model
for Any Resolutions, Scales, and Modalities
Arxiv 2025
Introducing AnySat
AnySat is a multimodal model based on joint embedding predictive architecture (JEPA) and resolution-adaptive spatial encoders, allowing us in a self-supervised manner to handle:
📏 Multiple Scales
From local to global observations
🔍 Various Resolutions
Spatial, spectral, and temporal
🛰️ Different Modalities
Multiple sensor combinations
Key Innovations
Shared Architecture
75% of parameters shared across all modalities and scales, enabling efficient multi-modal learning
Scale-Adaptive Design
Modified JEPA learning scheme with scale-adaptive spatial encoders for multi-resolution processing
Universal Compatibility
Handles data from 0.2m to 500m resolution, 3-12 channels, and areas from 0.3 to 2600 hectares
Datasets
We compile GeoPlex, a collection of multimodal datasets with varying characteristics to a single powerful model on these diverse datasets simultaneously. We argue quality and diversity over quantity.

Architecture

Scale-Adaptive JEPA Design
AnySat employs a novel Joint Embedding Predictive Architecture (JEPA) that adapts to multiple spatial scales and resolutions. Key components include:
- Resolution-adaptive spatial encoders
- Multi-modal fusion mechanism
- Scale-aware prediction heads
- Self-supervised learning framework
Results

Quick Start
import torch
import anysat
# Load the model
anysat_B = torch.hub.load('gastruc/anysat', 'anysat_B')
# Prepare your data
data = {
'naip':# NAIP single image B, 3, 24*H, 24*W
's2':# Sentinel-2 time series B, T1, 10, 3*H, 3*W
'alos':# ALOS-2 time series B, T2, 3, H, W
}
# Extract features
features = AnySat(data, patch_size=20) # patch_size in meters
Bibtex
@article{astruc2024anysat,
title={{AnySat: An Earth} Observation Model for Any Resolutions, Scales, and Modalities},
author={Astruc, Guillaume and Gonthier, Nicolas and Mallet, Clement and Landrieu, Loic},
journal={arXiv preprint arXiv:2412.14123},
year={2024}
}
Acknowledgements
This work was granted access to the HPC resources of IDRIS under the allocations AD011014719 and AD011014286R1 made by GENCI. We thank Jordi Inglada, Antoine Labatie, Dimitris Samaras, Yohann Perron, Vincent Lepetit for inspiring discussions and valuable feedback.