AnySat: An Earth Observation Model for Any Resolutions, Scales, and Modalities

Guillaume Astruc, Nicolas Gonthier, Clement Mallet, Loic Landrieu

For more details and results, please check out our github and project page.

Abstract

AnySat is a versatile Earth Observation model designed to handle diverse data across resolutions, scales, and modalities. Using a scale-adaptive joint embedding predictive architecture (JEPA), AnySat can train in a self-supervised manner on highly heterogeneous datasets.

We train a single AnySat model on GeoPlex, a collection of 5 multimodal datasets spanning 11 sensors with varying characteristics. In fine-tuning or linear probing, AnySat achieves SOTA or near-SOTA performance for land cover segmentation, crop type classification, change detection, tree species identification, and flood mapping.

Key Features

• 🌍 Versatile Model:Handles diverse datasets with resolutions spanning 3–11 channels, tiles ranging from 0.3 to 2600 hectares, and any combination of 11 sensors.
• 🚀 Simple to Use: Install and download AnySat with a single line of code, select your desired modalities and patch size, and immediately generate rich features.
• 🦋 Flexible Task Adaptation: Supports fine-tuning and linear probing for tasks like tile-wise classification and semantic segmentation.
• 🧑‍🎓 Multi-dataset Training: Trains a single model across multiple datasets with varying characteristics.

🚀  Quickstart

Check out our demo notebook or huggingface page for more details.

Install and load Anysat

import torch

AnySat = torch.hub.load('gastruc/anysat', 'anysat', pretrained=True, flash_attn=False)

Set flash_attn=True if you have flash-attn module installed. It is not required and only impacts memory and speed.

Format your data

Arrange your data in a dictionary with any of the following keys:

Dataset	Description	Tensor Size	Channels	Resolution
aerial	Single date tensor	Bx4xHxW	RGB, NiR	0.2m
aerial-flair	Single date tensor	Bx5xHxW	RGB, NiR, Elevation	0.2m
spot	Single date tensor	Bx3xHxW	RGB	1m
naip	Single date tensor	Bx4xHxW	RGB	1.25m
s2	Time series tensor	BxTx10xHxW	B2, B3, B4, B5, B6, B7, B8, B8a, B11, B12	10m
s1-asc	Time series tensor	BxTx2xHxW	VV, VH	10m
s1	Time series tensor	BxTx3xHxW	VV, VH, Ratio	10m
alos	Time series tensor	BxTx3xHxW	HH, HV, Ratio	30m
l7	Time series tensor	BxTx6xHxW	B1, B2, B3, B4, B5, B7	30m
l8	Time series tensor	BxTx11xHxW	B8, B1, B2, B3, B4, B5, B6, B7, B9, B10, B11	10m
modis	Time series tensor	BxTx7xHxW	B1, B2, B3, B4, B5, B6, B7	250m

Note that time series requires a _dates companion tensor containing the day of the year: 01/01 = 0, 31/12=364.

Example Input for a tile of 60x60m and a batch size of B:

data = {
    "aerial": ... #Tensor of size [B, 4, 300, 300] : 4 channels, 300x300 pixels at 20cm res
    "spot": ... #Tensor of size [B, 3, 60, 60]: 3 channels, 60x60 pixels at 1m res
    "s2": ... #Tensor of size [B, 12, 10, 6, 6] : 12 dates, 10 channels, 6x6 pixels at 10m res
    "s2_dates":  ... #Tensor of size [B, 12] : 12 dates
}

Ensure that the spatial extent of each modality multiplied by its resolution is consistent.

Extract Features

Decide on:

• Patch size (in m, must be a multiple of 10): adjust according to the scale of your tiles and GPU memory. In general, avoid having more than 1024 patches per tile.
• Output type: Choose between:
  • 'tile': Single vector per tile
  • 'patch': A vector per patch
  • 'dense': A vector per sub-patch
  • 'all': a tuple with all three outputs

The sub patches are 1x1 pixels for time series and 10x10 pixels for VHR images. If using output='dense', specify the output_modality.

Example use:

features = AnySat(data, scale=10, output='tile') #tensor of size [D,]
features = AnySat(data, scale=10, output='patch') #tensor of size [D,6,6]
features = AnySat(data, scale=20, output='patch') #tensor of size [D,3,3]
features = AnySat(data, scale=20, output='dense', output_modality='aerial') #tensor of size [D,30,30]

Explanation for the size of the dense map: d=10 for 'aerial' which has a 0.2m resolution, the sub-patches are 2x2 m.

Advanced Installation

Install from source

# clone project
git clone https://github.com/gastruc/anysat
cd anysat

# [OPTIONAL] create conda environment
conda create -n anysat python=3.9
conda activate anysat

# install requirements
pip install -r requirements.txt

# Create data folder where you can put your datasets
mkdir data
# Create logs folder
mkdir logs

Run Locally

To load the model locally, you can use the following code:

from hubconf import AnySat

AnySat = AnySat.from_pretrained('base', flash_attn=False) #Set flash_attn=True if you have flash-attn module installed
#For now, only base is available.
#device = "cuda" If you want to run on GPU default is cpu

Every experience of the paper has its config file. Feel free to explore configs/exp folder.

# Run AnySat pretraining on GeoPlex
python src/train.py exp=GeoPlex_AnySAT

# Run AnySat finetuning on BraDD-S1TS
python src/train.py exp=BraDD_AnySAT_FT

# Run AnySat linear probing on BraDD-S1TS
python src/train.py exp=BraDD_AnySAT_LP

Supported Datasets

Our implementation already supports 9 datasets:

GeoPlex Datasets

1. TreeSatAI-TS

   • Description: Multimodal dataset for tree species identification.
   • Extent: 50,381 tiles covering 180 km² with multi-label annotations across 20 classes.
   • Modalities: VHR images (0.2 m), Sentinel-2 time series, Sentinel-1 time series.
   • Tasks: Tree species classification.

2. PASTIS-HD

   • Description: Crop mapping dataset with delineated agricultural parcels.
   • Extent: 2,433 tiles covering 3986 km² with annotations across 18 crop types.
   • Modalities: SPOT6/7 VHR imagery (1.5 m), Sentinel-2 time series, Sentinel-1 time series.
   • Tasks: Classification, semantic segmentation, panoptic segmentation.

3. FLAIR

   • Description: Land cover dataset combining VHR aerial imagery with Sentinel-2 time series.
   • Extent: 77,762 tiles covering 815 km² with annotations across 13 land cover classes.
   • Modalities: VHR images (0.2 m), Sentinel-2 time series.
   • Tasks: Land cover mapping.

4. PLANTED

   • Description: Global forest dataset for tree species identification.
   • Extent: 1,346,662 tiles covering 33,120 km² with annotations across 40 classes.
   • Modalities: Sentinel-2, Landsat-7, MODIS, Sentinel-1, ALOS-2.
   • Tasks: Tree species classification.

5. S2NAIP-URBAN

   • Description: Urban dataset with high-resolution imagery and time series data.
   • Extent: 515,270 tiles covering 211,063 km² with NAIP, Sentinel-2, Sentinel-1, and Landsat-8/9 data.
   • Modalities: NAIP (1.25 m), Sentinel-2 time series, Sentinel-1 time series, Landsat-8/9.
   • Tasks: Pretraining only (no official labels).

External Evaluation Datasets

1. BraDD-S1TS

   • Description: Change detection dataset for deforestation in the Amazon rainforest.
   • Extent: 13,234 tiles with Sentinel-1 time series.
   • Tasks: Change detection (deforestation segmentation).

2. SICKLE

   • Description: Multimodal crop mapping dataset from India.
   • Extent: 34,848 tiles with Sentinel-1, Sentinel-2, and Landsat-8 time series.
   • Tasks: Crop type classification (paddy/non-paddy).

3. TimeSen2Crop

   • Description: Crop mapping dataset from Slovenia.
   • Extent: 1,212,224 single-pixel Sentinel-2 time series.
   • Tasks: Crop type classification.

4. Sen1Flood11

   • Description: Flood mapping dataset with global scope.
   • Extent: 4.8K Sentinel-1/2 time series.
   • Tasks: Flood classification (flooded/ not flooded).

Reference

Please use the following 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}
}