BiliSakura/SATLAS-transformers

SatlasPretrain Transformers Models

Hugging Face–compatible checkpoints from the official SatlasPretrain foundation models (ICCV 2023). Each subfolder is a standalone model repo layout (config.json, model.safetensors, preprocessor, and remote code) for feature extraction on satellite and aerial imagery.

Model Description

These models are remote sensing encoders pretrained on the SatlasPretrain dataset covering Sentinel-2, Landsat 8/9, and high-resolution aerial imagery.

This collection bundles 19 checkpoints spanning:

• Architectures: Swin-v2-Base, Swin-v2-Tiny, ResNet50, ResNet152
• Input sources: Sentinel-2, Landsat, aerial
• Modes: single-image (SI) and multi-image temporal max-pooling (MI)
• Band configs: RGB (3ch), multispectral (9ch), Landsat all-bands (11ch)

All folders ship self-contained remote code (modeling_satlaspretrain.py, processor, pipeline) and load with trust_remote_code=True.

Developed by: Allen AI / Satlas
Packaged for Hugging Face by: BiliSakura
License (weights): ODC-BY
Original paper: SatlasPretrain: A Large-Scale Dataset for Remote Sensing Image Understanding

Available checkpoints (19 models)

Folder	Source	Backbone	Mode	Bands
satlaspretrain-sentinel2-swinb-si-rgb	Sentinel-2	Swin-B	SI	RGB
satlaspretrain-sentinel2-swinb-mi-rgb	Sentinel-2	Swin-B	MI	RGB
satlaspretrain-sentinel2-swinb-si-ms	Sentinel-2	Swin-B	SI	MS (9ch)
satlaspretrain-sentinel2-swinb-mi-ms	Sentinel-2	Swin-B	MI	MS (9ch)
satlaspretrain-sentinel2-swint-si-rgb	Sentinel-2	Swin-T	SI	RGB
satlaspretrain-sentinel2-swint-mi-rgb	Sentinel-2	Swin-T	MI	RGB
satlaspretrain-sentinel2-swint-si-ms	Sentinel-2	Swin-T	SI	MS (9ch)
satlaspretrain-sentinel2-swint-mi-ms	Sentinel-2	Swin-T	MI	MS (9ch)
satlaspretrain-sentinel2-resnet50-si-rgb	Sentinel-2	ResNet50	SI	RGB
satlaspretrain-sentinel2-resnet50-mi-rgb	Sentinel-2	ResNet50	MI	RGB
satlaspretrain-sentinel2-resnet50-mi-ms	Sentinel-2	ResNet50	MI	MS (9ch)
satlaspretrain-sentinel2-resnet152-si-rgb	Sentinel-2	ResNet152	SI	RGB
satlaspretrain-sentinel2-resnet152-si-ms	Sentinel-2	ResNet152	SI	MS (9ch)
satlaspretrain-sentinel2-resnet152-mi-rgb	Sentinel-2	ResNet152	MI	RGB
satlaspretrain-sentinel2-resnet152-mi-ms	Sentinel-2	ResNet152	MI	MS (9ch)
satlaspretrain-landsat-swinb-si	Landsat 8/9	Swin-B	SI	All (11ch)
satlaspretrain-landsat-swinb-mi	Landsat 8/9	Swin-B	MI	All (11ch)
satlaspretrain-aerial-swinb-si	Aerial	Swin-B	SI	RGB
satlaspretrain-aerial-swinb-mi	Aerial	Swin-B	MI	RGB

Usage

Processors default to do_resize: false. Pass tensors or arrays at native resolution; sensor-specific normalization is still applied.

from transformers import pipeline
import torch

model_dir = "satlaspretrain-sentinel2-swinb-si-rgb"
pipe = pipeline(
    task="satlaspretrain-feature-extraction",
    model=model_dir,
    trust_remote_code=True,
)

# Sentinel-2 RGB at native size, values in [0, 1]
x = torch.rand(1, 3, 512, 512)
features = pipe(x, pool=True, return_tensors=True)
print(features.shape)

# Dense feature map
tokens = pipe(x, pool=False, return_tensors=True)

Opt in to resize using the reference size in preprocessor_config.json:

features = pipe(x, pool=True, return_tensors=True, image_processor_kwargs={"do_resize": True})

Standard image-feature-extraction also works:

pipe = pipeline(
    task="image-feature-extraction",
    model=model_dir,
    trust_remote_code=True,
)

Normalization

The bundled image processor applies sensor-specific normalization:

• Sentinel-2 / aerial RGB: divide by 255
• Landsat: (x - 4000) / 16320, clipped to [0, 1]

See Normalization.md for details.