BiliSakura / AnySat-transformers

AnySat Transformers Models

Hugging Face–compatible checkpoints converted from the official AnySat GeoPlex-pretrained weights. Each subfolder is a standalone model repo ( config.json , model.safetensors , processor, pipeline, and remote code) for multimodal Earth observation feature extraction.

Model Description

AnySat is a scale-adaptive, multimodal Earth Observation foundation model supporting 11 sensor modalities (aerial, Sentinel-1/2, Landsat, MODIS, etc.) at variable resolutions. The anysat-base checkpoint is the published base encoder (768-dim, 6 layers, 12 heads).

All hyperparameters live in config.json and preprocessor_config.json — Python remote code does not embed global defaults.

Developed by: AnySat Authors
Converted for Hugging Face by: BiliSakura
License (weights): MIT
Original paper: AnySat: One Earth Observation Model for Many Resolutions, Scales, and Modalities (CVPR 2025 Highlight)

Available checkpoints

Legacy source: models/raw/AnySat_full.pth (see conversion_manifest.json ).

Usage

Inputs must be normalized (data - mean) / std per modality when normalize=True in the processor. Set normalization_stats in preprocessor_config.json for your dataset statistics.

Custom pipeline (recommended)

from transformers import pipeline
import numpy as np

MODEL = "/path/to/AnySat-transformers/anysat-base"

pipe = pipeline(
    task="anysat-feature-extraction",
    model=MODEL,
    trust_remote_code=True,
)

# Sentinel-2 time series: (B, T, C, H, W)
s2 = np.random.randn(1, 4, 10, 32, 32).astype(np.float32)
dates = np.array([[100, 120, 140, 160]])

# Tile features (CLS token)
features = pipe(s2=s2, s2_dates=dates, patch_size=10, pool=True, return_tensors=True)
print(features.shape)  # torch.Size([1, 768])

# Patch features
patch_features = pipe(s2=s2, s2_dates=dates, patch_size=10, output="patch", return_tensors=True)
print(patch_features.shape)  # torch.Size([1, 1024, 768])

Multimodal example

pipe = pipeline(task="anysat-feature-extraction", model=MODEL, trust_remote_code=True)

s2 = np.random.randn(1, 4, 10, 32, 32).astype(np.float32)
s1 = np.random.randn(1, 4, 3, 32, 32).astype(np.float32)
features = pipe(
    s2=s2,
    s2_dates=np.array([[100, 120, 140, 160]]),
    s1=s1,
    s1_dates=np.array([[100, 120, 140, 160]]),
    patch_size=10,
    pool=True,
    return_tensors=True,
)

Built-in ImageFeatureExtractionPipeline

AnySatModel returns BaseModelOutputWithPooling , so the registered image-feature-extraction task also works:

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

Supported modalities

Time-series modalities require {modality}_dates (day-of-year, 0–364). Images must be square. patch_size is in meters (internally divided by 10).

Output modes