radiance-mae-base-v1

Radiance MAE Base v1

Radiance is an open-weight ViT-Base Masked Autoencoder (MAE) foundation model for Sentinel-1 SAR imagery from Stratus Labs. Pretrained self-supervised on the SSL4EO-S12 Sentinel-1 GRD split for 100 epochs, then intended to be fine-tuned for downstream Earth-observation tasks (flood mapping, crop type, deforestation, urban change, etc.).

Radiance is the SAR sibling to Nocturne (Stratus Labs' bioacoustic classifier). Same lab, same open license, different modality.

Model

Training

• Dataset: SSL4EO-S12 Sentinel-1 GRD split (globally sampled SAR patches, dB-normalised with S1_MEAN = [-12.577, -20.265], S1_STD = [5.176, 5.870])
• Epochs: 100
• Batch size: 64
• Optimiser: AdamW, lr = 1.5e-4, wd = 0.05, cosine schedule
• Precision: bf16 autocast
• Hardware: 1× NVIDIA GB10 (DGX Spark) — ~15 days wall-clock

Final training loss decreased monotonically across all 100 epochs. The epoch_099.pt checkpoint published here is the final-epoch weights.

Usage

import torch
from radiance.model import RadianceMAE  # from the training repo; also included in this repo as model.py

ckpt = torch.load("epoch_099.pt", map_location="cpu")
model = RadianceMAE(img_size=224, patch_size=16, in_chans=2)
model.load_state_dict(ckpt["model"])
model.eval()

# Encode a 2-channel SAR patch (VV, VH) shape [B, 2, 224, 224]
with torch.no_grad():
    cls, patches = model.encode(sar_patch)   # cls: [B, 768]; patches: [B, 196, 768] in RASTER order
    # patches.transpose(1, 2).reshape(B, 768, 14, 14) is the feature grid for any dense head

Use encode() for downstream features, not forward_encoder(). MAE's random_masking returns the kept tokens in a random permutation even at mask_ratio = 0 (the pretrain decoder restores order with ids_restore). Reshaping forward_encoder output into the patch grid scrambles spatial positions; our first flood fine-tune shipped with exactly that bug (see the SEN1Floods11 card). encode() un-shuffles for you and is deterministic. Added to model.py 2026-09-06.

Files

Intended use

Pretrained backbone for downstream SAR tasks. First downstream head: stratus-labs/radiance-sen1floods11-v1 (v2, 2026-09-06) — binary flood segmentation, held-out test flood IoU 0.598, out-of-region (bolivia) 0.651, with per-chip breakdown on the card. Further heads (crop type, change detection) will be published as sibling stratus-labs/radiance-* repos.

Off-the-shelf, you can use Radiance as a feature extractor for any 2-channel SAR patch task by freezing the encoder and training a task head on top.

Limitations

• Input is Sentinel-1 GRD dB-scaled specifically — other SAR products (SLC, IW, Sentinel-2 optical) are out-of-distribution.
• 224×224 patch size is fixed for the released weights; other sizes would need re-pretrain.
• No radiometric calibration correction beyond the SSL4EO-S12 dB normalisation.
• One downstream benchmark so far (SEN1Floods11 flood segmentation: test IoU 0.598 with a light head, S1-only, center-crop protocol); no head-to-head against other SAR foundation models under a shared protocol yet.

Citation

If you use Radiance in research or a product, please cite:

@misc{stratus-labs-radiance-mae-base-v1,
  title = {Radiance MAE Base v1: an open Sentinel-1 SAR foundation model},
  author = {Stratus Labs},
  year = {2026},
  publisher = {HuggingFace},
  howpublished = {\url{https://huggingface.co/stratus-labs/radiance-mae-base-v1}},
}

License

CC-BY-4.0. Weights, code, and model card are free to use with attribution. SSL4EO-S12 dataset used for pretraining carries its own license — see wangyi111/SSL4EO-S12.

About Stratus Labs

Stratus Labs ships open-weight foundation models for the natural world. First release: Nocturne (non-bird bioacoustic classifier). Second: Radiance (SAR foundation model, this repo).

Research page

The write-up and every Radiance release, in one place: https://runstratus.com/research/radiance-mae-base-v1-sentinel1-sar