ltdung/limodenet

Attention-Free Compact Encoders for Information-Preserving
Onboard Satellite Image Restoration

Paper (arXiv) · Code (GitHub) · SENTRY project

 

LIMODENet is a compact (0.69M-parameter), softmax-/QKV-free convolutional backbone for restoring channel-degraded Earth-observation imagery, designed under the constraint that it must also run as a spiking network on neuromorphic accelerators (BrainChip Akida, Intel Loihi-2). We do not claim it is the best restorer available — two modern CNN restorers (NAFNet, Restormer) beat it on raw fidelity. What we show is that it is the best restorer that is verifiably deployable where the power budget actually is: it converts end-to-end to a spiking network with zero blocked operations, while the fidelity-winning competitors have 22–24 each.

Code: pip install git+https://github.com/ltdung/limodenet

import torch
from limodenet.recon import build_recon
from huggingface_hub import hf_hub_download

path = hf_hub_download("ltdung/limodenet", "limodenet-skip-ae-1db100q.pth")
model = build_recon("limodenet_skip")
model.load_state_dict(torch.load(path))
restored = model(degraded_images)   # (N, 3, 128, 128), values in [0, 1]

▶ Try it

• Interactive demo (Space): compare eight restorers on real 1 dB DVB-S2X images, run the spiking-conversion audit, measure the budget.
• Watch the spiking network: spikes, membrane potentials and the restoration building up over 8 timesteps, recorded from limodenet-spiking-skip-recon-t8.pth.

Checkpoints

File	Task	Params	Headline number
limodenet-nano-eurosat.pth	classification	0.39M	95.65 ± 0.23% top-1 (EuroSAT, scratch, 3-seed)
limodenet-small-eurosat.pth	classification	1.55M	96.93 ± 0.16%
limodenet-base-eurosat.pth	classification	2.73M	97.26 ± 0.17%
limodenet-tiny-eurosat-tinft.pth	classification	0.69M	98.26% (Tiny-ImageNet pretrained + fine-tuned)
limodenet-big-eurosat-tinft.pth	classification	5.80M	off-ladder legacy config, Tiny-ImageNet FT
limodenet-ae-1db100q-60ep.pth	restoration	0.77M	37.39 ± 0.14 dB PSNR, 1dB/100q DVB-S2X (this seed: 37.22)
limodenet-ae-1db100q.pth	restoration	0.77M	earlier 25-epoch run (this seed: 35.68 dB); kept for compatibility
limodenet-skip-ae-1db100q.pth	restoration	0.78M	38.07 ± 0.26 dB PSNR, same condition
limodenet-spiking-classifier-t8.pth	classification, spiking	0.69M	93.55 ± 0.54% top-1, T=8, ANN-init
limodenet-spiking-skip-recon-t8.pth	restoration, spiking	0.78M	35.95 ± 0.05 dB PSNR, T=8, zero blocked ops
baseline-cnn-ae-1db100q.pth	restoration, comparison	0.75M	CNN-AE bottleneck peer (this seed: 35.84 dB)
baseline-unet-1db100q.pth	restoration, comparison	0.75M	U-Net skip reference (this seed: 36.70 dB)
baseline-dncnn-1db100q.pth	restoration, comparison	0.74M	DnCNN (this seed: 35.49 dB)
baseline-nafnet-lite-1db100q.pth	restoration, comparison	0.75M	NAFNet-lite, unconstrained (this seed: 38.34 dB)
baseline-restormer-lite-1db100q.pth	restoration, comparison	0.77M	Restormer-lite, unconstrained (this seed: 39.49 dB)

All classification numbers are on EuroSAT (10-class RGB land-use). All restoration numbers are on 1 dB $E_s/N_0$, JPEG quality 100 DVB-S2X-degraded EuroSAT, three-seed mean ± std unless noted. seed 0 of each 3-seed run is released; see the paper for the full 3-seed statistics.

The onnx/ folder holds ONNX exports of the eight restorers, used by the in-browser demo. They are checked against these PyTorch checkpoints by demo/tools/export_web.py (identical PSNR/SSIM on the demo's 30 samples; the spiking model, whose spikes are sensitive to float rounding, agrees within 0.02 dB).

Loading

# Classifier
from limodenet.family import build
model = build("tiny", num_classes=10)
model.load_state_dict(torch.load("limodenet-tiny-eurosat-tinft.pth"))

# Reconstructor
from limodenet.recon import build_recon
model = build_recon("limodenet_skip")   # or "limodenet" + limodenet-ae-1db100q-60ep.pth
model.load_state_dict(torch.load("limodenet-skip-ae-1db100q.pth"))

# Comparison restorers (Table tab:modern)
from limodenet.baselines import build_baseline
model = build_baseline("unet")          # cnn | unet | dncnn | nafnet | restormer
model.load_state_dict(torch.load("baseline-unet-1db100q.pth"))

# Spiking classifier (needs `pip install snntorch`)
from limodenet.snn import SpikingLIMODENet
model = SpikingLIMODENet(num_classes=10, T=8)
model.load_state_dict(torch.load("limodenet-spiking-classifier-t8.pth"))

# Spiking reconstructor
from limodenet.snn_recon import SpikingLIMODENetSkipRecon
model = SpikingLIMODENetSkipRecon(T=8)
model.load_state_dict(torch.load("limodenet-spiking-skip-recon-t8.pth"))

License

Weights: CC-BY-NC-4.0 (non-commercial). Code: MIT. See the GitHub repo's WEIGHTS_LICENSE.md for details, or contact the authors for a commercial license.

Part of the SENTRY project

LIMODENet is the foundational task of SENTRY, a project of the SIGCOM group at SnT, University of Luxembourg, funded by the Luxembourg National Research Fund (FNR), combining satellite communication networks with machine learning for Earth observation. Other tasks: github.com/ltdung/SnT-SENTRY

Citation

@article{le2026limodenet,
  title   = {LIMODENet: Attention-Free Compact Encoders for Information-Preserving
             Onboard Satellite Image Restoration},
  author  = {Le, Thanh-Dung and Ha, Vu Nguyen and
             Nguyen, Ti Ti and Chatzinotas, Symeon},
  journal = {arXiv preprint arXiv:2609.14690},
  year    = {2026}
}