README.md

metadata

library_name: pytorch
pipeline_tag: feature-extraction
tags:
  - synthetic-aperture-radar
  - sar
  - remote-sensing
  - self-supervised-learning
  - masked-image-modeling
  - jepa
  - vision-transformer

phi-JEPA: Where and What to Reconstruct in Physics-Informed Radar Pre-training?

This is the official model repository for phi-JEPA, a physics-informed masked autoencoder for self-supervised synthetic aperture radar (SAR) image pre-training.

Release status. This repository currently provides the pretrained checkpoint used in our main experiments. The pre-training code and downstream evaluation code are not included in the current release.

Model Description

Existing masked image modeling methods for SAR imagery commonly rely on random masking and low-level reconstruction targets. phi-JEPA introduces two SAR-oriented designs:

1. Scattering-Center-Aware Masked Modeling (SCM). A physical scattering prior reconstructed from attributed scattering centers guides masking toward target-related regions while maintaining a high overall masking ratio.
2. SAR-Specific Semantic Feature Prediction (SFP). A frozen SARCLIP ViT-B/16 image encoder provides patch-level semantic targets, shifting the pre-training objective from low-level appearance reconstruction to semantic representation learning.

The released ViT-B/16 checkpoint is pretrained on ATRNet-STAR and is intended for downstream SAR representation transfer.

Released Checkpoint

Model	Backbone	Pre-training dataset	Epochs	File	SHA-256
phi-JEPA	ViT-B/16	ATRNet-STAR	100	phi_jepa_vitb16_e100.pth	f94f048366528815c15a51a060ea27d923db65ae2c6434931121cfd7af72f095

Download

Using the Hugging Face CLI:

hf download kiki-orb/phi-JEPA \
  main_experiments/phi_jepa_vitb16_e100.pth \
  --local-dir ./checkpoints

Using Python:

from huggingface_hub import hf_hub_download

checkpoint_path = hf_hub_download(
    repo_id="kiki-orb/phi-JEPA",
    filename="main_experiments/phi_jepa_vitb16_e100.pth",
)
print(checkpoint_path)

Pre-training Configuration

Configuration	Value
Backbone	ViT-Base
Patch size	16 x 16
Pre-training dataset	ATRNet-STAR (SOC, 40 vehicle categories)
Optimizer	AdamW
Base learning rate	1e-3
Weight decay	0.05
Batch size	128
Pre-training epochs	100
Warmup	20 epochs
Learning-rate schedule	Cosine annealing
Overall masking ratio	80%
Target-region masking ratio	50%
Feature alignment objective	Cosine loss
Semantic teacher	Frozen SARCLIP ViT-B/16 image encoder