satforensics-newcastle/satforensics-dinov2-probe

SatForensics: DINOv2 + Linear Probe for Satellite Image Manipulation Detection

This repository contains the trained linear probe, feature scaler, and cached features for the SatForensics project — an MSc dissertation project at Newcastle University exploring foundation models as forensic priors for detecting AI-generated manipulations in satellite imagery.

What's in this repo

models/

• dinov2_linear_probe.pkl — Trained scikit-learn LogisticRegression
• dinov2_feature_scaler.pkl — Fitted StandardScaler (apply BEFORE the probe)

features/

• dinov2_features.npz — Cached DINOv2 features for the training/test set
  • features: shape (468480, 768), float32 (1,830 images × 256 patches)
  • labels: shape (468480,), int64 (0 = authentic, 1 = manipulated)
  • image_ids: shape (468480,), int64 (maps each patch to its source image)

results/

• rsfake_fake_per_image_scores.csv — Per-image scores on RSFAKE-1M rsinpaint (n=2000)
• rsfake_fake_diffusion_sat_per_image_scores.csv — Same for DiffusionSat (n=2000)
• rsfake_scores_and_stats*.csv — Scores joined with brightness/std/edge_density stats
• rsfake_heatmaps_*_qualitative.png — 6-sample heatmap visualisations
• rsfake_confound_scatter_*.png — Score vs image-statistics scatter plots

Training setup

• Backbone: DINOv2-Base ViT (facebookresearch/dinov2_vitb14), frozen
• Input preprocessing: Resize to 224×224 (bicubic), ImageNet normalisation
• Patch grid: 16×16 = 256 patches per image (14px patches)
• Training data: Airbus satellite imagery + CLIP-guided Stable Diffusion 2 + ControlNet inpainting (Chapman et al. 2025 pipeline)
• Split: by-image, seed=42, 70/15/15 → 1,281 train / 274 val / 275 test
• Probe: scikit-learn LogisticRegression
  • class_weight='balanced', solver='lbfgs', max_iter=1000, random_state=42

Headline results

Metric	Value
Pooled test patch AUROC (Airbus, n=70,400 test patches)	0.9593
RSFAKE rsinpaint top-10 mean (n=2,000)	0.84
RSFAKE rsinpaint % above 0.7 (max)	95.7%
RSFAKE DiffusionSat top-10 mean (n=2,000)	0.86
RSFAKE DiffusionSat % above 0.7 (max)	97.2%
Confound checks (brightness/std/edge_density)	All |r| ≤ 0.21

Related work

• Chapman et al. (2025) — Geospatially-Aware Multi-Modal Fusion for Satellite Image Manipulation Detection. VCIP 2025.
• Oquab et al. (2023) — DINOv2: Learning Robust Visual Features without Supervision. arXiv:2304.07193.

Project

This work is part of an MSc dissertation at Newcastle University (Project 22: SatForensics), supervised by Dr. Deepayan Bhowmik.

The cross-dataset evaluation notebook is at: https://github.com/Om-Ravindra-Patil/satforensics-crossdataset

License

The trained probe and cached features are released for research use only. The underlying training imagery is licensed separately by Airbus.