Satellite Frame Interpolation

Deep-learning pipeline for enhancing temporal resolution of geostationary satellite imagery.
Built by Aditya Karale & Om Pandey.
A 2-stage coarse-to-refined architecture for high-fidelity cloud motion synthesis in severe weather tracking.

Results

Validated on 70 unseen test triplets.

Metric	Model 1 (Coarse)	Model 1 + 2 (Refined)	Gain
SSIM ↑	0.886	0.900	+0.014
PSNR ↑	28.18 dB	29.59 dB	+1.41 dB
MSE ↓	0.0047	0.0042	−10.6%
MAE ↓	0.0368	0.0316	−14.1%

Architecture

Sequential refinement pipeline to minimize motion artifacts:

Frame t₀ ──┐
            ├──▶  Model 1: IFNet  ──▶  Coarse Frame  ──▶  Model 2: RefinementNet  ──▶  Final Frame
Frame t₁ ──┘

Model 1 — IFNet: Optical flow estimation + coarse intermediate frame synthesis.
Model 2 — RefinementNet: U-Net with CBAM attention; corrects residual artifacts in the coarse output.

Training

Three-stage pipeline. Each stage builds on the previous.

Stage	Focus	LR	Epochs	Status
1	Optical Flow Training	1e-4	100	✅ Done
2	Refinement Network	2e-4	50	✅ Done
3	End-to-End Fine-tuning	1e-5	50	✅ Done

Repository

├── models/
│   ├── ifnet.py              # IFNet optical flow architecture
│   └── refinement_unet.py    # U-Net + CBAM refinement network
├── scripts/
│   ├── ingest.py             # AWS S3 data ingestion
│   └── preprocess.py         # Normalization and augmentation
├── assets/
│   └── metrics_comparison.png
├── evaluate.py               # SSIM/PSNR/MSE/MAE evaluation suite
└── satellite_frame_interpolation.ipynb  # Full training + eval notebook