Socials

SatFrame AI

Geostationary Imagery · Deep Learning

PERFORMANCE METRICS

Does it work?

SSIM
0.0000
Coarse: 0.8863 → Refined: 0.9001 +0.0138
How similar the generated frame looks to the real one (higher = better)
PSNR
0.00 dB
Coarse: 28.18 → Refined: 29.59 +1.41 dB
Signal quality of the prediction (higher = better)
MSE
0.0000
Coarse: 0.0047 → Refined: 0.0042 -10.6%
Average pixel error (lower = better)
MAE
0.0000
Coarse: 0.0368 → Refined: 0.0316 -14.1%
Consistent pixel accuracy across the full frame (lower = better)
VISUAL PROOF

Show me proof

Drag to compare model output vs reality

Input Frame t0
Input Frame t₀
Actual t0.5
Predicted t0.5
Predicted t₀.₅ Actual t₀.₅
PERFORMANCE INDEX

How consistent is it?

PIPELINE DESIGN

How does it work?

Frame t₀
Frame t₁
IFNet
(Model 1)
300,943 params | Optical Flow Estimation | LR 1e-4 | 100 epochs
Coarse
Frame
RefinementNet
(Model 2)
8,139,698 params | U-Net + CBAM Attention | LR 2e-4 | 50 epochs
Final
Interpolated Frame

Stage 1: Optical Flow

  • Model IFNet
  • Epochs 100
  • Learning Rate 1e-4
  • Loss Flow Consistency

Stage 2: Refinement Network

  • Model RefinementNet
  • Epochs 50
  • Learning Rate 2e-4
  • Loss Charbonnier + Census

Stage 3: E2E Fine-tuning

  • Model Full Pipeline
  • Epochs 50
  • Learning Rate 1e-5
  • Loss Combined Loss
THE PROJECT

Who built this?

Built by Aditya & Om.

A two-stage deep learning pipeline that synthesizes intermediate satellite frames between two consecutive inputs — effectively doubling the temporal resolution of geostationary weather imagery without new hardware.

Trained on GOES-16. Tested on INSAT-3D.

  • Aditya Karale
    Frontend Dev, Data Collection & Data Ingestion, End-to-End Training Pipeline & Model Fine-tuning
  • Om Pandey
    Data Collection & Data Ingestion, Preprocessing Pipeline & ML Model Architecture
View Repository