PS12 · ISRO BAH 2026

INSAT-3DS Temporal Super-Resolution Dashboard

Problem Statement 12 — Enhance the temporal resolution of INSAT-3DS TIR1 imagery from 30 min to 15 min (and optionally 7.5 min) using deep-learning-based optical flow frame interpolation.

---

🛰 Live Interactive Dashboard

▶ https://insat-3-ds-temporal-super-resolution-lonh-7m5jx9kmz.vercel.app

Fully interactive — runs in any browser, no install needed.

Tab	What it shows
🖥 Side-by-side viewer	Original 30-min INSAT-3DS frames vs. AI-interpolated 15-min frames, animated with playback controls
🌀 Optical flow	Live vector field showing cloud motion estimated by the model between frames
📈 Training curves	Loss, SSIM, PSNR, and learning rate plotted over 100 training epochs
🗂 Source code browser	All 7 project files with syntax highlighting, copyable, in-browser
⚙ Config panel	Full hyperparameter, dataset, and inference configuration table

---

Deploy your own (Vercel — free, 1 click)

This dashboard is a pure static HTML file — no server, no Python, no build step.

1. Fork this repo
2. Go to vercel.com/new → Import the fork
3. Vercel auto-detects vercel.json → click Deploy
4. Done — public URL in ~30 seconds

vercel.json and .vercelignore are pre-configured in this repo.
The Python backend (backend/, model/, requirements.txt) is for local training only — Vercel ignores it automatically.

---

Project structure

INSAT3DS-Temporal-SuperResolution/
├── frontend/
│   └── index.html          ← Self-contained dashboard (the live site above)
├── backend/
│   ├── train.py            ← Training script (GOES-19 → model weights)
│   ├── dataset.py          ← GOES-19 (.nc) & INSAT-3DS (.h5) data loaders
│   ├── infer.py            ← Inference: generate synthetic .h5 frames
│   └── metrics.py          ← SSIM, PSNR, MSE, FSIM, motion centroid error
├── model/
│   └── model.py            ← FlowEstimator (RAFT-lite) + FrameSynthesizer (U-Net)
├── config.yaml             ← All hyperparameters and data paths
├── requirements.txt        ← Python dependencies (training only)
├── vercel.json             ← Static deployment config for Vercel
└── .vercelignore           ← Excludes Python backend from Vercel build

---

Quickstart (local training)

1. Install dependencies

pip install -r requirements.txt

2. Get training data (GOES-19)

# AWS S3 public bucket — no credentials needed
aws s3 sync s3://noaa-goes19/ABI-L1b-RadC/ /data/GOES19/ABI-L1b-RadC/ \
    --no-sign-request \
    --exclude "*" \
    --include "OR_ABI-L1b-RadC-M6C13_G19_s2024*"

3. Get inference data (INSAT-3DS)

Register at: https://mosdac.gov.in
Download:    Products → INSAT-3DS → L1B → TIR1 → 3SIMG_*_TIR1*.h5
Place in:    /data/INSAT3DS/TIR1/

4. Configure paths

Edit config.yaml:

data:
  train_root: /data/GOES19/ABI-L1b-RadC/
  insat_root: /data/INSAT3DS/TIR1/

5. Train

python backend/train.py --config config.yaml
# Checkpoints → ./checkpoints/best_model.pth
# TensorBoard → tensorboard --logdir logs/

6. Run inference on INSAT-3DS

# 15-min frames (one synthetic frame per 30-min pair)
python backend/infer.py --config config.yaml --t-factors 0.5

# 7.5-min frames (three synthetic frames per pair)
python backend/infer.py --config config.yaml --t-factors 0.25 0.5 0.75

# Output → ./outputs/synthetic_frames/3SIMG_AI_*_TIR1_*.h5

7. Open dashboard locally

open frontend/index.html   # macOS
# or double-click it — works in any browser, no server needed

---

Architecture

I₀ (T=00:00) ──┬──► FlowEstimator ──► F₀→ₜ, F₁→ₜ ──┬──► FrameSynthesizer ──► Iₜ (T=00:15)
               │     (RAFT-lite)                        │    (U-Net + visibility masks)
I₁ (T=00:30) ──┘                                       └── warp(I₀, F₀→ₜ), warp(I₁, F₁→ₜ)

Network 1 — FlowEstimator (RAFT-lite)

• Feature pyramid encoder (shared weights for I₀ and I₁)
• 4D correlation volume (patch matching between frames)
• GRU update block (4 iterative refinements)
• Outputs bi-directional flow fields F₀→₁ and F₁→₀

Network 2 — FrameSynthesizer (Super SloMo / U-Net)

• Backward-warps I₀ and I₁ to time t using scaled flow fields
• Predicts per-pixel visibility masks V₀, V₁ (occlusion handling)
• U-Net refinement with skip connections for fine detail
• Final blend: Iₜ = ((1−t)·V₀·I₀w + t·V₁·I₁w) / ((1−t)·V₀ + t·V₁)

Loss function

L = 0.84 × L1  +  0.12 × (1 − SSIM)  +  0.04 × Perceptual

---

Metrics

Metric	Description	Target
SSIM	Structural similarity ↑	> 0.90
PSNR	Peak signal-to-noise ↑	> 33 dB
MSE	Mean squared error ↓	< 0.005
FSIM	Feature similarity ↑	> 0.89
MCE	Cloud centroid error ↓	< 0.01

---

Domain transfer: GOES-19 → INSAT-3DS

Property	GOES-19 ABI Ch.13	INSAT-3DS TIR1
Wavelength	11.2 µm	10.2–11.2 µm
Spatial res.	2 km	4 km
Temporal res.	10 min	30 min
Coverage	GOES-East full disk	Indian region
Data format	NetCDF4 (.nc)	HDF5 (.h5)

The model trains on GOES-19 (abundant 10-min ground truth) and is applied to INSAT-3DS. The spectral overlap (~11 µm TIR in both) makes domain transfer viable; resolution differences are handled in normalised pixel space.

---

Citation

PS12 · ISRO Bharatiya Antariksh Hackathon 2026
Temporal Super-Resolution of INSAT-3DS TIR1 Imagery
Using Optical Flow Frame Interpolation

GitHub    : https://github.com/Apurba-06/INSAT3DS-Temporal-SuperResolution
Dashboard : https://insat-3-ds-temporal-super-resolution-lonh-7m5jx9kmz.vercel.app