Chase the Cloud Challenge '26 (ISRO)

MeghDrishti AI — Diffusion-based INSAT cloud nowcasting for Chase the Cloud Challenge '26 (ISRO). LINKS • GitHub: https://github.com/Dynamicsoumya/Meghdrishtiai • API health: https://meghdrishtiai.onrender.com/health • Rubric map: docs/CHALLENGE_COVERAGE.md • Eval report: docs/evaluation_report.md • Dataset decisions: docs/dataset_card.md WHAT WE BUILT • 6 historical multi-channel frames (VIS/TIR/MIR/WV, 3 hours) → ≥2 future frames (t+30, t+60) • Conditional residual latent diffusion + ConvLSTM temporal encoder • Baselines: Persistence, Optical Flow, ConvLSTM • Metrics: MAE, RMSE, PSNR, SSIM, Cloud IoU/CSI, motion error • Dashboard: Time Machine, Reveal Actual, motion vectors, Storm Development Index, Futures wall, Judge Demo REGION SCOPE Bay of Bengal East Coast crop (lat 10–22N, lon 80–95E) — active convection + compute-aware; challenge allows region selection. Pipeline supports other MOSDAC crops. DATA NOTE MOSDAC INSAT-3D/3DR/3DS path documented (preprocessing/download.py). Demo bootstrap uses documented synthetic INSAT-like sequences so reviewers can run without MOSDAC credentials. LOCAL SETUP python -m venv .venv .venv\Scripts\activate # Windows pip install -r requirements.txt python scripts/generate_synthetic_data.py python inference/predict.py uvicorn dashboard.backend.main:app --port 8000 cd dashboard/frontend && npm i && npm run dev → http://localhost:3000 DASHBOARD SHORTCUTS D = Judge Demo | C = Cinema Mode | / = Commands | Reveal Actual = AI vs real wipe SCIENCE FRAMING Predicts satellite cloud evolution during INSAT gaps. Storm Development Index is decision-support only — NOT a lightning/IMD warning product. ONE SENTENCE The moment my model successfully predicted where the clouds would move next was when the Time Machine scrubber crossed NOW and the AI forecast continued the same convective cell northeast — then Reveal Actual confirmed the core stayed aligned.