Gulf Watch compresses 40 years of Gulf of Mexico sea surface height (SSH) data into a real-time hurricane rapid-intensification prediction dashboard.
The core insight: Loop Current Eddy (LCE) separation events precede rapid intensification in Atlantic hurricanes. When a warm-core eddy detaches from the Loop Current, it deposits a reservoir of deep ocean heat directly in hurricane corridors — the same mechanism that fueled Katrina, Rita, and Ida.
We trained a two-stage spatial-temporal model on ECCO's 40-year Gulf simulation (~14,600 daily SSH fields). A CNN encoder extracts spatial structure from each daily SSH anomaly field, then an LSTM classifier reads 30-day sequences of those features to predict LCE separation probability at t+7 and t+30 days — before a hurricane even forms. Labels are self-supervised: no external storm catalog, just physics (SSH anomaly > 17 cm + sharp LC-zone gradient drop).
The dashboard lets you drag an SST warming slider (+0 to +4°C) and watch the RI heatmap intensify and events-per-year counter climb — making abstract climate risk tangible to a non-technical audience.
Stack: PyTorch CNN+LSTM on Databricks, FastAPI serving endpoint, Next.js + deck.gl frontend, ECCO Gulf SSH dataset (Scripps/UCSD).
Log in or sign up for Devpost to join the conversation.