MekongMamba: A Multi-Phase System for Compound Flood EarlyWarning in the Mekong Delta

The Mekong Delta is home to 18 million people and produces over 90\% of Vietnam's rice exports, yet no operational system delivers commune-level flood warnings at the 24-48 hour lead times needed for effective evacuation. We propose MekongMamba, a multi-phase system built on two properties of the delta: first, several days of discharge travel time from the Kratie gauge (Cambodia) make the riverine boundary effectively observed at a 48-hour horizon; second, astronomical tides are analytically predictable. Phase 1 pairs two small sequence models that turn shelf weather into coastal surge and fuse that surge with tides, observed levels, upstream discharge, and numerical weather prediction (NWP) rainfall into calibrated probabilistic 48-hour water levels at the delta's stations. Phase 2 converts station forecasts into 100m probabilistic inundation maps via a height-above-nearest-drainage surface sharpened by per-cell flood-onset statistics mined from ten years of Sentinel-1 imagery. Phase 3 is a graph model using the delta's sluice-gate and canal topology to refine the inundation maps. We estimate that the deployable system totals under 5M parameters, which can be deployed and operated on CPU hardware that Vietnam's National Centre for Hydro-Meteorological Forecasting (NCHMF) already operates, and delivers warnings through an alert network reaching over 3,000 communes.