Our paper “Multimodal Deep Learning for Flood Detection: Integrating Water Level Sensors, Rainfall, Cameras, and Satellite Imagery” has been accepted for publication in Future Generation Computer Systems (Elsevier), in the special issue on Distributed Intelligence for Natural Disaster Management.
The paper presents an end-to-end multimodal AI framework that fuses hydrometer, rainfall, camera, and satellite data in a single temporal modeling pipeline. Visual observations are converted into quantitative time-series flood indicators so the model can learn jointly from physical and visual signals, and a reasoning layer enforces logical flood-stage progression. Multimodal integration improves classification accuracy by over 16 percentage points compared to unimodal baselines at the shortest forecast horizon, and the reasoning layer reduces stage oscillations by 25% while also improving accuracy. Validated on real-world field data without retraining, the model reliably distinguishes calm from flood conditions and escalates its forecast appropriately during a genuine flood event.
Authors: Ipshita Ahmed Moon, Jack Barbieri, Alfred Kalyanapu, and Susmit Shannigrahi (Tennessee Technological University).