TR174: HYDROSIGHT

NARENDRAN RAMASENDERAN CENTER FOR R&D OF IOT ASIA PACIFIC UNIVERSITY

Malaysia's flood warning system remains reactive: alerts are issued from single-source river gauge exceedances, at coarse spatial resolution, and typically only hours before inundation. HydroSight closes this gap with a near real-time, AI-driven national flood prediction and risk-monitoring platform that delivers district-level probabilistic forecasts at a 1-to-3-day lead time.

A modular seven-agent CrewAI pipeline orchestrates continuous dual-loop ingestion and fusion of four independent data sources: Sentinel-2 optical spectral indices (NDWI, MNDWI, NDVI, SWIR), Sentinel-1 SAR cloud-penetrating water segmentation, live water levels and trends from 530 Department of Irrigation and Drainage gauges, and dynamic weather forecasts. These resolve into a single 13-feature vector for each of Malaysia's 263 administrative districts. Prediction is performed by a weighted four-model ensemble Random Forest (30%), CNN-LSTM (30%), XGBoost (20%), Multilayer Perceptron (20%) bounded by a physics-based, terrain-scaled rainfall hazard floor that prevents the ensemble from under-forecasting extreme events outside its training distribution.

Retrospective backtesting on the December 2014, 2021 and 2024 Malaysian floods yielded event-independent generalisation at a mean AUC-ROC of 0.925, including 1.00 recall two days before the peak of the 2021 Klang Valley disaster. A prospective 38-day live field trial scoring 9,788 district-days returned a mean Brier score of 0.0871 against a climatological reference of 0.25, a Brier Skill Score of 0.65 with an independent append-only forecast ledger correlating at r = 0.910 with post-event observations.

Outputs reach the public through an interactive Leaflet.js dashboard combining district risk choropleths with FloodSENS satellite-verified inundation polygons, a server-side "what-if" rainfall simulator, and a Groq-hosted Llama 3.1 8B agent generating localised situation reports in English, Bahasa Melayu, Chinese and Tamil. The full system runs 24/7 on a single AWS EC2 instance in ap-southeast-5 at USD 2.40/day, national coverage at roughly the cost of a coffee.