HaorFloodAlert: Deseasonalized ML Ensemble for 72-Hour Flood Prediction in Bangladesh Haor Wetlands

2026-05-19 · Source: Artificial Intelligence · Field: Science & Research — Environmental Science & Earth Systems, Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, quick

Summary

HaorFloodAlert is a deseasonalized machine learning ensemble designed for 72-hour flash flood prediction in Bangladesh's haor wetlands, specifically for the Sunamganj Haor (approximately 8,000 km2). Developed to address the failure of existing riverine flood models in flat basin dynamics, it prevents damage to the annual boro rice harvest. The system incorporates an upstream Barak River Sentinel-1 SAR proxy from Silchar, Assam, providing about 36 hours of lead time, with Otsu-thresholded SAR change detection validating at 84-91 percent spatial match. A key innovation was identifying and correcting for temperature's seasonal "cheat code," which inflated accuracy by 6.9 pp. The operational ensemble, weighted RF 0.5625 + XGBoost 0.4375, achieves 89.6 percent LOOCV accuracy, 87.5 percent recall, and 0.943 AUC-ROC on 77 real Sentinel-1 events. It also includes a three-tier alert pipeline and a BRRI-calibrated boro rice damage estimator.

Key takeaway

For hydrologists or disaster management teams developing flood early warning systems in complex, non-riverine basins, HaorFloodAlert demonstrates a critical approach. You should prioritize deseasonalizing input features to avoid misleading accuracy metrics and integrate upstream satellite SAR data for crucial lead time. Consider an ensemble of robust ML models like RF and XGBoost to achieve high accuracy and recall, enhancing your system's reliability for critical agricultural protection.

Key insights

Deseasonalized ML ensembles and SAR proxies accurately predict flash floods in complex, flat wetland environments.

Principles

• Deseasonalize features to prevent spurious accuracy inflation.
• Upstream SAR data provides critical lead time for flood prediction.
• Ensemble models improve predictive robustness.

Method

Combine Random Forest and XGBoost with deseasonalized features and an upstream Sentinel-1 SAR proxy for 72-hour flood probability forecasting. Validate with Otsu-thresholded SAR change detection.

In practice

• Implement deseasonalization for time-series ML models.
• Integrate satellite SAR data for early warning systems.
• Use ensemble weighting for improved model performance.

Topics

• Flash Flood Prediction
• Machine Learning Ensemble
• Sentinel-1 SAR
• Haor Wetlands
• Disaster Management
• Deseasonalization

Best for: AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• The next chapter in flood resilience: Open sourcing Google’s hydrology framework — Open-sourcing Google's AI hydrology framework democratizes advanced flood forecasting, enabling local adaptation and global resilience.
• Beyond MSE: Improving Precipitation Nowcasting with Multi-Quantile Regression — Multi-quantile regression improves precipitation nowcasting by reducing MSE and providing risk-sensitive heavy rainfall predictions.
• Doubly Outlier-Robust Online Infinite Hidden Markov Model — Doubly robust online iHMMs combine observation-space and state-space robustness to mitigate outlier sensitivity and spurious regime creation.
• EpiEvolve: Self-Evolving Agents for Streaming Pandemic Forecasting under Regime Shifts — LLM forecasters can adapt to streaming data and regime shifts via memory-based self-evolution, not parameter updates.
• Hybrid NARX-LLM for Greenland Iceberg Discharge: Prompt-Driven Residual Correction — A hybrid NARX-LLM framework uses physics-informed prompts to correct iceberg discharge predictions, enhancing accuracy for complex climate dynamics.
• Advanced Flood Prediction with Physics-Guided Deep Learning: Combining UNet, FNO, and SAR/Optical Imagery — Embedding hydrodynamic principles into deep learning models significantly improves flood prediction accuracy and physical consistency.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.