Al-Rawas et al. (2026) Leveraging Machine Learning Flood Forecasting: A Multi-Dimensional Approach to Hydrological Predictive Modeling

Identification

• Journal: Water
• Year: 2026
• Date: 2026-01-12
• Authors: Ghazi Al-Rawas, Mohammad Reza Nikoo, Nasim Sadra, Malik Al-Wardy
• DOI: 10.3390/w18020192

Research Groups

• Department of Civil and Architectural Engineering, Sultan Qaboos University, Muscat, Oman
• School of Mathematical and Computational Sciences, Massey University, Palmerston North, New Zealand
• Center for Environmental Studies and Research, Sultan Qaboos University, Muscat, Oman

Short Summary

This study introduces an LSTM model with a customized loss function and wavelet decomposition to accurately predict extreme rainfall events in Oman's Al-Batina region, demonstrating superior performance over traditional models and quantifying predictive uncertainty using a Bayesian MCMC framework.

Objective

• To develop and evaluate a custom extreme loss function specifically designed for LSTM networks to improve accuracy in predicting extreme rainfall events in Oman’s Al-Batina region, without sacrificing overall model performance, and to quantify predictive uncertainty using a Bayesian MCMC framework.

Study Configuration

• Spatial Scale: Al-Batinah area, northeastern coast of Oman, approximately 7979 km².
• Temporal Scale: Data from 2006 to 2020, aggregated to monthly resolution. Lagged features at 1, 3, and 7 months/days; rolling averages of 7- and 30-day intervals.

Methodology and Data

• Models used:
  • Primary: Custom-built two-layer Long Short-Term Memory (LSTM) network with a weighted quantile loss function.
  • Baseline: Random Forest (RF), Support Vector Machine (SVM), Artificial Neural Network (ANN).
  • Hybrid: LSTM-RF ensemble.
  • Uncertainty Quantification: Metropolis–Hastings Markov Chain Monte Carlo (MCMC).
  • Feature Engineering: Discrete Wavelet Transform (DWT) using Daubechies-4 (db4) wavelet.
• Data sources:
  • Precipitation (Rainfall): In situ measurements from the Ministry of Regional Municipalities and Water Resources.
  • Land Surface Temperature (LST): MODIS/Terra MOD11A2 (NASA LP DAAC).
  • Vegetation Index (NDVI): MODIS/Terra MOD13A1 (NASA LP DAAC).
  • Elevation (DEM) and Derived Topography: NASADEM (NASA JPL).
  • Soil properties (sand, silt, clay fractions, bulk density, gravel content): Derived from FAO-HWSD, Saxton et al. (1986), Wieder et al. (2014), and SoilGrids dataset (ISRIC).
  • Proximity metrics: Distance to infrastructure (e.g., nearest road distance).
  • Derived variables: Water Stress Index (WSI), lagged features, rolling means.

Main Results

• The custom loss LSTM model achieved the best performance: Mean Absolute Error (MAE) = 0.0222 mm/day, Root Mean Squared Error (RMSE) = 0.1098 mm/day, Coefficient of Determination (R²) = 0.8068, and Symmetric Mean Absolute Percentage Error (SMAPE) = 7.62%.
• The LSTM + RF ensemble model also performed well (MAE = 0.0275 mm/day, RMSE = 0.1125 mm/day, R² = 0.7971, SMAPE = 13.65%).
• A Kruskal–Wallis H-test confirmed a highly significant difference in prediction errors among models (H = 14,021.12, p < 0.001), indicating the custom LSTM's superior performance.
• The custom weighted loss function significantly outperformed standard loss functions (MSE, MAE, Quantile Loss q=0.9) in forecasting rainfall, particularly for extreme events.
• Bayesian MCMC framework quantified predictive uncertainty, yielding a posterior mean for the standard deviation parameter (σ) of 0.118 with a 95% credible interval of [0.117, 0.120].
• Sensitivity analysis identified rainfall and Land Surface Temperature (LST) as the most influential predictors, while NDVI had a comparatively lower impact.
• Wavelet decomposition provided multi-scale insights, revealing high-frequency variations (short-term disturbances) and long-term trends (e.g., constant decrease in water stress, changing rainfall patterns).

Contributions

• Development and evaluation of a novel custom extreme loss function for LSTM networks, specifically designed to prioritize accurate prediction of rare, high-impact extreme rainfall events in arid/semi-arid regions.
• Integration of multi-scale feature extraction using wavelet decomposition with dynamic (rainfall, LST, NDVI) and static (soil type, topography, infrastructure proximity) environmental variables.
• Quantification of predictive uncertainty using a Bayesian MCMC framework, providing probabilistic outputs and credible intervals for improved decision-making in flood risk management.
• Demonstrated superior performance of the custom LSTM model compared to traditional machine learning algorithms (RF, SVM, ANN) and an LSTM-RF ensemble, particularly in capturing extreme rainfall events.

Funding

• Sultan Qaboos University (SQU)
• Diwan of Royal Court
• His Majesty’s (HM) grant number SR/DVC/CESR/22/01

Citation

@article{AlRawas2026Leveraging,
  author = {Al-Rawas, Ghazi and Nikoo, Mohammad Reza and Sadra, Nasim and Al-Wardy, Malik},
  title = {Leveraging Machine Learning Flood Forecasting: A Multi-Dimensional Approach to Hydrological Predictive Modeling},
  journal = {Water},
  year = {2026},
  doi = {10.3390/w18020192},
  url = {https://doi.org/10.3390/w18020192}
}