Latif et al. (2026) Novel advances in real-time pluvial flash flood forecasting under climate change through combination of various machine learning models

Identification

• Journal: Theoretical and Applied Climatology
• Year: 2026
• Date: 2026-01-30
• Authors: Sarmad Dashti Latif, Wong Hei Man Anson, Ali Najah Ahmed, Ahmed El-Shafie
• DOI: 10.1007/s00704-026-06037-w

Research Groups

• Civil Engineering Department, College of Engineering, Komar University of Science and Technology, Sulaimany, Kurdistan Region, Iraq
• Department of Civil Engineering, Faculty of Engineering, University of Malaya (UM), Kuala Lumpur, Malaysia
• School of Engineering and Technology, Sunway University, Bandar Sunway, Petaling Jaya, Malaysia
• National Water and Energy Center, United Arab Emirates University, Al Ain, United Arab Emirates

Short Summary

This study advances real-time pluvial flash flood forecasting under climate change by evaluating and combining various machine learning models in Malaysia. It found that integrating a Random Forest model with a Long Short-Term Memory network achieved a higher prediction accuracy of 0.65 compared to a Support Vector Regression and LSTM combination.

Objective

• To investigate advancements in real-time pluvial flash flood forecasting (PFFF) to enhance flood mitigation capabilities in the context of climate change.
• To evaluate the latest flash flood (FF) prediction models developed in Malaysia and assess their effectiveness, accuracy, and applicability to local conditions.
• To investigate the use of Support Vector Regression (SVR) and Random Forest (RF) for FF prediction and analyze their performance and potential improvement methods, especially when combined with Long Short-Term Memory (LSTM).

Study Configuration

• Spatial Scale: Petaling Jaya, Petaling district, Selangor, Malaysia.
• Temporal Scale: Daily rainfall and precipitation data from January 2004 to December 2023 (20 years). Real-time forecasting.

Methodology and Data

• Models used: Support Vector Regression (SVR) with Radial Basis Function (RBF) kernel, Random Forest (RF), Long Short-Term Memory (LSTM). Hybrid models combining LSTM with SVR and LSTM with RF.
• Data sources: Rainfall and precipitation data from the Malaysian Meteorological Department. Historical flood event information gathered from old news reports.

Main Results

• Initial comparison showed Random Forest (RF) outperformed Support Vector Regression (SVR) across most metrics: RF accuracy was 0.5 compared to SVR's 0.25; RF F1 score was 0.5 compared to SVR's 0.1; RF Log Loss was 0.76 compared to SVR's 0.96; RF Matthews Correlation Coefficient (MCC) was 0.33 compared to SVR's 0. Both models had a Receiver Operating Characteristic Area Under Curve (ROC AUC) of 0.33.
• Applying 5-fold cross-validation significantly improved model performance:
  • RF accuracy increased by 0.25 to 0.8.
  • SVR accuracy increased by 0.20 to 0.45.
  • RF F1 score increased by 0.30 to 0.693.
  • SVR F1 score increased by 0.24 to 0.34.
  • RF ROC AUC improved by 0.50 to 0.767.
  • SVR ROC AUC improved by 0.10 to 0.433.
  • RF MCC increased by 0.18 to 0.515.
  • SVR MCC declined by 0.05 to -0.048.
• The combination of LSTM with RF achieved a higher flash flood prediction accuracy of 0.65.
• The combination of LSTM with SVR achieved a lower flash flood prediction accuracy of 0.47.
• The RF-LSTM model demonstrated higher sensitivity (0.60) and specificity (0.71) compared to the SVR-LSTM model (sensitivity 0.40, specificity 0.57).

Contributions

• Novel integration of two machine learning models (SVR and RF) with a deep learning model (LSTM) to enhance the accuracy of real-time pluvial flash flood prediction.
• Comprehensive evaluation of the effectiveness, accuracy, and applicability of these advanced machine learning techniques for flood forecasting in Malaysian local conditions.
• Validation of Random Forest's superior ability to capture complex interactions between weather, hydrology, and landscape for flash flood forecasting, especially when augmented with LSTM for temporal feature extraction.

Funding

Not applicable.

Citation

@article{Latif2026Novel,
  author = {Latif, Sarmad Dashti and Anson, Wong Hei Man and Ahmed, Ali Najah and El-Shafie, Ahmed},
  title = {Novel advances in real-time pluvial flash flood forecasting under climate change through combination of various machine learning models},
  journal = {Theoretical and Applied Climatology},
  year = {2026},
  doi = {10.1007/s00704-026-06037-w},
  url = {https://doi.org/10.1007/s00704-026-06037-w}
}