Sathiyamoorthy et al. (2026) STORM-Net for urban flood risk prediction: an AI-based spatiotemporal tracking and mapping approach

Identification

• Journal: Stochastic Environmental Research and Risk Assessment
• Year: 2026
• Date: 2026-04-01
• Authors: M Sathiyamoorthy, P. Subramanian, N. V. S. Sree Rathna Lakshmi, S. Hemalatha, L. Nagarajan, S. Sathish, M. Gnana Prakash
• DOI: 10.1007/s00477-026-03186-2

Research Groups

• Department of Computer Science and Engineering, Saveetha School of Engineering, Saveetha Institute of Medical and Technical Sciences, Chennai, India
• Department of Electronics and Communication Engineering, SRM Institute of Science and Technology, Ramapuram, Chennai, India
• Department of Electrical and Electronics Engineering, St. Joseph’s Institute of Technology, OMR, Chennai, India
• Department of Biotechnology, Aarupadai Veedu Institute of Technology, Vinayaka Missions Research Foundation (DU), Paiyanoor, Chengalpattu, India
• Department of Chemical Engineering, Sathyabama Institute of Science and Technology, Chennai, India
• Department of Information Technology, Sri Sairam Institute of Technology, Chennai, India

Short Summary

This paper proposes STORM-Net, a novel hybrid AI-based spatiotemporal deep learning model, for high-precision urban flood risk prediction. It integrates SAFER for intelligent feature elimination and BRAVE for adaptive attention scaling, achieving superior accuracy and computational efficiency compared to existing models across diverse datasets.

Objective

• To develop a sound spatiotemporal deep learning model (STORM-Net) that can effectively forecast urban flood risk by incorporating heterogeneous urban data, spatial interdependencies, and temporal dynamics for different geographic domains.
• To develop and deploy an intelligent feature reduction model (SAFER) using Sparse Autoencoder-guided Reinforcement Learning to remove redundant and less-informative features, thereby avoiding overfitting and enhancing model generalization.
• To introduce the BRAVE optimization module (Builder–Fairy Refined Attention Value Estimator) that learns and updates attention scaling factors during training, improving sensitivity to spatiotemporal flood features and predictive interpretability.
• To validate the performance of the hybrid STORM-Net model with robust experimental testing on real-world flood data, comparing it to traditional and hybrid current models in terms of accuracy, F1-score, and computational cost, aiming for at least 15–20% better performance.

Study Configuration

• Spatial Scale: City-scale urban areas, districts, specific buildings (fine-grained), major cities in India (e.g., Mumbai, Chennai, Kolkata), and over 40 weather stations across Australia.
• Temporal Scale: Hourly, daily, short-term, long-term, and seasonal monsoon rainfall patterns.

Methodology and Data

• Models used:
  • Proposed: STORM-Net (Spatiotemporal Tracking and Observation for Resilient Mapping), SAFER (Sparse Autoencoder-guided Feature Elimination with Reinforcement), BRAVE (Builder–Fairy Refined Attention Value Estimator).
  • Comparison: CNN, LSTM, GRU, BiLSTM, Transformer, ResNet+LSTM, BiLSTM+GRU, Transformer+GRU, CNN+LSTM, Random Forest, SVM, CapsNet, CatBoost, XGBoost, LightGBM, AdaBoost.
• Data sources:
  • Kaggle Urban Flood Prediction Dataset: Tabular data (50,000 instances, 21 numeric attributes) including rainfall (mm), water level, humidity, temperature, wind speed, barometric pressure, basin runoff, and binary flood labels.
  • UCI Rainfall in Australia Dataset: Historical weather information (over 145,000 records from 40+ stations) including daily rain (mm), temperature, evaporation, sunshine, wind gust direction and speed (km/h), humidity, pressure (hPa), cloud cover, and a binary 'RainTomorrow' target.
  • Multi-source data: Sensor networks, hydrological observation networks, digital elevation models (DEM), satellite imagery, remote sensing data, crowd-sourced reports.

Main Results

• STORM-Net achieved a classification accuracy of 98.9%, precision of 98.8%, recall of 98.7%, and an F1-score of 98.9%.
• The model registered the lowest validation loss of 0.28 and the shortest training time of 3.1 seconds (inference time of 2.5 seconds for Kaggle dataset).
• AUC-ROC scores were approximately 0.985 for the Kaggle dataset and 0.972 for the UCI dataset, demonstrating high discrimination performance.
• STORM-Net significantly outperformed baseline and hybrid models across all metrics, showing superior accuracy, efficiency, and generalization capacity.
• The model effectively simulated urban flood levels across major Indian cities, demonstrating spatially distributed vulnerability and interpretability.

Contributions

• Introduction of STORM-Net, a novel hybrid deep learning architecture that synergistically combines convolutional and recurrent networks with attention mechanisms for robust spatiotemporal urban flood risk prediction.
• Development of SAFER, a unique feature selection algorithm integrating sparse autoencoders and reinforcement learning to effectively eliminate redundant features and mitigate multicollinearity, enhancing model interpretability and efficiency.
• Creation of BRAVE, a hybridized optimization technique (Builder Optimization Algorithm and Superb Fairy-wren Optimization) for dynamic and context-sensitive attention scaling, which improves the model's ability to focus on critical spatiotemporal patterns and generalize across diverse conditions.
• Demonstrated state-of-the-art performance in accuracy, precision, recall, F1-score, AUC-ROC, and computational efficiency, significantly outperforming existing deep learning and hybrid models.
• Provided a robust and scalable solution for urban flood risk forecasting that addresses challenges of data heterogeneity, overfitting, and real-time operational requirements.

Funding

Not explicitly mentioned in the provided text.

Citation

@article{Sathiyamoorthy2026STORMNet,
  author = {Sathiyamoorthy, M and Subramanian, P. and Lakshmi, N. V. S. Sree Rathna and Hemalatha, S. and Nagarajan, L. and Sathish, S. and Prakash, M. Gnana},
  title = {STORM-Net for urban flood risk prediction: an AI-based spatiotemporal tracking and mapping approach},
  journal = {Stochastic Environmental Research and Risk Assessment},
  year = {2026},
  doi = {10.1007/s00477-026-03186-2},
  url = {https://doi.org/10.1007/s00477-026-03186-2}
}