Abstract
River flooding is a natural phenomenon that can have a devastating effect on human life and economic losses. There have been various approaches in studying river flooding; however, insufficient understanding and limited knowledge about flooding conditions hinder the development of prevention and control measures for this natural phenomenon. This paper entails a new approach for the prediction of water level in association with flood severity using the ensemble model. Our approach leverages the latest developments in the Internet of Things (IoT) and machine learning for the automated analysis of flood data that might be useful to prevent natural disasters. Research outcomes indicate that ensemble learning provides a more reliable tool to predict flood severity levels. The experimental results indicate that the ensemble learning using the Long-Short Term memory model and random forest outperformed individual models with a sensitivity, specificity and accuracy of 71.4%, 85.9%, 81.13%, respectively.
Introduction
River flooding remains one of the most catastrophic natural phenomena globally, yet accurate prediction is frequently hampered by the inherent complexity of environmental data. Traditional predictive models often struggle with "non-linearities," where minor fluctuations in variables like rainfall or snowmelt trigger disproportionately large changes in water levels.
Furthermore, flood datasets are notoriously imbalanced; normal water levels are recorded far more frequently than high-risk events. This imbalance often leads standard machine learning models to favor the majority class, failing to identify rare but critical flood risks. This research introduces a robust framework that combines Internet of Things (IoT) sensor data with ensemble machine learning to bridge this gap, providing a more effective early warning system.
Methodology
Data Acquisition and Pre-processing
The study utilized a comprehensive dataset from the Environment Agency, consisting of 4,214 flood samples collected over 31 years. The data included 11 environmental attributes, such as monsoonal rain, duration, snowmelt, and magnitude. These were categorized into three severity levels: Class 1 (Normal), Class 2 (Abnormal), and Class 3 (High-risk/Dangerous).
To refine the model, the researchers employed a multi-stage pipeline:
- Feature Selection: A Chi-square filtering approach was used to identify the 6 most statistically significant attributes.
- Data Visualization: Principal Component Analysis (PCA) and t-Distributed Stochastic Neighbourhood Embedding (t-SNE) were used to analyze data structure, revealing that flood classes were overlapping and not linearly separable.
- Model Development: The core of the framework is a "stacking" and "voting" ensemble approach, illustrated in the figure below.
The Ensemble Approach
The researchers tested several individual models, including Artificial Neural Networks (ANN), K-Nearest Neighbour (KNN), and Support Vector Machines (SVM). However, the most effective results were achieved by combining Long Short-Term Memory (LSTM) networks—which excel at capturing temporal dependencies—with Random Forests (RF), which are adept at handling non-linear decision boundaries. This combination effectively transformed several "weak" classifiers into a single "strong" ensemble model.
Results and Discussion
The ensemble models consistently outperformed individual classifiers across all performance metrics. The LSTM + RF ensemble emerged as the top performer, achieving an overall testing accuracy of 81.13% on unseen data.
Key insights from the results include:
- Resilience to Imbalance: While individual models performed well on "Normal" water levels, they often failed on "Dangerous" levels. The ensemble approach, utilizing bootstrapping and aggregation, proved significantly more resilient to class imbalance.
- Generalization: Although the LSTM model showed high sensitivity during training (0.925), it struggled to generalize in isolation. Pairing it with Random Forest "smoothed out" individual errors, leading to a more robust real-world application.
- Non-Linearity Handling: The t-SNE visualizations confirmed that flood data is a complex, non-linear "tangled mess" in low-dimensional space, justifying the use of deep learning and ensemble methods over simpler linear classifiers.
Conclusion and Future Work
This research demonstrates that ensemble learning, particularly the combination of LSTM and Random Forest, provides a reliable tool for predicting flood severity. While the model achieved a high specificity of 85.9%, class imbalance remains a persistent challenge for identifying high-risk events with perfect accuracy.
Future work will focus on integrating optimization algorithms, such as Particle Swarm Optimization (PSO), to fine-tune hyperparameters automatically. Additionally, the researchers aim to move beyond classification to regression-based models for predicting exact water levels and deploying these systems directly onto IoT edge devices for real-time, local processing.
Summary updated