Introduction
Urban infrastructure, particularly culverts, plays a vital role in flood mitigation. However, these structures are prone to blockage by debris such as vegetation and urban waste, which can lead to catastrophic flash flooding. Current management strategies typically rely on "visual blockage"—the percentage of the culvert opening obscured by debris.
A critical limitation of this approach is the disconnect between visual appearance and actual flow reduction, known as "hydraulic blockage." For instance, a culvert 80% covered by porous branches may still allow significant water passage, whereas a 20% blockage by solid plastic could severely impede flow. Historically, no quantifiable relationship has existed to translate visual data into hydraulic performance metrics. This research addresses this gap by proposing a deep learning-based pipeline to predict hydraulic blockage directly from images.
Methodology
The researchers developed an end-to-end machine learning pipeline designed to map high-dimensional visual features to a continuous hydraulic blockage percentage. The process involves three distinct stages:
- Deep Visual Feature Extraction: The system utilizes pre-trained Convolutional Neural Networks (CNNs)—specifically MobileNet, ResNet50, and EfficientNetB3—as feature extractors. By removing the final classification layers, these models transform raw culvert images into numerical feature vectors.
- Data Pre-processing: Extracted features undergo normalization via Standard Scalar transformation. This ensures that the data has a mean of zero and a standard deviation of one, which is essential for stable and efficient training of the subsequent regression model.
- ANN Regression: The normalized features are fed into an Artificial Neural Network (ANN). The study evaluated three ANN architectures (ANN1, ANN2, and ANN3) with depths ranging from 5 to 10 hidden layers to determine the optimal complexity for different CNN feature sets.
Experimental Setup and Datasets
Due to the scarcity of high-quality visual and hydraulic data from real-world flood events, the study utilized a controlled laboratory environment to build a foundational dataset.
- Hydrology-Lab Dataset (HD): This dataset was generated using a 12-meter flume equipped with scaled culvert models. It comprises 352 hydraulic data points across 173 unique blockage scenarios involving both vegetative and urban debris.
- Visual Hydrology-Lab Dataset (VHD): Synchronized with the hydraulic measurements, this dataset contains 352 images extracted from video footage, providing the visual ground truth for each hydraulic state.
The experimental setup included upstream and downstream point gauges to measure water levels, allowing for the precise calculation of hydraulic blockage based on flow capacity reduction relative to the visual state captured by the camera.
Results and Discussion
The study yielded several significant insights into the application of deep learning for hydrological monitoring:
Model Performance
The most effective configuration was the combination of MobileNet features with a 5-layer ANN (ANN1). This model achieved an $R^2$ score of 0.7855, demonstrating a strong correlation between visual features and hydraulic behavior. The Mean Absolute Error (MAE) was 4.67, indicating that the model's predictions were, on average, within approximately 5% of the actual measured hydraulic blockage.
Feature Complexity vs. Performance
An unexpected finding was that MobileNet, the simplest architecture tested, outperformed the deeper ResNet50 and EfficientNetB3 models. The authors hypothesize that the larger models extracted a high volume of redundant or uncorrelated features, introducing noise that hindered the regression task. This suggests that for specific infrastructure monitoring tasks, "less is more" regarding model depth.
Computational Efficiency
MobileNet also proved to be the most computationally efficient, with a total inference time of approximately 1.25 seconds per image. This high speed, combined with its lower memory footprint, makes the pipeline a strong candidate for deployment on "edge" devices—smart cameras capable of real-time processing at remote culvert sites.
Conclusion and Future Directions
This research provides empirical evidence that visual data contains sufficient information to predict hydraulic flow reduction, bridging a long-standing gap in flood management. However, several challenges remain for real-world implementation. The current models were trained in a controlled lab environment; real-world deployment will require models robust to varying lighting, weather conditions, and complex backgrounds. Future work should focus on collecting "in-the-wild" datasets from actual flood events and refining camera placement to minimize background interference. Despite these hurdles, this pipeline represents a significant step toward automated, data-driven flood risk assessment.
Summary updated