Culverts are essential yet often overlooked structures in our cities. These tunnels or pipes allow water to flow under roads, railways, or other infrastructure, playing a crucial role in water management. The UK alone is estimated to have over a million culverts. However, when these culverts become blocked by debris or trash, they can cause flooding. Monitoring them is a significant challenge, but new research suggests that AI, trained on CCTV data, could revolutionize this process.
Water management has been a core challenge for cities since ancient times. Effective sewage systems were among the key innovations that enabled urban expansion. However, in modern times, urban areas face an increasing risk of flooding, especially as climate change intensifies storms and rainfall. Many cities are experiencing more frequent floods, leading to costly damage.
While this new research may not completely solve urban flooding, it addresses a specific issue: detecting culvert blockages.
Culvert entrances typically feature screens (usually a set of bars) that are meant to stop debris from passing through. Ironically, however, these screens often become clogged themselves. And if they do get clogged, they become a significant flood risk. Municipalities deal with this issue in two ways. Either they have teams regularly clean up these entrances or they clean them up when they receive a flood warning.
Both of these approaches have downsides. Having teams regularly clean culverts is safe but wastes a lot of resources when there’s no blockage, while reactive cleaning could be too late or even dangerous for the cleaning crew.
Rory Smith and colleagues from the University of Bath believe there’s a better way of doing things.
AI, meet culverts
They developed a method that uses CCTV cameras to monitor trash screens continuously. Using CCTV cameras and image-based classification, this system can determine whether a culvert is blocked or unblocked, allowing for quicker, more efficient responses and reducing the risk of urban flooding.
The study focused on the Tongwynlais screen in Cardiff, where a CCTV camera takes regular images each morning, and additional images when water levels rise. Initially, they started with 755 images and after quality control, reduced the dataset to 577 images. Around 80% of these images showed blocked screens while 20% showed unblocked screens.
<!– Tag ID: zmescience_300x250_InContent_3
–>
Researchers explored three approaches for the data because with far more blocked images than unblocked ones, the model could become biased, leading to incorrect predictions. The approaches were:
- Using the original, imbalanced dataset — This model trained on all the images as they were.
- Undersampling — The number of blocked images was reduced to match the number of unblocked images, creating a more balanced dataset.
- Data augmentation — The number of unblocked images was artificially increased by adding noise to the images, thus expanding the dataset without requiring more real-world examples.
Of the three approaches, data augmentation performed the best, boosting the model’s accuracy by 8% and achieving an overall accuracy of 88%. This improvement highlights the potential of data augmentation as a solution for dealing with imbalanced datasets, a common issue in machine learning.
This approach can be scaled
While the study focused on a single trash screen in Cardiff, the implications are much broader. In the future, similar systems could be installed across entire cities, creating a real-time network for flood monitoring. However, several challenges need to be addressed before this vision can become a reality.
Each culvert is different, and the conditions affecting trash screens can vary widely. One solution could be to develop a database of pre-trained models for different types of culverts, allowing new installations to benefit from transfer learning — a process that adapts an existing model to a new location.
In the future, systems like these could form the backbone of an early warning network for urban flooding, providing real-time data to
authorities and helping them respond before disaster strikes. With continued research and development, this technology could revolutionize how we manage urban water infrastructure in the face of a changing climate.
Journal Reference: Rory Cornelius Smith et al, CCTV image‐based classification of blocked trash screens, Journal of Flood Risk Management (2024). DOI: 10.1111/jfr3.13038