Mohammad Nooshzadi Motlagh , Sierra Young, Kenneth Chapman , Evelynn Wilcox, François Birgand

Stormwater systems, as infrastructure draining urban water runoff into water bodies, are pivotal in preserving municipal functionality while they play an important hydrological and environmental role. As such, the ability to reliably monitor stormwater outflow in many locations could provide valuable information for water managers. However, in most cases stormwater outlets have not been designed or installed to facilitate the measurement of stormflow, leaving traditional contact-based velocity and water level measurement techniques ill-suited to capture highly variable and turbulent flows. We propose a non-contact alternative approach based on computer vision, capable of quantifying discharge from images and videos obtained from cameras facing the outlets. In variable lighting and on often ‘noisy’ images and videos, this approach came with its own set of challenges, and classical computer-vision techniques did not perform reliably and accurately. To solve these issues, we used the combination of computer vision and machine learning (CV-ML) techniques, using the round geometry of culverts to our advantage. In our approach, the water stage at the outlet is determined by calculating the difference between the height of the extracted shape of a round culvert and the height of the empty area above water. Then, using a checker board as a reference object, the measurements from images are converted into real-world measurements. Finally, as a first approximation using rating curves, the calculated water stage can be converted into discharge values. To evaluate the model's performance on stage measurements, two methods were considered. In the first, the uncertainty on measurements was assessed by comparing the culvert diameter with that of our CV-ML model calculated value. As a result, the model was on average capable of making measurements within $\pm 1$ cm approximately 80\% of the times. In the second method, we compared measurements from our model to those `visually' made from images obtained during a flow event. For this method also the model estimated 63\% of the stage values within $\pm 1$ cm and 96\% within $\pm 2$ cm. These results could be considered as satisfactory, especially considering the complexity of the field conditions.