Ensuring reliable water level measurement for flooding: A redundancy-based approach with pressure transducer and computer vision

Abstract

Fluid level measurement is essential in many fields, including industrial and civil sectors, especially for urban flood detection, where there is a high risk of mortality and economic losses. However, although contact-based methods that employ pressure transducers can achieve a high degree of precision, they are susceptible to damage from direct contact with the fluid. This study adopts a redundancy-based approach that combines pressure transducer measurements with computer vision to provide enhanced reliability and reduce the risk of sensor failures. Our approach entails training a deep-learning model that uses pressure sensor data to mitigate this potential risk of damage and avoid the need for manually annotating sets of images. The results show that the pressure transducer has high accuracy, with a mean absolute error (MAE) of 1.21 cm, and that the computer vision model which is trained on pressure sensor data, achieves a comparable MAE of 6.67 cm. This approach also makes the system more robust and includes a dependable backup measurement method in case the primary sensor fails. Furthermore, the model trained on the sensor data led to results that were very similar to those trained directly on ground-truth data.