Deep Learning and object detection for water level measurement using patterned visual markers

Abstract

Flooding is one of the most impactful natural disasters, causing significant losses and prompting extensive research into monitoring water levels in urban streams. Current technologies rely on pressure and ultrasonic sensors, which, while accurate, can be susceptible to damage from floods and are often costly. As an alternative, ground camera approaches offer a low-cost solution; however, most of these methods use raw images from the water stream and are sensitive to environmental factors. We address this gap with a dataset comprising a visual marker with black bars indicating the water level, which we refer to as "barcode panel". We employed various deep learning algorithms to predict the water level and compared their performance. The proposed approach was evaluated using classic classification and error metrics. The models demonstrated accuracy in detecting the water level. These promising results provide important insights for practical applications and future studies.