A Water Level Measurement Approach Based on YOLOv5s

Area of Science:

• Hydrology
• Computer Vision
• Deep Learning

Background:

• Traditional water gauge reading methods using image analysis suffer from poor adaptability and robustness in diverse environmental conditions.
• Limitations in existing techniques hinder accurate and reliable hydrological monitoring.

Purpose of the Study:

• To develop and validate a novel water level measurement method using deep learning to address the shortcomings of current image-based approaches.
• To enhance the accuracy and robustness of water level readings in various challenging scenarios.

Main Methods:

• A deep learning model (YOLOv5s, a convolutional neural network) was employed to extract critical areas from video images, including the water gauge and scale markings.
• Image processing techniques were utilized to precisely identify the water surface line and calculate the water level elevation.

Main Results:

• The proposed method demonstrated a low systematic error of only 7.7 mm in field tests.
• High accuracy was maintained across various conditions: 95% (daylight), 98% (night infrared), 97% (strong light), 91% (rainfall), and 90% (dirty gauge).
• The method showed notable performance even with transparent water bodies (45% accuracy).

Conclusions:

• The deep learning-based water level measurement method exhibits strong performance and robustness across different environmental conditions.
• This approach offers a reliable and effective solution for hydrological monitoring, providing a valuable reference for applying deep learning in the field.