Search

Traditional discharge monitoring usually relies on measuring flow velocity and cross-section area with various velocimeters or remote-sensing approaches. However, the topography of mountain streams in remote sites largely hinders the applicability of velocity–area methods. Here, we present a method to continuously monitor mountain stream discharge using a low-cost commercial camera and deep learning algorithm. A procedure of automated image categorization and discharge classification was developed to extract information on flow patterns and volumes from high-frequency red–green–blue (RGB) images with deep convolutional neural networks (CNNs). The method was tested at a small, steep, natural stream reach in southern China. Reference discharge data were acquired from a V-shaped weir and ultrasonic flowmeter installed a few meters downstream of the camera system. Results show that the discharge-relevant stream features implicitly embedded in RGB information can be effectively recognized and retrieved by CNN to achieve satisfactory performance in discharge measurement. Coupling between CNNs and traditional machine learning models (e.g., support vector machine and random forest) can potentially synthesize individual models' diverse merits and improve generalization performance. Besides, proper image pre-processing and categorization are critical for enhancing the robustness and applicability of the method under environmental disturbances (e.g., weather and vegetation on river banks). Our study highlights the usefulness of deep learning in analyzing complex flow images and tracking flow changes over time, which provides a reliable and flexible alternative apparatus for continuous discharge monitoring of rocky mountain streams. [ABSTRACT FROM AUTHOR]

Traditional discharge monitoring usually relies on measuring flow velocity and cross-section area with various velocimeters or remote-sensing approaches. However, the topography of mountain streams in remote sites largely hinders the applicability of velocity-area methods. We here present a method to continuously monitor mountain stream discharge using a low-cost commercial camera and deep learning algorithm. A procedure of automated image categorization and discharge classification was developed to extract information on flow patterns and volumes from high-frequency red–green–blue (RGB) images with deep convolutional neural networks (CNNs). The method was tested at a small, steep, natural stream reach in southern China. Reference discharge data was acquired from a V-shaped weir and ultrasonic flowmeter installed a few meters downstream of the camera system. Results show that the discharge-relevant stream features implicitly embedded in RGB information can be effectively recognized and retrieved by CNN to achieve satisfactory accuracy in discharge measurement. Coupling CNN and traditional machine learning models (e.g., support vector machine and random forest) can potentially synthesize individual models' diverse merits and improve generalization performance. Besides, proper image pre-processing and categorization are critical for enhancing the robustness and applicability of the method under environmental disturbances (e.g., weather and vegetation on river banks). Our study highlights the usefulness of deep learning in analyzing complex flow images and tracking flow changes over time, which provides a reliable and flexible alternative apparatus for continuous discharge monitoring of rocky mountain streams. [ABSTRACT FROM AUTHOR]