Application of Deep Learning for Imaging‐Based Stream Gaging. Issue 11 (1st November 2021)

• Record Type: Journal Article
• Authors: Vanden Boomen, Ryan L.; Yu, Zeyun; Liao, Qian
• Is Part Of: Water resources research. Volume 57:Issue 11(2021)
• Page Start: n/a
• Page End: n/a
• Publication Date: 2021-11-01
• Subjects: Machine learning -- river stage -- deep neural network -- image regression model -- stream flow monitoring; Hydrology -- Periodicals; 333.91
• DOI: 10.1029/2021WR029980
• Languages: English
• ISSNs: 0043-1397

Abstract

Abstract: Stream gages are critically important for measuring stream flow in water resources management. Stream gages monitor and record flow and water stage within some water body. The United States Geological Survey maintains a network of stream gages across the country. Many of these sites are also equipped with webcams to gather real‐time river information, especially in flood flows. Remotely measuring flow discharge using particle image tracking has been researched intensively. This study demonstrates a process for training a deep neural network that will utilize the webcam and stream gage data to generate a water stage prediction based on webcam captured river images. This study presents the experiments on stream gages located at the Clear Creek in Iowa, Auglaize River in Ohio, and Milwaukee River in Wisconsin. The process outlined utilizes transfer learning and well‐known image classification models as a basis for a generalized river stage regression model. Across the training, validation, and deployment experiments, the developed process shows success in creating an accurate model for these testing sites of different perspective settings. The results of this study show confidence in future studies utilizing ground‐level remote stream gaging with Machine Learning‐enabled image regression. Key Points: It was demonstrated that the stage of river flow can be directly estimated from static images with a supervised machine learning approach Transfer learning with pretrained weight inherited from ImageNet is effective in extracting image features relevant to river stage An image‐classifier‐based deep neural network regression model is able to predict river stage values in continuous domain