Advancing Camera-Based Monitoring for Operational Hydrologic Applications

Camera-based monitoring systems have enormous—yet currently untapped—potential to collect continuous observations for hydrologic monitoring by extracting measurements from imagery and video. Noncontact optical measurement systems can complement and, in some cases, replace conventional sensing technology. While computer vision methods can provide promising and accurate results, practical barriers exist to the widespread implementation and operationalization of optical stream gauging. Common challenges include variable environmental conditions, image and video data management, and systems integration. Deep learning methods can potentially segment water from other objects in images under varying conditions, thus making optical gauging methods more generalizable. Further, edge and cloud computing technologies can be designed and leveraged to overcome data management and processing challenges associated with imagery and video. This project will demonstrate how operational requirements for integrating low-cost cameras and computing infrastructure into existing hydrologic monitoring networks can be met, along with evaluating the benefits of cameras for continuous monitoring and prediction. Specifically, this 3-year project will:

1. Demonstrate the value and generalizability of existing computer vision and deep learning-based approaches for processing images;
2. Demonstrate how cameras can be reliably integrated into gage operations to extract meaningful hydrologic data; and
3. Demonstrate how serverless and edge computing architectures can enable strategic scaling of 1 and 2 to many sites.

This project will develop and deploy operational, camera-based monitoring stations within an existing instrumented watershed in Logan, UT, and at selected sites in Columbia, SC. We will use AI-based image processing tools for water level monitoring and channel width estimation and apply existing image velocimetry methods to imagery collected at several (2-3) strategically placed cameras in both UT and SC. We will evaluate serverless computing, such as Amazon Web Services, for on-demand, cloud-based image processing. We will also optimize a subset of algorithms for processing imagery at the edge using low-power, low-cost computing devices, such as the Raspberry Pi or the NVIDIA Jetson Nano.

By determining the requirements for integrating low-cost cameras into existing hydrologic monitoring networks, this project will directly support operationalizing existing—but currently isolated—state-of-the-art image analysis and computing tools into a deployable camera-based monitoring station that instrumented networks can readily adapt. This will advance the state of practice for innovative hydroinformatics applications and contribute to the National Water Information System (NWIS) Modernization Program. In this 3-year project, we expect to generate multiple peer-reviewed articles, publish code, documentation, and resulting monitoring data, and mentor several graduate students while building new collaborations across institutions. This project will also result in a new set of recommendations for scaling computations required for processing video and image data and generate a greater understanding of the utility and accuracy of using existing AI-based and computer vision tools for processing imagery in the context of hydrologic monitoring.