Application of time-lapse imagery and machine learning to improve stream discharge monitoring
Abstract
Time-lapse imagery of streams and rivers can provide hydrologists with greater insight into watershed dynamics, particularly when site visits are biased towards baseflow conditions. Ground-based imagery is also rich in quantitative information that can improve streamflow monitoring programs. In this study we focused on the use of image analysis and machine learning to fill data gaps that could result from stream gauge failure or equipment removal due to budget constraints. Using ground-based timelapse images captured for a documentary watershed project (Forsberg and Farrell; http://plattebasintimelapse.com), we analyzed scalar features (e.g., whole-image statistics, surface water features above and below a weir in the image scene) to test the capabilities of image-based statistical machine learning models of stream stage and discharge. The camera site was located at a United States Geological Survey (USGS) gauging station on the North Platte River (Nebraska and Wyoming border, USA; https://waterdata.usgs.gov/wy/nwis/uv?site_no=06674500) and captured images between 2011 and 2020. For initial "best case" model runs, 30% of the data (13,879 daytime images collected hourly, each paired with stream stage and discharge data from the USGS gauge) were randomly selected and used to train machine learning models. The models were then tested using the remaining 70% of the data (n = 28,180), with results filtered for noise with a Kalman filter. The optimal Support Vector Regression (SVR), Random Forest Regression (RFR), and Multilayer Perceptron (MLP) models had root mean squared error (RMSE) of 12.3 to 14.7 m³/s , or <7% of the observed range in discharge (0.2 to 224 m³/s ). More specific test cases were also evaluated for 2015, 2016, and 2017. When using 5,000 observations before and after each year of interest (i.e., training set of 10,000 observations for each year), RMSE for the SVR, RFR, and MLP models with noise filtering was in the range of 6.46 to 22.34 cm and 7.86 to 44.15 m³/s for stage and discharge, respectively. The greatest errors were for the 2016 water year, where a high flow event far exceeded the highest flows captured in the training dataset. Overall, errors are generally larger than desired for direct measurement of stream stage and discharge, but results suggest that imagery provides quantitative information to fill year-long data gaps. The tested statistical machine learning models gave sufficiently precise discharge for the purpose of calibrating traditional surface water or groundwater models.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2020
- Bibcode:
- 2020AGUFMH166.0008C
- Keywords:
-
- 1805 Computational hydrology;
- HYDROLOGY;
- 1847 Modeling;
- HYDROLOGY;
- 1942 Machine learning;
- INFORMATICS;
- 1952 Modeling;
- INFORMATICS