Camera-based Water Stage and Discharge Prediction with Machine Learning

Abstract. Time-lapse imagery of streams and rivers provide new qualitative insights into hydrologic conditions at stream gauges, especially when site visits are biased toward baseflow or fair-weather conditions. Imagery from fixed, ground-based cameras is also rich in quantitative information that can improve streamflow monitoring. For instance, time-lapse imagery may be valuable for filling data gaps when sensors fail and/or during lapses in funding for monitoring programs. In this study, we automated the analysis of time-lapse imagery from a single camera at a single location, then built and tested machine learning models using programmatically calculable scalar image features to fill data gaps in stream gauge records. Time-lapse images were taken with a fixed, ground-based camera that is part of a documentary watershed imaging project (https://plattebasintimelapse.com/). Features were extracted from 40,000+ daylight images taken at one-hour intervals from 2012 to 2019. The algorithms removed dawn and dusk images that were too dark for feature extraction. The image features were merged with United States Geological Survey (USGS) stage and discharge data (i.e., response variables) from the site based on image capture times and USGS timestamps. We then developed a workflow to identify a suitable feature set to build machine learning models with a randomly selected training set of 30 % of the images with the remaining 70 % for a test set. Predictions were generated from Multi-layer Perceptron (MLP), Random Forest Regression (RFR), and Support Vector Regression (SVR) models. A Kalman filter was applied to the predictions to remove noise. Error metrics were calculated, including Nash-Sutcliffe Efficiency (NSE), Prediction Bias (PBIAS), RMSE-Standard Deviation Ratio (RSR), and an alternative metric that accounted for seasonal runoff. After suitable features were identified, the dataset was divided into test sets of simulated data gaps for 2015, 2016, and 2017. The training sets for each gap were features from contiguous images and sensor readings before and after the gaps. NSE for the year-long gap predictions ranged from 0.63 to 0.90 for discharge and 0.47 to 0.90 for stage. The predictions for 2015 and 2017 displayed lower prediction errors than for 2016. The 2016 discharge was significantly higher than training data, which could explain the poorer performance. First and second half-year test sets were created for 2016 along with MLP models from before/after training sets for each of the gaps that held discharge measurements similar to those in the gaps. The half-year gap models' predictions improved NSE, PBIAS and RSR. The results show it is possible to extract features from images taken with the downstream facing camera to build machine learning models that produce accurate stage and discharge predictions. The methods employed should be transferrable to other sites with ground-based cameras.