Automatic water-level class estimation from repeated crowd-based photos of streams

Citizen science projects engage the public in monitoring the environment and can collect useful data. One example is the CrowdWater project, in which stream levels are observed and compared to reference photos taken at an earlier time to obtain stream level class data. However, crowd-based observations are uncertain and require data quality control. Therefore, we used a deep learning model to estimate the water-level class for photos taken by citizen scientists at different times for the same stream and compared different options for model training. The models had a root mean square error (R) of 0.5 classes or better for all but four of the 385 sites for which the model was trained. Low water levels were in general predicted better than high water levels (R of 0.6 vs 1.0 classes). The study thus highlights the potential of human-computer interaction for data collection and quality control in citizen science projects.