Improving hydrological models with the assimilation of citizen science data

Small streams boost the health and integrity of downstream rivers and provide refuge to many freshwater species. Therefore, accurate spatial representations of streamflow and the characterization of species habitats are needed, which can primarily be accomplished through data collection and/or computer modeling. Environmental agencies play a key role of providing reliable hydrologic data; however, they are often challenged by shrinking budgets and the growing number of monitoring needs. Given the spatial mismatch of observed data and small watersheds/headwaters, citizens with smart phones can act as a potentially rich data source for hydrologic observations. We discuss how to use CrowdHydrology to improve a hydrologic model of the Boyne River, a small watershed in northern Michigan. CrowdHydrology is a citizen science program that collects water stage height and stream temperature. Participants provided measurements—text messages sent to a server—at four sites with different but growing levels of participation over time. We tested whether stream stage and temperature observations, measured by citizens, improved the performance of a Soil and Water Assessment Tool (SWAT) model of the Boyne River. Measured observations were integrated into the model using a data assimilation approach. This framework allowed us to integrate observation error, track the temporal variability of model parameters, and simulate daily streamflow and stream temperature across the watershed. Measures of model performance included the modified Nash-Sutcliffe efficiency (E_f-mod), refined index of agreement (d_r), and relative bias (Bias). Four years of streamflow assimilation show that performance metrics evolved from poor values (E_f-mod = -0.71, d_r = 0.45, Bias = -0.32) to acceptable levels (E_f-mod = 0.53, d_r = 0.74, Bias = 0.04) at locations with richer datasets. We observed a similar trend during the assimilation of streamflow temperature. Results show that while model performance increases as sequential periods of citizen science data are assimilated, acceptable performance metrics arise in locations of higher levels of participation.