Multi-source Data Integration under a Deep Learning Framework to Improve Streamflow Forecast Ability

Integrating available observations with the forecasting model, typically under a data assimilation framework, is well-known to improve the forecast ability. However, data assimilation requires delicate choices including the assimilation scheme, bias correction, setting up the covariance matrix, etc.. Each different type of observations, e.g., daily, monthly, or running averages, would incur different assimilation strategies. The development of Deep Learning (DL) networks demonstrates a promising solution to this problem benefiting from their strong flexibility to absorb data and distill information. In this study, we set up an efficient time series DL framework to flexibly assimilate multiple types of observations at different spatio-temporal scales. The results indicate that the proposed DL framework can easily achieve observation integration tasks which are hard to implement traditionally, and significantly improve the streamflow forecast performance. Moreover, the performance of this data-driven framework can provide hydrologic insights since the integration of different observed variables compensates the ignored hydrologic processes in the original model. This further supports our view that the data-driven framework could be a complementary avenue for the knowledge distillation process.