Assessing the reliability of a physical-based model and a convolutional neural network in an ungauged watershed for daily streamflow calculation: a case study in southern Portugal

Main conclusions:

"Considering the sub-basin modeling, the 1D-CNN model demonstrated a better performance than MOHID-Land when considering the daily values and the wet period. The MOHID-Land model showed a better performance in estimating streamflow values during dry periods and for a monthly analysis".

"Although the results were considered from satisfactory to very good in all the steps taken during the validation process, the generation of negative values by the 1D-CNN is of concern. In that sense, the model presented here should be a target of improvement in future applications. In turn, MOHID-Land model revealed a lower performance for daily streamflow estimation, but its physical basis contributes to avoiding unpredictable and incomprehensible results.
Finally, it is worth noting that neural network models are developed and trained for present and/or past conditions, and their application to future scenarios can be limited. Also, the prediction of events that go beyond the observations can be problematic. This limitation is mainly related to its lack of capacity to absorb information about future conditions in cases where neural networks were not prepared to be forced by variables that include the impact of those future changes.
Nonetheless, the changes in future conditions can be easily imposed in physically based models, with the main problems being: (i) the detail of the characterization of future conditions, that most of the time is too coarse for the detail adopted on physical models; and (ii) the high computational time needed to run long-term simulations, usually performed in analysis of future scenarios".