The role of model complexity in streamflow projections on Brazilian catchments

The complexity of hydrological models can significantly influence the accuracy of streamflow predictions. While more complex models may seem advantageous due to their robustness, previous research found that simpler models can often yield comparable or even superior results for some applications, particularly if they are able to adequately represent key hydrological processes and features. Here, we investigated the impact of model complexity on the streamflow projection over the century in Brazil, a continental-scale country that presents hydrological and landscape heterogeneity. We employed projected climate change data from 10 models of the CLIMBra dataset over 735 Brazilian catchments, forced by the CMIP6-based SSP2-4.5 and SSP5-8.5 scenarios. We used five hydrological models representing a full range of model complexity: 1. Functional forms, 2. Grunsky method, 3. HYMOD model, 4. MISDC model, and 5. a regional model based on the Long Short-Term Memory (LSTM) algorithm. On a daily basis (models 3 to 5), when comparing the traditional models with the LSTM, we found that LSTM overperformed HYMOD and MISDC, with a median KGE of 0.72. The MISDC presented the worst performance in daily predictions. All the models were evaluated on a long-term basis, with KGE ranging from 0.62 (Grunsky method) to 0.85 (LSTM model). The conventional hydrological models, HYMOD and MISDC presented KGE of 0.78 and 0.80, showing great suitability for Brazilian catchments, but with the disadvantage of needing local parametrization. It is also worth noting that performance metrics were improved in all cases from a daily to a long-term basis, due to the longer timescale. Regarding the streamflow over the century, when comparing an ensemble mean of climate projections, different models estimated, on average, changes from -21% to +15% in long-term mean daily streamflow. Simpler models projected a slight increase in the mean streamflow, while more complex models projected greater changes. We also found some spatial patterns of variation according to the model complexity, with greater differences in the arid catchments (where the KGEs were lower). We further discuss model complexity and performance in view of climate models' inherent uncertainties. The comparative performance presented in our study showed that while complexity can enhance performance, some simpler models can show similar outputs and might be preferred for some applications.