Assessment of climate change impacts on Brazilian catchments using a regional deep learning approach

The potential impacts of climate change in catchment hydrology are still unknown in most of the world and it is not different in Brazil. Conducting an integrated analysis of catchments based on similarity groups allows us to extract conclusions and observations about the overall controls of hydrological behavior, but also considering the specific and distinctive characteristics of each of these groups. This approach enables us to identify and comprehend the primary features influencing hydrological behavior within each distinct group, increasing hydrologic predictability and knowledge of catchments’ functioning, which is essential to better understand the impacts of climate change. In this study, we investigate the possible shifts in Brazilian catchment hydrology behavior in response to a changing climate. Employing a regional approach of the Long Short-Term Memory (LSTM) to understand and predict streamflow across 735 catchments of six hydrologically similar groups in Brazil, we simulated streamflow throughout the 21st century. This simulation utilized a multi-model ensemble comprising 19 bias-corrected Global Climate Models (GCMs) from the sixth phase of the Coupled Model Intercomparison Project (CMIP6), driven by intermediate and high-emission scenarios (SSP245 and SSP585). Our results show that the regional LSTM outperforms the conventional hydrological modeling (NSE≈0.60), underscoring the reliability of deep learning to estimate streamflow with simplified input. Interestingly, we found that substantial variations in projected temperatures across scenarios do not necessarily correspond to significant differences in projected streamflow. Moreover, changes in precipitation and temperature may not exert proportional impacts on streamflow. Further, we will investigate the dynamics of transitions between catchment groups. This innovative approach to assess the impacts of climate change enhances the reliability of projected streamflow trajectories, a critical consideration given the uncertainties associated with CMIP6 models. Furthermore, this study holds potential utility in developing strategies to mitigate the impacts of climate change on Brazilian water resources.