Combining process-based model and machine learning to predict hydrological regimes in floodplain wetlands under climate change

Climate change has remarkable effects on the hydrological regime in floodplain wetlands. However, the hydrodynamic simulation under climate change conditions suffers from the uncertainty of the downstream water level boundary. In this study, an effective approach was developed by integrating an advanced machine learning algorithm and a process-based watershed hydrological model to predict the lake level under climate change. The developed approach was validated in a vital conservation wetland in the Ramsar Convention, the Poyang Lake floodplain wetland in China and utilized to simulate the hydrological regime in climate change. We found that the Long Short–term Memory (LSTM) performed relatively well in the simulation of water level, with coefficients of determination (R2) of 0.94 and 0.83 in the training and testing periods, respectively. The mean cumulative flooding time in the future period was slightly less than that in the baseline period, with a decreased range of 1.01 ∼ 12.05 days/year. The highest and mean water levels rose significantly in the wet season, with a risen range of 1.01 ∼ 3.3 and 1.56 ∼ 2.44 m, respectively, and the appearance of extreme flood events became more frequent. The maximum, minimum, and mean inflows for Poyang Lake in the dry season exhibited a remarkable reduction trend, especially for the minimum inflow, with a reduced range of 59.37%∼79.22%, indicating that more serious extreme drought events would appear in Poyang Lake. The developed approach has good performance for water level simulation and extreme runoff prediction, thus providing a viable way for future water resources management in floodplain wetlands.