Surface and sub-surface flow estimation at high temporal resolution using deep neural networks

Recent intensification in climate change have resulted in the rise of hydrological extreme events. Hydrological processes modelling at high temporal resolution is required to better understand flow patterns at catchment scale. A physically-based model called Hydrological Simulated Program-FORTRAN and two deep learning-based models were implemented to model surface runoff and sub-surface flow in the tropical humid headwater catchment of Houay Pano, northern Lao PDR. One deep learning model consisted of only one long short-term memory (simple LSTM), whereas the other model simulated processes in each hydrological response unit (HRU) by defining one separate LSTM for each HRU (HRU-based LSTM). The models use time-series data and two-dimensional spatial data to predict surface and sub-surface flows at 6-minute time step. The simple LSTM model outperformed the other models on surface runoff prediction, whereas HRU-based LSTM model better predicted sub-surface flow. This study demonstrates the performance of a deep learning model when simulating hydrological cycle with high temporal resolution.

The paper is an achievement of the GET–iEES Paris–UNIST collaboration and was published in the Journal of Hydrology. 50 days’ free access is available here.