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

Abstract

Recent intensification in climate change have resulted in the rise of hydrological extreme events. This demands modeling of hydrological processes at high temporal resolution to better understand flow patterns in catchments. To model surface and sub-surface flows in a catchment we utilized a physically based model called Hydrological Simulated Program-FORTRAN and two deep learning-based models. 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 environmental time-series data and two-dimensional spatial data to predict surface and sub-surface flows at 6-minute time step simultaneously. We tested our models in a tropical humid headwater catchment in northern Lao PDR and compared their performances. Our results showed that the simple LSTM model outperformed the other models on surface runoff prediction with the lowest MSE (7.4e−5 m3 s−1), whereas HRU-based LSTM model better predicted patterns and slopes in sub-surface flow in comparison with the other models by having the smallest MSE value (3.2e−4 m3 s−1). This study demonstrated the performance of a deep learning model when simulating hydrological cycle with high temporal resolution.