Deciphering Nonlinear Hydrological Process by a Coupled Deep Learning and Physical Based Model in Southern Tibetan Plateau

Abstract

Abstract Interpretability of deep learning (DL) poses a significant challenge in hydrology modeling, particularly under the complex and frigid conditions of the Tibetan Plateau (TP), which further restricts its application. In this study, we developed a cascade‐style hybrid modeling framework by integrating the Variable Infiltration Capacity (VIC) model with a two‐dimensional grid long short‐term memory (termed VIC‐LSTM), and incorporated a dual‐layer probe for training and investigating non‐linear hydrological processes. Our objective was to explore the framework's potential for enhancing hydrological simulation accuracy and expanding interpretability. This framework was adopted for the Yarlung Zangbo River Basin (above the Lazi guaged station) in the TP. The results of the VIC‐LSTM demonstrated its effectiveness, achieving a simulated daily streamflow NSE of 0.78 compared to 0.69 for the VIC model during the training period. Moreover, the probe experiments, aided by remote‐sensed images, successfully deciphered the timing signals of snowmelt and glacier melt, with ablation and duration time errors within 2 weeks. The average errors for snowmelt and glacier melt were approximately 7 and 10 days, respectively. The combined spatial and temporal feature quantification indicated that snowmelt and glacier melt contributed 23.7% and 7.7% to the total streamflow, respectively. Additionally, VIC‐LSTM identified that snowmelt signals generally preceded glacier melt signals. These findings underscore the potential of VIC‐LSTM to interpret hydrological processes, increasing confidence in using DL approaches. Therefore, it offers a new perspective for quantifying water resources and understanding hydrological processes in high‐elevation, glacier‐snow coexisting basins.

Keywords

• deep learning
• hybrid modeling
• hydrological process
• snowmelt runoff
• Tibetan Plateau