Global sensitivity analysis with deep learning-based surrogate models for unraveling key parameters and processes governing redox zonation in riparian zone

The riparian zone constitutes an intricate redox environment, giving rise to distinct redox zones characterized by dissolved oxygen (DO), nitrate, manganese dioxide (Mn (IV)), iron hydroxide (Fe(III)), and sulfate. Processes encompassing river fluctuation, groundwater flow, microbial growth and death, as well as solute reactive transport, exert substantial influences on the spatiotemporal zonation of these redox zones in the riparian zone. Nonetheless, understanding remains elusive regarding how these processes govern the thickness of these zones. In this study, we built a one-dimensional (1-D) model which is adapted from Zhu et al. (2023) that integrates these processes to simulate the temporal variation of the thicknesses of different redox zones in riparian zone. Sobol’s sensitivity analysis method and the hierarchical sensitivity analysis framework based on Bayesian Networks (BNs) were then employed to quantify the sensitivities of the 17 selected parameters and the four generalized processes governing redox zonation, respectively. To mitigate the computational costs, multi-layer perceptron (MLP) was employed to establish the surrogate models, efficiently predicting model outputs for sensitivity analysis. The results of sensitivity analysis highlight the groundwater flow and reactive transport processes as the most important processes affecting thickness variations of the redox zones. Parameters such as the initial hydraulic conductivity (Ks0) and DOC concentration (CDOC) play a predominant role in governing the thickness variations. Parameters linked to river fluctuation and microbe growth processes exhibit very limited effects on redox zonation. These findings offer valuable insights into the factors controlling redox zones in the riparian environment. [Download paper here]

Recommended citation: Zhejiong Yu, Heng Dai, Jing Yang, Yonghui Zhu, Songhu Yuan. "Global sensitivity analysis with deep learning-based surrogate models for unraveling key parameters and processes governing redox zonation in riparian zone." Journal of Hydrology. 2024, 638: 131442. https://www.sciencedirect.com/science/article/pii/S0022169424008370