[A-GE28] Energy-Environment-Water Nexus and Sustainable Development

[AGE28-P05] Understanding karst hydrological processes via interpretable machine learning

Huiqing Hao¹、*Yonghong Hao Hao¹ (1.Tianjin Normal University)

キーワード：Karst hydrological processes, Spring discharge, Explainable machine learning , LSTM, SHAP , Niangziguan Springs

A karst spring record can be used to manifest the integrated effects of spatiotemporal varying hydrologic processes. Understanding the mechanism of karst hydrological processes including groundwater recharge, convergence, and discharge is vital to agricultural irrigation, urban water supply, and regional water resources sustainable development. Behaviors of karst springs are manifestations of spatial and temporal dynamics involving multi-hydrogeological processes, including precipitation, surface water runoff, infiltration, groundwater flow, and anthropogenic activities. These dynamic processes are usually nonlinear and nonstationary. In this study, we couple the Shapley Additive exPlanation (SHAP) with a Long Short-Term Memory (LSTM) recurrent neural network to produce a explainable deep learning model to explore the precipitation driven spring discharge mechanism and to predict spatial-temporal behaviors of karst springs. The model is then used to study Niangziguan Springs catchment, China, and the results show that the precipitation infiltration volume of each catchment subregion is the primary factor driving the spring discharge, and the precipitation over the 12-month period has the most significant effect on the spring discharge. We classify the precipitation-driven spring discharge at the catchment into three patterns according to each subregion's landform and karst aquifer characteristics based on the SHAP analysis. Firstly, hydraulic connections between the surface formation and karst aquifer are copious in the well-developed river valleys. Since valleys cut through sediments and reach karst aquifers, moderate- and low-intensity precipitation leads to fewer recharges to the spring than high-intensity ones. Secondly, in subregions where the karst aquifer is deep, and the mountain valley slopes are steep, heavy precipitation becomes surface runoff, leaving only a tiny fraction for spring recharges. As a result, the karst aquifer is mainly recharged by moderate to light precipitation. Thirdly, in the karst exposed and groundwater discharge areas, the groundwater level is the primary factor dictating precipitation and spring discharge processes. When the groundwater level is low, precipitation can contribute to the spring discharge irrespective of its intensity. Only heavy precipitation can recharge the spring discharge when the groundwater level is high. The active time length (i.e., the effective time) of precipitation/ spring discharge is identified as 12 months for the subregion in the catchment with uneven effectiveness. In the first type, precipitation within 6–9 months contributes significantly to the spring discharge. In the second type, contributions of precipitation to the spring discharge increase gradually within 12 months. The disparity could stem from spatiotemporal features of the recharge-discharge in the heterogeneous karst aquifers of the catchment during groundwater converging and diverging flows. The specific cases of the spring discharge maximum and medium represent the major hydrological processes of the karst catchment. Quantifying the contribution from precipitation to the maximum and medium spring discharge enhances our understanding of hydrological processes in the karst catchment. When the spring discharge maximum occurs, the precipitation at each subregion may not necessarily reaches the maximum. Effective high precipitation in most subregions is responsible for the maximum spring discharge. The medium spring discharge signifies that most precipitation at subregions approximately is the average value, leading to the medium spring discharge.