Causal interpretation for groundwater exploitation strategy in a coastal aquifer

Machine learning models (ML), as a collection of nonparametric or semiparametric estimation methods, can successfully encode the distribution of the problems into its trainable parameters based on observation data. However, the distributions of hydrological variables may change suddenly under complex environmental conditions, leading to biased estimates when using ML models. This work is the first attempt to solve this issue using structural causal models (SCM). Specifically, two SCM were constructed based on hydrological conditions and monitoring data. Then the Propensity Score estimator and the Double Machine Learning estimator were employed to estimate the causal effects of four treatments on the mean Cl^- concentration (MCL) in the coastal aquifer. The results showed that pumping groundwater from area A1 or increasing the river level directly leads to a decrease in MCL, while pumping area A3 directly leads to an increase in MCL. Moreover, MCL can be effectively controlled by cooperative-treatment strategies. Finally, two practical exploitation strategies are derived. In the planting month, it should increase groundwater pumping from area A1, limit groundwater pumping from A2, and prohibit groundwater pumping from A3. For the normal month, it is proposed to increase the height of the rubber dam to raise the river level and reduce groundwater pumping from A1 and A2.