Variable sensitivity analysis in groundwater level projections under climate change adopting a hybrid machine learning algorithm

Studies on climate change have largely overlooked the delayed response of Ground Water Levels (GWL) to atmospheric conditions. This gap is critical because fluctuations in GWL can lead to hazards like land subsidence. This study addresses the issue by identifying optimal delay times for key variables, which improves GWL projection accuracy. The input data process consists of introducing meteorological and hydrological variables in the form of 42 combinations. Meteorological data under climate change scenarios were obtained by downscaling outputs from the General Circulation Models (GCMs) within the Shared Socio-economic Pathways (SSP) scenarios. So far, no similar study has attempted to rank such a wide array of delay time combinations. This study improves hybrid Random Forest and Genetic Algorithm (RF-GA) projections by introducing the best combination of input variables. The investigation assessed the performance of both the conventional Random Forest (RF) and the RF-GA in simulating groundwater fluctuations. The variable sensitivity analysis results indicated that watershed discharge holds a higher Variable Importance (VI) compared to meteorological variables. The findings in the validation section also demonstrated that the RF-GA outperformed an RF that runs on default hyperparameters. Temperature and evaporation show a 3 and 2-month delay time, respectively. It was discovered that precipitation was the only variable with two possible delay times of 2 and 4-month. Also, combinations with many and few variables performed poorly. The projection results indicate an increase of 6.8 and 7.1 cm in the average GWL in the Silakhor plain under the low-emission SSP1-2.6 and high-emission SSP5-8.5 scenarios, respectively.