A flexible multi-scale approach for downscaling GRACE-derived groundwater storage anomaly using LightGBM and random forest in the Tashk-Bakhtegan Basin, Iran

Abstract

Study region: Tasht-Bakhtegan Basin, Iran Study focus: The main objectives of this study are to reconstruct and downscale GRACE data from a coarse resolution of 1-degree to a finer resolution of 1-km. This was accomplished using a robust and flexible multi-scale approach, leveraging machine learning algorithms, specifically random forest and LightGBM. The models were meticulously calibrated and thoroughly evaluated across various spatial scales. Additionally, the study examined the lag effects of influential covariates in the downscaling process, further enhancing model accuracy.New hydrological insights for the region The multi-scale calibration of the models provided new insights into the relationship between terrestrial water storage anomalies (TWSa) and various environmental and hydrological factors. It was found that precipitation and land surface temperature (LST) were the most influential covariates in the reconstruction and downscaling process. Specifically, precipitation with a two-month delay, LST with a three-month delay, and evapotranspiration with an eight-month delay exhibited the highest correlations with TWSa. These findings offer valuable insights into the temporal influence of key hydrological variables on TWSa within the region, shedding light on how delayed responses of precipitation, LST, and evapotranspiration affect groundwater storage. This enhances the understanding of the underlying dynamics governing hydrological variability in the study area.

Keywords

• GRACE
• Total water storage
• LightGBM
• Clustering