Improving Model Performance Through Indirect Integration of Human Impacts in Data-Driven Hydrologic Models

Abstract

Increasing water demand driven by fast-growing global population and climate change impact are threatening the groundwater resources. As a result, major aquifer systems in the world are undergoing depletion, The Central Valley aquifers is an examples. The paucity of hydrologic data, specifically groundwater pimping data contributes to model uncertainty which limits our capacity to better assess terrestrial water storage trends and manage water resources. This study aimed at improving hydrologic model performance by using multi-sourced datasets: in-situ, remote sensing, and model-based data in a highly stressed aquifer system where very limited water use or groundwater pumping data exists. Machine learning-based models estimating groundwater level anomaly (GWLA) were constructed using predictor data such as Terrestrial Water Storage Anomaly (TWSA), precipitation, soil moisture, stream discharge (Q), Land Surface Temperature (LST), Normalized Difference Vegetation Index (NDVI), plant canopy water, evaporation, transpiration, texture of aquifer material, clay and aquifer thickness and Vertical Ground Displacement (VGD) data from Continuous Global Positioning Systems (CGPS). In the study area, ground deformation is highly associated with groundwater pumping and/or recharge. Thus, the VGD data, assumed as indirect measures of groundwater pumping were used as one of the predictor variables in the models. Two model sets, one that includes the VGD as predictor while the other set with no VGD data as predictor variable, were trained and evaluated using boosted regression tree technique. Each model set consists of 23 individual models representing unique groundwater wells in the study area. The models run at a monthly time scale with a time span ranging from 2002-2017. The results indicated models that include VGD data as predictor generally performed better than models constructed without the VGD data. Considering the VGD data, the average statistical measures between simulated and observed groundwater level anomaly were 658.7 mm, 0.94, and 0.88 for Root MEAN Square Error (RMSE), R, and Nash-Sutcliffe Efficiency (NSE), respectively. While without the VGD data, the statistical measures between simulated and observed were 1261.3 mm, 0.91, and 0.80 for RMSE, R, and NSE, respectively. Among the predictor variables, the VGD data were the primary influential predictor variable followed by TWSA, soil moisture, and stream discharge. Models trained with the VGD as predictor performed better for groundwater wells located in higher depletion zones where ground subsidence is significant. Models trained without the VGD data showed slightly better performance for groundwater wells located in low depletion areas where rate of groundwater pumping is modest.