Machine Learning Applications For Downscaling Groundwater Storage Changes Integrating Satellite Gravimetry And Other Observations

Download or read book Machine Learning Applications for Downscaling Groundwater Storage Changes Integrating Satellite Gravimetry and Other Observations written by Vibhor Agarwal and published by . This book was released on 2021 with total page 182 pages. Available in PDF, EPUB and Kindle. Book excerpt: Anthropogenic excessive groundwater depletion (GWD) is a major problem affecting numerous regions in the world that depend on these precious water resources for drinking, irrigation, industrial and urban needs. Climate change is thought to further exacerbate scarcity and degrade the quality of these freshwater resources globally. The Gravity Recovery and Climate Experiment (GRACE) and its successor, GRACE Follow- On (GRACE-FO) twin-satellite gravimetry missions, have been observing the global temporal variations in Terrestrial Water Storage (TWS) for almost two decades at monthly sampling and spatial resolution longer than 333 km (half-wavelength). Innovative methodologies have enabled the retrieval of Groundwater Storage (GWS) anomalies in the world’s large climate-stressed aquifers, or disaggregated the signal from satellite gravimetry observed total TWS by removing the surface hydrologic signals via simulated or assimilated hydrologic model output data, or via hydrologic observations. However, uncertainties during the disaggregation process coupled with the limited spatial resolution (666 km grids) of GRACE/GRACE-FO estimated GWS have limited the use of such data for local-scale assessment of GWS variations and for practical applications of water resources management. In this research, we develop and leverage Machine Learning (ML) approach to estimate decadal or longer GW variations for the Central Valley (CV) in California, USA, and North China Plain (NCP) in China, to a local scale (5 km). These two study regions are among the regions in the world, largely dependent on GW for agricultural irrigation and other usages and are currently undergoing severe GWD due primarily to anthropogenic activities and plausibly exacerbated by an increasingly warmer Earth. First, we developed and implemented the robust Artificial Neural Network (ANN) and Random Forest (RF) ML modeling framework in the Central Valley (CV) to study the severe GWD problem using GRACE-derived TWS and other hydrological data as input variables and GW level (GWL) as output over the entire study region. RF ML model showed that GRACE-derived TWS is the most important predictor variable, and we concluded that the RF model is a better choice to model the GWS variations over the CV, as compared to ANN. We then internally validated our modeled results using the continuously available in-situ GWL data over Oct 2002-Sep 2016 used to build the ML model. It is well-known that excessive anthropogenic GW pumping has directly led to severe land subsidence in the CV. Thus, we compare the ML downscaled decadal GWL changes with independently determined geodetic land subsidence observations, including Global Positioning System (GPS) vertical land displacement data and Cryosat-2 radar altimeter observed land surface subsidence data, which shows overall good correlations. In addition, we estimated the inelastic storage coefficient (Skv), an important aquifer mechanical property, during the extended drought period of 2011-2015 for the southern portion of the San Joaquin (SJ) valley, which is located south of the Sacramento-SJ River Delta and is part of the CV. The CV lost 30 km3 of GWS during our study period of Oct 2002-Sep 2016 and showed the maximum rate of GWS loss during the severe droughts. Our analysis provides an overall holistic understanding of the spatiotemporal variations of GWD in the region. Similarly, we developed and implemented the RF model to study GWD in the North China Plain (NCP) located in China during 2005-2013. Here the in-situ GWL data display better spatiotemporal coverage than the GWL data in the Central Valley (CV), California, USA. We constructed separate RF models for shallow and deeper wells, and while the shallow wells show good accuracy, the accuracy of the deeper wells can be improved further with the selection of more complex input patterns in a future study. RF modeling result shows that the GRACE-derived TWSA input data are the most important variable for both the ML models. We generated high-resolution GWS trend maps at 5 km resolution during the study period of 2005-2013, which further validate that the rate of GWD is dependent on irrigation intensity, as well as closeness to the cities and industrial activity, similar to what had been concluded by previous studies. The NCP region shows a rapid GWS decline for deeper wells as compared to the decline observed in the shallow wells. The overall rate of GWS loss for the NCP is -5.4 ± 1.0 km3 yr-1 during 2005-2013. In conclusion, the ML modeling approaches performed well in predicting the complex GWL at high spatial resolutions in both study regions, Central Valley, California, USA, and North China Plain, China. The modeled results are consistent with past studies. This study employs more data and covered larger areas of two distinctly climatically and geographically different aquifers in the world. It is anticipated that the ML-based modeling approach is applicable to holistically quantify the GWD in the world’s large aquifers, especially where in-situ GW wells are sparse or unavailable.