GSA Connects 2021 in Portland, Oregon

Paper No. 71-9
Presentation Time: 10:45 AM


SMITH, Ryan1, MAJUMDAR, Sayantan1, HASAN, Md Fahim1, BUTLER Jr., J.J.2, LAKSHMI, Venkataraman3, CONWAY, Brian D.4 and RIGBY, James R.5, (1)Geological Engineering, Missouri S&T, 1400 N Bishop St, Rolla, MO 65409-6531, (2)Kansas Geological Survey, University of Kansas, 1930 Constant Ave, Lawrence, KS 66047-3724, (3)Engineering Systems and Environment, University of Virginia, P.O. Box 400224, Charlottesville, VA 22904-4224, (4)Arizona Department of Water Resources, 1110 W. Washington Ave, Suite 310, Phoenix, AZ 85007, (5)U.S. Geological Survey, Lower Mississippi-Gulf Water Science Center, 640 Grassmere Park, Nashville, TN 37211

Changes in groundwater storage due to excessive withdrawals threatens food, water and energy security. Globally, groundwater withdrawals are one of the primary drivers creating changes in storage. However, very few regions of the world monitor groundwater withdrawals at the local scale necessary to implement sustainable management solutions. Multiple satellite sensors are sensitive to different elements of groundwater flux, yet integrating these sensors to predict withdrawals is a challenge due to a lack of water use data, and complexities in the theory relating sensor measurement to withdrawals. Machine learning methods have been shown to be effective in relating earth science observations to various processes and are a promising technique for improving our ability to model and monitor groundwater systems. Here, we present a machine learning framework for integrating multiple satellite sensors and derivative data products related to water use. Using primarily data from Landsat, Sentinel-1, GRACE, GRACE-FO, and MODIS, as well as products derived from sensors on these satellites, we are able to estimate groundwater withdrawals at the local (1-5 km resolution) scale. The satellite products measure different components of the water balance at resolutions ranging from 10 m to 100s of km, and are integrated using data driven algorithms including random forests and neural networks. We are currently implementing these models in portions of three regional aquifer systems: the Mississippi Alluvial Aquifer in Mississippi, Arkansas, Missouri, and Louisiana; the High Plains Aquifer, in Kansas; and the Basin and Range Aquifer System, in Arizona. Complexity in irrigation practices and the aquifer itself results in varying model quality with r2 ranging from 0.5 to 0.9 across these study areas. Ultimately, we plan to extend these withdrawal estimates to the conterminous United States, producing high resolution, annual estimates of groundwater use.