GSA Annual Meeting in Phoenix, Arizona, USA - 2019

Paper No. 17-4
Presentation Time: 9:05 AM

ESTIMATING GROUNDWATER EXTRACTION WITH INTEGRATED SATELLITE DATASETS AND MACHINE LEARNING


SMITH, Ryan G.1, MAJUMDAR, Sayantan2, OYLER, Lindi1, BUTLER Jr., James J.3 and LAKSHMI, Venkataraman4, (1)Geosciences and Geological and Petroleum Engineering, Missouri University of Science and Technology, 1400 N Bishop, 129 McNutt Hall, Rolla, MO 65409, (2)Geosciences and Geological and Petroleum Engineering, Missouri University of Science and Technology, Missouri S&T, Rolla, MO 65409, (3)Kansas Geological Survey, University of Kansas, 1930 Constant Ave, Lawrence, KS 66047, (4)Engineering Systems and Environment, University of Virginia, P.O. Box 400224, Charlottesville, VA 22904-4224

Groundwater is a key component of the water cycle, and monitoring groundwater withdrawals is a crucial step in water resources management and mitigating storage loss. However, very few regions of the U.S. and world monitor their groundwater withdrawals at the local scale necessary to implement sustainable management solutions. In this study, we use data from various satellite sensors that measure different components of the water balance at resolutions ranging from 10 m to 100s of km. Since the measurements these satellites collect are related in different ways to the water balance, combining them can result in a powerful monitoring tool. However, due to their varying resolutions and the complexity of combining them in a water balance, approaches to use them to estimate groundwater withdrawals have been limited. Here we overcome these limitations using a hybrid water balance/machine learning approach. Using this approach, we estimate groundwater withdrawals at ~1 km resolution. Our study area is the High Plains Aquifer in Kansas, where withdrawals have depleted the aquifer substantially.

We use a combination of evapotranspiration estimates from MODIS, precipitation estimates from TRMM, GPM and PRISM, and crop type and land use estimates from NASS as proxies for water demand, as well as surface water extent maps with weighted spatial filters to estimate quantity of and proximity to surface water. When combined in a machine learning framework, these variables can predict groundwater withdrawals at the local scale. We use the random forests approach to assess the impact of each variable on the outcome and determine the mechanisms driving groundwater demand, and find that land use as well as proxies for surface water availability are strong predictors of groundwater demand. This method can be used to estimate groundwater withdrawals in regions with limited data, or to fill in temporal or spatial gaps in areas with moderate data. Implementing this method to estimate groundwater could enable water managers to proactively implement sustainable solutions.