Machine-Learning Predictions of Groundwater Specific Conductance in the Mississippi Alluvial Plain, South-Central United States, With Evaluation of Regional Geophysical Aerial Electromagnetic Data as Explanatory Variables

Scientific Investigations Report 2023-5099
Water Availability and Use Science Program
By:  and 

Links

Abstract

The Mississippi Alluvial Plain, located in the south-central United States, is undergoing long-term groundwater-level declines within the surficial Mississippi River Valley alluvial aquifer (hereinafter referred to as “alluvial aquifer”), which has raised concerns about future groundwater availability. In some parts of the alluvial aquifer, groundwater availability for common uses such as irrigation, public supply, and domestic use is limited by quality (for example, high salinity) rather than quantity of water stored in the aquifer. The Mississippi Alluvial Plain region has an abundance of water-quality measurements in the alluvial aquifer and deeper aquifers; however, large areas lack direct measurements of salinity to evaluate regional groundwater availability. Statistical models can interpolate between wells to fill in spatial data gaps. In 2021, the U.S. Geological Survey trained two boosted regression tree (BRT) machine-learning models on specific conductance data available between 1942 and 2020 to predict spatially continuous surfaces of groundwater salinity at multiple depths for the alluvial aquifer and deeper aquifers. Well construction information, water levels, and surficial variables such as geomorphology and soils were included as explanatory variables in this baseline model. Additionally, subsurface electrical resistivity data from the first aquifer-wide aerial electromagnetic (AEM) survey for the region were incorporated to create a geophysical model. This work expands on prior BRT salinity predictions of the alluvial aquifer and extends predictions south to the Gulf of Mexico, where groundwater salinity is high. AEM survey data were not available for the southern extent of the alluvial aquifer at the time of modeling. A BRT model was trained without (baseline) and with (geophysical) AEM variables to test the ability of the models to predict salinity where explanatory data are missing and response data are sparse. Additionally, model sensitivity to AEM survey data was evaluated to better understand how AEM variables influence specific conductance predictions. Model performance was improved with the addition of geophysical data, which added three-dimensional information, thereby improving salinity predictions at depth. Groundwater specific conductance predictions can help inform other geophysical investigations in the southern extent of the study area, where high groundwater specific conductance can obfuscate changes in aquifer sediment resistivity and could limit groundwater resources for agricultural, public supply, and domestic uses.

Suggested Citation

Killian, C.D., and Knierim, K.J., 2023, Machine-learning predictions of groundwater specific conductance in the Mississippi Alluvial Plain, south-central United States, with evaluation of regional geophysical aerial electromagnetic data as explanatory variables: U.S. Geological Survey Scientific Investigations Report 2023–5099, 36 p., 1 pl., https://doi.org/10.3133/sir20235099.

ISSN: 2328-0328 (online)

Study Area

Table of Contents

  • Acknowledgments
  • Abstract
  • Introduction
  • Methods
  • Results
  • Discussion
  • Summary
  • References Cited
Publication type Report
Publication Subtype USGS Numbered Series
Title Machine-learning predictions of groundwater specific conductance in the Mississippi Alluvial Plain, south-central United States, with evaluation of regional geophysical aerial electromagnetic data as explanatory variables
Series title Scientific Investigations Report
Series number 2023-5099
DOI 10.3133/sir20235099
Year Published 2023
Language English
Publisher U.S. Geological Survey
Publisher location Reston, VA
Contributing office(s) Lower Mississippi-Gulf Water Science Center
Description Report: viii, 36 p., 1 Plate: 33.04 × 37.14 inches; Dataset; Data Release
Country United States
Online Only (Y/N) Y
Google Analytic Metrics Metrics page
Additional publication details