Boosted regression trees (BRT), a type of ensemble-tree machine-learning method, were used to predict specific conductance concentration at multiple depths throughout the Mississippi River Valley alluvial aquifer (MRVA) and underlying aquifers. Groundwater from the MRVA, coincident with the Mississippi Alluvial Plain (MAP), is a vital resource for agriculture and drinking-water supplies in the central United States. Water availability can be limited in some areas of the aquifer by high concentrations of salinity, measured as specific conductance. Two models were created to test the incorporation of datasets from a regional aerial electromagnetic (AEM) survey and evaluate model performance. Explanatory variables for the BRT models included attributes associated with well location and construction, surficial variables (such as hydrologic position and recharge), and variables from the AEM survey of the aquifer. This data release provides the R scripts to tune and reproduce the BRT models and final prediction rasters. For a full description of modeling workflow and final model selection see: Killian, C.D. and Knierim, K.J., (2022) Machine learning predictions of groundwater specific conductance in the Mississippi Alluvial Plain, United States with evaluation of geophysical and regional aerial electromagnetic data as predictor variables, US Geological Survey Scientific Investigations Report XXXX.

Boosted regression trees (BRT), a type of ensemble-tree machine-learning method, were used to predict specific conductance concentration at multiple depths throughout the Mississippi River Valley alluvial aquifer (MRVA) and underlying aquifers. Groundwater from the MRVA, coincident with the Mississippi Alluvial Plain (MAP), is a vital resource for agriculture and drinking-water supplies in the central United States. Water availability can be limited in some areas of the aquifer by high concentrations of salinity, measured as specific conductance. Two models were created to test the incorporation of datasets from a regional aerial electromagnetic (AEM) survey and evaluate model performance. Explanatory variables for the BRT models included attributes associated with well location and construction, surficial variables (such as hydrologic position and recharge), and variables from the AEM survey of the aquifer. This data release provides the R scripts to tune and reproduce the BRT models and final prediction rasters. For a full description of modeling workflow and final model selection see: Killian, C.D. and Knierim, K.J., (2022) Machine learning predictions of groundwater specific conductance in the Mississippi Alluvial Plain, United States with evaluation of geophysical and regional aerial electromagnetic data as predictor variables, US Geological Survey Scientific Investigations Report XXXX.

A multiple machine-learning model (Asquith and Killian, 2024) implementing Cubist and Random Forest regressions was used to predict monthly mean groundwater levels through time for the available years described in the metadata for the Mississippi River Valley alluvial aquifer (MRVA). The MRVA is the surficial aquifer of the Mississippi Alluvial Plain (MAP), located in the south-central United States. Employing two machine-learning techniques offered the opportunity to generate model and statistical error and covariance between them to estimate total uncertainty. Potentiometric surface predictions were made at the 1-kilometer grid scale using the National Hydrogeologic Grid (Clark and others, 2018). Results produced by the mmlMRVAgen1 software have been condensed into thirteen themes each containing a multi-banded GeoTIFF raster with 504 layers, corresponding to each month for the available years described in the metadata. The themes include the final predicted monthly water-level altitudes, in feet North American Vertical Datum of 1988 (NAVD88), for the study area (pol), which were computed by pooling through weighted-mean averaging by cell the even and odd year predictions for that month. The depth to water was predicted in feet (nhgd2w), utilizing the NHG cell altitude as the land surface datum. Model errors were evaluated using both the normal error (modnorerr) in standard deviations of feet and the polynomial-density-quantile4 distribution (PDQ4)-error model (modpdqerr) without the inclusion of land-surface variation of the NHG. The equivalent standard deviations of these error models were calculated both with and without the inclusion of land-surface variation of the NHG (norerr, norerrnhg, pdqerr, pdqerrnhg). The lower and upper bounds (in feet) of the 90-percent prediction limits for both model error forms were computed (norlwr, norupr, pdqlwr, pdqupr). Lastly, the ratio of model error to total error (modtotrat) was also computed. Complementing each of the GeoTIFFs are `.json` extensions to each file. These provide additional multi-band support information. This double-file representation stems from the native GeoTIFF drivers within the terra R package underpinning the operations. Overall, the model objects created by the mmlMRVAgen1 from about 156,000 water-level records for about 58,000 wells report (1) a normalized Nash−Sutcliffe Efficiency (NNSE) of about 0.997, (2) a root-mean-square error (RMSE) of about 4.15 feet, and (3) a bias prior to computing the NNSE and RMSE of about 0.0963 feet before its subsequent removal (see mmlMRVAgen1 software diagnostics associated with "MRVA_MML_CONSTANTS"). The model objects also report for the 156,000 water-level records (1) a mean percent ratio of model error to total error of about 69.2 percent and (2) a mean width of about 12.05 feet for the 90-percent prediction bounds from the PDQ4 error framework (see mmlMRVAgen1 software diagnostics associated with "genMML/03step.R"). The model objects were used in post-model creation to predict each of the rasters provided in this data release. (Note, the results herein are associated with the "April 21, 2024" model run, see mmlMRVAgen1/model_archive/README.md.) For a full description of covariate assemblage and hydrograph modeling, see Asquith and Killian (2022) (covMRVAgen1 software). For a full description of multiple machine-learning modeling, see Asquith and Killian (2024) (mmlMRVAgen1 software).

A multiple machine-learning model (Asquith and Killian, 2024) implementing Cubist and Random Forest regressions was used to predict monthly mean groundwater levels through time for the available years described in the metadata for the Mississippi River Valley alluvial aquifer (MRVA). The MRVA is the surficial aquifer of the Mississippi Alluvial Plain (MAP), located in the south-central United States. Employing two machine-learning techniques offered the opportunity to generate model and statistical error and covariance between them to estimate total uncertainty. Potentiometric surface predictions were made at the 1-kilometer grid scale using the National Hydrogeologic Grid (Clark and others, 2018). Results produced by the mmlMRVAgen1 software have been condensed into thirteen themes each containing a multi-banded GeoTIFF raster with 504 layers, corresponding to each month for the available years described in the metadata. The themes include the final predicted monthly water-level altitudes, in feet North American Vertical Datum of 1988 (NAVD88), for the study area (pol), which were computed by pooling through weighted-mean averaging by cell the even and odd year predictions for that month. The depth to water was predicted in feet (nhgd2w), utilizing the NHG cell altitude as the land surface datum. Model errors were evaluated using both the normal error (modnorerr) in standard deviations of feet and the polynomial-density-quantile4 distribution (PDQ4)-error model (modpdqerr) without the inclusion of land-surface variation of the NHG. The equivalent standard deviations of these error models were calculated both with and without the inclusion of land-surface variation of the NHG (norerr, norerrnhg, pdqerr, pdqerrnhg). The lower and upper bounds (in feet) of the 90-percent prediction limits for both model error forms were computed (norlwr, norupr, pdqlwr, pdqupr). Lastly, the ratio of model error to total error (modtotrat) was also computed. Complementing each of the GeoTIFFs are `.json` extensions to each file. These provide additional multi-band support information. This double-file representation stems from the native GeoTIFF drivers within the terra R package underpinning the operations. Overall, the model objects created by the mmlMRVAgen1 from about 156,000 water-level records for about 58,000 wells report (1) a normalized Nash−Sutcliffe Efficiency (NNSE) of about 0.997, (2) a root-mean-square error (RMSE) of about 4.15 feet, and (3) a bias prior to computing the NNSE and RMSE of about 0.0963 feet before its subsequent removal (see mmlMRVAgen1 software diagnostics associated with "MRVA_MML_CONSTANTS"). The model objects also report for the 156,000 water-level records (1) a mean percent ratio of model error to total error of about 69.2 percent and (2) a mean width of about 12.05 feet for the 90-percent prediction bounds from the PDQ4 error framework (see mmlMRVAgen1 software diagnostics associated with "genMML/03step.R"). The model objects were used in post-model creation to predict each of the rasters provided in this data release. (Note, the results herein are associated with the "April 21, 2024" model run, see mmlMRVAgen1/model_archive/README.md.) For a full description of covariate assemblage and hydrograph modeling, see Asquith and Killian (2022) (covMRVAgen1 software). For a full description of multiple machine-learning modeling, see Asquith and Killian (2024) (mmlMRVAgen1 software).

Groundwater is a vital resource in the Mississippi embayment of the central United States. An innovative approach using machine learning (ML) was employed to predict groundwater salinity—including specific conductance (SC), total dissolved solids (TDS), and chloride (Cl) concentrations—across three drinking-water aquifers of the Mississippi embayment. A ML approach was used because it accommodates a large and diverse set of explanatory variables, does not assume monotonic relations between predictors and response data, and results can be extrapolated to areas of the aquifer not sampled. These aspects of ML allowed potential drivers and sources of high salinity water that have been hypothesized in other studies to be included as explanatory variables. The ML approach integrated output from a groundwater-flow model and water-quality data to predict salinity, and the approach can be applied to other aquifers to provide context for the long-term availability of groundwater resources. The Mississippi embayment includes two principal regional aquifer systems; the surficial aquifer system, dominated by the Quaternary Mississippi River Valley Alluvial aquifer (MRVA), and the Mississippi embayment aquifer system, which includes deeper Tertiary aquifers and confining units. Based on the distribution of groundwater use for drinking water, the modeling focused on the MRVA, middle Claiborne aquifer (MCAQ), and lower Claiborne aquifer (LCAQ). Boosted regression tree (BRT) models (Elith and others, 2008; Kuhn and Johnson, 2013) were developed to predict SC and Cl to 1-kilometer (km) raster grid cells of the National Hydrologic Grid (Clark and others, 2018) for 7 aquifer layers (1 MRVA, 4 MCAQ, 2 LCAQ) following the hydrogeologic framework of Hart and others (2008). TDS maps were created using the correlation between SC and TDS. Explanatory variables for the BRT models included attributes associated with well location and construction, surficial variables (such as soils and land use), and variables extracted from a MODFLOW groundwater flow model for the Mississippi embayment (Haugh and others, 2020a; Haugh and others, 2020b). Prediction intervals were calculated for SC and Cl by bootstrapping raster-cell predictions following methods from Ransom and others (2017). For a full description of modeling workflow and final model selection see Knierim and others (2020).

Groundwater is a vital resource in the Mississippi embayment of the central United States. An innovative approach using machine learning (ML) was employed to predict groundwater salinity—including specific conductance (SC), total dissolved solids (TDS), and chloride (Cl) concentrations—across three drinking-water aquifers of the Mississippi embayment. A ML approach was used because it accommodates a large and diverse set of explanatory variables, does not assume monotonic relations between predictors and response data, and results can be extrapolated to areas of the aquifer not sampled. These aspects of ML allowed potential drivers and sources of high salinity water that have been hypothesized in other studies to be included as explanatory variables. The ML approach integrated output from a groundwater-flow model and water-quality data to predict salinity, and the approach can be applied to other aquifers to provide context for the long-term availability of groundwater resources. The Mississippi embayment includes two principal regional aquifer systems; the surficial aquifer system, dominated by the Quaternary Mississippi River Valley Alluvial aquifer (MRVA), and the Mississippi embayment aquifer system, which includes deeper Tertiary aquifers and confining units. Based on the distribution of groundwater use for drinking water, the modeling focused on the MRVA, middle Claiborne aquifer (MCAQ), and lower Claiborne aquifer (LCAQ). Boosted regression tree (BRT) models (Elith and others, 2008; Kuhn and Johnson, 2013) were developed to predict SC and Cl to 1-kilometer (km) raster grid cells of the National Hydrologic Grid (Clark and others, 2018) for 7 aquifer layers (1 MRVA, 4 MCAQ, 2 LCAQ) following the hydrogeologic framework of Hart and others (2008). TDS maps were created using the correlation between SC and TDS. Explanatory variables for the BRT models included attributes associated with well location and construction, surficial variables (such as soils and land use), and variables extracted from a MODFLOW groundwater flow model for the Mississippi embayment (Haugh and others, 2020a; Haugh and others, 2020b). Prediction intervals were calculated for SC and Cl by bootstrapping raster-cell predictions following methods from Ransom and others (2017). For a full description of modeling workflow and final model selection see Knierim and others (2020).

Groundwater is a vital resource to the Mississippi embayment region of the central United States. Regional and integrated assessments of water availability that link physical flow models and water quality in principal aquifer systems provide context for the long-term availability of these water resources. An innovative approach using machine learning was employed to predict groundwater pH across drinking water aquifers of the Mississippi embayment. The region includes two principal regional aquifer systems; the Mississippi River Valley alluvial (MRVA) aquifer and the Mississippi embayment aquifer system that includes several regional aquifers and confining units. Based on the distribution of groundwater use for drinking water, the modeling effort was focused on the MRVA, Middle Claiborne aquifer (MCAQ), and Lower Claiborne aquifer (LCAQ)of the Mississippi embayment aquifer system. Boosted regression tree (BRT) models (Elith and others, 2008; Kuhn and Johnson, 2013) were used to predict pH to 1-km raster grid cells of the National Hydrologic Grid (Clark and others, 2018). Predictions were made for 7 aquifer layers (1 MRVA, 4 MCAQ, 2 LCAQ) following the hydrogeologic framework used in a regional groundwater flow model (Hart and others, 2008). Explanatory variables for the BRT models included attributes associated with well position and construction, surficial variables, and variables extracted from a MODFLOW groundwater flow model for the MISE (Haugh and others, 2020a,b). For a full description of modeling workflow see Knierim and others (2020).

Groundwater is a vital resource to the Mississippi embayment region of the central United States. Regional and integrated assessments of water availability that link physical flow models and water quality in principal aquifer systems provide context for the long-term availability of these water resources. An innovative approach using machine learning was employed to predict groundwater pH across drinking water aquifers of the Mississippi embayment. The region includes two principal regional aquifer systems; the Mississippi River Valley alluvial (MRVA) aquifer and the Mississippi embayment aquifer system that includes several regional aquifers and confining units. Based on the distribution of groundwater use for drinking water, the modeling effort was focused on the MRVA, Middle Claiborne aquifer (MCAQ), and Lower Claiborne aquifer (LCAQ)of the Mississippi embayment aquifer system. Boosted regression tree (BRT) models (Elith and others, 2008; Kuhn and Johnson, 2013) were used to predict pH to 1-km raster grid cells of the National Hydrologic Grid (Clark and others, 2018). Predictions were made for 7 aquifer layers (1 MRVA, 4 MCAQ, 2 LCAQ) following the hydrogeologic framework used in a regional groundwater flow model (Hart and others, 2008). Explanatory variables for the BRT models included attributes associated with well position and construction, surficial variables, and variables extracted from a MODFLOW groundwater flow model for the MISE (Haugh and others, 2020a,b). For a full description of modeling workflow see Knierim and others (2020).

Groundwater is a vital resource to the Mississippi embayment region of the central United States. Regional and integrated assessments of water availability that link physical flow models and water quality in principal aquifer systems provide context for the long-term availability of these water resources. An innovative approach using machine learning was employed to predict groundwater pH across drinking water aquifers of the Mississippi embayment. The region includes two principal regional aquifer systems; the Mississippi River Valley alluvial (MRVA) aquifer and the Mississippi embayment aquifer system that includes several regional aquifers and confining units. Based on the distribution of groundwater use for drinking water, the modeling effort was focused on the MRVA, Middle Claiborne aquifer (MCAQ), and Lower Claiborne aquifer (LCAQ)of the Mississippi embayment aquifer system. Boosted regression tree (BRT) models (Elith and others, 2008; Kuhn and Johnson, 2013) were used to predict pH to 1-km raster grid cells of the National Hydrologic Grid (Clark and others, 2018). Predictions were made for 7 aquifer layers (1 MRVA, 4 MCAQ, 2 LCAQ) following the hydrogeologic framework used in a regional groundwater flow model (Hart and others, 2008). Explanatory variables for the BRT models included attributes associated with well position and construction, surficial variables, and variables extracted from a MODFLOW groundwater flow model for the MISE (Haugh and others, 2020a,b). For a full description of modeling workflow see Knierim and others (2020).

Groundwater is a vital resource in the Mississippi embayment of the central United States. An innovative approach using machine learning (ML) was employed to predict groundwater salinity—including specific conductance (SC), total dissolved solids (TDS), and chloride (Cl) concentrations—across three drinking-water aquifers of the Mississippi embayment. A ML approach was used because it accommodates a large and diverse set of explanatory variables, does not assume monotonic relations between predictors and response data, and results can be extrapolated to areas of the aquifer not sampled. These aspects of ML allowed potential drivers and sources of high salinity water that have been hypothesized in other studies to be included as explanatory variables. The ML approach integrated output from a groundwater-flow model and water-quality data to predict salinity, and the approach can be applied to other aquifers to provide context for the long-term availability of groundwater resources. The Mississippi embayment includes two principal regional aquifer systems; the surficial aquifer system, dominated by the Quaternary Mississippi River Valley Alluvial aquifer (MRVA), and the Mississippi embayment aquifer system, which includes deeper Tertiary aquifers and confining units. Based on the distribution of groundwater use for drinking water, the modeling focused on the MRVA, middle Claiborne aquifer (MCAQ), and lower Claiborne aquifer (LCAQ). Boosted regression tree (BRT) models (Elith and others, 2008; Kuhn and Johnson, 2013) were developed to predict SC and Cl to 1-kilometer (km) raster grid cells of the National Hydrologic Grid (Clark and others, 2018) for 7 aquifer layers (1 MRVA, 4 MCAQ, 2 LCAQ) following the hydrogeologic framework of Hart and others (2008). TDS maps were created using the correlation between SC and TDS. Explanatory variables for the BRT models included attributes associated with well location and construction, surficial variables (such as soils and land use), and variables extracted from a MODFLOW groundwater flow model for the Mississippi embayment (Haugh and others, 2020a; Haugh and others, 2020b). Prediction intervals were calculated for SC and Cl by bootstrapping raster-cell predictions following methods from Ransom and others (2017). For a full description of modeling workflow and final model selection see Knierim and others (2020).