Purpose and Scope

This report documents the development and use of machine learning methods to model and map manganese and 1,4-dioxane in groundwater underlying Long Island. The models were developed to investigate the use of machine learning methods with available data to characterize the occurrence and distribution of contaminants in groundwater within the Long Island aquifer system. The models are considered preliminary in the sense that they are an initial effort at developing these kinds of models specifically for Long Island. The models were based on only a selected fraction of the available data that potentially could be used for these purposes, in terms of both concentration data in water from wells and explanatory variables relevant to Long Island groundwater quality. The models provide an initial depiction of predictions for the selected contaminants in groundwater underlying Long Island. Four models were developed for manganese, predicting the probability of detection and of concentrations exceeding three thresholds relevant for drinking-water quality. The models for manganese extend across the entire Long Island area. One model was developed for 1,4-dioxane, predicting the probability of detection. The 1,4-dioxane model is limited geographically to Suffolk County (fig. 1), because 1,4-dioxane data were available only for that area in this study. Similarly, the models are limited to the upper glacial and Magothy aquifers of the Long Island aquifer system because limited data were available for the deeper Lloyd aquifer. The models also represent discrete snapshots in time, based on the temporal distribution of water-quality data. For manganese, the time period represented is a 20-year period centered on 2010, whereas for 1,4-dioxane, only 2018 is represented.

This report describes data compilation, explanatory variable processing, model development and evaluation, and model applications to make predictions across the study area. Predictions concerning manganese and 1,4-dioxane concentrations were made at a 500-square-foot (ft²) resolution for five depth horizons within the Long Island aquifer system, of which three are shown in this report (all predictions are provided in DeSimone, 2023). Thus, the predictions depict the three-dimensional distribution of manganese and 1,4-dioxane in the aquifer system. Model limitations also are discussed, including suggestions for model improvement by the inclusion of additional data and testing.

Study Area

Long Island extends eastwards from the southernmost tip of New York State, roughly parallel to the Connecticut coast (fig. 1). Long Island is about 1,400 square miles (mi²) in area and 120 miles (mi) long, with a maximum width of about 23 mi. The island is bounded by Long Island Sound to the north and the Atlantic Ocean on the south and east, and it is divided politically into four counties. Kings and Queens Counties, at the western end, are boroughs of New York City and contain nearly two-thirds of the island’s total population of 8 million (U.S. Census Bureau, 2021). These counties use drinking water provided by the New York City supply system from reservoirs in upstate New York. Nassau and Suffolk Counties, which occupy most of the land area of the island, rely on public or domestic wells located on the island to supply their combined population of 2.9 million (U.S. Census Bureau, 2021). More than 400 million gallons per day (Mgal/d) are pumped from public supply wells, and an additional 15 Mgal/d is withdrawn by domestic (private household) wells, on an average annual basis (Kenny and others, 2009; Long Island Commission for Aquifer Protection, 2019a).

Land use and population density generally vary from west to east on Long Island along a gradient of decreasing development. High- and medium-intensity developed land use occupies most of Kings and Queens Counties (fig. 2; U.S. Geological Survey, 2021a, b). Nassau County contains mostly medium-to-low intensity developed land use and developed open space, along with some high-intensity developed, forest, and other undeveloped land uses/land cover. Land use in Suffolk County varies from high-, medium-, and low-intensity developed land use and developed open space in the west to forest, agriculture, and predominantly low-intensity developed land use/land cover in the east. Population density ranged from less than 1,000 people per hectare in parts of Suffolk and Nassau Counties to more than 16,000 people per hectare in parts of Kings and Queens Counties in 2010 (Falcone, 2016).

The aquifer system underlying Long Island is predominantly composed of glaciofluvial, glaciolacustrine, deltaic, and morainal sediments of Pleistocene and Cretaceous age, which form a seaward thickening wedge over crystalline bedrock and reach a total thickness of nearly 2,000 feet (ft; Walter and others, 2020b). The upper glacial aquifer and the Magothy aquifer and its associated aquifers (Monmouth greensand and Jameco aquifer, collectively referred to as the “Magothy aquifer” in this report) are the primary sources of drinking water supplies on Long Island (fig. 3). The low-permeability Gardiners Clay, which is of limited areal extent, separates the upper glacial and Magothy aquifer mainly along the southern shore of Long Island. The Lloyd aquifer, not included this study, underlies the Magothy aquifer and is separated from the Magothy aquifer by the Raritan confining unit (Walter and others, 2020b). Lithologically, the aquifer system sediments range from coarse-grained sand and gravel to silt and clay. Glacial moraines, composed of more poorly sorted and less permeable sediments than the fluvial and deltaic deposits, extend in an east-west trending band on the northern half of the island and into the two forks on the eastern end (Walter and Finkelstein, 2020; Walter and others, 2020b). Lignite, a low-grade coal, is common as interstitial particles and as interbedded layers and lenses in the middle parts of the Magothy aquifer (Walter and Finkelstein, 2020).

In the upper glacial aquifer, groundwater is recharged across the landscape from precipitation and groundwater flow is towards freshwater streams and the coasts (fig. 3). A regional groundwater divide extends in the east-west direction along the center of the island. The Magothy and Lloyd aquifers are recharged by downward vertical flow in bands along this groundwater divide, which represent about one-quarter of the island area for recharge to the Magothy aquifer and only about 1 percent of the island area for recharge to the Lloyd aquifer (Walter and others, 2020b). Groundwater in the upper glacial aquifer is generally young (less than 25 years residence time), except near the shore where flow is upward from the Magothy aquifer to discharge; groundwater in the Magothy aquifer is generally tens to hundreds of years old; and groundwater in the Lloyd aquifer is generally thousands of years old, based on groundwater-flow model simulations (Buxton and Modica, 1992). In the vertical dimension, the freshwater aquifers are bounded by salty groundwater (fig. 3); pumping has moved the saltwater interface inland along in parts of Kings and Queens Counties and along the southwestern shore of the Nassau and Suffolk Counties (Stumm and others, 2020).

Manganese and 1,4-Dioxane Concentrations

Data on manganese and 1,4-dioxane concentrations in groundwater underlying Long Island consisted of well-sample data from the USGS National Water Information System (NWIS) database, the U.S. Environmental Protection Agency (EPA) Safe Drinking Water Information (SDWIS) database, and the Suffolk County Water Authority (SCWA). The NWIS and SDWIS data were compiled as part of a national data aggregation by the USGS, in support of multiple water-quality investigations of the NAWQA program (Erickson and others, 2019). A total of 910 samples were used for manganese models (fig. 4A). Most of the manganese data (66 percent) were from the SCWA, and the remaining data were equally from NWIS and SDWIS (17 percent each). Manganese data extended across all of Long Island (fig. 4A). A total of 553 well samples were used for 1,4-dioxane, all located in Suffolk County and all from the SCWA (fig. 4B). The samples were primarily from public supply wells; the manganese data also included about 10 percent samples from monitoring wells and about 1 percent samples from supply wells of other types, including domestic wells. Both manganese and 1,4-dioxane data were denser in the western part of the island than in the east, reflecting the distribution of public supply wells across the island.

All manganese and 1,4-dioxane data were from samples representing individual wells and were of untreated groundwater. Sample collection dates for manganese data ranged from 1999 to 2018; however, only 10 samples were collected prior to 2008 and 60 percent (including all SCWA samples) were collected in 2010. All samples for 1,4-dioxane were collected in 2018. Samples for manganese were analyzed primarily by inductively coupled plasma-mass spectrometry (EPA method 200.8 or USGS method PLM43; U.S. Environmental Protection Agency, 1994a; Faires, 1992) and secondarily by inductively coupled plasma-atomic emission spectrometry (EPA method 200.7 or USGS method PLA11, U.S. Environmental Protection Agency, 1994b; Fishman, 1993). Analytical method information was not available for SDWIS samples for manganese. All samples for 1,4-dioxane were analyzed by solid-phase extraction and gas chromatograph/mass spectrometry (EPA method 522; Munch and Grimmett, 2008). Quality-control data specific to the manganese or 1,4-dioxane analytical data were not reviewed as part of this study. Manganese analysis was conducted on filtered samples for SCWA and NWIS data and on whole-water (unfiltered) samples for SDWIS; inclusion of unfiltered analysis for SDWIS data introduced an additional source of variability into the manganese dataset but this variability was considered acceptable in order to increase the sample size. Analysis for 1,4-dioxane samples was conducted on whole water (unfiltered) samples.

Prediction Points

The points at which predictions were made for manganese and 1,4-dioxane across the study area corresponded spatially to the central locations of the Long Island groundwater-flow model grid of Walter and others (2020a, b). This model has a three-dimensional grid of 500-×500-ft cell size in the horizontal dimension, rotated 18 degrees counterclockwise with respect to north, and 25 layers of variable thickness in the vertical dimension. Layering in the vertical dimension reflects the dimensions of the aquifer and confining unit layers (Walter and others, 2020b). Predictions were made for the land surface area of Long Island (excluding several small islands in Queens and Suffolk counties) plus a 1.24-mi buffer area at five depth horizons that correspond to 5 of the 25 model layers: the shallow upper glacial aquifer (flow model grid layer 1), the deep upper glacial aquifer (layer 3), the top of the Magothy aquifer (layer 5), the middle of the Magothy aquifer (layer 14), and the bottom of the Magothy aquifer (layer 23). Prediction points were the vertical midpoints of the model grid layers. For the uppermost model layer (layer 1), the simulated water table rather than the layer top (which was set equal to the land surface) was used to determine the layer midpoint to use for predictions. For the upper glacial aquifer (layers 1 and 3), the depth horizons at which the predictions are made are relatively flat lying but vary in altitude; for the Magothy aquifer (layers 5, 14, and 23), the depth horizons of predictions slope downwards towards the southern coast (Walter and others, 2020b, fig. 20). For mapping, prediction points were converted to raster data layers using a raster representation of the flow model grid.

Explanatory Variables

Thirty-two explanatory variables were provided as input to the machine learning models (table 1). The variables described position on the landscape with respect to surface-water flow, position in the groundwater-flow system, groundwater recharge and flow, aquifer characteristics, the predicted geochemical conditions of pH and low dissolved oxygen, land-surface and water-table altitudes, land use, soil characteristics, population density, and the presence or absence of public water supply and sewers. Variables that represented surface characteristics, such as land use, were the same for all depths at single horizontal location. Variables that represented subsurface characteristics, such as aquifer texture variables or pH, varied with depth. Categorical information, such the presence or absence of public water, were described by binary variables indicting true or false. The variables, their sources, and the methods used to attribute them to wells or prediction points are described in the following sections.

Model Selection and Performance

The final XGBoost models developed for manganese and 1,4-dioxane are described in terms of their hyperparameters in table 3 and their performance metrics in table 4. The model objects, input data, and output data are documented in DeSimone (2023). For simplicity during this preliminary study, all five final models were selected by maximizing accuracy, as described previously, during hyperparameter tuning (model training). Accuracy in this context is the proportion of correct predictions, both positive (correctly predicting a concentration above the threshold) and negative (correctly predicting a concentration below the threshold), where positive predictions are instances in which the predicted probability is greater than 0.5. Use of accuracy as the performance metric to maximize during model training is reflected in the high accuracy metrics for all five models (table 4). The high accuracy values also may indicate that the models may overfit the training data.

For classification models based on datasets in which the negative instances greatly outnumber the positive instances, accuracy is an overly optimistic measure of model performance (Kuhn and Johnson, 2013). High accuracies can be achieved by simply predicting the majority class because correct prediction of the minority class, which is often of greater interest, contributes little to overall accuracy (Maloof, 2003; Sun and others, 2009). The kappa statistic is an alternative measure of the agreement between predicted and observed classes that takes into account the degree of agreement that would result because of chance (Viera and Garrett, 2005; Kuhn and Johnson, 2013). Kappa values range from 0 (no agreement) to 1 (perfect agreement). The kappa values for the testing data subsets for models were within the ranges of agreement described as follows (Viera and Garrett, 2005): substantial agreement (from 0.61 to 0.80) for the 1.4-dioxane model and moderate agreement (from 0.41 to 0.60) for the manganese models. Sensitivity is the proportion of correctly predicted positive instances (here, detections or concentrations greater than the threshold); its inverse, specificity, is the proportion of correctly predicted negative instances. In the manganese and 1,4-dioxane models, sensitivity was less than specificity for all models in both training and testing data subsets; among the manganese models, sensitivity was low for the three models predicting the probability of manganese concentrations greater than 50, 150, and 300 µg/L (table 4).

Sensitivity, specificity, kappa, and accuracy all are dependent on the probability threshold (0.5 in this study) used to classify model output into positive and negative instances. The AUC performance statistic describes the capability of a classification model to distinguish between positive and negative instances across all possible probability thresholds (James and others, 2013). AUC values are not biased towards the minority or majority class and are relatively insensitive to class imbalance (unequal data distribution among classes) except when data are highly skewed (Fawcett, 2006; Branco and others, 2016). AUC is the area under a curve that plots the true positive rate (correct classification of positive instances, the same as sensitivity) against the false positive rate (incorrectly classifying negative instances as positive) as the probability threshold for classification changes. AUC values range from 0 to 1, with a value of 1 indicating perfect capability to distinguish between positive and negative instances, a value of 0.5 indicting a capability no better than chance, and a value of 0 indicating perfect incorrect classification. AUC values for the testing data subsets ranged from 0.714 to 0.904 for manganese and AUC was 0.951 for 1,4-dioxane (table 4), indicating predictive skill relative to no-information, chance predictions.

Changing the probability threshold for identifying positive instances is one approach for addressing class imbalance in predictive models; other approaches include altering the input datasets or modeling process through oversampling, undersampling, or weighting (Maloof, 2003; Sun and others, 2009; Branco and others, 2016; Krawczyk, 2016). Altering the input datasets or modeling process to address imbalance was beyond the scope of this study. However, changing the probability threshold is a postprocessing step that can be easily applied to model output. This approach was applied to the output from manganese models at the 50, 150, and 300 µg/L concentration thresholds. Using a probability threshold of 0.3 rather than 0.5 to identify positive predicted instances increased the training data subset sensitivity to 0.977 (from 0.852) for the 50 µg/L manganese model, to 0.949 (from 0.718) for 150 µg/L model, and to 0.714 (from 0.619) for the 300 µg/L model. For the testing data subsets, sensitivity increased to 0.545 (from 0.318) for the 50 µg/L model and to 0.556 (from 0.333) for the 150 µg/L model; testing data sensitivity for the 300 µg/L model (0.400) did not change.

It is important to note, however, that the small sample size of the testing data subsets (181 or 182 samples), combined with the low frequency of positive instances, meant that the numbers of positive instances in testing data subsets for the models were small (22, 9, and 5 positive instances for the 50, 150, and 300 µg/L models, respectively). Thus, a small change in the number of positive instances correctly or incorrectly predicted results in a relatively large change in sensitivity for the models (for example, a change in the prediction of one of the five positive instances in the 300 µg/L model would change the sensitivity by 0.2). The models also have random components, both in the model-development algorithm and in the training/testing data partitioning, and this randomness also introduces instability in model-performance metrics when sample sizes are small. A conclusion from this analysis, and from the model-performance metrics for these models generally, is that a larger dataset or the use of methods to address class imbalance, such as the methods described previously, would be needed to better predict manganese concentrations for the higher concentration thresholds.

Explanatory Variables in the Models

The final models retained 23 to 31 of the 32 explanatory variables that were provided as input to the model development process. In this section, the explanatory variables that were the most important are described and their influence on model predictions is discussed. Patterns of explanatory variable importance and influence that are interpretable and consistent with prior understanding provide confidence in model predictions because this consistency indicates that the models are capturing and simulating patterns that are reflective of underlying physical, chemical, or biological processes. However, it should be noted that explanatory variables may be used as surrogates for other factors or processes that are not included in the model, or their influence may reflect interactions with other variables that are not readily discernable.

The top 10 most influential variables in each model are shown in figure 5, where influence is described by the XGBoost gain metric; gain quantifies the fractional contribution of the variable to model predictions (Chen and others, 2020). The complete listing of variables in each final model and their importance rankings are provided in appendix 1. The direction of influence of these variables is shown by partial dependence plots (PDPs; fig. 6). A PDP shows the average change in the predicted probability as a variable changes in value when all other variables in the model are held at their observed values (Greenwell, 2017). For simplicity, PDPs are shown for the top 10 ranked variables for the 1,4-dioxane model and for only one of the manganese models, the model predicting the probability of manganese concentrations greater than or equal to 10 µg/L (manganese detection). Note that PDPs do not capture interactions between variables and can be difficult to interpret when relations are complex and not monotonic, especially for lower-ranked explanatory variables.

Depth or vertical location in the Long Island aquifer system, captured by the variable Well_depth, was among the top three most important variables in all but two of the models and was among the top 10 in all models (fig. 5). The direction of influence of Well_depth also was the same in all models; the predicted probability of manganese concentrations above thresholds or 1,4-dioxane detection decreased with increasing well depth. For 1,4-dioxane, the predicted probability of detection started decreasing at a deeper depth (about 400 ft; fig. 6M), than for manganese (about 200 ft; fig. 6B), possibly reflecting the greater prevalence of 1,4-dioxane deeper in the Magothy aquifer (Long Island Commission for Aquifer Protection, 2019b) compared with manganese.

Land-surface and water-table altitudes also were among the top ranked variables in the manganese and 1,4-dioxane models (fig. 5, Land_surface and WaterTable variables). For these variables, however, the directions of influence in the manganese models and in the 1,4-dioxane model were not the same. The predicted probability of 1,4-dioxane detection was greater at higher water-table and land-surface altitudes (fig. 6K–L), whereas the predicted probability of manganese concentrations above thresholds decreased at higher land-surface altitudes, and when it was among the top 20 variables in the manganese models, at higher water-table altitudes (fig. 5A; the PDP for the WaterTable variable was not among the top 10 for the manganese model at the 10 µg/L threshold and is not shown). These relations point to the more likely presence of higher manganese concentrations in groundwater beneath low-lying areas and near the coast of Long Island, and the more likely detection of 1,4-dioxane in the central, upland areas of the island.

Some of the most important variables in the manganese models described location on the landscape with respect to surface-water flow and location within the groundwater-flow system. These include thickness-weighted, simulated groundwater flux (GWFlux_thkwtd), which was among the top five variables in all four manganese models, and the two MOHP variables (MOHP_DSD1, MOHP_LP2) that describe land-surface position with respect to streams (fig. 5A–D). The GWFlux_thkwtd and MOHP variables were not among the top 10 most influential variables in the 1,4-dioxane model. Instead, three of the four variables representing land use, LandUse_Resi, LandUse_Undev, and LandUse_Urban, were the fourth, sixth, and seventh most influential variables, respectively, in the 1,4-dioxane model (fig. 5E). In contrast, a land use variable was among the top 10 (ranked sixth) in only one of the manganese models. The contrasting importance of flow-related variables on manganese concentrations as compared to the importance of land-use variables on 1,4-dioxane detection is consistent with the primarily geogenic sources of manganese as opposed to the manmade sources of 1,4-dioxane.

The redox condition variable DissOxy_lt1, the probability of low dissolved oxygen (dissolved oxygen concentration less than 1 mg/L), was ranked ninth, third, and fifth in the three lower threshold manganese models (10, 50, and 150 µg/L, respectively) and eleventh in the 300-µg/L threshold manganese model and in the 1,4-dioxane model. Conceptually, low dissolved oxygen (anoxic redox condition) is positively associated with dissolved manganese in groundwater, because manganese is more soluble in anoxic water, and that is the direction of influence of the DissOxy_lt1 variable on the probability of manganese detection, at concentrations greater than 10 µg/L. The direction of influence of the variable in the 1,4-dioxane model was the opposite; the probability of 1,4-dioxane detection was less with increasing probability of low dissolved oxygen. This relation may reflect the less likely detection of 1,4-dioxane in old, confined groundwater that was recharged before 1,4-dioxane use that is also anoxic, rather than any direct influence of redox condition, especially if 1,4-dioxane could be degraded under aerobic conditions (Adamson and others, 2015). The influence of DissOxy_lt1, the probability of low dissolved oxygen, in the three manganese models at thresholds other than 10 µg/L, was either not monotonic (50 and 300 µg/L models) or was inversely related to the probability of manganese concentrations greater than the model threshold (150 µg/L). These complex relations may result in part because of interactions among explanatory variables that were not investigated as part of this study.

Mapped Predictions

The predicted probability of manganese detected at concentrations greater than or equal to 10 µg/L, manganese concentrations greater than the SMCL of 50 µg/L, and 1,4-dioxane detected at concentrations greater than or equal to 0.07 µg/L are shown in figures 7 to 11. Predicted probabilities are shown at two depth horizons in the upper glacial aquifer (figs. 7–8), and at three depth horizons in the Magothy aquifer (top, middle, and bottom; figs. 9–11). Predictions are not shown for manganese concentrations greater than 150 and 300 µg/L thresholds because of their infrequent occurrence and the low sensitivity of the models at those thresholds. Mapped predictions are available in tagged image file format (tiff) in DeSimone (2023).

Predicted detections of manganese (concentrations greater than or equal to 10 µg/L, probability greater than 0.5) were distributed throughout much of the top layer of the upper glacial aquifer and were predicted with higher probability along the southern shore and in the central part of the island (fig. 7). In the bottom layer of the upper glacial aquifer and in the top and middle layers of the underlying Magothy aquifer (figs. 8–10), predicted manganese detections were more limited to the southern half of the island. This distribution of manganese detections is consistent with the occurrence of groundwater in which oxygen has been depleted over long flow paths along the southern shore or through contact with reducing material such as lignite or pyrite (Brown and others, 1999; Walter and Finkelstein, 2020), and the increased solubility of manganese under reducing conditions. Manganese concentrations greater than the SMCL of 50 µg/L were predicted infrequently in any of the aquifer layers (figs. 7–11).

Detections of 1,4-dioxane (concentrations greater than or equal to 0.07 µg/L, probability greater than 0.5) were predicted across most of the western, more highly developed half of Suffolk County. The areas in which 1,4-dioxane was predicted to occur, in both the upper glacial aquifer layers and in the top and middle of the Magothy aquifer, were similar and extended from the northern shore to near the southern shore of the island in these aquifer layers (figs. 7–10). In the bottom layer of the Magothy aquifer (fig. 11), the area in which 1,4-dioxane was predicted likely to be detected was smaller and was limited to the northwestern part of Suffolk County near the northern shore.

Model Limitations

The machine learning models developed in this study are considered preliminary and their interpretations are otherwise limited for various reasons, as discussed in this section. The models developed for this study were based on a selected fraction of the data potentially available for use in modeling manganese and 1,4-dioxane concentrations in groundwater underlying Long Island. The relatively small sample size and sparse distribution of samples in some parts of the aquifer system place limits on the accuracy that the models can achieve. Class imbalance (unequal data distribution among classes) in the model training data, especially when combined with the small sample size, also limited the capability of the models to correctly predict high concentrations of manganese and 1,4-dioxane (model sensitivity). These limitations could be addressed by expanding the datasets used to build the models and by use of more sophisticated modeling methods than used here to address class imbalance. In addition, it is anticipated that additional explanatory variables, specific to Long Island, such as historical land use and improved resolution of groundwater-flow and source information, would also likely improve model accuracy, sensitivity, and overall model performance.

The degree to which model results accurately described the entire aquifer system was also potentially affected by the representativeness of the well dataset. One example of such a potential effect is the possible effect of well type. Nearly all of the concentration data used to build the models in this study were collected from public supply wells, which may shape the water-quality dataset in various ways (Suffolk County Government, 2015). Public supply wells open to unconsolidated deposits typically have long well screens (for example, more than 10 ft). Water withdrawn from a long-screened well may be a mixture of groundwater from parts of the aquifer with differing geochemical conditions. This introduces variability that may make it more difficult to predict a redox-sensitive constituent like manganese, especially when redox conditions are affected by heterogeneous aquifer sediment textures (Walter, 1997; Brown and others, 2019; Walter and Finkelstein, 2020). Additionally, the high pumping rates (for example, hundreds or thousands of gallons per minute) at some supply wells may alter flow directions so that their samples reflect water quality over a broader area and from shallower depths than samples from wells with lower pumping rates or with no active withdrawals. Expansion of the well dataset to include data from monitoring wells could provide insight into the representativeness of models built primarily on public supply well data such as those of the present study. Expanding the dataset to include monitoring wells could perhaps also fill some of the spatial gaps in sample density, especially for 1,4-dioxane, and thereby better define the conditions associated with 1,4-dioxane detection across Long Island.

Temporal variability in manganese or 1,4-dioxane concentrations was not considered extensively in developing the models described here. Constituent concentrations in groundwater are not static over time, even for constituents like manganese that originate primarily from geogenic sources (Degnan and others, 2020). Manganese concentration data were collected from a 20-year time period (1999 to 2018), over which concentrations may have varied because of changes in the groundwater-flow system, seasonal effects, or locally, manmade sources such as compositing facilities (Suffolk County Department of Health Services, 2016). 1,4-Dioxane data were all collected in 2018, but even data from a single year such as these data represent a large time period (multiple years) in source characteristics because of the varying groundwater-travel times from sources to the sampling locations. Explanatory variables that varied over time were represented by a single year, centrally located within the 20-year time period, or averaged over multiple years. For contaminants that vary over time, time averaging of explanatory variables and temporal discord between the explanatory variables and the sample data may reduce the meaningful, process-based information provided by the explanatory variables. For contaminants that vary over short time periods (for example, weeks or months), use of single point-in-time concentration data may introduce variability that cannot be accurately predicted using static, long-term-averaged explanatory variables and cannot be well characterized at the regional scale.

Finally, spatial resolution places limitations on model accuracy as well as limitations on intended model use. For example, some explanatory variables, such as predicted pH, the predicted probability of low dissolved oxygen, and simulated groundwater residence time, were obtained from regional groundwater-flow models. The spatial resolution of these variables is relatively coarse compared to the size of Long Island, and model predictions reflect the spatial resolution of the explanatory variables. Finer scale data would be more useful for the Long Island machine learning models and could increase model accuracy. Moreover, although predictions presented in this report are at a relatively fine grid cell resolution of 500 ft², the model results are intended to represent regional patterns across Long Island rather than to predict concentrations at specific point locations.