A comparison of data-driven groundwater vulnerability assessment methods.

Increasing availability of geo-environmental data has promoted the use of statistical methods to assess groundwater vulnerability. Nitrate is a widespread anthropogenic contaminant in groundwater and its occurrence can be used to identify aquifer settings vulnerable to contamination. In this study, multivariate Weights of Evidence (WofE) and Logistic Regression (LR) methods, where the response variable is binary, were used to evaluate the role and importance of a number of explanatory variables associated with nitrate sources and occurrence in groundwater in the Milan District (central part of the Po Plain, Italy). The results of these models have been used to map the spatial variation of groundwater vulnerability to nitrate in the region, and we compare the similarities and differences of their spatial patterns and associated explanatory variables. We modify the standard WofE method used in previous groundwater vulnerability studies to a form analogous to that used in LR; this provides a framework to compare the results of both models and reduces the effect of sampling bias on the results of the standard WofE model. In addition, a nonlinear Generalized Additive Model has been used to extend the LR analysis. Both approaches improved discrimination of the standard WofE and LR models, as measured by the c-statistic. Groundwater vulnerability probability outputs, based on rank-order classification of the respective model results, were similar in spatial patterns and identified similar strong explanatory variables associated with nitrate source (population density as a proxy for sewage systems and septic sources) and nitrate occurrence (groundwater depth).

Influence of threshold value in the use of statistical methods for groundwater vulnerability assessment.

Statistical techniques can be used in groundwater pollution problems to determine the relationships among observed contamination (impacted wells representing an occurrence of what has to be predicted), environmental factors that may influence it and the potential contamination sources. Determination of a threshold concentration to discriminate between impacted or non impacted wells represents a key issue in the application of these techniques. In this work the effects on groundwater vulnerability assessment by statistical methods due to the use of different threshold values have been evaluated. The study area (Province of Milan, northern Italy) is about 2000 km(2) and groundwater nitrate concentration is constantly monitored by a net of about 300 wells. Along with different predictor factors three different threshold values of nitrate concentration have been considered to perform the vulnerability assessment of the shallow unconfined aquifer. The likelihood ratio model has been chosen to analyze the spatial distribution of the vulnerable areas. The reliability of the three final vulnerability maps has been tested showing that all maps identify a general positive trend relating mean nitrate concentration in the wells and vulnerability classes the same wells belong to. Then using the kappa coefficient the influence of the different threshold values has been evaluated comparing the spatial distribution of the resulting vulnerability classes in each map. The use of different threshold does not determine different vulnerability assessment if results are analyzed on a broad scale, even if the smaller threshold value gives the poorest performance in terms of reliability. On the contrary, the spatial distribution of a detailed vulnerability assessment is strongly influenced by the selected threshold used to identify the occurrences, suggesting that there is a strong relationship among the number of identified occurrences, the scale of the maps representing the predictor factors and the model efficiency in discriminating different vulnerable areas.

Spatial and statistical assessment of factors influencing nitrate contamination in groundwater.

The weights of evidence (WofE) modeling technique has been used to analyze both natural and anthropogenic factors influencing the occurrence of high nitrate concentrations in groundwater resources located in the central part of the Po Plain (Northern Italy). The proposed methodology applied in the Lodi District combines measurements of nitrate concentrations, carried out by means of a monitoring net of 69 wells, with spatial data representing both categorical and numerical variables. These variables describe either potential sources of nitrate and the relative ease with which it may migrate towards groundwater. They include population density, nitrogen fertilizer loading, groundwater recharge, soil protective capacity, vadose zone permeability, groundwater depth, and saturated zone permeability. Once conditional dependence problems among factors have been solved and validation tests performed, the statistical approach has highlighted negative and positive correlations between geoenvironmental factors and nitrate concentration in groundwater. These results have been achieved analysing the calculated statistical parameters (weights, contrasts, normalized contrasts) of each class by which each factor has been previously subdivided. This has permitted to outline: the overall influence each factor has on the presence/absence of nitrate; the range of their values mostly influencing this presence/absence; the most and least critical combination of factor classes existing in each specific zone; areas where the influence of impacting factor classes is reduced by the presence of not impacting factor classes. This last aspect could represent an important support for a correct land use management to preserve groundwater quality.