3.1. Calibration and Validation of DRASTIC Weights
As shown in
Table 5, RMSE for aquifer vulnerability without calibration was 0.70. The RMSE for aquifer vulnerability with calibrated DRASTIC parameters using Bi-GA was 0.57. RMSE for Bi-GA might be decreased (the lower RMSE, the better performance) because calibrated DRASTIC weights using Bi-GA maintained the ratios of original DRASTIC weights. Previous studies did not maintain the ratios of original DRASTIC weights to improve just performance evaluation of DRASTIC. However, if the ratios of original DRASTIC weights are not maintained, the number of degrees of freedom of the DRASTIC index (result scores for aquifer vulnerability) would be increased by calibrating DRASTIC weights. Further, physical properties for aquifer vulnerability could potentially be ignored.
For validation of the results by using Bi-GA, accuracy assessment was computed with 35 wells with nitrate levels > 2 ppm. As shown in the
Table 6, total accuracies of uncalibrated DRASTIC and calibrated DRASTIC were 34% (=12/35) and 46% (=16/35), respectively. Thus, the results of accuracy assessment indicate calibrated DRASTIC predicted aquifer vulnerability areas contaminated by human activities more accurately than uncalibrated DRASTIC.
DRASTIC, an overlay and index GIS model, does not compute nitrate concentrations in aquifers, rather it predicts aquifer vulnerability classes from very high vulnerability to very low vulnerability. This study assumed nitrate concentrations greater than 2 ppm were caused by human activities and over 2 ppm of nitrate concentrations should typically be detected in “High” and “Very high” vulnerability classes. Thus, the greater the proportion of nitrate detections >2 ppm in “High” and “Very high” vulnerability areas, the better the prediction of aquifer vulnerability. If “High” and “Very high” vulnerability areas as a percentage are larger than number of nitrate detections >2 ppm as a percentage, the model performance should be regarded as poor, which would be overestimated by DRASTIC. This is captured in the concept of a detection ratio (percent of nitrate detections >2 ppm to percent of “Very high” and “High” vulnerability areas) with larger detection ratios indicating better prediction used to evaluate model performance in this study.
3.2. Aquifer Vulnerability Mapping
An aquifer vulnerability map without calibrating DRASTIC weights was created using DRASTIC (
Figure 4a). Aquifer vulnerability indices were classified into five classes: 0–0.2 (“Very low”), 0.2–0.4 (“Low”), 0.4–0.6 (“Moderate”), 0.6–0.8 (“High”), and 0.8–1.0 (“Very high”). As shown in
Figure 4a and
Table 7, 10.6% of the aquifer systems in the UWRW were within in “Very low” vulnerability class, and 60.4% of the area was estimated as “Low”, 25.8% within “Moderate” vulnerability class, 3.0% within “High” vulnerability class, and 0.2% within “Very high” vulnerability class.
The aquifer vulnerability results (
Table 7) without calibration of DRASTIC weights were validated with the observed nitrate concentrations in wells. The results showed that approximately 35.3% of nitrate detections >2 ppm were within “High” and “Very high” vulnerability areas (represent 3.2% of vulnerability area) as simulated by DRASTIC. Moreover, 60.3% of the nitrate detections were within the “Moderate” vulnerability class (25.8% of area), 3.4% of the nitrate detections were within the “Low” vulnerability class (60.4% of area), and 0.9% of the nitrate detections were within the “Low” vulnerability class (10.6% of area) (
Table 7).
An aquifer vulnerability map with calibrated DRASTIC using Bi-GA was produced (
Figure 4b). As shown in
Figure 4b and
Table 8, 9.6% of the aquifer systems in the UWRW was within the “Very low” vulnerability class, and 60.1% of the area was estimated as “Low”, 26.9% within the “Moderate” vulnerability class, 3.2% within the “High” vulnerability class, and 0.2% within the “Very high” vulnerability class.
The aquifer vulnerability results (
Table 8) from calibrated DRASTIC were validated with the well database. The results showed that approximately 42.2% of nitrate detections >2 ppm were within “High” and “Very high” vulnerability areas (represent 3.4% of vulnerability area) as simulated by DRASTIC. Moreover, 53.4% of the nitrate detections were within the “Moderate” vulnerability class (26.9% of area), and 4.3% of the nitrate detections were within the “Low” vulnerability class (60.1% of area). In aquifer vulnerability assessment, nitrates in wells >2 ppm were not detected within the “Very low” vulnerability class (9.6% of area) (
Table 8). These results indicated that aquifer vulnerability assessment using DRASTIC with Bi-GA better predicted nitrate detections than DRASTIC without calibration.
Very high and high vulnerability areas were located along the stream and river because those areas include highly permeable alluvium, sand, and gravel. Further, depth to water is shallow and the vadose zone media includes gravel, sand, and peat. According to the land use map, very high and high vulnerability classes include areas that are mainly near streams, agricultural fields, and urban areas because fertilizer and urban organic wastes from these areas infiltrate toward aquifers. Based on the topography map, very high and high vulnerability classes are mainly observed in lowland areas where it is common to find agricultural lands that receive fertilizers and urban complexes that contribute in various ways to pollution.
This study assumed that nitrate concentrations in wells >2 ppm should have not been detected in “Low” and “Very low” vulnerability areas. However, five nitrate concentrations in wells >2 ppm (4.3% of total nitrate detections >2 ppm) were found in “Low” and “Very low” vulnerability areas. Four main reasons (groundwater age, point sources, wells that have failed, and groundwater flow) explain why detections may have occurred in “Low” and “Very low” vulnerability areas as these factors are not considered in DRASTIC.
GIS-based overlay and index models such as DRASTIC can be affected by data resolution and accuracy [
48]. Navulur [
16] used three models (i.e., DRASTIC, SEEPAGE, and combined DRASTIC and NLEAP (Nitrate Leaching and Economic Analysis)) to estimate aquifer vulnerability of groundwater systems in Indiana using a GIS environment at a 1:250,000 scale. The data scale used in Navulur’s [
16] study was coarse (1:250,000) for field scale simulations. However, in this study, high resolution data (1:24,000) were used by data preprocessing of recharge (R), aquifer media (A), soil media (S), topography (T), and impact of vadose zone media (I) maps.
As shown in Navulur’s [
16] results for all of Indiana, the result of DRASTIC shows 80.7% of nitrate detections in wells >2 ppm are within “High” and “Very high” vulnerability areas (represent 24.8% of area) as predicted by DRASTIC. For SEEPAGE, 60.5% of nitrate detections in wells >2 ppm are within “High” and “Very high” vulnerability areas (28.6% of area). The result of the combined DRASTIC and NLEAP indicate 91.8% of nitrate detections in wells >2 ppm are within “High” and “Very high” vulnerability areas (56.9% of area).
Compared with Navulur’s [
16] study, the results presented herein had approximately 42.2% of nitrate detections in wells >2 ppm within “High” and “Very high” (3.4% of area) vulnerability areas as predicted by DRASTIC with high resolution data. Detection ratio (% of nitrate detections to % of vulnerability areas with larger detection ratio indicating better prediction) for “High” and “Very high” areas from Navulur’s [
16] study results in a value of 3.3 for DRASTIC, 2.1 for SEEPAGE, and 1.6 for the combined DRASTIC and NLEAP. In contrast to the three models from Navulur’s [
16] results, the results presented herein provide a value of 12.4. Thus, the detection ratio results indicate that DRASTIC with high resolution data may estimate areas of “High” and “Very high” vulnerability classes more accurately than models with coarse resolution data (
Table 9).
3.3. Potential Groundwater Monitoring and Management Sites
The Gi* statistic method was used to determine potential groundwater monitoring and management sites. Three ranges of z-scores (1.65–1.96, 1.96–2.58, and >2.58) indicate potential groundwater monitoring and management sites (hotspots). Hotspots were predicted based on the z-score with statistical significance using the Gi* statistic method. The Gi* statistic method identifies statistically significant spatial clusters of high values (high vulnerability areas) and low values (low vulnerability areas). The Gi* statistic method returns a z-score and the higher the z-score, the stronger the intensity of the clustering.
In
Table 10 and
Figure 5, z-scores of hotspot analysis maps to identify potential groundwater monitoring sites were estimated using calibrated DRASTIC by Bi-GA. Higher z-scores and red color (potential vulnerability areas) in the maps (
Table 10 and
Figure 5) indicate hotspots which suggest priority areas for groundwater monitoring and management. The portion of the study area with a z-score ≥ 1.65 for Bi-GA is 19.9% (percentage of study area, 6944 km
2), suggesting areas where groundwater monitoring and BMPs for groundwater quality might be considered. In
Figure 5, hotspot areas (z-score ≥ 1.65) were located along the stream and river because those areas include highly permeable alluvium, sand, and gravel. Further, depth to water is shallow. These areas would be priorities for groundwater protection.