- Program Areas
- Installation Energy and Water
- Environmental Restoration
- Munitions Response
- Resource Conservation and Resiliency
- Weapons Systems and Platforms
Demonstration and Validation of GTS Long-Term Monitoring Optimization Software at Military and Government Sites
Philip Hunter | HQ AFCEE/BCE
Objectives of the Demonstration
Groundwater contamination is common at Department of Defense (DoD) sites and large monitoring networks comprising dozens, hundreds, or even thousands of wells are in place at many facilities, as required for long-term monitoring (LTM).
The objective of this ESTCP project was to demonstrate and validate use of the Geostatistical Temporal-Spatial (GTS) groundwater optimization software which offers a set of tools for long-term monitoring optimization (LTMO).
The GTS software demonstrated in this project consists of five major modules:
- Prepare imports analytical and water-level data, imports site boundaries and shape file overlays, and enables data management via (a) an internal SQLite database, (b) creation of analysis variables, and (c) identification of outliers.
- Explore allows for basic statistical exploration via data summaries and graphs, analysis and ranking of contaminants based on optimization potential, and identification and analysis of multiple vertical aquifer horizons.
- Baseline displays initial groundwater monitoring network status, fits nonlinear baseline trends via locally weighted quadratic regression (LWQR), displays trend maps, builds spatial models via bandwidth selection, computes and displays potentiometric surfaces, and constructs and displays concentration-based plume base maps using quantile local regression (QLR).
- Optimize allows for both temporal and spatial optimization. Temporal optimization in GTS consists of two components: (1) temporal variograms applied to groups of wells and (2) iterative thinning of individual wells. More than one temporal optimization method allows for flexible handling of the kinds of data available at different installations. Spatial optimization within GTS consists of (1) searching for statistical redundancy via mathematical optimization using the GTSmart algorithm; (2) determining optimal network size with the aid of cost-accuracy trade-off curves; and (3) assessing whether new wells should be added and where (i.e., network adequacy).
- Predict allows import and comparison of new sampling data against previously estimated trends and maps. Two options include trend flagging and plume flagging to identify potentially anomalous new values.
To support the Optimize module, GTS also includes a separate stand-alone Excel spreadsheet Cost Comparison Calculator to realistically calculate the financial benefits of implementing a GTS-optimized sampling program, as well as return on investment (ROI).
The three demonstration sites were:
- Air Force Plant 44 Site, Tucson, AZ (AFP44 site)
- Former Nebraska Ordnance Plant Site, Mead, NE (NOP site)
- Fernald DOE Site, Ross, OH (Fernald site)
Key results of the project are listed below:
- The GTS software was found to be easy to use and navigate by the testers and mid-level site analysts, even though none of these users was formally trained on the software. Because GTS v1.0 represents a major overhaul and upgrade to the previous beta version, with a software architecture that was completely redesigned, a significant number of software bugs, logic flaws, and glitches were encountered during both internal and external testing of the software. By the end of project, no significant bugs or software errors remained.
- Graphical outputs in GTS were found to be quite helpful and attractive to users. These, combined with the unique exploratory data tools built in the software, were rated as one of its strong points.
- GTS was found to be effective as an optimization tool. Significant degrees of redundancy were identified at each demonstration site. The iterative thinning function recommended reductions in sampling frequency ranging from 50 to 75% across the three demonstration sites, while the GTSmart algorithm found degrees of spatial redundancy ranging from 16 to 40%. Further, GTS was run successfully at larger sites having more than 200 distinct well locations.
- Of the temporal optimization tools, iterative thinning was found to be superior in performance to temporal variograms. The variograms were easily computed, but yielded poor to mixed results. Overall, the results did not enable reliable or replicable estimates of optimal sampling intervals, since few variogram ranges (denoting points of optimality) could be identified at the test sites.
- A goal of this project was to enable users to perform water-level-aided spatial mapping as an option in GTS. Internal testing of this feature led to mixed results and a decision not to include it in v1.0 of the software. However, as a by-product of this testing, GTS now includes the ability to create potentiometric surface maps of groundwater levels. Users found this to be a useful tool and visualization feature.
- When the input data sets were essentially equivalent, GTS optimization results were shown to be highly reproducible when comparing results from expert users and independent mid-level analysts. Except for the Fernald site, where the input data sets substantially differed, the optimized sampling intervals were identical on a site-wide basis at the other demonstration sites and differed only slightly when broken down by aquifer zone. Spatially, the levels of redundancy found using the same COCs were very similar at both the AFP44 and NOP sites. Further, a locational analysis of which wells were flagged as redundant showed statistically significant similarity in common locations and spatial proximity.
- The trend and plume flagging tools in GTS were shown to be reasonably effective in flagging potentially anomalous measurements from a reserved subset of data from each demonstration site. And, because the reserved data sets were collected “close in time” to the historical data—being observations from the next year of sampling—the projected (i.e., extrapolated) trends and plumes successfully predicted (i.e., bounded) over 90% of the new measurements. Nevertheless, the trend and plume flagging features may be too sensitive in flagging anomalies; further investigation indicated that perhaps only 30% of the trend anomalies and 65% of the plume anomalies were values actually deserving further investigation or verification.
The network adequacy function successfully located areas of substantial mapping uncertainty at each demonstration site, and recommended coordinate locations for the siting of new wells. Because GTS cannot determine whether a suggested new location coincides with a physical obstruction or is unfeasible for other reasons, users were able to successfully override specific locations and to document those decisions visually on a post-plot of both existing and recommended locations.
Based on application of GTS v1.0 to the three demonstration sites during this project, the software has certain limitations that could be mitigated by future improvements. These include:
- GTS requires a number of input fields in ASCII text format in order to create a sufficient analysis database. Some users may find the directions for importing data and creating or augmenting databases within GTS more complicated than need be. The software would be improved if this process were streamlined and simplified.
- GTS does not offer sophisticated handling of radiochemical data, particularly measurements recorded with non-positive values (i.e., zeros or negatives). These data must first be converted to positive values, unless they represent non-detects with a known, positive detection or reporting limit. GTS could be improved by allowing a specific option for radiochemical data.
- Optimized sampling intervals from temporal variograms in GTS often do not match the optimized sampling intervals from iterative thinning using the same data. Further improvements to the temporal variogram algorithm may be needed, especially to account for sites with spatial trends that are actively changing over time.
- Cost-accuracy trade-off curves in GTS are not interactive. Although the bias limits can be adjusted by the user, the spatial optimization must be completely rerun each time those limits are changed, in order to see the impact of the revised limits and to generate a new optimal network. The software could be improved by combining the current trade-off curves into a single, weighted curve that would allow for interactive selection of different sampling plans by the user.
- There is no way in GTS v1.0 to batch print graphics. Since a GTS analysis typically generates a large number of statistical graphics, users may be frustrated with the inability to document graphical results outside the application. The software could be improved by enabling an option to do batch printing to popular image formats.
- The mathematical optimization algorithm in GTS is not a true genetic algorithm wherein portions of the binary string “DNA” representing alternate network configurations are allowed to “mate,” “mutate,” and create “offspring.” Instead, GTS does a “smart search” through the space of potential network configurations, selecting for testing only those strings with interwell spacing comparable to the full network. The software might be improved by incorporating a true genetic algorithmic search.
- The Prepare module may identify too many data records as “outliers” at some sites, necessitating needless user review and override. GTS could be revised and streamlined by combining the temporal and spatial outlier searches into a single, improved algorithm that better accounts for local trend fluctuations.
- “Time slices” in GTS—discrete, non-overlapping periods of sampling—are computed automatically, but are not adjustable by the user. The software could be improved by allowing user input to define or adjust time slices to accord with site-specific remedial events or histories.
- The Predict module readily identifies anomalous future measurements, but may be too sensitive in flagging anomalies. GTS could be revised with improved trend and plume flagging routines to better avoid flagging non-anomalous values.
Estimated total cost savings compared to the baseline monitoring program ranged from 39 to 45%, with ROI ranging between 4 and 6 months. The software and users guide will be available free for use by the public.