Bayesian geostatistical concepts for the design of surveys of subsurface materials

MARSSIM survey designs are more efficient for surface soils than had previously been used because MARSSIM makes provision for data from historical site assessments, scoping surveys and characterization surveys to be incorporated in the sample design process during survey unit classification. The result is that the most survey effort is made in class 1 survey units where the probability of contamination is actually the highest. By using a statistical survey design, the number of samples required for final status surveys using MARSSIM are usually lower by a factor of 3 or more compared to previous methods. The key to these gains in efficiency was to incorporate information from professional judgment and historical data into the design of an objective statistical survey plan. Clearly, in the case of subsurface sampling even further gains in statistical design efficiency are desirable to achieve practical survey plans. A more guided or adaptive technique of sampling will be needed. The number of samples required in a survey can also be reduced by increasing the information available by other means than simply taking more direct measurements. This can be done by (1) increasing the information available from professional knowledge of site processes, historical data, pollutant transport etc. and (2) making more efficient use of the hard data that is already available by the use of more advanced statistical methods. The major problem with implementing (I) is to incorporate such information into the decision-making process in an objective way. Implementing (2) requires more advanced statistical methods that essentially add information by increasing the complexity of the model assumed for the data. The difficulties associated with detecting and quantifying subsurface residual radioactivity are such that a method is needed by which more "soft information" (i.e., data other than the results of specific measurements at a location within the survey unit) can be incorporated into the process so that fewer hard data points will be required. Bayesian methods are used in statistics to make use of information available from prior knowledge. Geostatistical methods are used to make more efficient use of spatial data. These procedures are being incorporated into the SADA software being developed at the University of Tennessee.