Probabilistic assessment of potential leachate leakage from livestock mortality burial pits: A supervised classification approach using a Gaussian mixture model (GMM) fitted to a groundwater quality monitoring dataset

Abstract

After a severe epidemic of foot-and-mouth disease (FMD) in 2010–2011 in South Korea, more than 3 million livestock carcasses were promptly disposed of in a large number of on-site livestock mortality burial pits (approximately 44,000 sites) over the country. There has been significant concern regarding the potential leakage of carcass leachate from burial pits into underlying groundwater. To detect leakage, we monitored three chemical parameters (NH4+-N, Cl−, and EC) of groundwater from monitoring wells downgradient of burial pits (n=274) in 2011. The monitored data were applied as the prediction set to a supervised classification scheme using the Gaussian mixture model (GMM) which involves chemical analysis of both the leachate effluent and background groundwater (as the training set). The GMM was tested to the different data distributions of the training set and resulted in statistically accurate models (with 10-fold CV error <16%) that allocate the probabilistic leachate leakage (i.e., the posterior probability) to each burial pit in the prediction set. However, the overall likelihoods tended to be underestimated due to the uncertainty mainly associated with leachate contamination in background groundwater. Therefore, the best-fit GMM for the bivariate distribution of NH4+-N and Cl− was tuned by redefining the probability density function (pdf) of background groundwater only using a Gaussian component fitted to the distribution at the lowest concentration levels, which predicted the leakage more precisely. Consequently, according to the cutoff (p=0.5) of the probability, we concluded that leachate leakage occurred in 49% (n=133) of the burial pits. This study suggests that the burial method for the disposal of livestock carcasses requires careful consideration on the site selection and pit design to prevent leakage, and also demonstrates that GMM is very flexible for the model tuning in the supervised classification scheme.